IR-WSANet: A lightweight network with frequency-spatial joint optimization for infrared small target detection Authors: Zhan Jin, Haonan Chen, Jiaxiang Chen Published: 2026 ABSTRACT Infrared Small Target Detection (IRSTD) holds significant application value in military operations, early warning and surveillance, aerospace, and other fields. However, traditional detection methods face challenges related to small target pixel sizes, strong background noise, and sparse infrared image features, resulting in insufficient accuracy and robustness. This paper proposes IR-WSANet, a lightweight network based on an improved YOLOv10, which enhances the detection performance of infrared small targets through a frequency-spatial joint optimization strategy. Firstly, the discrete wavelet transform convolution (DWaveletConv) is introduced into the backbone network, and the fusion of high-frequency details and low-frequency semantics is enhanced by multi-band feature decomposition to suppress noise interference. Secondly, we designed a cooperative module (POS-SHSA) that integrates POSConvEmbedding with a partial channel single-head self-attention mechanism (SHSA), which combines local spatial features and global context information to improve the positioning accuracy of small targets. Experiments verify the effectiveness of the model on SIDD and HIT-UAV datasets: the mAP of IR-WSANet on SIDD-City, SIDD-Mountain and HIT-UAV datasets reaches 97.2%, 82.6% and 82.8%, respectively, which is 2.8% to 14.1% higher than the baseline YOLOv10, and the highest F1 score was improved to 14.8%, while maintaining low computing cost (27.9 GFLOPs) and real-time performance (42.8 FPS). The results show that IR-WSANet significantly improves the detection performance of infrared small targets in complex scenes through the combination of frequency domain filtering enhancement and space-channel dual attention mechanism.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-56213-1. Introduction Infrared Small Target Detection (IRSTD) is crucial for applications in military operations1, early warning systems, surveillance2, and aerospace3. Its potential stems from its ability to detect small targets in challenging environments. In these contexts, infrared imaging equipment is usually widely used in the detection and recognition of small targets at long distances, especially in scenarios involving unmanned aerial vehicles, automobiles, and pedestrians against diverse backgrounds. Infrared small targets typically occupy only a few to ten pixels, posing significant detection challenges due to limited pixel representation. Additionally, the long shooting distances, environmental interferences (such as clouds and smog), and the distinctive image noise associated with infrared images (including unique feature sparsity and a low signal-to-noise ratio) make detecting small targets in infrared images more difficult than in ordinary images4. The primary objective of the IRSTD task is to accurately determine the presence of a target, enabling subsequent processing based on the detected target. The early research methods for IRSTD are primarily classified as filtering-based methods5, local contrast-based methods5,6, and deep learning-based methods7,8. These research approaches are crucial for enhancing the efficiency and accuracy of IRSTD tasks. However, most of these methods rely on manually designed feature templates, which often struggle to meet task requirements due to the challenges posed by complex backgrounds and high noise levels in infrared images. In recent years, filtering-based methods have increasingly focused on optimizing filter kernels and improving robustness. He et al. (2024) proposed an algorithm based on variance difference weighted three-layer local contrast measurement (VDWTLCM) to address the problem of high false alarm rates in IRSTD under strong clutter interference, significantly improving detection performance9. Cui et al. (2025) integrate steering-kernel-guided saliency with weighted guided image filtering to produce edge-preserving background estimates, effectively suppressing complex clutter and reducing false alarms in infrared small-target detection10. IRSTD research based on deep learning methods begins with the investigation of CNN and R-CNN networks11, and some research results have been obtained. In follow-up research on deep learning methods, the model design at this stage emphasizes the characteristics of small targets, such as low contrast and small size. Recent graph and transformer-based studies have advanced infrared small and dim target detection by explicitly modeling long-range dependencies and exploiting structured background priors. Aibibu et al. (2023) proposed an efficient fusion network to address challenges related to small target size, low contrast, and occlusion in IRSTD12. Lin et al. (2023) introduced IR-TransDet, a Transformer tailored for infrared imagery that employs multi-scale feature aggregation and self-attention to strengthen cross-scale semantic interactions13. Lin et al. (2024) developed CS-ViG-UNet, which embeds cycle-shift vision-graph convolutions within a U-Net backbone to capture topological target–neighborhood relations and improve the discrimination of faint targets in complex scenes14. Lin et al. (2025) presented BPGA, integrating background-prior guidance with vision-graph attention to suppress structured clutter and reduce false alarms15. A recent state-space plus self-attention framework (2025) enables joint spectral–temporal fusion for dual-band and dynamic scenarios, improving stability on moving platforms16. Zhang et al. (2025) designed a multi-layer, multi-patch, multi-sensitivity, multi-dimensional attention network that expands the receptive field and enhances multi-scale feature processing17. Recent multimodal and lightweight approaches provide complementary strategies to enhance IR-WSANet. Prompt-driven unsupervised pre-training converts pseudo-masks from large vision–language models into kernel-level supervision, thereby improving query-based localization in low-annotation settings18. Vision-only dynamic reconstruction methods19 extract geometric and temporal priors from monocular streams, enabling robust separation of static and dynamic elements for moving platforms. Region-to-point geo-localization20 demonstrates that hierarchical semantic aggregation improves cross-scale spatial matching, which is particularly effective for disambiguating small aerial targets. Multimodal multiscale aggregation methods21 show that lightweight cross-modal fusion can significantly improve recall and localization accuracy for small objects. Integrating these approaches—self-supervised prompts, dynamic priors, hierarchical semantics, and efficient multimodal fusion—can directly complement IR-WSANet’s frequency-spatial optimization to improve robustness and generalization in low-data settings. In interactive research on IRSTD tasks, particularly those that use filtering methods, adapting to complex detection backgrounds poses significant challenges22,23. Conversely, in research methods based on deep learning, processing characteristic noise is difficult24. Building on this type of interactive research, we propose IR-WSANet, which integrates wavelet transform convolution with a single-head self-attention module25 and a position-coding module (POS-SHSA), all based on the YOLOv10 network. In the real-time improved IR-WSANet model, we replace the traditional convolution module with a discrete wavelet transform convolution module (DWaveletConv) within the YOLOv10 backbone. This modification addresses the issues of redundant feature extraction and limited computational efficiency in infrared image target detection26,27. In this architecture, we draw inspiration from the parallel branch concept of Conv2d, fusing features from different frequency bands and feeding them into the subsequent network. Utilizing the discrete wavelet transform (DWT)28, the input feature map is decomposed to generate low-frequency approximations (LL) and high-frequency detail (LH, HL, HH) subbands. Each subband’s features are then processed through convolution operations, and a position-aware convolutional embedding layer, referred to as POSConvEmbedding, is incorporated. Following these improvements, IR-WSANet offers a high-precision, low-latency solution for infrared small target detection through frequency-spatial joint optimization and detection head reconstruction, particularly in practical application scenarios such as unmanned aerial vehicles. Learnable position coding enhances the positional perception capabilities of the IR-WSANet network for detecting small targets in infrared images, thereby improving detection sensitivity. In this paper, lightweight emphasizes a practical accuracy–cost trade-off: IR-WSANet is