LIR-Mamba: consolidating YOLO and selective SSM with global-local scanning for robust infrared small target detection under laser interference Authors: Tianyi Chen, Xiangyi Hu, Jiafen Wang Published: 2026 ABSTRACT Detecting dim and small infrared targets under high-energy laser jamming is a critical prerequisite for modern electro-optical defense systems. However, intense glare and speckle noise severely disrupt spatial distributions, making robust detection on resource-constrained edge platforms highly challenging. Existing convolutional neural networks lack the global receptive field required to suppress large-scale jamming, while Vision Transformers incur prohibitive quadratic computational costs. Although emerging State Space Models (SSMs) offer linear complexity, their standard 1D scanning strategies inherently disrupt the 2D local spatial continuity of infrared images, causing fine-grained features of tiny targets to dissipate in strong background noise. To resolve these critical bottlenecks, we propose LIR-Mamba, a novel lightweight detection framework that synergizes the real-time efficiency of YOLO with the global modeling prowess of selective SSMs. At its core, we introduce an innovative Global-Local Selective Scanning 2D (GLSS2D) mechanism. By enforcing local scanning paths within spatial windows alongside global traversal, GLSS2D effectively bridges the gap between 1D sequence modeling and 2D spatial locality, firmly preventing the dispersion of local hot-spot features. Extensive experiments demonstrate that LIR-Mamba achieves a superior mAP@50 of 96.3% at 66 frames per second, significantly outperforming mainstream detectors such as YOLOv11 and RT-DETR in both accuracy and anti-jamming robustness. The proposed method provides competitive performance for real-time target detection in complex electro-optical countermeasure environments. Introduction In the fields of modern precision guidance, early warning reconnaissance, and electro-optical countermeasures, infrared sensing systems are the core means to achieve all-weather target acquisition and situational awareness1. However, with the rapid development of electro-optical jamming technologies, adversaries often utilize high-energy laser weapons to implement suppressive jamming against infrared imaging systems. This unstructured laser jamming forms high-intensity saturated spots and large-area scattering noise on the focal plane of the detector, causing a drastic decline in the contrast of the infrared image, and the target features are either submerged or morphologically distorted2. Therefore, developing a lightweight detection model that guarantees high-precision feature extraction while satisfying the low-power and low-latency constraints of missile-borne or airborne edge computing devices has become a research hotspot at the intersection of computer vision and electro-optical engineering. Current infrared small target detection algorithms are primarily divided into model-driven traditional methods and data-driven deep learning methods. Traditional methods based on filtering and the human visual system perform reasonably well when processing simple backgrounds, but they are highly prone to generating false alarms under complex backgrounds such as intense laser jamming3. Convolutional neural networks represented by the YOLO series dominate real-time tasks due to their excellent inference speed and local feature extraction capabilities. However, limited by the local receptive field of convolutional kernels, such models struggle to capture global contextual information. When faced with large-scale laser spot occlusion or high-density speckle noise, they often suffer from missed detections due to a lack of global semantic understanding4. On the other hand, architectures like Swin Transformer and DETR utilize self-attention mechanisms to achieve global contextual modeling, significantly improving detection accuracy in complex scenes. However, their computational complexity grows quadratically with image resolution, and the massive memory footprint and computational overhead make them difficult to deploy in resource-constrained embedded infrared detection systems5. In recent years, structured state space models (SSMs), particularly the Mamba architecture6, have provided a new perspective for balancing detection accuracy and efficiency with their linear complexity and robust sequence modeling capabilities. Although Mamba has shown tremendous potential in various vision tasks, existing visual Mamba methods7,8 typically adopt standard cross-scanning or four-way scanning mechanisms when processing 2D infrared images. While global connections are established, this 1D traversal strategy partially disrupts the inherent spatial continuity of the 2D plane. For infrared small targets, which often occupy only a few pixels and lack texture and color information, detection primarily relies on local contrast and hot-spot aggregation features. The spatial discretization caused by standard scanning can dilute weak target signals in long sequences by strong jamming noise, thereby degrading detection performance. To address these limitations, this paper proposes a lightweight detection framework, LIR-Mamba. The core idea is to embed a visual state space module with a local enhancement mechanism into a single-stage YOLO architecture. Specifically, the efficient cross-stage partial structure of YOLO is retained, but customized Mamba modules are integrated into the backbone and neck networks to achieve a global receptive field at a low computational cost. Crucially, to mitigate the destruction of locality caused by standard scanning, a locally enhanced feature scanning strategy is designed. This strategy strengthens the aggregation of local neighborhood information before state space evolution, achieving an organic unification of global jamming suppression and local target enhancement. The main contributions of this paper are summarized as follows:We propose the LIR-Mamba framework, which deeply fuses the linear-complexity state space model with a real-time object detector. This significantly enhances the feature representation capability of the model in intense laser jamming environments while substantially reducing the number of parameters and floating-point operations.We innovatively introduce SSM-based Mamba modules into the backbone and neck networks of the YOLO architecture, enabling the model to break through local window limitations and effectively capture global contextual dependencies. This resolves the inadequate feature fusion capabilities of traditional CNNs when dealing with large-area laser spots.We propose a global-local hybrid scanning strategy to address the loss of local hot-spot features caused by standard four-way scanning in existing visual SSMs. By enhancing local feature interaction along the scanning path, this mechanism effectively compensates for the deficiencies of SSMs in modeling fine-grained local features, significantly