Multi-scale feature and local detail enhancement fused low-illumination target detection method and system

Abstract

The invention provides a low-illumination target detection method and system fusing multi-scale features and local detail enhancement. The system comprises a feature extraction network based on a multi-pooling pyramid and cross-stage double-mixed attention, a dynamic detail semantic fusion pyramid network and double groups of detection heads. The feature extraction network based on the multi-pooling pyramid and the cross-stage double-mixed attention comprises a convolutional layer, a C2PSA module, an MPSPPF module and a CSP-EDHAN module; the dynamic detail semantic fusion pyramid network is used for fusing shallow high-frequency details and deep semantic information through top-down and bottom-up multi-scale feature fusion and introducing a surface detail fusion module into multiple scales, so as to output a fused feature map; the double groups of detection heads adopt a decoupling detection branch design, and the position and the category of a target are directly predicted on a fused feature map. The method can remarkably improve the precision and robustness of target detection in a low-illumination environment, and is suitable for the fields of night monitoring, automatic driving, security and protection and the like.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a low-light target detection method and system that integrates multi-scale features and local detail enhancement. Background Technology

[0002] Object detection technology has wide applications in computer vision, including video surveillance, autonomous driving, and intelligent security. However, in low-light environments, traditional object detection methods often fail to achieve ideal detection results due to insufficient illumination, severe noise interference, and loss of detail. Existing low-light object detection methods mainly fall into two categories: one type enhances the brightness or denoises the image before detection, but this easily introduces artifacts or loss of detail; the other type directly improves the network architecture, enabling the model to extract richer feature information from low-light images, thereby improving detection performance, but it still has shortcomings in multi-scale feature fusion and detail preservation.

[0003] In existing technologies, multi-scale feature fusion methods such as FPN and PAN have limited ability to detect small targets and details in low-light scenes, and are prone to losing shallow high-frequency detail information during the fusion process. Furthermore, the noise characteristics and brightness distribution of low-light images differ significantly, making it difficult for conventional convolutional feature extraction networks to simultaneously suppress noise and preserve details. Therefore, a target detection model that can effectively fuse multi-scale features, enhance local details, and suppress noise interference under low-light conditions is needed to improve detection accuracy and robustness. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a low-light target detection method and system that integrates multi-scale features and local detail enhancement, in order to solve the problems in the prior art that the shallow high-frequency detail information is easily lost during the fusion process and that it is difficult to simultaneously achieve noise suppression and detail preservation.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a low-light target detection method that integrates multi-scale features and local detail enhancement, comprising the following steps:

[0006] S1: After the low-light image is input into the low-light target detection network model, it first enters the feature extraction network based on multi-pooling pyramid and cross-stage dual hybrid attention to obtain multi-scale, multi-channel feature maps.

[0007] S2: The feature maps are then fused at different levels through a dynamic detail semantic fusion pyramid network;

[0008] S3: The image after feature fusion is input into the dual-group detection head to obtain the target detection result.

[0009] In a preferred embodiment, in step S1, the feature extraction network based on multi-pooling pyramid and cross-stage dual hybrid attention includes multiple convolutional layers, a C2PSA module, an MPSPPF module, and a CSP-EDHAN module for extracting multi-level feature information. The MPSPPF module introduces a parallel pooling branch and a gradient enhancement concatenation part. The parallel pooling branch includes two pooling operations: max pooling (MaxPool2d) and average pooling (AvgPool2d). The pooled outputs are weighted and fused to preserve the global structure while ensuring rich detail features. Furthermore, the MPSPPF module employs multiple sequential pooling operations. During forward propagation, the same pooling operation is performed on the pooling results of the previous layer, forming a deeper pyramid structure and allowing the network to construct a more comprehensive feature hierarchy. The CSP-EDHAN module consists of multiple ECAP modules connected sequentially to convolutional layers. It adopts a dual-path feature interaction structure: one path preserves the original feature information, and the other path is used for deep feature extraction and enhancement. When the two paths are fused, channel attention and spatial attention mechanisms are introduced to highlight key information and suppress redundant features.

