TinyDark-YOLO for adaptive and lightweight object detection in low-light conditions Authors: Xuefang Zheng, Lifang Chen, Jiechi Zhang Published: 2026 ABSTRACT Low-light object detection frequently suffers from weak contrast, high noise, and poor performance on small objects. Existing methods often adopt image enhancement to improve detection performance, which results in substantial computational resource consumption. To address these challenges, we propose TinyDark-YOLO, an adaptive and efficient low-light object detection algorithm based on YOLO11. First, we design an Adaptive Gamma Enhancement (AGE) module to adaptively adjust the brightness of input images. Second, we introduce an Attention-based Intra-scale Feature Interaction (AIFI) module to construct global context modeling on high-level features and capture long-range spatial dependencies, improving the detection accuracy of small objects. Finally, we propose a Lite Efficient Head (LEHead) to reduce computational overhead while enhancing the model’s ability to capture weak features in dark environments. We conduct validation and ablation experiments on the ExDark dataset. The results demonstrate that the proposed model achieves an mAP@0.5 of 67.20% on the ExDark dataset, yielding an improvement of 2.39% over the baseline YOLO11n, while reducing computational complexity from 6.3 to 5.8 GFLOPs. TinyDark-YOLO also achieves competitive performance on the DarkFace dataset. Introduction Object detection in low-light environments is an important research direction in computer vision, with far-reaching applications in autonomous driving, intelligent surveillance, drone inspection. With advances in deep learning, object detection algorithms have reached maturity under normal illumination. Classic detectors including Faster R-CNN1, SSD2 and the YOLO series3–8 achieve competitive performance and attain a desirable trade-off between detection accuracy and inference speed. However, in low-light scenarios including nighttime, tunnels, cloudy days, and dim indoor environments, insufficient illumination results in weak contrast, missing fine-grained details, and pervasive noise, significantly impairing the performance of traditional detection models. With expanding applications in autonomous systems and intelligent surveillance, improving detection robustness in low-light conditions has become a pivotal research direction for bridging the gap between laboratory performance and real-world utility. Current low-light object detection methods can be generally classified into four categories: image enhancement-based methods, end-to-end enhancement-detection collaborative methods, detector structure optimization methods, and lightweight low-light detection models. The first category refers to image enhancement-based detection methods. These methods preprocess images through enhancement algorithms such as RetinexNet9, Zero-DCE10, EnlightenGAN11, SCI12, and RetinexFormer13 to improve brightness and contrast. Object detection is then conducted on the processed images. With more distinct feature information available, detection models can identify and locate targets with higher accuracy. This method is seldom used, as there is a mismatch between the optimization goals of image enhancement and object detection. The visual improvement achieved by enhancement cannot always be converted into effective features for detection models, which limits the further improvement of detection performance. The second category refers to end-to-end enhancement-detection collaborative methods. Cascading image enhancement and object detection modules enables collaborative optimization through joint training, effectively improving detection performance. Jiang et al.14 developed the LOL-YOLO model by cascading SCI12 with YOLOv8n, which improves detection accuracy while effectively controlling computational cost. Liu et al.15 proposed the IA-YOLO model for adverse weather conditions, which predicts image enhancement parameters via a lightweight neural network and cascades the image enhancement module with YOLOv3, boosting detection performance in adverse weather. Yin et al.16 presented the PE-YOLO low-light object detection framework, which decomposes images into four components of different resolutions using a Laplacian pyramid, and designs a detail processing module and a low-frequency enhancement filter to enhance component at each scale separately, further improving detection performance.While these end-to-end frameworks enable joint optimization of enhancement and detection, the additional enhancement modules greatly increase the overall computational load. The third category refers to detector structure optimization methods. Rather than relying on external image enhancement, these methods achieve superior low-light detection performance by optimizing the detector’s internal network. Luo et al.17 proposed VnDark, which constructs a feature enhancement sub-network via an iterative light enhancement curve mechanism and adopts a U-shaped encoder-decoder to extract distinct edge features from nighttime images, thereby improving the detection accuracy of objects in night scenes. Han et al.18 proposed the 3L-YOLO model adopting Switchable Atrous Convolution(SAC)and a multi-scale attention fusion mechanism, which greatly improves detection accuracy for multi-scale targets. Sun et al.19 improved the YOLO11 framework by introducing the Feature Fusion Attention Network (FFA-Net) and a triple attention mechanism, which effectively enhances the detection performance in low-light environments. Most detector optimization schemes only adjust network structures and lack dedicated designs for low-light scenarios, resulting in limited improvement in detection performance. The fourth category refers to lightweight detection models. Such methods simplify network architectures to reduce model parameters and computational overhead, thereby enabling efficient deployment on edge devices.Hui et al.20 proposed the WSA-YOLO model by introducing a lightweight weighted spatial attention mechanism into YOLOv7. It enhances object-region features at low computational cost while preserving detection accuracy. Zhuo et al.21 presented SCL-YOLOv11, a lightweight low-light detection model. It optimizes feature extraction modules to cut computational consumption and deliver reliable performance in low-light environments. Lightweight models simplify network layers and feature extraction branches to reduce parameters and computational overhead, and may show insufficient robustness in complex low-light environments. Although current methods perform adequately in low-light environments, complex scenarios still induce frequent missed and false detections. Detection accuracy degrades further when dealing with blurred or small objects in extreme darkness. We propose an adaptive low-light object detection algorithm named TinyDark-YOLO, which is based on YOLO11. The main contributions are as follows: (1)We develop TinyDark-YOLO, an adaptive end-to-end framework designed for low-light object detection. By adaptively learning optimal Gamma coefficient for different low-light images and utilizing a Lite Efficient Head, it achieves a superior balance between detection accuracy and computational efficiency. (2)We design an Adaptive Gamma Enhancement (AGE) module in the YOLO11 backbone to restore structural and detailed information of objects in low-light scenarios. Meanwhile, the Attention-based Intra-scale Feature Interaction (AIFI) module is introduced to model global contextual relations from high-level features. The two modules work together to enhance small object detection performance in low-light environments. (3)We design a Lite Efficient Head (LEHead), which employs a lightweight feature enhancement branch based on two-layer 3 × 3 group convolution for feature maps of each detection scale. LEHead achieves significant gains in inference speed and resource efficiency without compromising detection robustness. Related work Learnable illumination-aware enhancement technology Low-light images often suffer from insufficient brightness and missing details, which degrade the performance of object detection models. As a mainstream method for low-light image processing, illumination-aware enhancement technology adjusts brightness and restores image quality according to the illumination distribution of images. Adaptive gamma enhancement is a widely adopted method of learnable illumination-aware enhancement. Different from traditional gamma transformation