improving the detection accuracy of dim and small infrared targets. We propose the LIR-Mamba framework, which deeply fuses the linear-complexity state space model with a real-time object detector. This significantly enhances the feature representation capability of the model in intense laser jamming environments while substantially reducing the number of parameters and floating-point operations. We propose the LIR-Mamba framework, which deeply fuses the linear-complexity state space model with a real-time object detector. This significantly enhances the feature representation capability of the model in intense laser jamming environments while substantially reducing the number of parameters and floating-point operations. We innovatively introduce SSM-based Mamba modules into the backbone and neck networks of the YOLO architecture, enabling the model to break through local window limitations and effectively capture global contextual dependencies. This resolves the inadequate feature fusion capabilities of traditional CNNs when dealing with large-area laser spots. We innovatively introduce SSM-based Mamba modules into the backbone and neck networks of the YOLO architecture, enabling the model to break through local window limitations and effectively capture global contextual dependencies. This resolves the inadequate feature fusion capabilities of traditional CNNs when dealing with large-area laser spots. We propose a global-local hybrid scanning strategy to address the loss of local hot-spot features caused by standard four-way scanning in existing visual SSMs. By enhancing local feature interaction along the scanning path, this mechanism effectively compensates for the deficiencies of SSMs in modeling fine-grained local features, significantly improving the detection accuracy of dim and small infrared targets. We propose a global-local hybrid scanning strategy to address the loss of local hot-spot features caused by standard four-way scanning in existing visual SSMs. By enhancing local feature interaction along the scanning path, this mechanism effectively compensates for the deficiencies of SSMs in modeling fine-grained local features, significantly improving the detection accuracy of dim and small infrared targets. Related work This section reviews the literature relevant to this study, outlines the evolutionary trajectory of infrared small target detection technologies, and analyzes the current technical bottlenecks. Traditional and two-stage infrared detection methods Early infrared small target detection primarily relied on prior assumptions about the physical properties of targets and backgrounds. Filtering-based methods, such as the Top-Hat transform9,10 and its variants11–13, enhance targets by suppressing the background but tend to leave high-frequency noise in complex scenes. Methods based on the human visual system (HVS) utilize local contrast differences to highlight targets. The Local Contrast Measure (LCM) algorithm proposed by Chen et al.14 is a typical representative, but when confronted with high-intensity laser spots, the high contrast at the spot edges often leads to severe false alarms. Furthermore, methods based on low-rank sparse recovery, such as the Infrared Patch-Image (IPI) model proposed by Gao et al.15, separate targets from the background using the low-rank property of the background and the sparsity of the target. Although highly accurate, the iterative optimization process is computationally expensive and struggles to meet the real-time requirements of electro-optical countermeasure environments. Deep learning-based detection methods With the advancement of deep learning, data-driven methods have gradually become the mainstream16–18. Based on the network architecture, they can be categorized into CNN-based and Transformer-based methods. In terms of CNNs, the Asymmetric Contextual Modulation (ACM) module proposed by Dai et al.19 effectively fused deep semantics with shallow details, improving the detection rate of small targets. DNA-Net, proposed by Li et al.20, designed dense nested interactive modules to achieve multiple feature enhancements for point targets. To meet real-time requirements, the YOLO series has been widely applied in the infrared domain. Sun et al.21 improved YOLOv5 by introducing an attention mechanism to address the issue of missing infrared textures. However, the inherent local inductive biases of CNNs limit their ability to suppress large-scale laser jamming. To address global modeling issues, Transformers were introduced to ISTD tasks. The IRSTD-Transformer proposed by Liu et al.22 utilized self-attention mechanisms to capture long-range dependencies, significantly reducing the false alarm rate. UIU-Net, proposed by Wu et al.23, achieved SOTA performance in complex backgrounds through a nested U-shaped structure combined with multi-level attention. Despite their superior performance, the massive parameter count and computational burden of Transformers make them difficult to deploy on edge devices with limited computing power, prompting researchers to seek more efficient architectures. Fig. 1The overall architecture of the proposed LIR-Mamba model. The overall architecture of the proposed LIR-Mamba model. Mamba models in visual recognition and detection Recently, structured state space models (SSMs), particularly the Mamba architecture, have emerged as a powerful paradigm for breaking through the quadratic computational limits of Vision Transformers. However, adapting 1D sequence modeling to 2D image planes necessitates specialized scanning strategies, which have evolved across three distinct structural paradigms. The first paradigm relies on direct 1D flattened sequence serialization, represented by bidirectional formulations like Vision Mamba (Vim)7, as well as specific dense task networks such as U-Mamba24 for biomedical image segmentation and MedMamba25 for medical classification. Although these frameworks achieve competitive efficiency by linearizing image tokens, their straightforward vectorization completely strips away the multi-dimensional spatial topology. This spatial fracture is highly detrimental to low-observable infrared small targets, whose tracking relies heavily on localized spatial intensity gradients and fine-grained patch geometries. The second paradigm introduces global multi-directional cross-scanning to capture multi-dimensional semantic associations, as pioneered by VMamba’s SS2D mechanism8. This holistic approach has been widely extended to multi-dimensional data modeling in Mamba-ND26 and hierarchical UNet structures in Swin-UMamba27 for dense semantic prediction. While these methods excel at modeling long-range contextual relationships, their exhaustive full-frame scanning paths inevitably extend the sequential distance between adjacent pixels of a tiny target to hundreds of time steps. Under intense electro-optical countermeasures, this global dilation causes weak sub-pixel target signals to be easily diluted