[0010] In a preferred embodiment, in step S2, the Dynamic Detail Semantic Fusion Pyramid Network includes multiple convolutional layers, upsampling, a C3K2 module, an ADown module, and a surface detail fusion module. It employs a bidirectional multi-scale feature fusion structure with both top-down and bottom-up approaches. The top-down path upsamples high-level feature maps and concatenates them with adjacent low-level feature maps to transfer global semantic information from high-level features to low-level feature maps. The bottom-up path downsamples low-level fused feature maps and concatenates them with adjacent high-level feature maps to transfer local detail information from low-level features to high-level feature maps. Simultaneously, a surface detail fusion module (SDFM) is introduced at the outputs of feature maps at three different scales (P3, P4, and P5). The SDFM module uses an attention mechanism to weightedly fuse shallow high-frequency detail information with deep semantic information, thereby enhancing semantic expression while preserving detail features. Finally, it outputs multi-scale enhanced features for dual-group detection heads to predict targets.

[0011] In a preferred embodiment, in step S3, the dual-group detection head includes group convolution and ordinary convolution, and adopts a decoupled detection branch design to directly predict the target on the fused feature map;

[0012] The design of the two consecutive groups of convolutions and their decoupled detection branches in the dual-group detection head includes:

[0013] 1) Continuous group convolution: First, two groups of convolutions are introduced to extract richer feature representations, and their corresponding outputs... The formula is as follows:

[0014]

[0015] in, This indicates a group convolution operation. , This indicates that each group has an independent convolutional kernel, BN(*) represents batch normalization, and SiLU(*) represents Sigmoid linear unit activation. Indicates the output;

[0016] 2) Decoupled detection branch design: The network head is split into completely independent regression and classification branches, with each branch using its own convolutional layers and parameters; the corresponding formulas are as follows:

[0017] Regression branch ,

[0018] Classification branches ,

[0019] Output splicing ,

[0020] in, This represents a 1×1 convolution kernel for the regression branch. This represents a 1×1 convolution kernel for the classification branch.

[0021] In a preferred embodiment, the ECAP module includes a three-level cascaded EDConv feature extraction unit, a feature concatenation and recalibration structure, and parallel channel attention and spatial attention mechanisms. The three-level cascaded EDConv feature extraction unit is used to extract multi-scale, multi-granularity feature information step-by-step. The feature concatenation and recalibration structure is used to fuse features from different levels and adjust the importance of each channel through weight allocation. The channel attention mechanism is used to highlight key information and suppress redundant features in the channel dimension, and the spatial attention mechanism is used to enhance the response of salient regions in the spatial dimension. The fused and attention-enhanced features are added to the input features through residual connections, thereby outputting enhanced features that contain both detailed and semantic information.

[0022] In a preferred embodiment, the EDConv module is composed of depthwise separable convolutional units and pointwise convolutional units connected sequentially. The depthwise separable convolutions are used to reduce computational complexity while maintaining spatial feature resolution, and the pointwise convolutions are used to perform feature mixing and information interaction in the channel dimension. The EDConv module also includes a shortcut branch for matching the size and number of channels of the input features and fusing them with the convolutional output of the main branch element-wise at the corresponding positions to achieve residual connections of features. The fused features are output after standard convolutional layers and batch normalization, thereby improving the training stability and convergence speed of the network while ensuring feature representation ability.

[0023] The present invention also provides a low-light target detection system that integrates multi-scale features and local detail enhancement, and runs the aforementioned low-light target detection method that integrates multi-scale features and local detail enhancement.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1. This invention constructs a feature extraction module based on multi-pooling pyramid and cross-stage dual hybrid attention, which includes a carefully designed MPSPPF module, CSP-EDHAN module, as well as convolutional layers and C2PSA module, effectively suppressing the interference of background and random noise, while reducing the high sensitivity of images to noise in low-light scenes, and further suppressing the loss of detail information;

[0026] 2. This invention constructs a dynamic detail semantic fusion pyramid network. By introducing a surface detail fusion module (SDFM) at three scales (P3, P4, and P5), the global semantic information of high-level features and the local detail information of low-level features are weighted and fused through an attention mechanism. The output is a multi-scale feature that combines detail preservation and semantic enhancement for the detection head to predict targets.