with fixed correction coefficients, this method dynamically calculates gamma values pixel by pixel or in local regions based on regional brightness features, so as to balance global brightness and preserve local details. With the advancement of deep learning, data-driven approaches are capable of learning illumination transformation and image enhancement strategies automatically, achieving better adaptability to complex low-light scenarios. Debnath et al.22 proposed a low-light image enhancement method combining gamma coefficient learning and attention-based encoder-decoder network. The optimal gamma coefficients are obtained via network training, which promotes the evolution of gamma correction towards a learnable paradigm. At present, various learnable designs have been extensively applied in computer vision. Xue et al.23 adopted a learnable frequency-aware interaction module and multi-expert fusion mechanism to improve the model robustness against illumination variation, object occlusion and other complex interferences. In large vision model tasks, learnable visual prompts and prompt pruning strategies enable efficient fine-tuning of pre-trained models24. For Mirror Detection25,26, researchers utilize learnable depth calibration, dynamic prompt strategies and feature decoupling modules to make models adapt to feature distributions of various scenarios. Some scholars also proposed a recognition framework based on learnable nearest centroids27, which explores feature patterns in a data-driven manner and effectively improves the generalization and interpretability of algorithms. The above concepts, learnable architectures, dynamic feature interaction and adaptive modules provide valuable references for low-light image enhancement. In this paper, a learnable adaptive gamma enhancement module is designed. It adopts a lightweight neural network to predict gamma coefficients dynamically and realize pixel-level illumination correction. This method can effectively enhance low-light images while meeting the practical requirements of object detection tasks. Feature interaction and attention enhancement methods An important approach to boosting the accuracy of low-light object detection is to strengthen feature representation and explore inter-feature correlations. Feature interaction and attention mechanisms can guide models to focus on valid information and suppress background noise, which have been widely adopted in various visual tasks and have become a mainstream research direction in low-light detection. Low-light images generally suffer from degradation problems such as blurred details and mixed noise. Various feature interactions can connect local textures and global semantic information, making up for the insufficient representation capability of a single feature branch. Pan et al.28 proposed MedCTM, a hybrid CNN-Transformer-Mamba network, which divides feature learning into three modules: local feature extraction, long-range dependency modeling and cross-branch feature interaction. A bidirectional channel interaction attention is designed to enable dynamic information interaction between CNN and Mamba branches, achieving excellent detection performance. The attention mechanism assigns adaptive weights to make the model focus on target regions and reduce the negative impacts caused by irrelevant backgrounds and noise. For ship detection under complex backgrounds29, researchers introduce an improved Transformer structure and build a four-scale residual feature fusion network, which effectively solves the problems of missed and false detection of small ship in complex scenarios. In addition, methods based on clustering and prototype learning also provide new ideas for feature optimization30,31. Reconstructing Transformer cross-attention can enhance feature discrimination and algorithm interpretability, and meet the requirements of unified modeling for multiple tasks. For scenarios with poor image quality and severe noise interference32, heterogeneous experts and hierarchical perception networks model texture details and global structures separately via multi-branch architectures, which greatly improves the model’s anti-interference capability. Overall, feature interaction and attention enhancement methods can significantly optimize model feature representation. In this paper, an attention-based feature interaction module is introduced into the YOLO11 framework to adapt to low-light scenarios, which enhances feature representation. Lightweight model design How to compress model parameters and cut computational overhead while guaranteeing detection accuracy has become an important research direction for low-light object detection algorithms. Current lightweight detection methods are mainly developed from module simplification, network structure optimization and feature fusion strategies.Targeting industrial defect detection, Pan et al.33 designed a task-gated detection head to decouple classification and regression via lightweight architecture.The model only has 2.63 M parameters and realizes model slimming while retaining noise robustness and boundary awareness. To meet real-time detection requirements, YOLO-ROC34 adopts hierarchical channel compression, reducing parameters to 0.89 M and GFLOPs to 2.6, which balances small object accuracy and inference speed. RT-DSAFDet35, constructed with three components including flexible attention feature extraction, dynamic scale-aware fusion, and efficient spatial downsampling, achieves 1.8 M parameters and 4.6G FLOPs, boosting multi-scale detection performance with significantly reduced computational cost. The above researches verify that embedding efficient attention mechanisms and dynamic fusion structures into lightweight networks is an effective solution to balance detection accuracy, inference speed and model size, which provides references for lightweight detector design under low-light and complex scenarios. With the attention-based feature interaction module introduced to strengthen feature representation, we utilize group convolution to build a lightweight efficient detection head, further reducing model parameters and computational complexity. Method TinyDark-YOLO is based on YOLO11. In its backbone, we design an AGE module to solve the problems of insufficient brightness and poor object discriminability in low-light images. The AGE module adaptively predicts Gamma coefficient through a lightweight neural network, then performs illumination correction on the low-light input image based on the predicted coefficient. To address the missed detection of small objects in extremely low-light environments, we replace the original SPPF module in the backbone with the AIFI module from RT-DETR. This simplifies the spatial pyramid pooling structure and strengthens feature representation through global context interaction, achieving more accurate detection of multi-scale objects. We design the detect head LEHead using two-layer 3 × 3 group convolution to improve the feature representation of blurred, weak-textured, and low-contrast objects in low-light environments. LEHead enhances semantic feature representation, improving detection accuracy while lowering computational complexity. The structure diagram of TinyDark-YOLO is shown in Fig. 1. Fig. 1The structure diagram of TinyDark-YOLO.The baseline model of TinyDark-YOLO is YOLO11. The TinyDark-YOLO detection network is composed of three main components: Backbone, Neck, and Head. In the backbone, the AGE and AIFI modules are introduced to enhance the model’s detection capability in low-light environments. LEHead is proposed to reduce computational cost while preserving low‑light detection accuracy. The structure diagram of TinyDark-YOLO.The baseline model of TinyDark-YOLO is YOLO11. The TinyDark-YOLO detection network is composed of three main components: Backbone, Neck, and Head. In the backbone, the AGE and AIFI modules are introduced to enhance the model’s detection capability in low-light environments. LEHead is proposed to reduce computational cost while preserving low‑light detection accuracy. Adaptive gamma enhancement module (AGE) Inspired by IA-YOLO, we redesign the first convolutional layer of the YOLO11 backbone and replace it with AGE module. AGE module improves target discriminability through adaptive brightness enhancement, enhancing object detection accuracy. In AGE module, we design a lightweight neural network named CNN-Gamma as a Gamma coefficient predictor, CNN-Gamma can adaptively predict the Gamma coefficient