designed to maximize detection accuracy per unit of computation and per parameter, while keeping absolute model size and runtime suitable for real-time deployment on embedded GPU hardware. The innovations presented in this paper are primarily reflected in the following aspects: Frequency Domain-spatial Domain Joint Optimization for Feature Extraction: This study addresses the challenge of high-frequency details being easily interfered with by noise and the inadequacy of low-frequency semantic information in infrared small target detection. We propose an improved backbone network method based on Discrete Wavelet Transform Convolution (DWaveletConv). By employing multi-band decomposition (LL, LH, HL, HH sub-bands) and cross-channel fusion, we enhance the edge feature extraction capability for high-frequency targets while simultaneously suppressing background noise.Location-Aware and Global Context Cooperative Attention Mechanism: A cooperative module, termed Position-Aware Single-Head Self-Attention (POS-SHSA), integrates location-aware coded convolution (POSConvEmbedding) with a partial channel single-head self-attention mechanism (SHSA). This design enhances local spatial sensitivity through learnable position coding while simultaneously modeling global dependencies across certain channels using the single-head self-attention mechanism. The computational cost is significantly reduced, with the parameter count decreasing by 27% compared to multi-head attention. This approach successfully balances the accuracy of small target positioning with the suppression of background interference.Engineering Optimization of a Lightweight Real-Time Detection Network: Based on YOLOv10, we have reconstructed the four-level feature pyramid and implemented an adaptive weighted fusion strategy. By utilizing dynamic resolution input and optimizing depth separable convolution, we have managed to control the computational complexity at 27.9 GFLOPs while achieving a reasoning speed of 42.8 FPS. This has been accomplished without compromising detection accuracy, as evidenced by a 14.1% improvement in mean Average Precision (mAP) on the HIT-UAV dataset. These results significantly outperform existing YOLO series models and Transformer models, such as RT-DETR.The verification scheme of strong robustness across various dataset. The model’s performance is systematically evaluated on complex scene dataset, including SIDD and HIT-UAV. Additionally, a dynamic label allocation strategy and a multi-scale loss function for infrared small targets are proposed. These innovations effectively address the issues of false detection and missed detection that are prevalent in traditional methods, particularly under conditions of strong noise and low contrast. Furthermore, the model exhibits enhanced adaptability in practical scenarios, such as UAV surveillance. Frequency Domain-spatial Domain Joint Optimization for Feature Extraction: This study addresses the challenge of high-frequency details being easily interfered with by noise and the inadequacy of low-frequency semantic information in infrared small target detection. We propose an improved backbone network method based on Discrete Wavelet Transform Convolution (DWaveletConv). By employing multi-band decomposition (LL, LH, HL, HH sub-bands) and cross-channel fusion, we enhance the edge feature extraction capability for high-frequency targets while simultaneously suppressing background noise. Frequency Domain-spatial Domain Joint Optimization for Feature Extraction: This study addresses the challenge of high-frequency details being easily interfered with by noise and the inadequacy of low-frequency semantic information in infrared small target detection. We propose an improved backbone network method based on Discrete Wavelet Transform Convolution (DWaveletConv). By employing multi-band decomposition (LL, LH, HL, HH sub-bands) and cross-channel fusion, we enhance the edge feature extraction capability for high-frequency targets while simultaneously suppressing background noise. Location-Aware and Global Context Cooperative Attention Mechanism: A cooperative module, termed Position-Aware Single-Head Self-Attention (POS-SHSA), integrates location-aware coded convolution (POSConvEmbedding) with a partial channel single-head self-attention mechanism (SHSA). This design enhances local spatial sensitivity through learnable position coding while simultaneously modeling global dependencies across certain channels using the single-head self-attention mechanism. The computational cost is significantly reduced, with the parameter count decreasing by 27% compared to multi-head attention. This approach successfully balances the accuracy of small target positioning with the suppression of background interference. Location-Aware and Global Context Cooperative Attention Mechanism: A cooperative module, termed Position-Aware Single-Head Self-Attention (POS-SHSA), integrates location-aware coded convolution (POSConvEmbedding) with a partial channel single-head self-attention mechanism (SHSA). This design enhances local spatial sensitivity through learnable position coding while simultaneously modeling global dependencies across certain channels using the single-head self-attention mechanism. The computational cost is significantly reduced, with the parameter count decreasing by 27% compared to multi-head attention. This approach successfully balances the accuracy of small target positioning with the suppression of background interference. Engineering Optimization of a Lightweight Real-Time Detection Network: Based on YOLOv10, we have reconstructed the four-level feature pyramid and implemented an adaptive weighted fusion strategy. By utilizing dynamic resolution input and optimizing depth separable convolution, we have managed to control the computational complexity at 27.9 GFLOPs while achieving a reasoning speed of 42.8 FPS. This has been accomplished without compromising detection accuracy, as evidenced by a 14.1% improvement in mean Average Precision (mAP) on the HIT-UAV dataset. These results significantly outperform existing YOLO series models and Transformer models, such as RT-DETR. Engineering Optimization of a Lightweight Real-Time Detection Network: Based on YOLOv10, we have reconstructed the four-level feature pyramid and implemented an adaptive weighted fusion strategy. By utilizing dynamic resolution input and optimizing depth separable convolution, we have managed to control the computational complexity at 27.9 GFLOPs while achieving a reasoning speed of 42.8 FPS. This has been accomplished without compromising detection accuracy, as evidenced by a 14.1% improvement in mean Average Precision (mAP) on the HIT-UAV dataset. These results significantly outperform existing YOLO series models and Transformer models, such as RT-DETR. The verification scheme of strong robustness across various dataset. The model’s performance is systematically evaluated on complex scene dataset, including SIDD and HIT-UAV. Additionally, a dynamic label allocation strategy and a multi-scale loss function for infrared small targets are proposed. These innovations effectively address the issues of false detection and missed detection that are prevalent in traditional methods, particularly under conditions of strong noise and low contrast. Furthermore, the model exhibits enhanced adaptability in practical scenarios, such as UAV surveillance. The verification scheme of strong robustness across various dataset. The model’s performance is systematically evaluated on complex scene dataset, including SIDD and HIT-UAV. Additionally, a dynamic label allocation strategy and a multi-scale loss function for infrared small targets are proposed. These innovations effectively address the issues of false detection and missed detection that are prevalent in traditional methods, particularly under conditions of strong noise and low contrast. Furthermore, the model exhibits enhanced adaptability in practical scenarios, such as UAV surveillance. Materials and methods The model is trained, validated, and tested on the SIDD dataset29 and HIT-UAV dataset30 without utilizing pre-trained weights. The experimental results demonstrate the advantages of IR-WSANet in IRSTD tasks. Datasets SIDD dataset The SIDD dataset was derived from the Anti-UAV benchmark31, which consists of 318 pairs of high-quality RGB-T videos (25 fps) captured by various drone models (mainly DJI and Parrot) under diverse illumination and background conditions. For this study, we extracted only the infrared modality and categorized the IR images into four scene types: city, mountain, ocean, and aerial. We focus on two representative scenarios—city (dense man-made structures) and mountain (comprising forests, slopes, and sparse infrastructure). The dataset division follows the Anti-UAV