[0027] 3. This invention constructs a dual-group detection head, introducing two consecutive groups of convolution to extract richer feature representations. At the same time, it adopts a decoupled detection branch design to enhance the model's perception of image details and target positions in low-light environments, enabling the model to extract and identify detailed features and locate targets more accurately when dealing with complex scenes. Attached Figure Description

[0028] Figure 1 A flowchart illustrating the implementation of the low-light target detection model that integrates multi-scale features and local detail enhancement, as provided in this embodiment of the invention.

[0029] Figure 2 This is an architecture diagram of the low-light target detection network model in an embodiment of the present invention;

[0030] Figure 3This is a schematic diagram of the feature extraction network based on multi-pooling pyramid and cross-stage dual hybrid attention in an embodiment of the present invention;

[0031] Figure 4 This is a schematic diagram of the MPSPPF module in an embodiment of the present invention;

[0032] Figure 5 This is a schematic diagram of the CSP-EDHAN module in an embodiment of the present invention;

[0033] Figure 6 This is a schematic diagram of the structure of the dynamic detail semantic fusion pyramid network in an embodiment of the present invention;

[0034] Figure 7 This is a schematic diagram of the dual-group detection head in an embodiment of the present invention;

[0035] Figure 8 The image below is a comparison of a randomly obtained first low-light image and the target detection result in an embodiment of the present invention. (a) is a correctly labeled low-light image, (b) is a target detection effect image of the prior art, and (c) is a target detection effect image of the present invention. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0037] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0038] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0039] like Figure 1-8 As shown, this embodiment provides a low-light target detection model that integrates multi-scale features and local detail enhancement, including the following steps:

[0040] S1: After the low-light image is input into the low-light target detection network model, it first enters the feature extraction network based on multi-pooling pyramid and cross-stage dual hybrid attention to obtain multi-scale, multi-channel feature maps.

[0041] S2: The feature maps are then fused at different levels through a dynamic detail semantic fusion pyramid network;

[0042] S3: The image after feature fusion is input into the dual-group detection head to obtain the target detection result.

[0043] Specifically, the structure diagram of the feature extraction network based on multi-pooling pyramid and cross-stage dual hybrid attention is as follows: Figure 3 As shown, the model first extracts features from the input image step by step through multiple convolutions to obtain basic low-level features. Then, it introduces multiple CSP_EDHAN modules to effectively improve feature representation and reduce redundant computation. Next, it uses the MPSPPF module to perform multi-scale feature fusion to capture contextual information under different receptive fields. Finally, it uses the C2PSA module to fuse channel and spatial attention information to output a high-quality feature map that combines global semantics and local details.

[0044] The structure diagram of the MPSPPF module is as follows: Figure 4 As shown, the system includes an input convolutional layer, three sets of parallel pooling units, a feature concatenation unit, and an output convolutional layer connected in sequence. Each set of parallel pooling units consists of max pooling and average pooling, with the max pooling ratio being 0.6 and the average pooling ratio being 0.4. The outputs of the two pooling units are added element-wise and used as the input to the next set of pooling units. The outputs of the three sets of pooling units and the self-enhanced input features are fed into the feature concatenation unit for fusion. The fusion result is then used by the output convolutional layer to generate the output features.

[0045] The structural diagram of the CSP_EDHAN module is as follows: Figure 5 As shown, it includes an input convolutional layer, a feature segmentation unit, several cascaded ECAP modules, a feature concatenation unit, and an output convolutional layer connected in sequence. The input convolutional layer is used to perform convolutional transformation on the input features. The feature segmentation unit divides the input features into two branches along the channel dimension. One branch is directly passed to the feature concatenation unit, and the other branch is sequentially input to N cascaded ECAP modules for processing. The features processed by the N ECAP modules are fused with the directly passed features along the channel dimension in the feature concatenation unit. The fused result is then output as a feature by the output convolutional layer generation module.

[0046] The ECAP module employs a cascaded structure consisting of three EDConv layers, each capable of independently extracting features. Subsequently, the features from the three layers are concatenated and fused at multiple scales using 1×1 convolutions, followed by Batch Norm and SiLU activation for feature recalibration. Finally, weak edge and texture information is enhanced through channel and spatial attention, and overall robustness is improved through random depth regularization.