based on the brightness of the input image. The details of AGE module are as follows: The input image with arbitrary resolution is downsampled to 256 × 256 via bilinear interpolation to reduce computational cost.The downsampled image is fed into CNN-Gamma to predict the Gamma coefficient.Perform gamma transformation on the original-resolution image with the predicted gamma coefficient to enhance its brightness.The enhanced image, as the output of module AGE, is used for subsequent feature extraction. The input image with arbitrary resolution is downsampled to 256 × 256 via bilinear interpolation to reduce computational cost. The input image with arbitrary resolution is downsampled to 256 × 256 via bilinear interpolation to reduce computational cost. The downsampled image is fed into CNN-Gamma to predict the Gamma coefficient. The downsampled image is fed into CNN-Gamma to predict the Gamma coefficient. Perform gamma transformation on the original-resolution image with the predicted gamma coefficient to enhance its brightness. Perform gamma transformation on the original-resolution image with the predicted gamma coefficient to enhance its brightness. The enhanced image, as the output of module AGE, is used for subsequent feature extraction. The enhanced image, as the output of module AGE, is used for subsequent feature extraction. As a pixel-wise filter, Gamma correction is essentially a power transformation operation, as shown in Eq. (1). The mapping function of Gamma transformation is differentiable with respect to both the input image and the Gamma correction, supporting back-propagation for network training. CNN-Gamma updates its own parameters according to gradients from back-propagation, and adjusts the predicted Gamma coefficient to make the enhanced image more favorable for detection. After repeated iterations, CNN-Gamma automatically learns optimal enhancement coefficient for different low-light images under the weak supervision of detection loss, achieving the goal of optimizing detection performance with image enhancement.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{O}\left(x,y\right)={\left(\frac{\boldsymbol{I}(x,y)}{255}\right)}^{\gamma\:}\times\:255$$\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}$$\:\boldsymbol{I}(\mathrm{x},\mathrm{y})$$\end{document} is the pixel value of original image at coordinates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(\mathrm{x},\mathrm{y})$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{O}\left(\mathrm{x},\mathrm{y}\right)$$\end{document} is the pixel value after Gamma transformation, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:\:$$\end{document}is the Gamma coefficient. As shown in Fig. 2, CNN-Gamma is composed of five convolutional blocks and two fully-connected layers. Each convolutional block includes a 3 × 3 convolution with stride 2 and a Leaky ReLU. The final fully-connected layer outputs the predicted Gamma coefficient. The output channel of these five convolutional layers are 16, 32, 32, 32 and 32, respectively. The CNN-Gamma network contains only 165 K parameters. Fig. 2The structure diagram of AGE module. The lightweight neural network CNN-Gamma is employed to predict the Gamma coefficient from the resized image. Gamma correction is then performed on the input image with the predicted coefficient, followed by a 3 × 3 convolution. The resulting feature is taken as the output of the AGE module and fed into the subsequent layer of TinyDark-YOLO. The structure diagram of AGE module. The lightweight neural network CNN-Gamma is employed to predict the Gamma coefficient from the resized image. Gamma correction is then performed on the input image with the predicted coefficient, followed by a 3 × 3 convolution. The resulting feature is taken as the output of the AGE module and fed into the subsequent layer of TinyDark-YOLO. In AGE module, the predicted Gamma coefficient is constrained within [0.01,0.99]. The Sigmoid activation function is used to map raw network outputs to the valid interval. In addition, value constraints are imposed to avoid extreme parameters. This design effectively prevents excessive noise enhancement and image detail loss, and ensures stable Gamma prediction throughout the training process. Attention-based intra-scale feature interaction module (AIFI) In low-light environments, YOLO11 still faces challenges in detecting small objects and edge-blurred objects. To enhance the attention of YOLO11 on small and blurred objects in low-light conditions, we introduce the AIFI module to achieve dynamic fusion of local and global information and optimize the correlated feature expression of targets. The AIFI module is a key component in the RT-DETR36 model. It only operates on the highest semantic feature layer, enabling intra-scale feature interactions through a single-scale Transformer encoder. It uses the Multi-Head Self-Attention mechanism (MHSA)37 to encode the highest feature layer, which effectively reduces the computational complexity. Compared with low-level feature maps, high-level feature maps contain richer semantic information. Performing self-attention operations on high-level feature maps can effectively capture the global correlations between target entities, which helps subsequent modules locate and recognize targets. The structure of the AIFI module is shown in Fig. 3, and its execution process is as follows: Input: The high-level feature map S from the backbone, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{S}\in\:{\mathbb{R}}^{\mathrm{B}\times\:\mathrm{C}\times\:\mathrm{H}\times\:\mathrm{W}}$$\end{document}, Input: The high-level feature map S from the backbone, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{S}\in\:{\mathbb{R}}^{\mathrm{B}\times\:\mathrm{C}\times\:\mathrm{H}\times\:\mathrm{W}}$$\end{document}, Input: The high-level feature map S from the backbone, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{S}\in\:{\mathbb{R}}^{\mathrm{B}\times\:\mathrm{C}\times\:\mathrm{H}\times\:\mathrm{W}}$$\end{document}, where B is the batch size, C is the number of channels, and H/W is the height/width of the feature map. 2)Serialization and Positional Encoding Embedding: Flatten the 2D feature map S into a 1D sequence and rearrange its dimensions, then add 2D sine-cosine positional encodings to preserve spatial position information. 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{X}}_{\mathrm{s}\mathrm{e}\mathrm{q}}=\mathrm{P}\mathrm{e}\mathrm{r}\mathrm{m}\mathrm{u}\mathrm{t}\mathrm{e}\left(\mathrm{F}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\left(\boldsymbol{S}\right)\right)\:\:\:\:\:\:\:{\quad\boldsymbol{X}}_{\mathrm{s}\mathrm{e}\mathrm{q}}\in\:{\mathbb{R}}^{B\times\:N\times\:C}$$\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}$$\:\boldsymbol{Q}=\boldsymbol{K}={\boldsymbol{X}}_{\mathrm{s}\mathrm{e}\mathrm{q}}+{\boldsymbol{P}}_{\mathrm{e}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{d}},\:\:\boldsymbol{V}={\boldsymbol{X}}_{\mathrm{s}\mathrm{e}\mathrm{q}}$$\end{document} 2)Serialization and Positional Encoding Embedding: Flatten the 2D feature map S into a 1D sequence and rearrange its dimensions, then add 2D sine-cosine positional encodings to preserve spatial position information. Serialization and Positional Encoding Embedding: Flatten the 2D feature map S into a 1D sequence and rearrange its dimensions, then add 2D sine-cosine positional encodings to preserve spatial position information. Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{X}}_{\mathrm{s}\mathrm{e}\mathrm{q}}$$\end{document} denotes the flattened and rearranged feature sequence, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N=H\times\:W$$\end{document} is sequence length, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{F}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\left(\right)$$\end{document} is flatten operation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{P}\mathrm{e}\mathrm{r}\mathrm{m}\mathrm{u}\mathrm{t}\mathrm{e}\left(\right)\:$$\end{document}is dimension rearrangement, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{Q},\boldsymbol{K},\boldsymbol{V}$$\end{document} are the query, key, and value matrices, respectively, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{P}}_{\mathrm{e}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{d}}$$\end{document} is the spatial position information. 