dataset standards, with training and validation sets drawn from non-overlapping segments of the same videos, while the test set is completely independent32. The mountain subset contains 2,146 images featuring extremely small targets, with a median bounding box size of approximately 15 × 9 pixels and a median area of 140 pixels². In contrast, the city subset comprises 1,088 images with relatively larger targets, showing a median bounding box size of about 31 × 18 pixels and a median area of approximately 589 pixels². To prevent data leakage across splits, all frames from the same video sequence are assigned to a single subset. For both subsets, the train, validation, and test splits follow an 8:1:1 ratio (city: 870:109:109; mountain: 1,716:215:215), with an average of one annotated target per image. These statistics confirm that SIDD predominantly represents a small-object detection scenario, which directly motivates the model design choices and evaluation protocol adopted in this study. Representative examples of the city and mountain subsets are shown in Fig. 1. Fig. 1The SIDD dataset: (a) City environment in the SIDD dataset; (b) Mountain environment in the SIDD dataset. The SIDD dataset: (a) City environment in the SIDD dataset; (b) Mountain environment in the SIDD dataset. HIT-UAV dataset The HIT-UAV dataset is a collection of high-altitude infrared thermal image sets from unmanned aerial vehicles. The data covers various locations, including schools, parking lots, roads, and amusement parks, reflecting the infrared imaging characteristics in complex real-world environments. It includes people, bicycles, cars, and other types of vehicles to support research on diverse object recognition tasks12. During data collection, the drone operated at altitudes ranging from 60 to 130 m to simulate different detection conditions. The camera’s field of view was varied between 30 and 90 degrees to investigate the impact of different fields of view on target detection, with images taken during the day and at night30. Evaluating algorithm performance across varying light intensities has expanded the drone’s application range in low-altitude flight conditions. The HIT-UAV dataset includes 2,866 infrared images and 24,899 annotated targets, with a data segmentation ratio of 8:1:1, partitioned into train: val: test sets of 2,292:287:287 images. The train set contains 17,628 targets (8.78 per image), while the validation and test sets show similar densities (8.6 and 8.4 per image, respectively). The median bounding-box width and height are 20 × 24 pixels, respectively, with a median area of 459 pixels², indicating substantial scale diversity. The content of the HIT-UAV datasets is shown in Fig. 2. Fig. 2The HIT-UAV dataset. The HIT-UAV dataset. Research method In this chapter, we give a detailed overview of the overall structure of IR-WSANet and related improvement modules. Overall network structure In IR-WSANet, the network integrates discrete wavelet transform convolution, a single-head self-attention mechanism, and position-aware coding convolution to improve the performance of small-target detection in infrared images. In the IR-WSANet architecture, we combine filtering with deep learning, moving away from traditional attention mechanisms. Instead, we innovatively combine a single-head self-attention mechanism with position aware convolution, thereby enhancing feature extraction and target detection in infrared images while maintaining a low parameter count. The overall architecture is shown in Fig. 3. The design of IR-WSANet is based on the classic YOLO series and consists of three components: Backbone, Neck, and Head33. It effectively fuses multi-scale features while remaining lightweight, resulting in a 2.8% improvement in the mean Average Precision (mAP) of the IR-WSANet network on the same dataset, achieved through an enhanced channel-fusion mechanism. In the backbone network of IR-WSANet, we introduce a Haar wavelet transform layer to replace the original standard convolution layer. This layer divides the input features into four sub-bands (LL, LH, HL, HH) using discrete wavelet transform34(DWT) decomposition. The low-frequency components in the LL sub-band are enhanced through a 3 × 3 convolution, while the high-frequency components in the HH sub-band are preserved via channel splicing35,36. Through the wavelet transform and applying filtering and denoising techniques to the four sub-bands, we further enhance the feature extraction capabilities of the backbone network for infrared images. Fig. 3The IR-WSANet overall network structure. The IR-WSANet overall network structure. In the head structure of IR-WSANet, a spatial-channel dual-feature fusion mechanism is proposed, which includes POSConvEmbedding and a single-head self-attention mechanism (SHSA)37. POSConvEmbedding utilizes depth-separable convolution to construct a position-sensitive feature map. After the backbone network enhances the overall feature information, it further strengthens the network’s ability to capture local spatial information and retain more detailed data. This enhancement is particularly crucial for small target detection tasks. The SHSA module realizes cross-channel attention through GroupNorm and dynamic weight allocation, applying self-attention to select channels. This approach reduces computational costs while compensating for the model’s ability to capture global features, effectively combining local feature extraction with global modeling capabilities. The global attention mechanism and local position coding create a two-path output through adaptive weighted fusion, expressed by the following formula:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{{out}} = \alpha F_{{POS}} + \left( {1 - \alpha } \right)F_{{SHSA}}$$\end{document} Among these \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\upalpha\:}$$\end{document}are learnable parameters that dynamically adjust the ratio of positional encoding to channel attention. IR-WSANet differs from the original YOLOv10 architecture33. We have constructed a four-level feature pyramid structure and realized feature fusion through upsampling and cross-layer connection (Concat)38. The low-level features (160 × 160) contain rich texture details, the middle-level features (80 × 80) fuse spatial semantic information, and the high-level features (40 × 40) carry global context information. A Pyramid Spatial Attention (PSA) module is introduced for feature recalibration, and a multi-scale attention map is generated using spatial pyramid pooling (SPPF). By combining multi-scale features and optimizing the attention mechanism33, the IR-WSANet network can efficiently handle small-target detection tasks and has significant practical engineering value. It offers an effective, robust solution for small-target detection in infrared images. DWaveletConv In IR-WSANet, we address the common problems of low resolution, low contrast, and noise interference in IRSTD tasks, inspired by reference34, we introduce discrete wavelet transform convolution into the Backbone structure of IR-WSANet, is shown in Fig. 4. Fig. 4Structure diagram of discrete wavelet transform. Structure diagram of discrete wavelet transform. To provide a systematic theoretical description of the DWaveletConv module, we briefly review the principles of multi-resolution analysis and the discrete wavelet transform (DWT). The wavelet transform is based on a set of scaling functions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varphi\:\left(t\right)\:$$\end{document}and wavelet functions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\psi\:\left(t\right)\:$$\end{document}, which form an orthogonal basis by translation and dilation:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi _{{j,k}} \left( t \right) = 2^{{j/2}} \varphi \left( {2^{j} t - k} \right)$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\psi _{{j,k}} \left( t \right) = 2^{{j/2}} \psi \left( {2^{j} t - k} \right)$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{j}\:$$\end{document}denotes the scale and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{k}\:$$\end{document}denotes the translation index. For a continuous signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{f}\left(\mathrm{t}\right)$$\end{document}, its wavelet coefficients are defined as the inner product \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{W}}_{\mathrm{f}}\left(\mathrm{a},\mathrm{b}\right)=〈\mathrm{f},{{\uppsi\:}}_{\mathrm{a},\mathrm{b}}〉\:$$\end{document}. In digital image processing, the DWT is realized through discrete filter banks according to Mallat’s multi-resolution analysis framework35. The Haar wavelet basis is used to construct a multi-scale feature analysis framework that enhances the network’s ability to characterize infrared targets