or completely overwritten by the massive accumulated gradients of large-area saturated laser spot noise. To recover localized proximity, the third paradigm incorporates windowed or structurally localized scanning patterns. This includes LocalMamba28 with windowed selective scans, EfficientVMamba29 using atrous selective sampling, PlainMamba30 utilizing non-directional multi-scale scanning, Devoted Local-Mamba31 optimizing local textures via biased windows, and Large-Window Mamba32 adapting receptive boundaries. Although these works successfully alleviate localized feature dissipation, their standalone block partitioning confines feature evolution to isolated sub-regions. Consequently, they lack the holistic contextual awareness required to semantically identify and suppress non-uniform optical artifacts that naturally extend across multiple window boundaries. To resolve these persistent bottlenecks, our proposed Global-Local Selective Scanning 2D (GLSS2D) breaks away from these singular trajectories by introducing a specialized Dual-Track Hybrid Parallel Architecture. Rather than enforcing a compromise between global context tracking8 and localized window constraints28, GLSS2D concurrently routes image tokens through parallel global axes and localized windowed zig-zag streams. This unified approach enables the model to interpret and filter large-scale laser noise at the macro-level while firmly clustering sub-pixel energy hotspots within the hidden states at the micro-level. Critically, since this multi-view path configuration is executed entirely via in-place token pointer permutation, it delivers high-fidelity spatial-contextual awareness with zero additional parameter overhead or network redundancy. Method LIR-Mamba architecture overview To address the challenges of detecting dim and small infrared targets in complex electromagnetic jamming environments, including non-uniform spot occlusion, weak signals, and limited edge computing resources, this paper proposes a lightweight detection framework named LIR-Mamba. The framework follows the single-stage detection paradigm and introduces targeted reconstruction in the feature extraction and fusion stages to better accommodate infrared imaging characteristics. As shown in Fig. 1, the overall architecture comprises an SSM-based backbone network, a multi-scale feature fusion neck network, and a decoupled detection head. In the feature extraction stage, the backbone network receives the infrared image input. Considering that infrared small targets often lack texture and are highly susceptible to loss during downsampling, we introduce a Vision Clue Merge module after the initial stem layer to replace standard strided convolutions. The design of this module is inspired by the Patch Merging operation in Swin Transformer33, aiming to expand the channel dimension while reducing resolution through spatial reorganization, thereby maximally preserving the spatial positional information of weak hot spots. Subsequently, the feature maps enter stacked Global-Local Mamba (GLM) blocks, which utilize linear computational complexity to capture long-range contextual dependencies in infrared images, such as distinguishing large-area laser scattering noise from actual point targets. To further enhance deep semantics, a parallel spatial attention cross-stage partial network (C2PSA) is integrated at the end of the backbone. This module introduces a spatial attention mechanism34 within the CSP bottleneck layer, effectively suppressing background clutter generated by high-energy lasers and increasing focus on high-radiation areas before features propagate to the next stage. The neck network constructs a Feature Pyramid Network (FPN) to aggregate multi-scale features for detecting targets of varying distances and sizes. We deploy GLM blocks within the Path Aggregation Network (PANet)35 to ensure cross-scale semantic consistency and global context awareness. This stage utilizes a Dynamic Content-Aware Reassembly Upsampling (CRU-Dynamic) module to replace standard nearest-neighbor interpolation. By employing a dynamic kernel generation mechanism to reassemble features based on local thermal radiation content, it effectively alleviates the aliasing effects and boundary blurring common at the edges of intense light spots. Finally, the enhanced multi-scale features are fed into a Decoupled Head36, which handles target classification and bounding box regression tasks separately through dual parallel branches and generates final detection results using the Non-Maximum Suppression (NMS) algorithm. Fig. 2Comparison of scanning mechanisms in visual state space models. (Top) The Traditional SS2D strategy employs four-way cross-scanning. (Bottom) The proposed Global-Local Selective Scanning (GLSS2D) strategy. Comparison of scanning mechanisms in visual state space models. (Top) The Traditional SS2D strategy employs four-way cross-scanning. (Bottom) The proposed Global-Local Selective Scanning (GLSS2D) strategy. Visual detection state space model State space model Structured state space models (SSMs) originate from continuous system control theory, mapping 1D input sequences to outputs via latent variables. The Mamba architecture introduces a Selective Scan mechanism6, allowing parameters to change dynamically with inputs, thereby enabling content-aware modeling with linear complexity. For discrete sequences of infrared image pixels, discretization is performed using the Zero-Order Hold (ZOH) principle. Specifically, a structured state space model is defined by linear ordinary differential equations:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & \mathbf{h}'(t) = \mathbf{A}\mathbf{h}(t) + \mathbf{B}x(t) \end{aligned}$$\end{document}2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & y(t) = \mathbf{C}\mathbf{h}(t) \end{aligned}$$\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}$$\mathbf{A} \in \mathbb {R}^{N \times N}$$\end{document} is the state transition matrix, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{B} \in \mathbb {R}^{N \times 1}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{C} \in \mathbb {R}^{1 \times N}$$\end{document} are projection parameters. In deep learning scenarios processing discrete data such as image pixel sequences, the ZOH principle must be applied to discretize the continuous system. Given a step size Δ, the discretized parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{\mathbf{A}}$$\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}$$\bar{\mathbf{B}}$$\end{document} are represented as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \bar{\mathbf{A}} = \exp (\Delta \mathbf{A}), \quad \bar{\mathbf{B}} = (\Delta \mathbf{A})^{-1}(\exp (\Delta \mathbf{A}) - \mathbf{I}) \cdot \Delta \mathbf{B} \end{aligned}$$\end{document}The final discrete recurrent formulation is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbf{h}_t = \bar{\mathbf{A}}\mathbf{h}_{t-1} + \bar{\mathbf{B}}x_t$$\end{document}, and the output