3)Single-Layer Transformer Self-Attention Interaction: Based on the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{Q},\boldsymbol{K},\boldsymbol{V}$$\end{document} matrices, self-attention computation is performed using a single-layer Transformer encoder to model long-range dependencies between features, achieve sufficient interaction of global contextual information, and obtain the enhanced feature sequence. 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{X}}_{\mathrm{a}\mathrm{t}\mathrm{t}}={\boldsymbol{X}}_{\mathrm{s}\mathrm{e}\mathrm{q}}+\mathrm{D}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{u}\mathrm{t}\left(\mathrm{M}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{i}\mathrm{H}\mathrm{e}\mathrm{a}\mathrm{d}\mathrm{A}\mathrm{t}\mathrm{t}\mathrm{n}(\boldsymbol{Q},\boldsymbol{K},\boldsymbol{V}\right))\:\:$$\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}$$\:{\boldsymbol{X}}_{\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}1}=\mathrm{L}\mathrm{a}\mathrm{y}\mathrm{e}\mathrm{r}\mathrm{N}\mathrm{o}\mathrm{r}\mathrm{m}\left({\boldsymbol{X}}_{\mathrm{a}\mathrm{t}\mathrm{t}}\right)$$\end{document} 3)Single-Layer Transformer Self-Attention Interaction: Based on the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{Q},\boldsymbol{K},\boldsymbol{V}$$\end{document} matrices, self-attention computation is performed using a single-layer Transformer encoder to model long-range dependencies between features, achieve sufficient interaction of global contextual information, and obtain the enhanced feature sequence. Single-Layer Transformer Self-Attention Interaction: Based on the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{Q},\boldsymbol{K},\boldsymbol{V}$$\end{document} matrices, self-attention computation is performed using a single-layer Transformer encoder to model long-range dependencies between features, achieve sufficient interaction of global contextual information, and obtain the enhanced feature sequence. where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{M}\mathrm{u}\mathrm{l}\mathrm{t}\mathrm{i}\mathrm{H}\mathrm{e}\mathrm{a}\mathrm{d}\mathrm{A}\mathrm{t}\mathrm{t}\mathrm{n}\left(\boldsymbol{Q},\boldsymbol{K},\boldsymbol{V}\right)=\mathrm{S}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}\left(\frac{{\boldsymbol{Q}\boldsymbol{K}}^{T}}{\sqrt{{d}_{k}}}\right)\boldsymbol{V}$$\end{document} is self-attention operation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{X}}_{\mathrm{a}\mathrm{t}\mathrm{t}}$$\end{document} represents the feature matrix after self-attention calculation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{D}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{u}\mathrm{t}\left(\right)$$\end{document} is the dropout operation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{L}\mathrm{a}\mathrm{y}\mathrm{e}\mathrm{r}\mathrm{N}\mathrm{o}\mathrm{r}\mathrm{m}\left(\right)$$\end{document} is layer normalization, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{X}}_{\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}1}$$\end{document} denotes the feature matrix after the first layer normalization. 4)Feature Purification and Shape Reconstruction: Apply the nonlinear transformation to the feature sequence via the Feed-Forward Network (FFN), achieve feature refinement by residual connections, and then restore the sequence-form feature map to the 2D feature map of original dimensions༚6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{X}}_{\mathrm{f}\mathrm{f}\mathrm{n}}={\boldsymbol{X}}_{\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}1}+\mathrm{D}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{u}\mathrm{t}\left(\mathrm{F}\mathrm{F}\mathrm{N}\left({\boldsymbol{X}}_{\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}1}\right)\right)$$\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}$$\:{\boldsymbol{X}}_{\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}2}=\mathrm{L}\mathrm{a}\mathrm{y}\mathrm{e}\mathrm{r}\mathrm{N}\mathrm{o}\mathrm{r}\mathrm{m}\left({\boldsymbol{X}}_{\mathrm{f}\mathrm{f}\mathrm{n}}\right)$$\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}$$\:\boldsymbol{F}=\mathrm{R}\mathrm{e}\mathrm{s}\mathrm{h}\mathrm{a}\mathrm{p}\mathrm{e}\left({\boldsymbol{X}}_{\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}2}\right)\:\:\:\:\:\quad\boldsymbol{F}\in\:{\mathbb{R}}^{B\times\:C\times\:H\times\:W}$$\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}$$\:{\boldsymbol{X}}_{\mathrm{f}\mathrm{f}\mathrm{n}}$$\end{document} denotes the feature matrix processed by feed-forward network, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{F}\mathrm{F}\mathrm{N}\left(\right)$$\end{document} is Feed-Forward Network with GELU activation function, for non-linear feature enhancement, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{X}}_{\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}2}\:$$\end{document}denotes the output feature matrix of the Transformer encoder, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{R}\mathrm{e}\mathrm{s}\mathrm{h}\mathrm{a}\mathrm{p}\mathrm{e}\left(\right)$$\end{document} perform dimension restoration from the sequence to the 2D feature map, make the output dimensions match the input feature map S, and F denotes the reshaped 2D feature map. 5)Output: The enhanced high-level feature F, it fuses global context information, providing high-quality semantic support for subsequent cross-scale feature fusion and target query generation. Fig. 3The structure of AIFI module based on Reference36. The input high-level feature map is flattened and rearranged. Single-layer Transformer self-attention is adopted for feature interaction to obtain enhanced feature sequences. Non-linear transformation of FFN combined with residual connections is used for feature refinement, after which the features are restored to their original dimensions. The structure of AIFI module based on Reference36. The input high-level feature map is flattened and rearranged. Single-layer Transformer self-attention is adopted for feature interaction to obtain enhanced feature sequences. Non-linear transformation of FFN combined with residual connections is used for feature refinement, after which the features are restored to their original dimensions. Lite efficient head(LEHead) In low-light environments, object textures suffer from severe degradation and feature responses are significantly attenuated. This inevitably leads to suboptimal detection accuracy and a high incidence of missed and false detections. To address these issues, we design the LEHead for low-light scenarios, as shown in Fig. 4. Unlike the original detection head of YOLO11, which performs independent convolution on the channel dimension using depthwise convolution (DWConv), LEHead achieves feature enhancement by two-layer 3 × 3 group convolution, enabling it to fully capture local contextual information and edge contour features of low-light objects. This effectively strengthens the discriminability of weak features and improves the model’s detection performance in extremely low-light and complex lighting conditions. Fig. 4Structure diagram of LEHead. LEHead is constructed on group convolution. The input feature map is divided into G groups according to channel number C, and each group independently performs a 3 × 3 convolution followed by feature concatenation. Features processed by two layers of group convolution are fed into classification and regression branches for category and bounding box prediction. Structure diagram of LEHead. LEHead is constructed on group convolution. The input feature map is divided into G groups according to channel number C, and each group independently performs a 3 × 3 convolution followed by feature concatenation. Features processed by two layers of group convolution are fed into classification and regression branches for category and bounding box prediction. The structure of grouped convolution is illustrated in Fig. 4, 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}$$\:\boldsymbol{X}_{\mathrm{i}\mathrm{n}}\in\:{\mathbb{R}}^{C\times\:H\times\:W}$$\end{document}is adaptively split into G groups according to C channels, that is G=C//16. Each group of feature maps independently performs a 3 × 3 convolution to learn intra-group associated features, while batch normalization (BN) and SiLU activation is used to suppress noise. Finally, the convolution results of each group are concatenated into a complete feature map, maintaining the same output dimension as the input. The feature mapping formula are shown in (9)–(10):9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{X}}_{\mathrm{o}\mathrm{u}\mathrm{t}}={f}_{2}{f}_{1}\left({\boldsymbol{X}}_{\mathrm{i}\mathrm{n}}\right)$$\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}$$\:{f}_{i}\left(\boldsymbol{X}\right)=\mathrm{G}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}(\boldsymbol{X},k=3,G=C//16,{C}_{\mathrm{i}\mathrm{n}}={C}_{\mathrm{o}\mathrm{u}\mathrm{t}}=C)+\mathrm{B}\mathrm{N}\left(\boldsymbol{X}\right)+\mathrm{S}\mathrm{i}\mathrm{L}\mathrm{u}\left(\boldsymbol{X}\right)\:\:\:\quad i=\mathrm{1,2}$$\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}$$\:{\boldsymbol{X}}_{\mathrm{i}\mathrm{n}}$$\end{document} 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}$$\:{\boldsymbol{X}}_{\mathrm{o}\mathrm{u}\mathrm{t}}$$\end{document} is the feature map enhanced by two-layer 3 × 3 group convolution,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:k$$\end{document} is the kernel size, G is the number of groups, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{\mathrm{i}\mathrm{n}}$$\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}$$\:{C}_{\mathrm{o}\mathrm{u}\mathrm{t}}$$\end{document} denote input and output channels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{1}/{f}_{2}$$\end{document} are 3 × 3 group convolutions, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{G}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}\left(\right)\:$$\end{document}is the group convolution;\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{B}\mathrm{N}\left(\right)$$\end{document} is batch normalization༛and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{S}\mathrm{i}\mathrm{L}\mathrm{u}\left(\right)$$\end{document} is the activation function. LEHead achieves feature enhancement by group convolutions. It effectively reduces computational cost while maintaining feature representation capability.The following formulas only compare theoretical the Floating Point Operations (FLOPs) of a single group convolution and standard convolution under identical parameters. This simplified comparison illustrates the lightweight property, excluding extra overhead from stacked layers, batch normalization, activation functions and prediction branches. The FLOPs of group convolution are shown as follows:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{\mathrm{G}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}}=k\times\:k\times\:{C}_{\mathrm{i}\mathrm{n}}\times\:H\times\:W\times\:{C}_{\mathrm{o}\mathrm{u}\mathrm{t}}/G$$\end{document} The FLOPs of standard convolution are as follows:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}}=k\times\:k\times\:{C}_{\mathrm{i}\mathrm{n}}\times\:H\times\:W\times\:{C}_{\mathrm{o}\mathrm{u}\mathrm{t}}$$\end{document} When group number\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\:G=16$$\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}$$\:{F}_{\mathrm{G}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}}=\frac{1}{16}{F}_{\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}}$$\end{document} The result indicates that the computational cost of a single group convolution is merely 1/16 that of standard convolution, which verifies the lightweight design of LEHead. The feature maps processed by two-layer group convolution are respectively fed into the decoupled classification branch and the regression branch. In the classification branch, the category feature mapping is realized with a 1 × 1 pointwise convolutional layer, and then the probability normalization is performed with the Sigmoid activation function to obtain the predicted probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{\mathrm{c}\mathrm{l}\mathrm{s}}$$\end{document} of the object category:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{\mathrm{c}\mathrm{l}\mathrm{s}}=\sigma\:\left({\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}}_{1\times\:1}\left({\boldsymbol{X}}_{\mathrm{o}\mathrm{u}\mathrm{t}}\right)\right)$$\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}$$\:{\boldsymbol{X}}_{\mathrm{o}\mathrm{u}\mathrm{t}}$$\end{document} is the feature map enhanced by the two-layer group convolution, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}}_{1\times\:1}\left(\right)$$\end{document} is pointwise convolution, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:\left(\right)$$\end{document} is Sigmoid activation function. The regression branch performs a 1 × 1 pointwise convolution on the input feature map to obtain the predicted features of the discrete distribution of the object bounding box. Then, through a decoding operation supervised by the Distribution Focal Loss (DFL), the integral of the discrete distribution is converted into continuous position offsets. By using the downsampling stride of the feature map and dynamic anchor boxes, the coordinates of the object bounding box at the original image scale are computed and output.15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{B}\boldsymbol{b}\boldsymbol{o}\boldsymbol{x}=\mathrm{D}\mathrm{F}\mathrm{L}\mathrm{D}\mathrm{e}\mathrm{c}\mathrm{o}\mathrm{d}\mathrm{e}\left({\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}}_{1\times\:1}\left({\boldsymbol{X}}_{\mathrm{o}\mathrm{u}\mathrm{t}}\right)\right)\cdot\:\boldsymbol{s}+\boldsymbol{A}$$\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}$$\:{\boldsymbol{X}}_{\mathrm{o}\mathrm{u}\mathrm{t}}\:$$\end{document}is the feature map enhanced by the two-layer group convolution, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}}_{1\times\:1}\left(\right)$$\end{document} is the pointwise convolution, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{D}\mathrm{F}\mathrm{L}\mathrm{D}\mathrm{e}\mathrm{c}\mathrm{o}\mathrm{d}\mathrm{e}\left(\right)$$\end{document} is the box coordinate decoding operation supervised by DFL, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{s}\:$$\end{document}is the downsampling stride of the feature map, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{A}$$\end{document} is the dynamically generated center coordinates of anchor boxes, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{B}\boldsymbol{b}\boldsymbol{o}\boldsymbol{x}$$\end{document} is the final output coordinates of object bounding box at the original image scale. Experiments Datasets We used the public dataset Exclusively-Dark-Image-Dataset (ExDark)38 to evaluate the performance of our model. ExDark is a professional low-light dataset released by the Loh team in 2019. It contains 7,363 low-light images, covering 10 different lighting conditions ranging from extremely dark to dimly lit environments, providing a comprehensive test platform for low-light object detection research. The dataset includes 12 common object categories: bicycle, boat, bottle, bus, car, cat, chair, cup, dog, motorcycle, people, and table.We set the random seed to 42 and split the dataset at a 7:1:2 ratio, resulting in 5,154 images for training, 736 for validation, and 1,473 for testing. All ablation and comparative experiments adopted this same split to eliminate the influence of random perturbations, thereby ensuring reliable and reproducible results. We also adopt the public dataset DarkFace39 to evaluate the robustness of TinyDark-YOLO. DarkFace dataset was released by Peking University in 2019, it is an authoritative dataset specifically for low-light and night-time face detection. All images in the DarkFace dataset are captured in real night-time scenes, covering complex environments such as teaching buildings, streets, bridges, and parks. The dataset contains 6,000 images with 50,399 annotated face boxes. For our experiments, this dataset was also split at a ratio of 7:1:2 with the random seed set to 42, obtaining 4200 training, 600 validation and 1200 test images. All subsequent experiments on the DarkFace dataset adopted this partition method. Experimental details For fair comparison, all experiments were conducted under same experimental conditions and training parameters. The experimental environment used in this study is shown in Table 1. During training, the input image size was set to 640 × 640, and the SGD optimizer was employed. Various data augmentation methods were applied, with Mosaic augmentation enabled throughout the training process. The model was trained from scratch without pre-trained weights, and the weights were initialized with He normal initialization. The confidence threshold was set to 0.25 and the NMS threshold to 0.7 for inference. Detailed experimental parameters are presented in Table 2. Table 1Experimental environment settings.EnvironmentSetOperating SystemWindows10CPUIntel(R) Core(TM) i5GPUNVIDIA