by fusing frequency-domain and spatial features. This approach allows the decomposition of signals across different scales, thereby capturing both coarse (low-frequency) and fine (high-frequency) structures. The processing results are shown in Fig. 5. For discrete signals (such as digital images), wavelet decomposition is implemented by multi-stage low-pass and high-pass filtering followed by downsampling. The discrete approximation and detail coefficients at scale \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{j}\:$$\end{document}can be written in terms of inner products with discrete scaling and wavelet atoms, or equivalently by filter-bank operations. Using the filter-bank notation, the one-dimensional updating equations are:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a_{{j + 1}} ~\left[ n \right] = \mathop \sum \limits_{k} h\left[ k \right]a_{j} \left[ {2n - k} \right]$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{{j + 1}} ~\left[ n \right] = \mathop \sum \limits_{k} g\left[ k \right]a_{j} \left[ {2n - k} \right]$$\end{document} Where\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:h\left[k\right]\:$$\end{document}and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:g\left[k\right]\:$$\end{document}denote the low-pass (scaling) and high-pass (wavelet) analysis filters, respectively;\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{{a}_{j}\right\}\:$$\end{document}are approximation coefficients and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{{d}_{j}\right\}\:$$\end{document}are detail coefficients. Fig. 5Visualization of wavelet processing. Visualization of wavelet processing. These coefficients are generated by convolving the signal with a filter, which decomposes the signal into frequency bands with varying frequencies. In the frequency domain, the input characteristics are decomposed into low-frequency (LL) and high-frequency (LH, HL, HH) components using a one-dimensional discrete wavelet transform (DWT). The expression is as follows:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{{LL}} = \left( {\left( {X*h_{r} } \right)_{{ \downarrow 2}} *h_{c}^{T} } \right)_{{ \downarrow 2}}$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{{LH}} = \left( {\left( {X*h_{r} } \right)_{{ \downarrow 2}} *g_{c}^{T} } \right)_{{ \downarrow 2}}$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{{HL}} = \left( {\left( {X*g_{r} } \right)_{{ \downarrow 2}} *h_{c}^{T} } \right)_{{ \downarrow 2}}$$\end{document}9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{{HH}} = \left( {\left( {X*g_{r} } \right)_{{ \downarrow 2}} *g_{c}^{T} } \right)_{{ \downarrow 2}}$$\end{document} Where\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\:\mathrm{*}\:\:$$\end{document}denotes 1-D (or separable 2-D) convolution and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\:\downarrow\:2\:\:$$\end{document}indicates downsampling by a factor of two in each direction. The LL subband captures coarse (low-frequency) content, while LH, HL, HH capture horizontal, vertical and diagonal detail (high-frequency) information respectively.The low-frequency component LL carries the primary energy information of the feature, while the high-frequency component contains the edge texture details. Channel concatenation is used to fuse the features across the four subbands in the channel dimension, and a 3 × 3 convolution kernel is used to interact across channels. Through multi-level decomposition, DWT divides the signal into multiple frequency levels, facilitating the analysis of target characteristics at different scales. This decomposition method provides a foundation for subsequent signal processing, particularly in infrared small-target detection. To address the high-frequency noise in infrared images, redundant information in the detail subband can be eliminated to enhance the clarity of the target signal. In our DWaveletConv module, we adopt the Haar wavelet (db1) due to its simplicity and computational efficiency. The time-domain forms of the Haar scaling (father) function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\upvarphi\:}\left(\mathrm{t}\right)\:$$\end{document}and wavelet (mother) function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\uppsi\:}\left(\mathrm{t}\right)\:$$\end{document}are:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi \left( t \right) = \left\{ {\begin{array}{*{20}c} {1,~0 \le t < 1} \\ {0,~otherwise} \\ \end{array} } \right.$$\end{document}11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\psi \left( t \right) = \left\{ {\begin{array}{*{20}c} {1,~0 \le t < \frac{1}{2}} \\ { - 1,~\frac{1}{2} \le t < 1} \\ {0,~otherwise} \\ \end{array} } \right.$$\end{document} The corresponding normalized analysis filter coefficients are:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h = \frac{1}{{\sqrt 2 }}\left[ {1,{{~}}1} \right]$$\end{document}13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g = \frac{1}{{\sqrt 2 }}\left[ {1, - 1} \right]$$\end{document} These filters satisfy orthogonality and perfect reconstruction conditions in the ideal (lossless) case. In the convolutional implementation, we perform a single-level DWT on the input feature map to obtain\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{LL},{X}_{LH},{X}_{HL},{X}_{HH}$$\end{document}.Each subband is processed by a lightweight convolutional block \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{LL}{\mathcal{F}}_{\boldsymbol{\cdot\:}}\left(\cdot\:\right)$$\end{document}(e.g., 3 × 3 conv + BN + activation). The processed subbands are then fused by channel concatenation and projection:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y = g\left( {\omega _{{LL}} {\mathcal{F}}_{{LL}} \left( {X_{{LL}} } \right)~\left| {\left| {~\omega _{{LH}} {\mathcal{F}}_{{LH}} \left( {X_{{LH}} } \right)~} \right|} \right|~\omega _{{HL}} {\mathcal{F}}_{{HL}} \left( {X_{{HL}} } \right)~||~\omega _{{HH}} {\mathcal{F}}_{{HH}} \left( {X_{{HH}} } \right)~} \right)$$\end{document} Where\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\left|\right|\:$$\end{document}denotes concatenation and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\:\omega\:}_{\boldsymbol{\cdot\:}\:}$$\end{document}are optional learnable scalar weights for adaptive emphasis. To address high-frequency noise commonly present in infrared images, we apply a soft-thresholding (or lightweight channel-wise gating) to the detail subbands prior to fusion. The soft-threshold operator is:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{d} = sign\left( d \right) \cdot max\left( {\left| d \right| - \lambda ,0} \right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\uplambda\:}\:\:$$\end{document}is a threshold that can be set empirically or learned during training (Donoho’s wavelet denoising theory provides theoretical justification)32. Through this operation and multi-level decomposition, DWaveletConv enhances edge features and target details while suppressing redundant high-frequency noise. The wavelet transform therefore plays a crucial role in edge detection and feature extraction for infrared images: low-frequency bands preserve the target’s primary energy and contrast, while detail bands concentrate edge and texture cues that are important for precise localization and shape characterization. Channel concatenation and subsequent 3 × 3 cross-channel convolution enable interaction among subbands and integration of multi-scale information, providing a robust foundation for downstream detection modules. POSConvembedding In the research of IRSTD, we address the problems of blurred spatial positioning and uneven thermal radiation distribution in infrared images by proposing a position-aware coded convolution. The POSConvEmbedding can capture hotspots in infrared images, compensating for the network’s limitations in capturing local features. To provide the network with explicit spatial position information, we adopt two complementary approaches: (1) fixed sinusoidal positional encodings in the Transformer architecture, and (2) learnable convolutional positional embeddings (or CoordConv-style coordinate channels) that inject spatial coordinates into feature maps. This combination enables the model to better locate small, sparse infrared targets, whose spatial extent often spans only a few pixels. The core concept of positional encoding is to assign a unique vector to each spatial location and to integrate it with the input features. The classic sinusoidal positional encoding for a single position index POSConv (e.g., a row or column index) and an embedding dimension (assumed even) is:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{PE}}_{{POS,~2i}} = sin\left( {\frac{{{\mathrm{POS}}}}{{10000^{{2i/d}} }}} \right)$$\end{document}17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$PE_{{POS,~2i + 