is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_t = \mathbf{C}\mathbf{h}_t$$\end{document}. Mamba8 introduces a selective mechanism that permits parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta , \mathbf{B}, \mathbf{C}$$\end{document} to vary dynamically with the input \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_t$$\end{document}, thus achieving content-aware modeling at a linear computational cost. Global-local selective scanning mechanism (GLSS2D) The fundamental limitation of applying 1D state space models to infrared images under intense laser jamming stems from the spatial disruption caused by the flattening process, wherein the non-causal nature of 2D spatial patterns conflicts with the causal processing framework of SSMs. As shown at the top of Fig. 2, traditional visual Mamba methods like VMamba8 typically employ a global cross-scanning strategy that traverses entire row or column axes. While this establishes a long-range receptive field, it inevitably elongates the sequence distance between vertically or horizontally adjacent pixels within local semantic regions. For scenarios detecting infrared small targets under intense laser interference, this disruption is particularly detrimental; adjacent pixels belonging to the same tiny target may be separated by hundreds of time steps in the flattened sequence. This causes the SSM to lose crucial high-frequency texture cues and edge details, which are vital for precisely detecting minuscule targets against intense laser backgrounds. To mitigate this discrepancy and bridge the adaptation gap between 1D sequence modeling and 2D locality, this paper proposes the Global-Local Selective Scanning 2D (GLSS2D) mechanism. As depicted in Fig. 2, unlike standard scanning which treats an image as a single sequence, GLSS2D introduces a window-based partitioning strategy before the state space evolution stage: specifically, the input feature map is partitioned spatially into multiple independent local windows, and the scanning trajectories are redesigned. First, all tokens within a specific local window are traversed to aggregate fine-grained neighborhood information before moving to the next window, ensuring that spatially adjacent tokens maintain continuity in the 1D sequence. This local traversal efficiently captures 2D dependencies related to the texture and geometry of small targets, allowing the hidden state to encode richer local contextual information. However, relying solely on local scanning would constrain the model’s ability to understand and capture features of tiny targets under more complex laser jamming environments, such as fluctuating laser intensities or cloud occlusions. Therefore, GLSS2D is designed as a hybrid mechanism that synchronously executes four complementary scanning paths: two global paths (horizontal and vertical) to capture global context at the overall environmental scale, and two local paths (based on the original window and its flipped version) to mine fine-grained object details. These distinct scanning sequences are processed in parallel through independent SSM heads. Notably, this window-based scanning strategy only involves the relocation of token indices without adding extra parameters or computational complexity to the model, providing a computationally efficient performance enhancement compared to standard methods. By integrating the aforementioned multi-view sequences, GLSS2D achieves an organic unification of global semantic understanding and local feature retention, significantly improving the model’s robustness in detecting dim and small infrared targets under intense laser jamming and scale variations. Global-local Mamba module Based on the proposed scanning strategy, we constructed the GLM block as the fundamental feature extraction and fusion unit, whose detailed structure is shown in Fig. 3. The GLM block receives input features, which first pass through Layer Normalization (LN) and linear projection, followed by Depthwise Convolution (DWConv). DWConv implicitly encodes local positional information via sliding window operations, which is critical for position-sensitive infrared small target detection. Fig. 3The detailed structure of the GLM module. (a) The overall module architecture. (b) The internal mechanism of the Global-Local Selective Scanning module. The detailed structure of the GLM module. (a) The overall module architecture. (b) The internal mechanism of the Global-Local Selective Scanning module. The core GLSS2D module (as shown in Fig. 3b) divides features into distinct scanning heads, each executing the global-local hybrid scanning. For each scanning direction, the input sequence is processed by the selective state space model:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x_i = \text {SSM}(x_i, \Delta _i, \mathbf{A}_i, \mathbf{B}_i, \mathbf{C}_i) \end{aligned}$$\end{document}Then, the output sequences from various directions are restored to 2D feature maps through inverse scanning and subsequently fused. Finally, the features are projected by an output linear layer and combined with the input via a residual connection, formulated as follows:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathbf{X}_{out} = \text {Linear}(\text {GLSS2D}(\text {DWConv}(\text {Linear}(\text {LN}(\mathbf{X}_{in}))))) + \mathbf{X}_{in} \end{aligned}$$\end{document}Through this design, the GLM block achieves the organic unification of global context awareness and local fine-grained feature mining within a single. Loss function To achieve precise target localization and classification, the training process of this model employs a multi-task loss function comprising classification loss37, bounding box regression loss38, and Distribution Focal Loss39:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathcal {L}_{total} = \lambda _{cls}\mathcal {L}_{cls} + \lambda _{box}\mathcal {L}_{box} + \lambda _{dfl}\mathcal {L}_{dfl} \end{aligned}$$\end{document} Fig. 4Visualization of training convergence and validation metrics. The first row displays the descending curves of box regression loss, classification loss, and distribution focal loss on the training set. The second row shows the corresponding loss curves on the validation set. The thirdfourth and fifth column illustrates the steady improvement and convergence of Precision, Recall, mAP@50, and mAP@50-95, indicating that the model achieves stable optimization without overfitting. Visualization of training convergence and validation metrics. The first row displays the descending curves of box regression loss, classification loss, and distribution focal loss on the training set. The second row shows the corresponding loss curves on the validation set. The thirdfourth and fifth column illustrates the steady improvement and convergence of Precision, Recall, mAP@50, and mAP@50-95, indicating that the model achieves stable optimization without overfitting. Experimental results and discussions Experimental setup Datasets The dataset