A low-light target detection method and device based on multi-scale dynamic fusion

Abstract

The application discloses a low-light target detection method and device based on multi-scale dynamic fusion, and constructs a low-light target detection network based on multi-scale dynamic fusion comprising a detection network and an improved low-light enhancer; the improved low-light enhancer effectively solves the problems that the target boundary becomes blurred and the feature is difficult to extract due to the existence of abnormal light source interference in a low-light environment and the large target scale difference through Laplacian pyramid decomposition, a multi-perception detail enhancement module MDEM and a multi-scale dynamic fusion module MDFM; a low-light target detection model is obtained by training the constructed low-light target detection network based on multi-scale dynamic fusion, and the optimal performance is evaluated and selected to perform target detection on a low-light image, so that a result image of target detection is obtained. The application solves the problems of abnormal light source interference and large target scale difference in a low-light image, and improves the accuracy of low-light target detection.

Description

Technical Field

[0001] This invention belongs to the field of low-light target detection technology, and in particular, it is a low-light target detection method and device based on multi-scale dynamic fusion. Background Technology

[0002] With the rapid development of deep learning technology, object detection, as one of the core tasks in computer vision, has made significant progress. Currently, two-stage object detection algorithms (such as Faster R-CNN) and single-stage object detection algorithms (such as YOLO and SSD) exhibit excellent performance in well-lit scenes and are widely used in autonomous driving, video surveillance, and security. However, in practical applications, low-light environments (such as nighttime streets, underground parking garages, and low-light indoor scenes) are very common, and these environments severely impact object detection performance. Low-light conditions bring several challenges to object detection: First, targets in low-light images often have significant scale differences; a single image may contain large, medium, and small targets simultaneously. However, existing low-light object detection algorithms lack the ability to perceive multi-scale features, making it difficult to simultaneously consider the feature representation of targets at different scales, thus leading to decreased detection accuracy. Second, low-light environments are often accompanied by abnormal light source interference (such as glare or halos), causing target boundaries to become blurred, thereby increasing false positives and false negatives in the algorithm. Therefore, how to enhance the robustness of target detection algorithms in low-light environments and improve detection accuracy has become a key issue that urgently needs to be addressed.

[0003] The patent application "A Target Detection Method in Low Light Based on Pyramid Enhancement Network" (patent application number: CN202310947059.0) from the Artificial Intelligence Research Institute of the Hefei Comprehensive National Science Center (Anhui Provincial Artificial Intelligence Laboratory) outlines the following steps: First, an image is decomposed into components of different resolutions using a Laplacian pyramid. Second, a detail processing module is constructed to enhance the details of the components output from the Laplacian pyramid at different scales. This includes global enhancement through contextual branches and texture enhancement through edge branches. The globally enhanced component and the edge-branch-processed component are then tensor-concatenated and processed through a convolutional layer to obtain the detail processing result. Third, a low-frequency filter is constructed to capture low-frequency semantics and block high-frequency noise, resulting in a low-frequency enhancement result. Fourth, the detail processing result and the low-frequency enhancement result are tensor-concatenated and their features are fused through a convolutional layer to obtain a convolutionally enhanced component. Fifth, the convolutionally enhanced component is reconstructed using a Laplacian pyramid to obtain an image with the same resolution as the original image. Sixth, the image is used as input to a target detector to obtain the detection result. While this method can improve the accuracy of low-light target detection, it is mainly due to the feature enhancer designed in the algorithm. It does not improve the target detection algorithm itself, and the introduction of the enhancer brings additional computation and parameters.

[0004] The patent application from Northwestern Polytechnical University, "A Method and System for High-Precision Target Detection under Low-Light Conditions" (patent application number: CN202411840084.X), involves the following steps: First, the extracted features of the low-light image are filtered multiple times to obtain multiple filtered features from high to low. Second, the extracted features obtained by upsampling the low-level filtered features and the high-level filtered features are added element-wise to obtain multiple fused features from high to low. Third, the extracted features of each fused feature, the filtered features of the corresponding layer, and the global fused features enhanced by max pooling are added element-wise to obtain high-level features, which are then sent to various detectors for detection. Although this method can improve the quality of the extracted feature map of the low-light image through denoising and enhancement, it does not specifically improve the target detection algorithm, thus limiting its feature extraction capability.