[0047] The EDConv module utilizes depthwise separable convolution to process the input features. First, depthwise convolution is performed independently in 3×3 layers on each channel, expanding the channel count to extract richer features. Subsequently, pointwise convolution compresses the expanded feature maps to the target number of channels. Furthermore, the EDConv module adds the original input to the convolved features via a separate branch. This not only helps prevent gradient vanishing through gradient backpropagation but also preserves information about weak features in low-light images.

[0048] The dynamic detail semantic fusion pyramid network structure diagram is as follows: Figure 6 As shown, the network includes multiple convolutional layers, upsampling, a C3K2 module, an ADown module, and a surface detail fusion module to fuse feature maps at different scales. The Dynamic Detail Semantic Fusion Pyramid network receives three feature maps P3, P4, and P5 from the feature extraction network. First, it upsamples the high-level features step-by-step through a top-down path and concatenates them with the low-level features, then fuses them through the C3K2 module to obtain high-resolution features. Next, it downsamples the low-level fused features using the ADown module through a bottom-up path, concatenates them with the high-level features, and fuses them through C3K2, progressively passing semantic information. At the outputs of P3, P4, and P5, SDFM modules are inserted respectively, fusing shallow details and deep semantics through an attention mechanism, ultimately outputting multi-scale enhanced features for the detection head. This design not only effectively suppresses background noise and illumination interference but also achieves precise alignment and complementary enhancement of low-level details, providing high-fidelity low-level feature representations for subsequent detection stages, thereby significantly improving the model's robustness and localization accuracy in complex scenes.

[0049] The dual-group detection head structure is as follows: Figure 7 As shown, the input features are first processed by grouping them by channel using the first layer of group convolution. Then, the group features are further extracted by the second layer of group convolution, which enhances the ability to capture local details under low brightness, reduces the amount of computation, and preserves the detailed features within the group. Subsequently, the output of the second layer of group convolution is divided into two parallel branches. Each branch uses ordinary two-dimensional convolution to extract features under different receptive fields. Finally, the convolution results of the two branches are concatenated in the channel dimension to achieve the fusion of multi-scale and multi-directional features, thereby improving the ability to recognize low-light targets.

[0050] In this embodiment, the feature extraction network based on multi-pooling pyramids and cross-stage dual hybrid attention first inputs the low-light image to be detected into a first conventional convolutional layer to downsample the input image, obtaining a first-stage feature map P1 / 2. The conventional convolutional layer can extract basic features such as low-level edges and textures while reducing resolution. Subsequently, the first-stage feature map is input into a second conventional convolutional layer for further downsampling, obtaining a second-stage feature map P2 / 4. This process can further compress the feature map size and extract more discriminative mid-to-low-level features. Next, P2 / 4 is input into the CSP_EDLAN module, which reduces the loss of low-level features such as texture, edges, and contours in low-light images through multi-granularity feature fusion and efficient feature extraction. Then, it is downsampled to P3 / 8 through a third conventional convolutional layer, and P3 / 8 is input into the CSP_EDLAN module to further enhance the multi-scale expressive capability of the features. Next, a fourth regular convolutional layer downsamples the data to P4 / 16, and this P4 / 16 is input to the CSP_EDLAN module to extract mid-to-high-level semantic features, improving the ability to perceive the overall shape and structure of the target. Then, a fifth regular convolutional layer downsamples the data to P5 / 32, and this P5 / 32 is input to the CSP_EDLAN module to extract high-level semantic features, enhancing the ability to distinguish targets in complex backgrounds. Subsequently, this P5 / 32 is input to the MPSPPF module, which uses multi-level hybrid parallel pooling and feature enhancement mechanisms to further suppress the loss of detailed information. Finally, the C2PSA module achieves channel and spatial attention fusion.