utilized in this experiment consists of two parts, totaling 2500 infrared small target samples. The first part is the publicly available NUDT-SIRST infrared small target dataset20, which contains 1327 real infrared images covering various complex backgrounds such as clouds, sea surfaces, and urban areas. This is primarily used to establish the model’s fundamental feature extraction capabilities for dim and small targets with low signal-to-clutter ratios. The second part is a self-constructed Laser-IR laser jamming dataset, originally containing 500 infrared images simulating electro-optical countermeasure scenarios. Since real laser suppressive jamming data is extremely difficult to obtain and is mainly concentrated in specific wavebands and intensities, we overlaid saturated light spots and speckle noise of varying intensities onto real backgrounds based on physical optical models to improve the network’s robustness in dynamic strong jamming environments and prevent overfitting. Furthermore, image enhancement operations such as random rotation, brightness jitter, and vertical/horizontal flipping were applied. This ultimately expanded the dataset to 1173 augmented images containing diffuse jamming and mixed jamming patterns. All images were manually annotated at the pixel level using the LabelImg software. During the annotation process, we strictly followed physical thermal radiation characteristics and eliminated high-brightness false noise generated by high-energy laser scattering to ensure label accuracy. The aggregate dataset utilized throughout all validation stages comprises 2500 independent infrared frames, strictly divided into an explicit training set (1500 images, 60%), validation set (500 images, 20%), and testing set (500 images, 20%). Both NUDT-SIRST and Laser-IR data components are distributed equally across these subsets to prevent distribution shifts. During the optimization cycle, frames are fully intermixed and dynamically shuffled at each epoch. The synthesized laser artifacts are mathematically derived from an optical physical model factoring in Airy disk diffraction envelopes, saturation tail cutoffs, and atmospheric scattering halos, effectively minimizing the simulation domain gap for modern field deployments. Table 1Quantitative and efficiency experimental results of the different detection models. Bold represents the best performance.MethodPrecisionRecallmAP@50mAP@50:95Params (M)FLOPs (G)Inference (ms)Faster R-CNN400.7950.8060.8990.52541.3183.624.9YOLOv8410.8660.8820.9310.59311.228.82.6YOLOv11420.8730.8980.9370.60412.610.41.9YOLOv12430.8540.8710.9280.58814.112.22.1YOLOv13440.9060.9140.9510.62916.514.82.5RT-DETR450.9210.9160.9530.63232.240.54.7DNA-Net200.8950.8900.9430.598 4.5 9.71.8UIU-Net230.9450.9340.9500.63925.534.45.2Mamba-YOLO460.9130.9080.9460.6214.8 8.2 1.4 LocalMamba280.9310.9220.9540.6355.28.91.7LIR-Mamba (Ours) 0.952 0.937 0.963 0.660 4.98.51.5 Table 1Quantitative and efficiency experimental results of the different detection models. Bold represents the best performance.MethodPrecisionRecallmAP@50mAP@50:95Params (M)FLOPs (G)Inference (ms)Faster R-CNN400.7950.8060.8990.52541.3183.624.9YOLOv8410.8660.8820.9310.59311.228.82.6YOLOv11420.8730.8980.9370.60412.610.41.9YOLOv12430.8540.8710.9280.58814.112.22.1YOLOv13440.9060.9140.9510.62916.514.82.5RT-DETR450.9210.9160.9530.63232.240.54.7DNA-Net200.8950.8900.9430.598 4.5 9.71.8UIU-Net230.9450.9340.9500.63925.534.45.2Mamba-YOLO460.9130.9080.9460.6214.8 8.2 1.4 LocalMamba280.9310.9220.9540.6355.28.91.7LIR-Mamba (Ours) 0.952 0.937 0.963 0.660 4.98.51.5 Quantitative and efficiency experimental results of the different detection models. Bold represents the best performance. Implementation details The model training in this study was conducted on a personal computer equipped with an NVIDIA RTX 4090 GPU and an Intel 14900KF CPU, utilizing the PyTorch deep learning framework. The software environment was built upon PyTorch version 2.1.0 and CUDA version 11.8. During the training phase, the Stochastic Gradient Descent (SGD) optimizer was employed, with the momentum parameter set to 0.937 and the weight decay coefficient set to 0.0005. The initial learning rate was set to 0.01, and dynamic decay was implemented through a cosine annealing strategy. Input images were uniformly resized to 640 × 640 pixels, and data augmentation methods such as Mosaic and Mixup were further introduced to expand the diversity of the training data. The model was trained for a total of 200 epochs with a batch size of 16. The convergence curves of the training and validation processes indicate steady optimization. As shown in Fig. 4, the training and validation loss curves gradually decrease, while Precision, Recall,mAP@50, and mAP@50-95 steadily improve, indicating stable convergence without obvious overfitting. The first to third columns of the first row demonstrate the descending curves of bounding box regression loss, classification loss, and distribution focal loss on the training set. The second row displays the corresponding loss curves on the validation set. The fourth and fifth columns illustrate the steady improvement and convergence of Precision, Recall, mAP@50, and mAP@50-95, indicating that the model achieved stable optimization without overfitting. Evaluation metrics The evaluation metrics adopt universal standards in the field of infrared small target detection: Precision, Recall, and mean Average Precision (mAP@50 and mAP@50-95), with the specific formulas as follows:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & \text {Precision} = \frac{\text {TP}}{\text {TP} + \text {FP}} \end{aligned}$$\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}$$\begin{aligned} & \text {Recall} = \frac{\text {TP}}{\text {TP} + \text {FN}} \end{aligned}$$\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}$$\begin{aligned} & \text {AP} = \int _{0}^{1} \text {Precision}(\text {Recall}) \, d(\text {Recall}) \end{aligned}$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & \text {mAP} = \frac{\sum _{i=1}^{N} \text {AP}_i}{N} \end{aligned}$$\end{document}where TP represents the number of true positive samples correctly predicted, FP denotes the number of false positives where the model incorrectly detects negative samples as positive, and FN refers to the number of false negatives where the model fails to detect positive samples. N represents the number of small target categories. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {AP}_i$$\end{document} indicates the average precision of the i-th category. mAP is the mean Average Precision calculated by measuring the area under the Precision-Recall curve. Quantitative comparative analysis To objectively evaluate the target detection accuracy of the proposed LIR-Mamba framework under strong laser interference scenarios, we compare LIR-Mamba with state-of-the-art mainstream detection algorithms, including the Transformer-based detector RT-DETR45, the two-stage detector Faster R-CNN40, single-stage detectors YOLOv841, YOLOv1142, YOLOv1243, and YOLOv1344, prevalent vision Mamba detection architectures Mamba-YOLO46 and LocalMamba28, as well as infrared small-target dedicated networks DNA-Net20 and UIU-Net23. As shown in Table 1, the experimental results across the convolutional models demonstrate distinct structural capabilities. Traditional two-stage layouts like Faster R-CNN struggle with low-observable sub-pixel points, yielding 0.899 mAP@50. Within the single-stage frameworks, the standard YOLOv8 and YOLOv11 baselines establish robust geometric profiles, scoring 93.1% and 93.7% mAP@50, respectively. The contemporary YOLOv13 extends this capacity further to hit 95.1% mAP@50 via dense multi-scale aggregation. Interestingly, YOLOv12 exhibits a minor localized dip (92.8%). This behavior is attributed to its highly optimized, texture-focused local attention layers designed for high-resolution natural objects; under intense laser countermeasures, these attention blocks can become over-sensitized by high-brightness background artifacts and scattering halos, slightly blurring out the faint, structureless sub-pixel targets. Concurrently, specialized models like DNA-Net (94.3%) and UIU-Net (95.0%) leverage deep nested linkages to preserve dim signatures but show minor limitations when managing widespread saturated occlusion zones. Evaluating the homogeneous state-space models reveals the critical impact of sequence parsing strategies. Basic Mamba-YOLO delivers a competitive 94.6% mAP@50, but its vanilla 1D sequence flattening breaks apart native 2D spatial contiguity, allowing sub-pixel target features to easily dissipate into the long-range sequence context. While LocalMamba mitigates this via windowed block partitioning to achieve 95.4% mAP@50, its isolated sub-window state evolution operates without cross-window context alignment, making it vulnerable to extensive atmospheric diffraction rings that cut across multiple window cells. The proposed LIR-Mamba firmly outbalances all comparative baselines, reaching a peak mAP@50 of 96.3% and a mAP@50 : 95 of 0.660. The principal reason behind this exceptional effectiveness is the architectural synergy of our GLSS2D mechanism. By deploying a dual-track parallel token stream, LIR-Mamba constructs wide macro-level contextual dependencies to interpret and filter out large-scale diffuse laser spot interference while simultaneously maintaining tight 2D spatial neighborhood contiguity within the local hidden states. This dual-track modeling successfully anchors the radiative signatures of weak sub-pixel objects, preventing feature dispersion under extreme electro-optical jamming environments. Qualitative visual analysis As shown in Fig. 5, qualitative visual comparisons intuitively reveal the robustness differences among various models under high-energy laser jamming scenarios. When facing complex backgrounds with large-area saturated light spots and cloud clutter, mainstream CNN detectors such as YOLOv8 and YOLOv11 expose obvious limitations: due to their over-reliance on local gradient features, these models are highly prone to confusing the background with the target, frequently misjudging the high-brightness edges of light spots or cloud textures as potential objects, thus leading to obvious false alarms and bounding box redundancy. Even RT-DETR, which introduces global attention, exhibits unstable drifting phenomena in localization under strong jamming, struggling to maintain keen capture of weak targets while thoroughly suppressing high-radiation background noise. Fig. 5Visual comparison between the proposed method and other mainstream models. Visual comparison between the proposed method and other mainstream models. In contrast, the proposed LIR-Mamba demonstrates exceptional anti-jamming perception capabilities. Benefiting from the global-local selective scanning mechanism, the model can effectively establish global contextual associations within the image, accurately identifying and filtering out large-scale laser spot jamming at the semantic level while precisely locking onto dim and small infrared targets that occupy extremely few pixels via local fine-grained features. Visual results clearly show that, whether in the high-saturation area at the center of the light spot or the blurry zones at the edges, LIR-Mamba generates compact, accurate, and unique detection bounding boxes without the missed detections or false alarms common in baseline models. This verifies the unique advantages of state space models in balancing global background suppression and local detail preservation. Efficiency analysis In modern early warning reconnaissance and electro-optical countermeasure edge computing scenarios, model efficiency is as critical as accuracy. Therefore, in this subsection, we comprehensively compare the parameters, computational complexity (GFLOPs), and inference time of the models. The data in Table 1 indicates that traditional models and Transformer-based architectures incur substantial computational overhead. The Faster R-CNN model demands up to 183.6 GFLOPs, while the attention-based RT-DETR possesses 32.2 M parameters, 40.5 GFLOPs of floating-point operations, and an inference latency of 4.7 ms. Among single-stage convolutional neural networks, although YOLOv11 retains a lightweight design, prevailing architectural scaling strategies induce a dramatic rise in computational cost. The infrared-specialized UIU-Net also suffers from excessive memory consumption. In stark contrast, vision state-space models exhibit remarkable lightweight structural properties. The baseline Mamba-YOLO and LocalMamba achieve outstanding runtime efficiency, with floating-point operations of only 8.2 GFLOPs and 8.9 GFLOPs, respectively. Our proposed LIR-Mamba achieves the optimal trade-off between computation and performance, delivering superior detection accuracy with merely 4.9 M parameters and 8.5 GFLOPs. Such favorable performance can be attributed to two core merits: First, the integrated visual state-space module features linear computational complexity, entirely eliminating the quadratic memory overhead inherent to Transformer self-attention layers. Second, the proposed GLSS2D spatial routing structure only performs in-place token pointer permutation. Without introducing extra matrix multiplications, redundant parameters, or hardware-level computational redundancy, this design enables high-fidelity alignment of global-local contextual features. Table 2Ablation results of the different module setups. Bold represents the best performance.No.Model setupVCMC2PSACRUGlobal scanLocal scanmAP@50mAP@50:95Param (M)FLOPs (G)1Baseline (CNN)0.9350.6016.711.62+ VCM Layer \checkmark 0.9390.6076.510.93+ C2PSA Attention \checkmark \checkmark 0.9430.6126.811.34+ CRU-Dynamic \checkmark \checkmark \checkmark 0.9460.6186.911.15+ GLM module(w/o GLSS2D) \checkmark \checkmark \checkmark \checkmark 0.9510.6244.98.56LIR-Mamba (Full) \checkmark \checkmark \checkmark \checkmark \checkmark 0.963 0.660 4.9 8.5 Table 2Ablation results of the different module setups. Bold represents the best performance.No.Model setupVCMC2PSACRUGlobal scanLocal scanmAP@50mAP@50:95Param (M)FLOPs (G)1Baseline (CNN)0.9350.6016.711.62+ VCM Layer \checkmark 0.9390.6076.510.93+ C2PSA Attention \checkmark \checkmark 0.9430.6126.811.34+ CRU-Dynamic \checkmark \checkmark \checkmark 0.9460.6186.911.15+ GLM module(w/o GLSS2D) \checkmark \checkmark \checkmark \checkmark 0.9510.6244.98.56LIR-Mamba (Full) \checkmark \checkmark \checkmark \checkmark \checkmark 0.963 0.660 4.9 8.5 Ablation results of the different module setups. Bold represents the best performance. Fig. 6Grad-CAM generated heatmaps for Tiny targets utilizing different module configurations. Grad-CAM generated heatmaps for Tiny targets utilizing different module configurations. Ablation experiments Module ablation experiments To verify the contribution of each individual component in the proposed LIR-Mamba, we conducted extensive ablation studies, and the quantitative comparison results are presented in Table 2. First, after embedding the Visual Clue Merging (VCM) module, the model accuracy increases to 0.939, while the floating-point operations drop from 11.6 G to 10.9 G. By structurally transferring low-level spatial matrices to the channel domain, the VCM module prevents the premature loss of subtle subpixel infrared hot-spot features. Next, C2PSA attention is introduced into the cross-stage bottleneck architecture to strengthen the model’s capacity to focus on spatial gradients of high-radiation regions and suppress clutter interference from background laser reflections, raising the model accuracy to 0.943. Furthermore, CRU-Dynamic is adopted to replace the nearest-neighbor interpolation algorithm, eliminating severe aliasing distortion and boundary blurring commonly observed at the edges of intense light spots, which steadily boosts the model accuracy to 0.946. Moreover, the heavy deep convolutional blocks are substituted with GLM modules (w/o SS2D), delivering two remarkable performance gains: the model accuracy rises substantially to 0.951, whereas the parameter count and computational cost plummet to 4.9 M and 8.5 G, respectively. This result verifies that constructing a global receptive field with linear complexity enables the network to distinguish global macroscopic interference features from discrete real point targets with extremely high structural efficiency. Finally, when the full GLSS2D strategy is activated, the model accuracy reaches 0.963, with the parameter count, floating-point operations and inference latency remaining unchanged. This prominent accuracy improvement intuitively illustrates the core functionality of the local window scanning branch, and demonstrates that preserving 2D spatial neighborhood continuity during one-dimensional sequence modeling is critical for capturing faint thermal infrared targets under heavy noise conditions. To intuitively reveal the significant differences in feature attention areas before and after the core component improvements, Fig. 6 utilizes Grad-CAM technology to visualize the heatmaps of feature responses under different configurations. Deep red highlighted areas in the figure indicate that the model has extremely high gradient weights and attention focus at that location, while cool-toned areas represent suppressed background information. Observing the comparison results, when the core backbone GLM module is removed (Fig. 6d), the network degrades to rely solely on local convolutional features. This causes it to completely fail to distinguish between high-energy laser spots and real targets, with thermal responses appearing chaotically diffused inside the light spot and at cloud edges, indicating that the model is severely misled by intense jamming. When the backbone is retained but the GLSS2D scanning mechanism is removed (Fig. 6c), although the large-area light spot is suppressed to some extent, the lack of enhanced aggregation of local neighborhood information causes the thermal activation of the target area to be insufficiently concentrated, accompanied by obvious background noise leakage. In contrast, the complete LIR-Mamba model (Fig. 6b) exhibits perfect anti-jamming focusing capabilities. Benefiting from the synergy of global context modeling and local fine-grained scanning, the heatmap accurately converges on the tiny UAV target, effectively filtering out the interference of high-brightness light spots and complex clouds. This proves the innovativeness and effectiveness of the proposed method in extracting robust tiny features in complex electro-optical countermeasure environments. Effectiveness of the GLM block at different positions We investigated the optimal placement strategy of the proposed GLM Block in Table 3. Compared to the baseline model, integrating the module solely into the backbone network or the neck network improved the mean Average Precision (mAP@50) by 1.7% and 1.4%, respectively. This confirms that the proposed GLM Block enhances the feature expression capabilities of both early features and deep semantic features. Meanwhile, fully integrating the GLM Block module into both the backbone and neck networks achieved the best performance, with a mAP@50 reaching 0.963. This indicates that the GLM Block module can jointly and effectively learn scale-adaptive and semantically diverse representations, which is beneficial for the task of detecting dim and small infrared targets under intense laser jamming. Table 3Ablation study on GLM block placement.MethodmAP@50Param. (M)FLOPs (G)Baseline0.9356.711.6Backbone Only0.9525.110.6Neck Only0.9495.59.9Both (Full) 0.963 4.9 8.5 Table 3Ablation study on GLM block placement.MethodmAP@50Param. (M)FLOPs (G)Baseline0.9356.711.6Backbone Only0.9525.110.6Neck Only0.9495.59.9Both (Full) 0.963 4.9 8.5 Ablation study on GLM block placement. Table 4Detailed ablation study of different scanning configurations and localized window sizes on the laser-IR dataset.ConfigurationScan strategy/Window setupmAP@50mAP@50:95Param (M)FLOPs (G)Standard paradigms1D Flattened Scan (Vim-style)70.9420.6094.98.5Standard SS2D Cross-Scan80.9510.6244.98.5Path deconstructionGlobal Paths Only (Horizontal + Vertical)0.9450.6154.98.5Local Paths Only (Window-tiled Zig-zag)0.9480.6204.98.5Window scale ΩGLSS2D (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Window size } 4 \times 4$$\end{document})0.9560.6324.98.5GLSS2D (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Window size } 8 \times 8$$\end{document}, Default) 0.963 0.660 4.9 8.5 GLSS2D (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Window size } 16 \times 16$$\end{document})0.9580.6394.98.5 Table 4Detailed ablation study of different scanning configurations and localized window sizes on the laser-IR dataset.ConfigurationScan strategy/Window setupmAP@50mAP@50:95Param (M)FLOPs (G)Standard paradigms1D Flattened Scan (Vim-style)70.9420.6094.98.5Standard SS2D Cross-Scan80.9510.6244.98.5Path