GeForce RTX 5060Development EnvironmentPycharmCompiler EnvironmentPython3.11Deep Learning FrameworkPytorch2.9.0 + cu128 Table 1Experimental environment settings.EnvironmentSetOperating SystemWindows10CPUIntel(R) Core(TM) i5GPUNVIDIA GeForce RTX 5060Development EnvironmentPycharmCompiler EnvironmentPython3.11Deep Learning FrameworkPytorch2.9.0 + cu128 Experimental environment settings. Table 2Experimental parameter settings.Parameter NameParameter ValueInput image size640 × 640Batch size32Training epochs200OptimizerSGDLearning rate settingFixed learning rate, initial value 0.01Data loading workers4Random seed0Mosaic augmentationclose_mosaic = 0 (enabled throughout training)Data augmentation typesHorizontal flip, Scale scaling, Random translation, HSV color transformationWeight initializationHe normal initializationPretrained weightsNot usedConfidence threshold0.25NMS threshold0.7 Table 2Experimental parameter settings.Parameter NameParameter ValueInput image size640 × 640Batch size32Training epochs200OptimizerSGDLearning rate settingFixed learning rate, initial value 0.01Data loading workers4Random seed0Mosaic augmentationclose_mosaic = 0 (enabled throughout training)Data augmentation typesHorizontal flip, Scale scaling, Random translation, HSV color transformationWeight initializationHe normal initializationPretrained weightsNot usedConfidence threshold0.25NMS threshold0.7 Experimental parameter settings. Evaluation metrics We adopt GFLOPs and parameters (Params) to assess the complexity of the model, and use precision (P), recall (R), and mean average precision (mAP) to evaluate the detection accuracy of the model. A smaller number of parameters indicates a more lightweight model, a lower GFLOPs means a lower computational complexity of the algorithm. Precision indicates the proportion of true positive samples among all samples predicted as positive by the model. A higher precision value indicates a lower false detection rate of the model. Recall represents the ratio of correctly predicted positive samples to all actual positive samples. Higher recall indicates fewer missed detections. The calculation formulas are shown in Equations (16)~(17).16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P=\frac{TP}{TP+FP}\times\:100\%$$\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}$$\:R=\frac{TP}{TP+FN}\times\:100\%$$\end{document} where TP(True Positive) is the number of positive samples correctly predicted by the model, FP(False Positive) is the number of negative samples incorrectly predicted as positive, and FN (False Negative) is the number of positive samples incorrectly predicted as negative. In object detection, multiple classes are usually detected, each with an average precision(AP), AP represents the average precision under different recall rates and corresponds to the area under the P-R curve. The mean AP of all classes is denoted as mAP, as shown in Equations (18)~(19). The model’s mAP is usually evaluated using two metrics: mAP@0.5 and mAP@0.5:0.95. A higher mAP indicates better overall detection accuracy.18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:AP={\int\:}_{0}^{1}P\left(R\right)dR$$\end{document}19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:mAP=\frac{\sum\:_{i=1}^{N}A{P}_{i}}{N}$$\end{document} Effectiveness experiments of the AGE module To verify the effectiveness of the AGE module, comparative experiments are conducted between fixed Gamma correction and adaptive Gamma correction. The experimental environment and parameters are shown in Tables 1 and 2. All network structures, training parameters, and datasets are kept consistent, and all models are trained from scratch to ensure a fair comparison. The evaluation metrics include Precision (P), Recall (R), mAP@0.5, and mAP@0.5:0.95 (see Table 3). Table 3Performance comparison between fixed Gamma correction and AGE.MethodP(%)R(%)mAP@0.5(%)mAP@0.5:0.95(%)AGE(Adaptive γ)73.74%57.76%65.42%39.54%γ = 0.270.41%55.84%65.19%39.52%γ = 0.572.04%57.38%65.03%38.97%γ = 0.872.00%58.17%64.83%38.90%γ = 1.570.35%58.29%63.82%38.11%γ = 2.066.13%56.17%61.38%36.62% Table 3Performance comparison between fixed Gamma correction and AGE.MethodP(%)R(%)mAP@0.5(%)mAP@0.5:0.95(%)AGE(Adaptive γ)73.74%57.76%65.42%39.54%γ = 0.270.41%55.84%65.19%39.52%γ = 0.572.04%57.38%65.03%38.97%γ = 0.872.00%58.17%64.83%38.90%γ = 1.570.35%58.29%63.82%38.11%γ = 2.066.13%56.17%61.38%36.62% Performance comparison between fixed Gamma correction and AGE. The results show that the AGE module significantly outperforms fixed Gamma correction. It dynamically adjusts parameters according to image illumination, achieving better precision and overall detection accuracy, and exhibits stronger adaptability than fixed parameter methods. In addition, detection performance degrades noticeably when γ > 1, which validates the rationality of constraining γ within [0.01,0.99] for low-light object detection. Ablation experiment To verify the effectiveness of AGE, AIFI and LEHead modules in our TinyDark-YOLO, we conduct ablation experiments on the ExDark dataset. Using YOLO11n as the baseline model, we evaluate the effectiveness of each module by adding the improved modules one by one. Table 4 presents the ablation results on the test set. Notably, integrating the AGE module alone elevates the mAP@0.5 from 64.81% to 65.42%, outperforming the baseline by 0.61%. Furthermore, the precision is boosted by 1.53%. These results demonstrate that the introduction of AGE enhances the model’s ability to perceive objects. Introducing the AIFI module alone improves the mAP@0.5 to 66.86%, outperforming the baseline by 2.05%. Meanwhile, the recall is improved by 2.29%. These results show that the attention-based feature interaction mechanism effectively captures global contextual information, improving object localization and recognition performance. Integrating LEHead alone elevates the mAP@0.5 from 64.81% to 66.18%, outperforming the baseline by 1.37%. Furthermore, the computational cost is reduced from 6.3 to 5.1 GFLOPs. These results suggests that the model’s detection capability under extremely low-light and complex illumination conditions is enhanced, with a significant decrease in computational cost. The combination of AGE and AIFI achieves a mAP@0.5 of 66.55%, representing an improvement of 1.74% over the baseline model, while recall is increased to 61.15%. Both mAP@0.5 and recall are significantly improved with a slight increase in computational cost. The combination of AGE and LEHead yields a mAP@0.5 of 66.24%, which is 1.43% higher than the baseline model. Meanwhile, computational cost is reduced by 1.1 GFLOPs and parameters is decreased by 3.9%. The combination of AIFI and LEHead improves mAP@0.5 and mAP@0.5:0.95 to 66.63% and 40.30% respectively, reduces computational cost by 0.9 GFLOPs and increases parameters by 0.36 M, achieving a balance between accuracy improvement and acceptable parameter increment. Ultimately, TinyDark-YOLO integrates the three modules, achieves a mAP@0.5 of 67.20% on the ExDark dataset, outperforming the baseline by 2.39%. In addition, precision, recall and mAP@0.5:0.95 are increased by 1.58%, 2.51% and 1.56% respectively, while computational cost is reduced by 0.5 GFLOPs. The ablation experiments show that the AGE, AIFI and LEHead are all effective and can work collaboratively. The joint use of these modules achieves superior performance, realizing a balanced improvement in performance metrics while maintaining the lightweight advantage of the model. Table 4Ablation experimental results of TinyDark-YOLO.AGEAIFILEHeadP(%)R(%)mAP@0.5(%)mAP@0.5:0.95(%)GFLOPsParams(M)---72.2158.4564.8138.956.32.58√--73.7457.7665.4239.546.72.75-√-73.2360.7466.8640.276.63.21--√71.0459.6166.1839.995.12.32√√-72.1761.1566.5540.247.03.38√-√72.0460.0366.2439.805.22.48-√√72.3559.7466.6340.305.42.94√√√73.7960.9667.2040.515.83.10 Table 4Ablation experimental results of TinyDark-YOLO.AGEAIFILEHeadP(%)R(%)mAP@0.5(%)mAP@0.5:0.95(%)GFLOPsParams(M)---72.2158.4564.8138.956.32.58√--73.7457.7665.4239.546.72.75-√-73.2360.7466.8640.276.63.21--√71.0459.6166.1839.995.12.32√√-72.1761.1566.5540.247.03.38√-√72.0460.0366.2439.805.22.48-√√72.3559.7466.6340.305.42.94√√√73.7960.9667.2040.515.83.10 Ablation experimental results of TinyDark-YOLO. Multiple experiment After validating the effectiveness of each improved module, we conduct multiple experiments to reduce the impact of random factors and evaluate model stability. The ExDark dataset is re-split using different random seeds. These experiments are performed on the