[0005] The patent application from Nanjing University of Posts and Telecommunications, entitled "A Low-Light Target Detection Method Based on Improved YOLOv8" (patent application number: CN202510474567.0), outlines the following steps: First, obtain the ExDark dataset; second, convert the ExDark dataset to a YOLO-compatible dataset; third, divide the converted YOLO-compatible ExDark dataset into training, validation, and test sets; fourth, construct an improved YOLOv8 low-light target detection network, which improves the feature extraction layer, feature fusion layer, and loss function of the YOLOv8 network; fifth, use the improved YOLOv8 low-light target detection network as the detection model, training and validating it using the training and validation sets to obtain the optimal detection model; sixth, use the optimal detection model from training, testing the improved YOLOv8 target detection network with images under low-light conditions as input. While this method improves the accuracy of low-light target detection and is more lightweight, it cannot differentiate between the difficulty levels of the samples.

[0006] As can be seen from the above, existing low-light target detection methods have not effectively solved the problems of feature extraction difficulties caused by large differences in target scale in low-light images and the blurring of target boundaries caused by interference from abnormal light sources. Summary of the Invention

[0007] To address the aforementioned deficiencies in existing technologies, the present invention aims to propose a low-light target detection method and apparatus based on multi-scale dynamic fusion. By applying a multi-sensory detail enhancement module to each multi-scale component obtained from Laplacian pyramid decomposition, the model can simultaneously enhance global semantic information and local detail information. This solves the problems of difficulty in feature extraction due to large differences in target scale in low-light images and blurred target boundaries caused by abnormal light source interference, thereby enhancing the expressive power of features and improving the accuracy of low-light target detection.

[0008] The present invention is achieved through the following technical solution.

[0009] One aspect of the present invention provides a low-light target detection method based on multi-scale dynamic fusion, comprising: A low-light target detection network based on multi-scale dynamic fusion is constructed, which includes a detection network and an improved low-light intensifier. The improved low-light enhancer is a multi-scale dynamic fusion enhancer, which includes a Laplacian pyramid, a multi-sensory detail enhancement module, and a multi-scale dynamic fusion module. The multi-sensory detail enhancement module is applied to each multi-scale component obtained by decomposing the Laplacian pyramid to enhance the feature representation capability. The constructed low-light target detection network based on multi-scale dynamic fusion was trained using the ExDark training set to obtain a low-light target detection model based on multi-scale dynamic fusion. The low-light target detection model based on multi-scale dynamic fusion was evaluated using the ExDark validation set. The model training parameters were adjusted based on the training logs and evaluation metrics to obtain an optimized model. The best-performing low-light target detection model based on multi-scale dynamic fusion was selected to perform target detection on the low-light image, and the target detection result image was obtained.

[0010] Preferably, an improved low-light enhancer is constructed to enhance feature representation capabilities, including: The Laplacian pyramid is used for multi-scale feature decomposition of input images. The multi-sensory detail enhancement module is used to enhance feature components at different resolutions; The multi-scale dynamic fusion module is used to adaptively weight and fuse the enhanced features, and reconstruct the enhanced feature map.

[0011] Preferably, the Laplacian pyramid performs multi-scale feature decomposition on the input image, including: Gaussian pyramids are used to perform multi-scale feature decomposition on the input image to obtain feature components at different resolutions. The Laplace pyramid was constructed based on the Gaussian pyramid, and the decomposed multi-scale features were obtained.

[0012] Preferably, the multi-sensory detail enhancement module includes a local edge enhancement branch and a global feature enhancement branch.

[0013] As a preferred approach, the local edge enhancement branch processing flow is as follows: Local edge enhancement branch utilization and The gradient information of the input features in the horizontal and vertical directions is calculated separately. The receptive field is further expanded by dilated convolution, and residual connections are introduced to reduce information loss.

[0014] As a preferred approach, the global feature enhancement branch processing flow is as follows: The global feature enhancement branch consists of six stacked depthwise separable convolutional layers, each followed by a ReLU activation function and a skip connection, with the channel dimension aligned through a depthwise separable convolution.