1}} = cos\left( {\frac{{POS}}{{10000^{{2i/d}} }}} \right)$$\end{document} For \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i=0,\cdots\:,\frac{d}{2}-1$$\end{document}. For two-dimensional positions\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\left(x,y\right)\:$$\end{document}, a standard construction is to compute separate encodings for\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\:x\:\:$$\end{document}and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\:\mathrm{y}\:\:$$\end{document}and concatenate them along the channel dimension:18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$PE\left( {x,y} \right) = \left[ {PE_{x} \left( x \right)~||~PE_{y} \left( y \right)} \right]$$\end{document} where\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\left|\right|\:$$\end{document}denotes concatenation. This scheme provides multi-scale (multi-frequency) positional cues: low-index dimensions encode slowly varying (coarse) position, and high-index dimensions encode fine-grained periodic position behavior. An alternative and often more practical method for vision tasks is learnable positional embedding:19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{{coord}} \left( {h,~\omega } \right) = \frac{\omega }{{W - 1}}$$\end{document}20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y_{{coord}} \left( {h,~\omega } \right) = \frac{h}{{H - 1}}$$\end{document} For each pixel location, where the image width is\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:W$$\end{document}and the height is H, these two coordinate maps are expanded to the same spatial size as the feature map and concatenated as two extra channels. The network then learns a position-aware representation through a POSConv layer:21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z = Conv_{{3 \times 3}} \left( {\left[ {X,~\alpha \cdot PE\left( {x,~y} \right),~\beta \cdot \left( {X_{{coord}} ,~Y_{{coord}} } \right)} \right]} \right)$$\end{document} Where X is the input feature map,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\alpha\:$$\end{document} and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\:\beta\:\:$$\end{document}are scalar or vector scaling factors (either fixed hyperparameters or learnable parameters), and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\left[\bullet\:\right]\:$$\end{document}denotes channel-wise concatenation. The\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\:Conv}_{3\times\:3}\:$$\end{document}block is typically followed by BatchNorm and a non-linear activation, producing the position-enhanced features used by subsequent layers. When the input is in flattened/sequential form \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:X\in\:{\mathbb{R}}^{B\times\:N\times\:C}\:$$\end{document}with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N=H\cdot\:W$$\end{document}, we reconstruct the spatial 4D tensor via transposition and reshape:22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F = reshape\left( {X^{T} } \right) \in {\mathbb{R}}^{{B \times C \times H \times W}}$$\end{document} After forming the positional maps (PE map and coordinate maps) at the same spatial resolution, the POSConv embedding is computed as:23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{{POS}} = Conv_{{3 \times 3}} \left( {\left[ {F,~\alpha \cdot PE\_map,~\beta \cdot Coord\_map} \right]} \right) \in {\mathbb{R}}^{{B \times C^{t} \times H^{t} \times W^{t} }}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:H{\prime\:}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:W{\prime\:}\:$$\end{document}are the output spatial dimensions after Conv and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:C{\prime\:}\:\:$$\end{document}is the output channel count. In implementation, these position vectors can be integrated with the original infrared image data as additional channels or combined with the convolution output at a specific stage of the feature map. This explicit position encoding can help the model accurately locate small targets, particularly in scenarios with significant background interference or low target contrast, thereby enhancing detection accuracy and stability. POS single-head self-attention In the research on small-target detection in infrared images, we introduce a novel POS Single-Head Self-Attention (POS-SHSA) mechanism into IR-WSANet. The SHSA module, combined with POSConvEmbedding, is designed to improve both detection performance and computational efficiency in complex infrared scenes by integrating positional encoding with single-head self-attention. The detailed structure of the SHSA module37 is shown in Fig. 6. Fig. 6SHSA module structure diagram. SHSA module structure diagram. The design inspiration for the SHSA comes from optimizing traditional attention mechanisms for high-resolution, small-target infrared scenarios. Let the input after POSConvEmbedding be \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:X\in\:{\mathbb{R}}^{N\times\:C}\:$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N=H\cdot\:W\:\:$$\end{document}is the flattened spatial size and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:C$$\end{document} is the number of channels. We compute linear projections for queries, keys, and values by learnable matrices\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{W}_{Q},{\:W}_{K},\:{W}_{V}\in\:{\mathbb{R}}^{C\times\:d}\:$$\end{document}:24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Q = X \cdot W_{Q} ,~~K = X \cdot W_{K} ,V = X \cdot W_{V}$$\end{document} The attention weights and output are given by the standard scaled dot-product:25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Attention\left( X \right) = V \cdot softmax\left( {\frac{{QK^{T} }}{{\sqrt {d_{k} } }}} \right)$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{k}\:\:$$\end{document}is the key dimension used for scaling we adopt \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{k}={C}_{p}$$\end{document}.To reduce the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left({N}^{2}\right)\:\:$$\end{document}spatial cost of full-channel attention, we split the channels into an attention partition and a residual partition, preserving local-detail information while computing global interactions on a small subset:26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X = \left[ {X_{{att}} ,~X_{{res}} } \right],~X_{{att}} \in {\mathbb{R}}^{{N \times C_{p} }} ,~X_{{res}} \in {\mathbb{R}}^{{N \times \left( {C - C_{p} } \right)}}$$\end{document} Compute attention only on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{att}\:\:$$\end{document}:27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{Y}_{{att}} = Attention\left( {X_{{att}} } \right)$$\end{document} A projection\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:Proj\left(\cdot\:\right)\:$$\end{document}(implemented as a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\times\:1\:$$\end{document}convolution) maps \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\sim}{Y}}_{att}\:$$\end{document}back to the target channel space and fuses with the residual channels. To let the network adaptively decide how much global information to incorporate, we use a learnable gate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:g\:$$\end{document}(scalar or per-channel) with sigmoid activation and a final \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\times\:1\:$$\end{document}conv for channel fusion:28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g = \sigma \left( \gamma \right),~Y = Conv_{{1 \times 1}} \left( {g~ \cdot ~Conact\left( {Proj\left( {\widetilde{{Y_{{att}} }}} \right),~X_{{res}} } \right)} \right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\upgamma\:}\:\:$$\end{document}is a learnable parameter (initialized near 0 so \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{g}={\upsigma\:}\left(0\right)=0.5\:$$\end{document}initially), ⊙denotes element-wise multiplication, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{a}\mathrm{c}\mathrm{t}\left(\cdot\:\right)\:\:$$\end{document}concatenates along the channel axis. This architecture preserves local information (through \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{res}\:$$\end{document}) while introducing selective global context (through \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\sim}{Y}}_{att}\:\:$$\end{document}and g); analytically the dominant attention cost is reduced from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left({N}^{2}d\right)\:$$\end{document}(full-channel) to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left({N}^{2}{C}_{p}\right)\:\:$$\end{document}(partial-channel), and including projection/fusion yields an overall additional cost on the order of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left({N}^{2}{C}_{p}+NC{C}_{p}\right)\:\:$$\end{document}, which