baseline YOLO11n and our TinyDark-YOLO. The results are presented as mean ± standard deviation in Table 5. TinyDark-YOLO outperforms the baseline across all metrics with a smaller standard deviations, demonstrating its good stability and robustness for detection tasks. Table 5Multiple experimental results of YOLO11n and TinyDark-YOLO.ModelSeedP(%)R(%)mAP@0.5(%)mAP@0.5:0.95(%)YOLO11n4272.21%58.45%64.81%38.95%2370.47%58.27%64.88%39.33%12370.07%60.73%65.15%40.68%45672.71%54.74%63.57%38.38%78971.94%58.36%64.61%39.03% 71.48%±1.15 58.11%±2.15 64.60%±0.61 39.27%±0.86 Ours(TinyDark-YOLO)4273.79%60.96%67.20%40.51%2373.51%58.88%66.37%40.08%12373.94%58.99%67.45%41.37%45671.89%59.26%66.12%39.91%78971.57%58.77%66.68%39.30% 72.94%±1.12 59.37%±0.91 66.76%±0.56 40.23%±0.77 The dataset is split with different random seeds for five runs. Bold values indicate the mean ± standard deviation of multiple experiments. Table 5Multiple experimental results of YOLO11n and TinyDark-YOLO.ModelSeedP(%)R(%)mAP@0.5(%)mAP@0.5:0.95(%)YOLO11n4272.21%58.45%64.81%38.95%2370.47%58.27%64.88%39.33%12370.07%60.73%65.15%40.68%45672.71%54.74%63.57%38.38%78971.94%58.36%64.61%39.03% 71.48%±1.15 58.11%±2.15 64.60%±0.61 39.27%±0.86 Ours(TinyDark-YOLO)4273.79%60.96%67.20%40.51%2373.51%58.88%66.37%40.08%12373.94%58.99%67.45%41.37%45671.89%59.26%66.12%39.91%78971.57%58.77%66.68%39.30% 72.94%±1.12 59.37%±0.91 66.76%±0.56 40.23%±0.77 The dataset is split with different random seeds for five runs. Bold values indicate the mean ± standard deviation of multiple experiments. Multiple experimental results of YOLO11n and TinyDark-YOLO. Ours (TinyDark-YOLO) The dataset is split with different random seeds for five runs. Bold values indicate the mean ± standard deviation of multiple experiments. Comparison experiment on ExDark On the ExDark dataset, We compare with general object detection algorithms such as YOLOv5n, YOLOv8n, YOLO26n, as well as advanced low-light object detection algorithms, including IA-YOLO15, Zero_DCE+YOLO11n10, SCI+YOLO11n12, Sun19, and SCL-YOLOv1121. All methods are retrained and inferred on this dataset under unified configurations. Consistent training epochs, data augmentation strategies and confidence thresholds are adopted for fair comparison, following the experimental settings in Tables 1 and 2.The performance of different models on the ExDark dataset is presented in Table 6. Table 6Performance of different models on the ExDark dataset.ModelP(%)R(%)mAP@0.5(%)mAP@0.5:0.95(%)GFLOPsParams(M)RT-DETR-L64.4154.1853.5733.42108.032.8YOLOv5n71.4756.0363.5537.747.12.51YOLOv8n73.2959.6365.9439.358.13.00YOLOv10n69.6756.4962.2937.976.52.27YOLO11n72.2158.4564.8138.956.32.58YOLO12n68.9955.3162.6537.835.82.51YOLO26n71.0254.4161.5036.975.22.38IA-YOLO1571.8456.3262.4937.9912.712.30Zero_DCE+YOLO11n1072.3055.6063.9038.5073.73.28SCI+YOLO11n1272.4458.6465.5339.208.32.58Sun1974.0056.2064.0038.9010.14.10SCL-YOLOv112174.9859.2364.5040.204.71.6Ours73.7960.9667.2040.515.83.10 Table 6Performance of different models on the ExDark dataset.ModelP(%)R(%)mAP@0.5(%)mAP@0.5:0.95(%)GFLOPsParams(M)RT-DETR-L64.4154.1853.5733.42108.032.8YOLOv5n71.4756.0363.5537.747.12.51YOLOv8n73.2959.6365.9439.358.13.00YOLOv10n69.6756.4962.2937.976.52.27YOLO11n72.2158.4564.8138.956.32.58YOLO12n68.9955.3162.6537.835.82.51YOLO26n71.0254.4161.5036.975.22.38IA-YOLO1571.8456.3262.4937.9912.712.30Zero_DCE+YOLO11n1072.3055.6063.9038.5073.73.28SCI+YOLO11n1272.4458.6465.5339.208.32.58Sun1974.0056.2064.0038.9010.14.10SCL-YOLOv112174.9859.2364.5040.204.71.6Ours73.7960.9667.2040.515.83.10 Performance of different models on the ExDark dataset. As shown in Table 6, RT-DETR-L has large Params with a computational cost of 108 GFLOPs, but its mAP@0.5 is only 53.57%, indicating that it is not suitable for low-light scenarios. Among YOLO-series models, YOLOv10n and YOLO26n have the fewest Params and low computational complexity, but their detection accuracy in low-light environments is the worst. Our model TinyDark-YOLO has a similar Params to YOLOv8n, with a computational cost of only 5.8 GFLOPs, which is decreased by 2.3 GFLOPs relative to YOLOv8n. Furthermore, it performs better than YOLOv8n in all indicators, the precision, recall, mAP@0.5, and mAP@0.5:0.95 are improved by 0.5%, 1.33%, 1.26%, and 1.16%, respectively. Compared with YOLOv5n, YOLO11n and YOLO12n, the mAP@0.5 of TinyDark-YOLO is improved by 3.65%, 2.39% and 4.55% respectively. Compared with the algorithm proposed by Sun21, the precision is reduced by 0.21%, but the mAP@0.5 is increased by 3.2% and the computational cost is reduced by 4.3 GFLOPs. Compared with IA-YOLO, Zero-DCE+YOLO11n and SCI+YOLO11n, TinyDark-YOLO can perform detection directly without cascading image enhancement modules. Its computational cost is reduced by 6.9 GFLOPs, 67.9 GFLOPs and 2.5 GFLOPs respectively, while mAP@0.5 is increased by 4.71%, 3.3% and 1.67% respectively. Experimental results show that TinyDark-YOLO achieves optimal detection accuracy while maintaining the lightweight property of the model, and it shows favorable effectiveness in low-light object detection tasks. As shown in Table 7, under the unified test setting of RTX5060, 640 × 640 input and PyTorch framework, TinyDark-YOLO achieves an FPS comparable to YOLO11n and YOLO12n and slightly higher than YOLOv8n. This demonstrates that our proposed TinyDark-YOLO balances favorable detection accuracy and fast inference speed.20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:FPS=\frac{1000}{{T}_{pre}+{T}_{i}+{T}_{post}}$$\end{document} Where FPSis in frames per second (f/s), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{pre}$$\end{document} is the preprocessing time, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{i}$$\end{document} is the inference time, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{post}$$\end{document} is the postprocessing time per image. Table 7Comparison of FPS of different models.ModelsDeviceFPSModel Size(MB)YOLOv8nNVIDIA GeForce RTX 5060394.66.0YOLO11n457.35.2YOLO12n461.15.2Ours415.06.2 Table 7Comparison of FPS of different models.ModelsDeviceFPSModel Size(MB)YOLOv8nNVIDIA GeForce RTX 5060394.66.0YOLO11n457.35.2YOLO12n461.15.2Ours415.06.2 Comparison of FPS of different models. To verify the generalizability of TinyDark-YOLO for object detection in low-light environments, we evaluate its detection performance on different categories. The detection accuracy (mAP@0.5) of different models on the 12 object categories of ExDark are shown in Table 8. Table 8Experimental results of category-wise comparison for different models on ExDark (mAP@0.5).Model ClassYOLOv5nYOLOv8nYOLOv10nYOLO11nYOLO12nYOLO26nIA-YOLOSunOursBicycle (%)74.3476.4075.0873.2672.3473.1072.6273.00 76.41 Boat (%)58.2362.3857.4262.5360.9858.8758.1663.90 65.31 Bottle (%)61.0264.6861.0565.0463.7162.7662.1355.00 68.17 Bus (%)80.9383.2385.3183.0182.0279.3979.1883.80 87.39 Car (%)73.0874.4672.2473.8570.6971.3672.0875.52 75.54 Cat (%)59.55 63.02 55.5158.5455.4456.0358.8359.8062.28Chair (%)50.6752.1149.5651.6448.4747.9350.1852.50 52.93 Cup (%)54.9658.8254.3159.2455.5256.3556.0352.00 59.81 Dog (%)67.16 70.12 64.6367.3466.2263.4663.4964.9067.34Motorbike (%)66.3968.6461.7466.7862.9260.5365.8164.40 69.27 People (%)67.8767.2465.0168.9766.3463.5463.78 71.20 68.68Table (%)48.3550.1645.5647.4947.1544.7347.5951.70 53.29 All (%)63.5565.9462.2964.8162.6561.5062.4964.00 67.20 Table 8Experimental results of category-wise comparison for different models on ExDark (mAP@0.5).Model ClassYOLOv5nYOLOv8nYOLOv10nYOLO11nYOLO12nYOLO26nIA-YOLOSunOursBicycle (%)74.3476.4075.0873.2672.3473.1072.6273.00 76.41 Boat (%)58.2362.3857.4262.5360.9858.8758.1663.90 65.31 Bottle (%)61.0264.6861.0565.0463.7162.7662.1355.00 68.17 Bus (%)80.9383.2385.3183.0182.0279.3979.1883.80 87.39 Car (%)73.0874.4672.2473.8570.6971.3672.0875.52 75.54 Cat (%)59.55 63.02 55.5158.5455.4456.0358.8359.8062.28Chair (%)50.6752.1149.5651.6448.4747.9350.1852.50 52.93 Cup (%)54.9658.8254.3159.2455.5256.3556.0352.00 59.81 Dog (%)67.16 70.12 64.6367.3466.2263.4663.4964.9067.34Motorbike (%)66.3968.6461.7466.7862.9260.5365.8164.40 69.27 People (%)67.8767.2465.0168.9766.3463.5463.78 71.20 68.68Table (%)48.3550.1645.5647.4947.1544.7347.5951.70 53.29 All (%)63.5565.9462.2964.8162.6561.5062.4964.00 67.20 Experimental results of category-wise comparison for different models on ExDark (mAP@0.5). TinyDark-YOLO achieves the best detection performance on 9 out of 12 categories, and ranks second in 2 categories. For small objects including Bottle, Cat and Cup, the detection accuracy of TinyDark-YOLO