[0051] In the Dynamic Detail Semantic Fusion Pyramid Network, the highest-level feature P5 / 32 of the feature extraction network based on multi-pooling pyramid and cross-stage dual hybrid attention is upsampled and concatenated with the mid-level feature P4 / 16 to obtain the first fused feature. This fusion process combines high-level semantic information with mid-level detail information, taking into account both global and local features. The first fused feature is input into the C3k2 module for convolutional fusion. This module improves the sufficiency of feature fusion through cross-convolution kernel feature interaction, resulting in an updated P4 feature. The updated P4 feature is then upsampled and concatenated with the shallow feature P3 / 8 to obtain the second fused feature. The shallow feature contains more edge and texture information, which is helpful for small object detection. The second fused feature is input into the C3k2 module to obtain the updated P3 feature, which serves as the small object feature map. Subsequently, the updated P3 feature is downsampled through the ADown module, which retains more detail information while reducing resolution. It is then concatenated with the updated P4 feature and input into the C3k2 module to obtain a new P4 feature, which serves as the mid-target feature map. The new P4 feature is then downsampled using the ADown module and concatenated with the P5 feature. This concatenation is then input into the C3k2 module to obtain a new P5 feature, which serves as the large target feature map. Next, P3, P4, and P5 are input into the SDFM module for feature enhancement. This module can enhance local detail representation in multi-scale features, improving the discernibility of target edges and textures under low-light conditions.

[0052] In the dual-head detection network, enhanced P3, P4, and P5 are input into the dual-head detection system to predict the category and bounding box coordinates of small, medium, and large targets, respectively. The dual-head structure can independently optimize detection performance at different scales, improving the accuracy and robustness of multi-scale detection. Finally, the multi-scale detection results are fused to output the final low-light target detection result.

[0053] The following simulation experiments will further illustrate the implementation effects of the present invention.

[0054] 1. Simulation conditions

[0055] The hardware configuration in this embodiment is an NVIDIA RTX3080 Ti (12 GB of VRAM). The code is written in Python 3.8 and uses PyTorch 2.0.0 as the deep learning framework. The model training consists of 300 epochs with a batch size of 8, an initial learning rate of 0.015, a momentum parameter of 0.937, and uses stochastic gradient descent (SGD) as the optimizer. All training samples are uniformly scaled to 640×640 resolution before being input into the network.

[0056] To evaluate the detection performance of the improved model in low-light environments, this paper uses the ExDark dataset, specifically designed for low-light vision tasks. This dataset contains 7363 low-light images, covering 12 target categories, and scenes including urban nightscapes, dimly lit interiors, vehicles, and pedestrians, realistically reflecting everyday low-light conditions. The dataset is randomly divided into training, validation, and test sets in an 8:1:1 ratio, corresponding to 5890, 736, and 737 images respectively. In this embodiment, correctly labeled low-light images are shown below. Figure 8 As shown in (a).

[0057] 2. Simulation Result Analysis

[0058] Figure 8 Image (b) shows the target detection effect of the existing technology. Figure 8 Image (c) shows the target detection effect of this invention. Figure 8 As can be seen, existing technologies not only misclassify irrelevant background noise as chairs but also result in missed detections. In contrast, this invention accurately detects and locates targets in low-light conditions, achieving more accurate classification and recognition. Furthermore, the detection confidence level of this invention is higher than that of existing technologies, enhancing the reliability of the detection results. These results demonstrate that, compared to existing technologies, this invention, when dealing with image blurring and the difficulty in distinguishing between background noise and the detected target in low-light environments, can not only accurately pinpoint the true location of the target in low-light conditions but also output the correct category label with higher confidence, exhibiting superior detection performance and accuracy.

Claims

1. A low-light target detection method that integrates multi-scale features and local detail enhancement, characterized in that, Includes the following steps: S1: After the low-light image is input into the low-light target detection network model, it first enters the feature extraction network based on multi-pooling pyramid and cross-stage dual hybrid attention to obtain multi-scale, multi-channel feature maps. S2: The feature maps are then fused at different levels through a dynamic detail semantic fusion pyramid network; S3: The image after feature fusion is input into the dual-group detection head to obtain the target detection result.

2. The low-light target detection method integrating multi-scale features and local detail enhancement according to claim 1, characterized in that, In step S1, the feature extraction network based on multi-pooling pyramid and cross-stage dual hybrid attention includes multiple convolutional layers, a C2PSA module, an MPSPPF module, and a CSP-EDHAN module, used to extract multi-level feature information. The MPSPPF module introduces a parallel pooling branch and a gradient enhancement concatenation part. The parallel pooling branch includes two pooling operations: max pooling (MaxPool2d) and average pooling (AvgPool2d). The pooled outputs are weighted and fused to preserve the global structure while ensuring rich detail features. Furthermore, the MPSPPF module employs multiple sequential pooling operations. During forward propagation, the same pooling operation is performed on the pooling results of the previous layer, forming a deeper pyramid structure, allowing the network to construct a more comprehensive feature hierarchy. The CSP-EDHAN module consists of multiple ECAP modules connected sequentially to convolutional layers. It adopts a dual-path feature interaction structure: one path preserves the original feature information, and the other path is used for deep feature extraction and enhancement. When the two paths are fused, channel attention and spatial attention mechanisms are introduced to highlight key information and suppress redundant features.