is substantially smaller when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{C}_{p}\ll\:C,{\:C}_{p}=max(16,C/4)\:$$\end{document}. In addition, the optimization of SHSA in computing efficiency and storage requirements warrants significant attention. In the traditional Multi-Head Self-Attention (MHSA) mechanism, operations such as Layer Normalization and data reshaping often consume substantial running time. These memory-intensive processes hinder the model’s execution efficiency on hardware. In contrast to traditional, SHSA is shown in Fig. 7, reduces the frequency and scope of these operations by confining attention computation to specific channels, thereby significantly decreasing memory access overhead39–42. This optimization enables SHSA to fully leverage the computational resources of GPU or CPU, particularly when processing high-resolution infrared images. The parameter quantities are presented in Table 1. Fig. 7Comparison of attention channels: (a) Full-channel attention mechanism ; (b) Partial-channel attention mechanism. Comparison of attention channels: (a) Full-channel attention mechanism ; (b) Partial-channel attention mechanism. Table 1Comparison of parameters under different channels.Single-Head AttentionParams(M)GPU(images/s)Top-1(%)(a) Full channel10.472475.1(b) Partial channel11.495175.21Tables may have a footer. Table 1Comparison of parameters under different channels.Single-Head AttentionParams(M)GPU(images/s)Top-1(%)(a) Full channel10.472475.1(b) Partial channel11.495175.21Tables may have a footer. Comparison of parameters under different channels. 1Tables may have a footer. SHSA captures long-range dependencies in images via an attention mechanism, enabling the model to comprehend the overall association between the target and the background while preserving the original channel information. This preservation ensures that the detailed features of small targets are not overlooked. The ability to consider both global and local features enables SHSA to excel at localization and recognition of small targets. In the POS-SHSA module, POSConvEmbedding is used as a position-coded convolution to model the spatial positioning of adjustments. The complete structure of the POS-SHSA module is shown in Fig. 8. This approach introduces positional information via lightweight convolution, addressing the lack of spatial awareness in self-attention. The output consists of enhanced features that maintain the same size as the original feature map. The SHSA module conducts single-head self-attention calculations based on these enhanced features, ultimately integrating global context with local positional information to form position-aware self-attention. The POS-to-SHSA serial connection works as follows: First, POSConv produces \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{POS}\in\:{\mathbb{R}}^{B\times\:C\times\:H\times\:W}\:$$\end{document} that encodes absolute and relative position cues. Second, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{POS}$$\end{document} is flattened to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{X}\in\:{\mathbb{R}}^{N\times\:C}\:$$\end{document} and split into \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{att}\:\:$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{res}$$\end{document}; Third, SHSA computes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\sim}{Y}}_{att}\:\:$$\end{document}on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{att}\:$$\end{document}, projects and gates the result, and fuses it with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{res\:}$$\end{document}to produce\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\:Y\:$$\end{document}of the same spatial size as the input feature map. This lightweight convolution-to-attention pipeline addresses the spatial ambiguity inherent in attention mechanisms while maintaining spatial resolution and limiting memory overhead. This is particularly important for high-resolution IRSTD scenes, where preserving tiny targets and fine textures is critical. Fig. 8Structure diagram of POS-SHSA module. Structure diagram of POS-SHSA module. Results In this chapter, we conduct a series of experiments to evaluate the model’s performance. We use detection accuracy, parameter size, and inference time as evaluation metrics. The model is trained, validated, and tested on the SIDD and HIT-UAV dataset without utilizing pre-trained weights. The experimental results demonstrate the advantages of IR-WSANet in IRSTD tasks. The verification experiments are divided into two types: comparison and ablation. The comparison experiments assess the overall effectiveness of the algorithm, while the ablation tests evaluate the contribution of each module. Experimental parameter settings The hardware configuration for the experiment includes a GPU (RTX3060 12 GB), CPU (Xeon (R) E5-2630 v4), 32 GB RAM, PyTorch 2.4.0 as the deep learning framework, and Python version 3.11. (1) SIDD Dataset: In the mountainous and city scenes of the SIDD dataset, we use uniform experimental parameters.The Adam optimizer is used in the training process, which consists of 200 epochs. The learning rate is set to 0.001, and the random gradient descent (SGD) momentum is set to 0.937, a value that ranges from 0 to 1 and influences the current update based on the previous gradient. The input image size is normalized to 640 × 640 pixels, and the batch size is set to 16. Detailed parameter settings are shown in Table 2: Table 2SIDD Dataset Parameter Settings.ParametersSetupEpochs200Batch Size16Imgsize640 × 640Learning Rate0.001OptimizerAdamMomentum0.937(SGD) Table 2SIDD Dataset Parameter Settings.ParametersSetupEpochs200Batch Size16Imgsize640 × 640Learning Rate0.001OptimizerAdamMomentum0.937(SGD) SIDD Dataset Parameter Settings. (2) HIT-UAV Dataset: In the HIT-UAV dataset, the parameter settings for model training are as follows. The AdamW optimizer is used to train for 300 epochs, with a learning rate set at 0.0001. The momentum of random gradient descent (SGD) is set to 0.9, the input image size is normalized to 640 × 640 pixels, and the batch size is set to 16. Detailed parameter settings are presented in Table 3. Table 3HIT-UAV Dataset Parameter Settings.ParametersSetupEpochs300Batch Size16Imgsize640 × 640Learning Rate0.0001OptimizerAdamWMomentum0.9(SGD) Table 3HIT-UAV Dataset Parameter Settings.ParametersSetupEpochs300Batch Size16Imgsize640 × 640Learning Rate0.0001OptimizerAdamWMomentum0.9(SGD) HIT-UAV Dataset Parameter Settings. Experimental Indicators In our research experiment based on the IRSTD task, we introduce comprehensive model evaluation metrics to evaluate the performance of the model. These metrics include Precision (P), Recall (R), F1-Score, Average Precision (AP), Mean Average Precision (mAP), Parameters, GFLOPs, and FPS. In the context of classification tasks, True Positive (TP), False Positive (FP), and False Negative (FN) are used to describe the samples correctly identified by the model, the samples incorrectly labeled as positive classes, and the samples that were not identified as positive classes, respectively. Precision (P) reflects the proportion of true positive results predicted by the model, and its formula is43:29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{Precision}} = \frac{{{\mathrm{TP}}}}{{{\mathrm{TP}} + {\mathrm{FP}}}}$$\end{document} The recall rate (R) represents the proportion of positive samples successfully found by the model out of all actual positive samples. The calculation method is as follows:30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}$$\end{document} The F1-Score is regarded as a comprehensive measure that considers both precision and recall rates. It utilizes the harmonic average of these two metrics to assess the model’s ability to balance them effectively. The formula is as follows:31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{F}}1 - {\mathrm{Score}} = \frac{{2 \times {\mathrm{P}} \times {\mathrm{R}}}}{{{\mathrm{P}} + {\mathrm{R}}}} = \frac{{2{\mathrm{TP}}}}{{2{\mathrm{TP}} + {\mathrm{FP}} + {\mathrm{FN}}}}$$\end{document} This index is particularly crucial for small target detection, as missed or false detections can significantly impact the overall effectiveness of the detection process. In addition, average precision (AP)44 and mean average precision (mAP) are higher-level metrics used to measure the model’s detection capability. For a specific category, AP indicates the model’s performance across various confidence thresholds by calculating the area under the precision-recall curve. mAP is the average of AP across all categories and is utilized to evaluate the overall detection accuracy of the model in multi-category tasks. The formula for its calculation is as