is increased by 3.13%, 3.74% and 0.57% respectively compared with the baseline model YOLO11n. For the Dog category, the detection accuracy of TinyDark-YOLO is equal to the baseline. In summary, TinyDark-YOLO achieves the best performance among the compared methods. Table 9 presents the APS, APM and APL metrics of different methods split by object scale on the ExDark dataset. Compared with YOLOv8n, YOLO11n and YOLO12n, our TinyDark-YOLO obtains the highest AP values on small, medium and large objects. Specifically, the proposed method achieves an APS of 16.98%, which is 2.37%, 2.53% and 3.64% higher than YOLOv8n, YOLO11n and YOLO12n respectively. Meanwhile, consistent accuracy improvements are also observed on medium and large objects, verifying the superiority of our method for multi-scale object detection in low-light conditions. Table 9AP results of different methods for objects with various sizes.ModelsAPSAPMAPLYOLOv8n14.61%24.93%44.64%YOLO11n14.45%24.10%44.79%YOLO12n13.34%22.35%43.49%Ours16.98%25.85%45.64% Table 9AP results of different methods for objects with various sizes.ModelsAPSAPMAPLYOLOv8n14.61%24.93%44.64%YOLO11n14.45%24.10%44.79%YOLO12n13.34%22.35%43.49%Ours16.98%25.85%45.64% AP results of different methods for objects with various sizes. Comparison experiment on DarkFace To further verify the robustness and generalizability of TinyDark-YOLO for object detection in low-light environments, we conduct experiments on the DarkFace dataset. All models are trained from scratch on this dataset, with consistent experimental settings and parameters listed in Tables 1 and 2.The results are shown in Table 10. Table 10Generalization experimental results on the DarkFace dataset.ModelP(%)R(%)mAP@0.5(%)mAP@0.5:0.95(%)yolov5n72.7748.8253.8922.74yolov8n72.9149.6155.0023.45yolov10n72.4148.8453.6722.68yolo11n73.2949.2655.4722.92yolo12n71.7748.5153.6822.87yolo26n71.5951.1556.5723.90IA-YOLO71.3249.5754.7822.13Zero_DCE+YOLO11n72.5449.7356.1022.96SCI+YOLO11n72.3948.8756.2822.77Ours75.1051.5557.8924.59 Table 10Generalization experimental results on the DarkFace dataset.ModelP(%)R(%)mAP@0.5(%)mAP@0.5:0.95(%)yolov5n72.7748.8253.8922.74yolov8n72.9149.6155.0023.45yolov10n72.4148.8453.6722.68yolo11n73.2949.2655.4722.92yolo12n71.7748.5153.6822.87yolo26n71.5951.1556.5723.90IA-YOLO71.3249.5754.7822.13Zero_DCE+YOLO11n72.5449.7356.1022.96SCI+YOLO11n72.3948.8756.2822.77Ours75.1051.5557.8924.59 Generalization experimental results on the DarkFace dataset. As shown in the Table 10, on the DarkFace dataset, TinyDark-YOLO reaches 75.1% precision, 51.55% recall, 57.89% mAP@0.5 and 24.59% mAP@0.5:0.95, outperforming the baseline by 1.81%, 2.29%, 2.42% and 1.67%, respectively, ranking first in all four metrics. Among other models, YOLO26n performs best among the YOLO-series models, with a mAP@0.5 of 56.57%. The other three models cascading image enhancement modules(IA-YOLO, Zero_DCE+YOLO11n, SCI+YOLO11n) achieve accuracy improvements over their base models. However, compared with our TinyDark-YOLO, they have significantly more parameters and higher computational cost, showing obvious efficiency disadvantages. Visualization analysis Figure 5 shows the precision-recall (P-R) curves of TinyDark-YOLO and the baseline YOLO11n on the ExDark dataset, which intuitively presents the performance improvement of TinyDark-YOLO on 12 categories. As illustrated in Fig. 5, 10 out of the 12 object categories achieve improved detection performance. The Dog category is on par with the baseline, while the People category decreases slightly by 0.29%. In particular, the mAP@0.5 of the Bus, Cat and Table is increased by 4.38%, 3.74% and 5.8% respectively compared with the baseline. These results show that TinyDark-YOLO can effectively improve object detection accuracy under low-light conditions. Fig. 5Comparison of P-R curves before and after model improvement(mAP@0.5) (a) P-R curves of the baseline model YOLO11n (b) P-R curves of the proposed TinyDark-YOLO. Comparison of P-R curves before and after model improvement(mAP@0.5) (a) P-R curves of the baseline model YOLO11n (b) P-R curves of the proposed TinyDark-YOLO. As shown in Fig. 6, we conducted predictions on the ExDark test set and selected representative images for result visualization. These visualizations clearly demonstrate the qualitative differences in detection performance among TinyDark‑YOLO, YOLO11n, and YOLO12n. Fig. 6Visualized detection results on the test set of ExDark. From left to right: ground-truth, YOLO11n, YOLO12n, and TinyDark-YOLO, with a confidence threshold of 0.25. The images are from the public ExDark dataset. Visualized detection results on the test set of ExDark. From left to right: ground-truth, YOLO11n, YOLO12n, and TinyDark-YOLO, with a confidence threshold of 0.25. The images are from the public ExDark dataset. In extremely low-light environments, as shown in the first and second rows of Fig. 6, TinyDark-YOLO exhibits stronger detection performance, effectively avoiding missed and false detections. In strong reflection and complex lighting environments, as shown in the third, fourth, and fifth rows of Fig. 6, TinyDark-YOLO detects blurred objects more accurately and effectively, improving detection performance for blurred objects. For small objects in low-light environments, as shown in the fifth, sixth, and seventh rows of Fig. 6, TinyDark-YOLO can more effectively detect small objects while maintaining high accuracy. In summary, compared with YOLO11n and YOLO12n, the proposed TinyDark-YOLO significantly improves the detection performance in various complex scenarios. It not only resolves the missed and false detection problems but also enhances the generalizability and practicality of the model in complex low-light environments. The detection performance for extremely dark environments, blurred objects, and small objects is significantly improved, validating the effectiveness of TinyDark-YOLO for low-light object detection tasks. Analysis of experimental limitations Although the TinyDark-YOLO model achieves notable performance in low-light scenarios, it still has several limitations in complex scenarios. The model is prone to false detections for objects with highly similar features, such as cats and dogs, bicycles and motorcycles, as illustrated in Fig. 7 (a). Meanwhile, missed detections frequently occur in cases of target occlusion and densely distributed targets, as shown in Fig. 7(b). We will conduct targeted optimizations in future research to mitigate the above problems, further improving the detection robustness and generalization ability of the model in complex practical environments. Fig. 7Failed detection examples. (a) Failed detection cases of objects with similar features. (b) Failed detection cases under dense and occluded scenarios. The images are from the public ExDark dataset. Failed detection examples. (a) Failed detection cases of objects with similar features. (b) Failed detection cases under dense and occluded scenarios. The images are from the public ExDark dataset. Conclusion To address the challenges of low detection accuracy, frequent false and missed detections, and inaccurate localization caused by degraded images in low-light environments, we propose TinyDark-YOLO, an adaptive low-light object detection algorithm that does not need to cascade image enhancement modules, with a simple structure and low computational cost. We design the AGE module in the backbone of YOLO11 to automatically perceives illumination changes in low-light environments. Furthermore, We introduce the AIFI module, which effectively enhances the structural and detailed features of small objects in low-light environments and improves the model’s small object detection performance under extremely low-light conditions. Finally, we design LEHead based on two-layer group convolution, which significantly reduces computational cost while ensuring detection accuracy in low-light scenarios. Experiments on the ExDark and DarkFace datasets show that TinyDark-YOLO achieves notably detection performance in low-light environments. Although TinyDark-YOLO achieves promising detection results, future work will focus on two aspects: first, conducting experiments on multiple datasets and real-world scenarios to further verify the model’s generalization capability; second, exploring feature interaction and attention enhancement approaches to alleviate false detections for similar-feature objects and missed detections under dense occlusion, thus improving the model’s robustness and practicability.