3. The low-light target detection method integrating multi-scale features and local detail enhancement according to claim 1, characterized in that, In step S2, the Dynamic Detail Semantic Fusion Pyramid Network includes multiple convolutional layers, upsampling, a C3K2 module, an ADown module, and a surface detail fusion module. It adopts a bidirectional multi-scale feature fusion structure with both top-down and bottom-up approaches. The top-down path upsamples high-level feature maps and concatenates them with adjacent low-level feature maps to transfer global semantic information from high-level features to low-level feature maps. The bottom-up path downsamples low-level fused feature maps and concatenates them with adjacent high-level feature maps to transfer local detail information from low-level features to high-level feature maps. Simultaneously, a surface detail fusion module (SDFM) is introduced at the outputs of feature maps at three different scales (P3, P4, and P5). The SDFM module uses an attention mechanism to weightedly fuse shallow high-frequency detail information with deep semantic information, thereby enhancing semantic expression while preserving detail features. Finally, it outputs multi-scale enhanced features for dual-group detection heads to predict targets.

4. The low-light target detection method integrating multi-scale features and local detail enhancement according to claim 1, characterized in that, In step S3, the dual-group detection head includes group convolution and ordinary convolution, and adopts a decoupled detection branch design to directly predict the target on the fused feature map; The design of the two consecutive groups of convolutions and their decoupled detection branches in the dual-group detection head includes: 1) Continuous group convolution: First, two groups of convolutions are introduced to extract richer feature representations, and their corresponding outputs... The formula is as follows: in, This indicates a group convolution operation. , This indicates that each group has an independent convolutional kernel, BN(*) represents batch normalization, and SiLU(*) represents Sigmoid linear unit activation. Indicates the output; 2) Decoupled detection branch design: The network head is split into completely independent regression and classification branches, with each branch using its own convolutional layers and parameters; the corresponding formulas are as follows: Regression branch , Classification branches , Output splicing , in, This represents a 1×1 convolution kernel for the regression branch. This represents a 1×1 convolution kernel for the classification branch.

5. The low-light target detection method integrating multi-scale features and local detail enhancement according to claim 2, characterized in that, The ECAP module includes a three-level cascaded EDConv feature extraction unit, a feature concatenation and recalibration structure, and parallel channel attention and spatial attention mechanisms. The three-level cascaded EDConv feature extraction unit extracts multi-scale, multi-granularity feature information step-by-step. The feature concatenation and recalibration structure fuses features from different levels and adjusts the importance of each channel through weight allocation. The channel attention mechanism highlights key information and suppresses redundant features in the channel dimension, while the spatial attention mechanism enhances the response of salient regions in the spatial dimension. The fused and attention-enhanced features are added to the input features through residual connections, thereby outputting enhanced features that combine detailed and semantic information.

6. The low-light target detection method according to claim 5, which integrates multi-scale features and local detail enhancement, is characterized in that... The EDConv module consists of depthwise separable convolutional units and pointwise convolutional units connected sequentially. The depthwise separable convolutions reduce computational complexity while maintaining spatial feature resolution, and the pointwise convolutions perform feature mixing and information interaction along the channel dimension. The EDConv module also includes a shortcut branch for matching the size and number of channels of the input features and fusing them element-wise with the convolutional output of the main branch at corresponding positions to achieve residual connections of the features. The fused features are then processed by standard convolutional layers and batch normalization before being output, thereby improving the training stability and convergence speed of the network while ensuring feature representation capabilities.

7. A low-light target detection system that integrates multi-scale features and local detail enhancement, characterized in that, Run the low-light target detection method that integrates multi-scale features and local detail enhancement as described in any one of claims 1-6.

Images