follows:32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{mAP}} = \frac{{\mathop \sum \nolimits_{{{\mathrm{i}} = 1}}^{{\mathrm{N}}} {\mathrm{AP}}_{{\mathrm{i}}} }}{{\mathrm{N}}}$$\end{document}33\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{AP}} = \mathop \smallint \limits_{0}^{1} {\mathrm{P}}\left( {\mathrm{r}} \right){\mathrm{~dr}}$$\end{document} In this context, N represent the total number of detection categories, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{A}\mathrm{P}}_{\mathrm{i}}\:$$\end{document}denotes the average precision of the first category. In the realm of small target detection within infrared images, the mAP effectively demonstrates the model’s ability to recognize various types of small targets. To further evaluate the computational cost of the model and the feasibility of its deployment, this paper uses parameter counts and GFLOPs as metrics. The parameter count refers to the total number of trainable parameters in the model, typically expressed in millions. A lower parameter count indicates a lighter model with reduced resource consumption. These metrics are particularly crucial in infrared detection tasks that demand high real-time performance. Additionally, FPS (frames per second) serves as an intuitive measure of reasoning speed, quantifying the number of images the model can process per second and is directly related to its responsiveness in real-world scenarios. Ablation experiment In this chapter, we conducted a series of ablation experiments to assess the contribution of each enhancement module to the model’s performance on IRSTD tasks. We sequentially integrated each module into the baseline YOLOv10s model. Based on the experimental parameter settings outlined in Table 3, we compared the results of the ablation experiments using the HIT-UAV dataset. The comparative results of the two new fusion modules—the wavelet transform DWaveletConv module and the POS-SHSA module—in the IR-WSANet model are presented in Table 4. Table 4Comparison of ablation experiments with HIT-UAV dataset.DWaveletConvPOS-SHSAPOSSHSA\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5}$$\end{document}(%)F1(%)GFLOPsParamsBaseline72.471.824.77.2√79.880.327.57.9√80.681.124.47.1√78.979.424.57.1√79.480.124.17.0√√82.682.427.98.0 Table 4Comparison of ablation experiments with HIT-UAV dataset.DWaveletConvPOS-SHSAPOSSHSA\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5}$$\end{document}(%)F1(%)GFLOPsParamsBaseline72.471.824.77.2√79.880.327.57.9√80.681.124.47.1√78.979.424.57.1√79.480.124.17.0√√82.682.427.98.0 Comparison of ablation experiments with HIT-UAV dataset. In the ablation study on the HIT-UAV dataset (Table 4; Fig. 9), we evaluate the contributions of DWaveletConv and POS-SHSA. Compared with the baseline (mAP 72.4%, F1 71.8%, 24.7 GFLOPs, 7.2 M params), DWaveletConv alone raises mAP to 79.8% (+ 10.2% relative) and F1 to 80.3% (+ 11.8% relative) at the cost of increased computation (27.5 GFLOPs, + 11.3%) and parameters (7.9 M). POS-SHSA alone improves mAP to 80.6% (+ 11.3%) and F1 to 81.1% (+ 13.0%), while slightly reducing computation (24.4 GFLOPs) and keeping the model lightweight (7.1 M params). Enabling only the POS or only the SHSA branch produces moderate gains (mAP 78.9% and 79.4%, respectively), indicating limited effectiveness when used in isolation. The full IR-WSANet, combining DWaveletConv and POS-SHSA, achieves the best performance (mAP 82.6%, F1 82.4%), with a modest increase in complexity to 27.9 GFLOPs (+ 13.0%) and 8.0 M parameters. This demonstrates a favorable trade-off between accuracy and cost for infrared small-target detection. Fig. 9Ablation module performance comparison chart. Ablation module performance comparison chart. Contrast experiment To validate the comprehensive performance of the improved model for the infrared small-target detection task, we compare it with current mainstream detection algorithms across three datasets: SIDD-City, SIDD-Mountain, and HIT-UAV. All models are trained from scratch, with an input resolution of 640 × 640. The experimental metrics include detection accuracy (Precision, Recall, AP, mAP, F1), computational complexity (GFLOPs), and real-time performance (FPS). SIDD-city experiment In Table 5, we show the model performance on the SIDD-City dataset. The improved model proposed in this paper outperforms the baselines on multiple metrics. Specifically, IR-WSANet attains a Precision of 99.8% and a Recall of 94.4%, representing a 1.0% improvement in Recall over YOLOv8 (Recall 93.4%). In terms of detection accuracy, IR-WSANet achieves \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5}:\:97.2{\%}\:$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5:0.95}:76.4{\%}$$\end{document}, which are higher than those of YOLOv10 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5}:\:94.7{\%}\:$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5:0.95}:\:75.2{\%}$$\end{document}) and exceeding YOLOv8’s \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5:0.95}:\:75.7{\%}\:$$\end{document}. The F1 score of the proposed model reaches 97.6%, surpassing all baseline methods and indicating improved balance between missed detections and false positives. Regarding efficiency, IR-WSANet has a GFLOPs of 27.9, which is comparable to YOLOv10 (24.8 GFLOPs). Despite a slightly higher computational cost, IR-WSANet achieves higher inference speed (42.8 FPS vs. 41.9 for YOLOv10), validating the method’s effective trade-off between accuracy and runtime. The parameter curves for the evaluated algorithms are shown in Fig. 10. Table 5Comparison of SIDD-city experiments.ModelPrecisionRecall \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5:0.95}$$\end{document} F1GFLOPsFPSParamsSSD96.782.293.474.188.931.432.426.3Faster R-CNN95.580.492.372.387.391.536.9137.1YOLO-v799.492.994.774.296.0105.252.437.2YOLO-v899.893.496.675.796.928.675.611.1YOLO-v1099.593.594.775.296.624.841.97.2YOLO-v1199.794.296.176.296.921.534.79.4YOLO-v1299.793.596.376.096.521.439.69.1YOLOE-v899.893.496.675.796.928.631.412.9YOLOE-v1199.794.296.176.296.721.531.011.0EfficientDet-D399.593.796.376.496.525.237.812.0RT-DETR99.494.193.974.596.757.433.632.0IR-WSANet99.894.497.276.497.627.942.88.0 Table 5Comparison of SIDD-city experiments.ModelPrecisionRecall \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5:0.95}$$\end{document} F1GFLOPsFPSParamsSSD96.782.293.474.188.931.432.426.3Faster R-CNN95.580.492.372.387.391.536.9137.1YOLO-v799.492.994.774.296.0105.252.437.2YOLO-v899.893.496.675.796.928.675.611.1YOLO-v1099.593.594.775.296.624.841.97.2YOLO-v1199.794.296.176.296.921.534.79.4YOLO-v1299.793.596.376.096.521.439.69.1YOLOE-v899.893.496.675.796.928.631.412.9YOLOE-v1199.794.296.176.296.721.531.011.0EfficientDet-D399.593.796.376.496.525.237.812.0RT-DETR99.494.193.974.596.757.433.632.0IR-WSANet99.894.497.276.497.627.942.88.0 Comparison of SIDD-city experiments. Fig. 10SIDD-city dataset model performance. SIDD-city dataset model performance. SIDD-mountain experiment In Table 6, we show the model performance on the SIDD-Mountain dataset. The proposed IR-WSANet achieves a Precision of 88.1% and a Recall of 74.8%. Compared with YOLOv11 (Precision 87.1%, Recall 73.6%), IR-WSANet attains a higher Precision (+ 1.0%) while its Recall is marginally higher (+ 1.2%). In terms of detection accuracy, IR-WSANet yields \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5}:$$\end{document}82.6% and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5:0.95}:$$\end{document}49.2%, outperforming YOLOv10 (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5}:$$\end{document}80.6%, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5:0.95}:$$\end{document} 48.3%) by + 2.0% and + 0.9%, respectively. The F1 score of IR-WSANet reaches 80.4%, which is + 0.9% higher than RT-DETR (79.5%), indicating a better balance between missed detections and false positives in complex mountainous scenes. Importantly, IR-WSANet achieves a higher real-time throughput (42.8 FPS) than RT-DETR (33.6 FPS), demonstrating superior runtime performance in challenging backgrounds. The parameter curves for the evaluated algorithms are shown in Fig. 11. Overall, these results indicate that IR-WSANet attains improved detection accuracy for mountain scenarios while maintaining competitive computational efficiency and runtime performance. Table 6Comparison of SIDD-mountain experiment.ModelPrecisionRecall \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5:0.95}$$\end{document} F1GFLOPsFPSParamsSSD84.566.379.447.674.331.432.426.3Faster R-CNN83.469.277.646.475.691.536.9137.1YOLO-v786.572.178.547.778.7105.252.437.2YOLO-v887.775.380.748.480.928.675.611.1YOLO-v1086.973.280.648.379.824.841.97.2YOLO-v1187.173.681.848.980.121.534.79.4YOLO-v1287.373.481.648.879.821.439.69.1YOLOE-v887.774.081.248.480.328.631.412.9YOLOE-v1187.973.681.848.980.121.531.011.0EfficientDet-D387.074.282.149.180.225.237.812.0RT-DETR87.572.881.348.779.557.433.632.0IR-WSANet88.174.882.649.280.427.942.88.0 