deconstructionGlobal Paths Only (Horizontal + Vertical)0.9450.6154.98.5Local Paths Only (Window-tiled Zig-zag)0.9480.6204.98.5Window scale ΩGLSS2D (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Window size } 4 \times 4$$\end{document})0.9560.6324.98.5GLSS2D (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Window size } 8 \times 8$$\end{document}, Default) 0.963 0.660 4.9 8.5 GLSS2D (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Window size } 16 \times 16$$\end{document})0.9580.6394.98.5 Detailed ablation study of different scanning configurations and localized window sizes on the laser-IR dataset. Table 5Physical deployment and efficiency benchmarking on the NVIDIA Jetson Orin NX (20W) platform.MethodEdge latency (ms)Edge throughput (FPS)Peak memory footprint (MB)YOLOv11 14.2 70.4 142 YOLOv1317.459.5173RT-DETR48.620.6645DNA-Net22.444.6210Mamba-YOLO19.850.5184LIR-Mamba (Ours) 16.3 61.3 168 Table 5Physical deployment and efficiency benchmarking on the NVIDIA Jetson Orin NX (20W) platform.MethodEdge latency (ms)Edge throughput (FPS)Peak memory footprint (MB)YOLOv11 14.2 70.4 142 YOLOv1317.459.5173RT-DETR48.620.6645DNA-Net22.444.6210Mamba-YOLO19.850.5184LIR-Mamba (Ours) 16.3 61.3 168 Physical deployment and efficiency benchmarking on the NVIDIA Jetson Orin NX (20W) platform. Ablation experiments on scanning strategies To comprehensively verify the structural validity of the proposed GLSS2D mechanism, we systematically isolate the behavioral contributions of its directional scan branches and evaluate the scaling impacts of its window partition hyperparameters. All experiments in this subsection are performed by holding the rest of the LIR-Mamba pipeline constant, ensuring that accuracy variations arise exclusively from the sequence encoding layout. The empirical results are detailed in Table 4. As observed from the tracking data, the sequential arrangement protocol significantly dictates infrared small target detection capacity under intense laser jamming. The standard 1D Flattened Scan generates the lowest metrics (0.942 mAP@50) due to its severe disruption of 2D geometric contiguity. Transitioning to full-frame Standard SS2D cross-scanning8 improves the baseline to 0.951 mAP@50 by restoring row-and-column connectivity, yet it still suffers from weak target signal dissipation when faced with large-area saturated spot occlusion. The path deconstruction block clearly validates our dual-track hybrid rationale. Deploying the ‘Global Paths Only’ layout provides a wide receptive field but drops the local contrast clustering ability (0.945 mAP@50), whereas executing the ‘Local Paths Only’ configuration retains sharp point profiles but fails to model macroscopic environmental cloud clutter (0.948 mAP@50). Combining both paths symmetrically yields a substantial accuracy surge, proving their strong conceptual synergy. Lastly, evaluating the spatial window scale Ω reveals an optimal peak at a medium setting (8 × 8). A tightly constrained window size (4 × 4) fragments the sequence too aggressively, limiting the state variable’s capacity to establish meaningful internal context. Conversely, expanding the window boundaries to an overly broad layout (16 × 16) increases the scanning path length excessively, causing sub-pixel thermal target signals to be diluted by adjacent laser scattering boundaries within the same sub-sequence. Setting the partition scale at a balanced 8 × 8 configuration ensures the most robust optimization layout, successfully anchoring sub-pixel objects against severe electromagnetic interference patterns with zero computational penalty. Performance tests on embedded edge platform To further verify the applicability of the proposed LIR-Mamba on airborne computing devices, we conduct edge-adaptation experiments on the NVIDIA Jetson Orin NX edge computing platform, with experimental results listed in Table 5. As can be observed from the table of experimental results, the inference speed of LIR-Mamba reaches 61.3 FPS on the Jetson Orin NX platform, significantly outperforming the Transformer-based RT-DETR model, which only achieves 20.6 FPS. Such frame rate satisfies the standard real-time deployment requirements of missile-borne and airborne optoelectronic interception systems. Furthermore, benefited from the hardware-oriented parallel scan optimization strategy, the proposed model achieves inference performance comparable to pure convolutional networks while delivering remarkably superior anti-interference detection capability. Discussion Extensive experiments mentioned above demonstrate that LIR-Mamba sets a new technical benchmark for lightweight infrared small target detection under intense laser jamming. The core advantage of our method lies in the innovative combination of the real-time YOLO architecture with the linear global modeling capability of state space models. The proposed GLSS2D mechanism successfully resolves the dilemma of balancing global context and local details, a bottleneck that has persistently challenged traditional convolutional neural networks and vision transformers. However, we note a minor limitation: the current Laser-IR simulation protocol primarily models the macro-level energy envelope via superimposed saturated profiles and localized scattering halos . While this effectively captures the dominant operational interference patterns that disrupt image contrast, it simplifies certain low-level non-linear optical corner cases encountered in real field deployments, such as high-frequency diffraction rings and directional sensor register saturation tails. Relying solely on single-frame spatial modeling may leave a narrow margin of ambiguity under these extreme multi-spectral artifacts. Future work will expand the simulation pipeline by incorporating advanced wave-optics propagation models. Concurrently, we plan to introduce spatiotemporal context constraints, utilizing the distinct motion continuity of real airborne and missile targets against the random flashing or bleeding characteristics of complex laser artifacts to construct a full-time, robust dynamic anti-jamming system. Conclusions In this paper, we proposed the LIR-Mamba lightweight detection framework for the task of infrared small target detection under laser jamming. By deeply integrating a linear-complexity state space model with the real-time YOLO architecture, and innovatively designing a global-local selective scanning mechanism, this paper effectively resolves the issues of inadequate feature extraction capabilities and limited computational resources faced by traditional algorithms under intense jamming. Experimental results demonstrate that LIR-Mamba significantly improves the detection accuracy and robustness for tiny targets under laser spot occlusion and low signal-to-clutter ratio environments while ensuring low-latency inference, providing robust technical support for the intelligent upgrade of next-generation anti-jamming electro-optical detection systems.