[0015] Preferably, the multi-scale dynamic fusion module adaptively weights and fuses the enhanced features, and reconstructs the enhanced feature map, including: The high-resolution and low-resolution feature maps of adjacent scales, which are enhanced by the multi-sensory detail enhancement module, are processed by two convolution operations and a sigmoid activation function in sequence. The fused features are obtained by adding adjacent resolution features element by element. The fused features are divided into M blocks along the channel dimension and input into the hybrid attention block for dynamic fusion. The dynamically fused features are then stitched together along the channel dimension to obtain the final fused features.

[0016] As a preferred approach, the hybrid attention block processing flow is as follows: Using a self-attention mechanism, attention weights are dynamically generated through queries, keys, and values; A channel attention mechanism is introduced to adaptively adjust the weights of different channels; A spatial attention mechanism is introduced to generate a two-dimensional attention map through convolution operations, enabling the network to focus on the target contour and edge texture regions.

[0017] As a preferred method, YOLOv3 is used for target detection.

[0018] Preferably, the constructed low-light target detection network based on multi-scale dynamic fusion is trained using the ExDark training set, including: The annotation information in the ExDark dataset is converted into the YOLO format, which is convenient for training; the annotation information of each target after conversion includes the target's category and the target's location in the image; A low-light target detection network based on multi-scale dynamic fusion was trained using the ExDark dataset; Record the training log, which contains the loss value and AP metric for each iteration during training, to obtain a low-light target detection model based on multi-scale dynamic fusion.

[0019] Another aspect of the present invention provides a low-light target detection device based on multi-scale dynamic fusion of the method, comprising: The building module is used to construct a low-light target detection network based on multi-scale dynamic fusion, which includes a detection network and an improved low-light intensifier. Through the Laplacian pyramid, multi-sensory detail enhancement module and multi-scale dynamic fusion module in the improved low-light intensifier, the multi-sensory detail enhancement module is applied to each multi-scale component obtained by Laplacian pyramid decomposition to enhance the feature representation capability. The training module is used to train the constructed low-light target detection network based on multi-scale dynamic fusion using the ExDark training set, so as to obtain the low-light target detection model based on multi-scale dynamic fusion. The evaluation and optimization module is used to evaluate the low-light target detection model based on multi-scale dynamic fusion using the ExDark validation set, and adjust the model training parameters according to the training logs and evaluation metrics to obtain the optimized model. The detection module is used to select the best-performing low-light target detection model based on multi-scale dynamic fusion to perform target detection on low-light images and obtain the target detection result image.

[0020] In another aspect, the present invention provides an electronic device, including a processor and a memory; The memory is used to store computer programs, the computer programs including program instructions; The processor is used to call the program stored in the memory to execute the low-light target detection method based on multi-scale dynamic fusion.

[0021] The present invention, by adopting the above technical solution, has the following beneficial effects: 1. In order to improve the network's ability to extract features from targets of different scales and to address the problem of large differences in target scale in low-light target detection, a multi-sensory detail enhancement module is applied to each multi-scale component obtained from the Laplacian pyramid decomposition, so that the model can simultaneously enhance global semantic information and local detail information.

[0022] 2. To further address issues such as blurred target boundaries caused by abnormal light source interference in low-light images, a local edge enhancement branch was designed in the multi-sensory detail enhancement module. This branch extracts gradient information in the horizontal and vertical directions of the image, thereby capturing edge features more accurately. Even when the texture becomes blurred due to insufficient lighting, key structural information can still be extracted through the relative changes in local gray levels. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, do not constitute an undue limitation of the invention. In the drawings: Figure 1 A flowchart of the low-light target detection method based on multi-scale dynamic fusion provided by the present invention; Figure 2 The present invention provides a structural diagram of a low-light target detection algorithm based on multi-scale dynamic fusion; Figure 3 A structural diagram of the multi-sensory detail enhancement module designed for this invention; Figure 4 This is a structural diagram of the multi-scale dynamic fusion module and hybrid attention block designed for this invention; Figures 5(a) and 5(b) show the detection results of the YOLOv3 algorithm and the low-light target detection algorithm based on multi-scale dynamic fusion of the present invention, respectively. Figure 6 This is a schematic diagram of a low-light target detection device based on multi-scale dynamic fusion, as shown in an embodiment of the present invention. Figure 7 This is a schematic diagram of the electronic device structure shown in an embodiment of the present invention. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions of the present invention are used to explain the present invention, but are not intended to limit the present invention.