Table 6Comparison of SIDD-mountain experiment.ModelPrecisionRecall \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5:0.95}$$\end{document} F1GFLOPsFPSParamsSSD84.566.379.447.674.331.432.426.3Faster R-CNN83.469.277.646.475.691.536.9137.1YOLO-v786.572.178.547.778.7105.252.437.2YOLO-v887.775.380.748.480.928.675.611.1YOLO-v1086.973.280.648.379.824.841.97.2YOLO-v1187.173.681.848.980.121.534.79.4YOLO-v1287.373.481.648.879.821.439.69.1YOLOE-v887.774.081.248.480.328.631.412.9YOLOE-v1187.973.681.848.980.121.531.011.0EfficientDet-D387.074.282.149.180.225.237.812.0RT-DETR87.572.881.348.779.557.433.632.0IR-WSANet88.174.882.649.280.427.942.88.0 Comparison of SIDD-mountain experiment. Fig. 11SIDD-mountain dataset model performance. SIDD-mountain dataset model performance. HIT-UAV experiment The improved model, when applied to the HIT-UAV infrared dataset, further demonstrates its multi-target detection capabilities (Table 7). The model achieves an accuracy rate of 89.4% (Precision) and a Recall rate of 77.8%, outperforming PARE-YOLO, which has a Precision of 89.2% and a Recall of 78.0%. Additionally, the mAP scores of the model are significantly higher than those of YOLOv10, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5}:\:$$\end{document}82.8% and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5:0.95}$$\end{document}: 54.8%, compared to YOLOv10’s \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5}$$\end{document}: 72.7% and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5:0.95}$$\end{document}: 48.3%. The F1 score of 83.2% is 2.1% higher than PARE-YOLO’s 81.1%, indicating that the model exhibits superior overall performance in complex multi-objective scenarios. Furthermore, the model’s GFLOPs (27.9) are lower than those of PARE-YOLO (45.1), and its FPS rate of 42.8 is close to meeting real-time requirements, thereby validating the effectiveness of its lightweight design. Under the HIT-UAV dataset, the parameter curves for various algorithms are shown in Fig. 12. Table 7Comparison of HIT-UAV experiment.ModelPrecisionRecall \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5:0.95}$$\end{document} F1GFLOPsFPSParamsSSD81.468.370.943.574.231.432.426.3Faster R-CNN78.165.869.841.571.491.536.9137.1YOLO-v781.869.771.645.875.3105.252.437.2YOLO-v883.672.773.149.277.828.675.611.1YOLO-v1088.770.172.748.378.324.841.97.2YOLO-v1188.874.678.751.181.021.534.79.4YOLO-v1288.373.981.851.780.521.439.69.1YOLOE-v886.372.773.149.277.828.631.412.9YOLOE-v1188.874.678.751.181.221.531.011.0EfficientDet-D388.675.479.351.781.525.237.812.0PARE-YOLO4589.278.081.352.781.145.145.926.3IR-WSANet89.477.882.854.883.227.942.88.0 Table 7Comparison of HIT-UAV experiment.ModelPrecisionRecall \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5:0.95}$$\end{document} F1GFLOPsFPSParamsSSD81.468.370.943.574.231.432.426.3Faster R-CNN78.165.869.841.571.491.536.9137.1YOLO-v781.869.771.645.875.3105.252.437.2YOLO-v883.672.773.149.277.828.675.611.1YOLO-v1088.770.172.748.378.324.841.97.2YOLO-v1188.874.678.751.181.021.534.79.4YOLO-v1288.373.981.851.780.521.439.69.1YOLOE-v886.372.773.149.277.828.631.412.9YOLOE-v1188.874.678.751.181.221.531.011.0EfficientDet-D388.675.479.351.781.525.237.812.0PARE-YOLO4589.278.081.352.781.145.145.926.3IR-WSANet89.477.882.854.883.227.942.88.0 Comparison of HIT-UAV experiment. Fig. 12HIT-UAV dataset model performance. HIT-UAV dataset model performance. Table 8Detailed experimental comparison of HIT-UAV dataset.ModelPersonCarBicycleOther VehiclesDontCare \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5}$$\end{document} YOLO-v1088.995.981.164.731.772.7IR-WSANet92.598.591.487.843.382.8 Table 8Detailed experimental comparison of HIT-UAV dataset.ModelPersonCarBicycleOther VehiclesDontCare \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{m}{\mathbf{A}\mathbf{P}}_{0.5}$$\end{document} YOLO-v1088.995.981.164.731.772.7IR-WSANet92.598.591.487.843.382.8 Detailed experimental comparison of HIT-UAV dataset. Discussion Based on the comparative experiments presented above, the improved model proposed in this paper shows significant advantages in multiple dataset. The accuracy rates for the SIDD-City, SIDD-mountain, and HIT-UAV dataset are 76.4%, 49.2%, and 54.8%, respectively \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{m}\mathrm{A}\mathrm{P}}_{0.5:0.95}$$\end{document}, which represent an increase of 1.2% to 6.5% over the baseline model. In the lightweight IR-WSANet, GFLOPs (27.9) are significantly lower than those of YOLOv8 (28.6) and PARE-YOLO (45.1). Furthermore, with only 8.0 million parameters, it maintains a more compact model size compared to YOLOv8 (11.1 M) and is significantly smaller than PARE-YOLO (26.3 M). This balance between the number of parameters and computational efficiency demonstrates the model’s adaptability to complex scenes. Future work will focus on optimizing the model’s real-time performance (FPS) and on exploring a dynamic-resolution input strategy to further enhance its deployment capabilities. The comparison between the IR-WSANet and the baseline model YOLOv10 using each dataset is shown in Fig. 13. In Fig. 14, we show the detection performance of IR-WSANet compared to the baseline YOLOv10 on small targets in the IRSTD task, demonstrating its effectiveness. Fig. 13Comparison diagram of models for each dataset. Comparison diagram of models for each dataset. Fig. 14Visualization results of the model in SIDD-City, SIDD-Mountain, HIT-UAV dataset. Visualization results of the model in SIDD-City, SIDD-Mountain, HIT-UAV dataset. As shown in Fig. 14, IR-WSANet achieves an optimal balance of accuracy, efficiency, and speed. It performs exceptionally well on Precision, Recall, and mAP metrics while maintaining low computational requirements and high throughput. These attributes make it an ideal solution for detecting small and dense targets in IRSTD task scenarios that require real-time processing and limited computing resources. This paper compares the detection performance of IR-WSANet against a baseline model across three training datasets, as depicted in Fig. 14. The missed detection rate is significantly lower than that of the baseline model, YOLOv10. All targets that went undetected by the baseline model are successfully detected by IR-WSANet, and the distribution diagram from the model’s CAM46,47 is shown in Fig. 15. Overall, IR-WSANet shows excellent detection accuracy and robustness in infrared scenes, clearly outperforming YOLOv10 and other models, and achieving higher detection accuracy than the PARE-YOLO model referenced in45 on the HIT-UAV dataset. Fig. 15IR-WSANet CAM distribution. IR-WSANet CAM distribution. Conclusions In this paper, we propose an enhanced lightweight network, IR-WSANet, based on YOLOv10, that addresses the challenges of fuzzy features, noise interference, and localization difficulties in infrared small-target detection. By employing frequency domain decomposition and integrating a discrete wavelet transform convolution module (DWaveletConv), we effectively enhance the extraction of high-frequency detail features while minimizing the impact of background noise. Additionally, the collaborative design of POSConvEmbedding and a single-head self-attention mechanism (SHSA) enables efficient integration of local spatial information and global contextual features, thereby significantly improving the positioning accuracy of small targets. To better assess the model’s efficiency, we report module-level complexity metrics—parameters (Params) and giga floating-point operations per second (GFLOPs)—by measuring the incremental cost introduced by each core module (DWaveletConv, POSConvEmbedding, and SHSA). This breakdown clarifies the relative cost of each design choice and provides guidance for potential lightweighting, such as reducing the channel dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:{\mathrm{C}}_{\mathrm{p}}\:$$\end{document}or replacing standard 3 × 3 convolutions with depthwise separable variants. Comparative experiments conducted on the SIDD and HIT-UAV datasets show that they outperform mainstream detection models, such as YOLOv8 and PARE-YOLO45, in key performance metrics, including mAP and F1 score, while maintaining manageable model complexity to satisfy real-time requirements. Ablation studies further validate the collaborative optimization effects of each module, revealing that the joint deployment of DWaveletConv and POS-SHSA enhances mAP by 14.1%. Future research will focus on optimizing the model’s dynamic-resolution input strategy to further reduce computational costs and explore its potential applications in multimodal sensor fusion scenarios. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1