[0025] Reference Figure 1 This invention provides a low-light target detection method based on multi-scale dynamic fusion, which specifically includes the following steps: Step 101: Construct a low-light target detection network based on multi-scale dynamic fusion, including a detection network and an improved low-light intensifier.

[0026] Step S11: Construct an improved low-light enhancer structure to enhance feature representation capabilities.

[0027] Reference Figure 2 The original low-light image is input into a low-light enhancer to enhance its feature representation capabilities, and then the enhanced features are input into the YOLOv3 detection network for target detection.

[0028] like Figure 2 As shown, the low-light enhancer mainly consists of three components: the Laplacian pyramid is used to perform multi-scale feature decomposition on the input image; the Multi-Perception Detail Enhancement (MDEM) module is introduced to enhance feature components at different resolutions; and the Multi-Scale Dynamic Fusion (MDFM) module is used to adaptively weight and fuse the enhanced features to reconstruct the enhanced feature map.

[0029] The improved low-light enhancer enhances feature representation capabilities, and the specific steps are as follows: (a) The Laplacian pyramid performs multi-scale feature decomposition on the input image.

[0030] First, the input image Perform uniform scaling to obtain Subsequently, a Gaussian pyramid is used for multi-scale decomposition to obtain feature components at different resolutions. This process is represented as: (1) in, This indicates an image downsampling operation. This indicates that a convolution kernel size of [size missing] is used. The Gaussian convolution kernel obtained the first The output of the first layer of the Gaussian pyramid. Then, a four-layered Laplace pyramid is constructed based on the Gaussian pyramid. Specifically, the first layer of the Laplace pyramid... Layer output Output from the Gaussian pyramid of this layer Output of the Gaussian pyramid above The difference between the upsampled results is represented by the following formula: (2) in, This indicates an image upsampling operation.

[0031] (b) Input the multi-scale features into the Multi-Perception Detail Enhancement (MDEM) module to enhance the feature components at different resolutions.

[0032] Multi-sensory detail enhancement module structure such as Figure 3 As shown, it includes the local edge enhancement branch LEEB and the global feature enhancement branch GFEB.

[0033] The specific process for LEEB processing of local edge enhancement branches is as follows: First, use the input features to... and The gradient information in the horizontal and vertical directions of the input features is calculated separately; then, the receptive field is further expanded by dilated convolution; finally, residual connections are introduced to reduce information loss. This process is represented as follows: (3) in, and These represent gradient information in the horizontal and vertical directions, respectively. This represents ordinary convolution. This indicates dilated convolution.

[0034] The specific process of the Global Feature Enhancement (GFEB) branch is as follows: The global feature enhancement branch consists of six stacked depthwise separable convolutions. To accelerate convergence and stabilize gradient propagation, each convolutional layer is followed by a ReLU activation function and a skip connection. Finally, the channel dimensions are aligned using a depthwise separable convolution.

[0035] (c) The multi-scale dynamic fusion module performs adaptive weighted fusion of the enhanced features and reconstructs the enhanced feature map.

[0036] The structures of the Multi-Scale Dynamic Fusion Module (MDFM) and the Hybrid Attention Block (HAB) are as follows: Figure 4 As shown, the processing flow of the multi-scale dynamic fusion module is as follows: Input feature maps of adjacent scales (High resolution) and (Low resolution). First, the input is processed by two convolution operations and the Sigmoid activation function. Then, the features of adjacent resolutions are added element by element to obtain the fused features. The process is shown in the following formula.

[0037] (4) Subsequently, the fused features The channel is uniformly divided into M blocks along its dimension. The m-th block... Includes the original feature map within the channel index range The characteristics within are expressed mathematically as follows.

[0038] (5) in, , This indicates the number of channels. Then, The inputs are then processed by the Hybrid Attention Block (HAB).

[0039] The hybrid attention block first utilizes a self-attention mechanism to dynamically generate attention weights through queries, keys, and values. Then, a channel attention mechanism is introduced to adaptively adjust the weights of different channels, thereby emphasizing feature channels more relevant to the object detection task. Finally, a spatial attention mechanism is introduced to generate a two-dimensional attention map through convolution operations, allowing the network to focus on regions such as object contours and edge textures. This process can be represented as follows: (6) in, , and These represent the query, key, and value in self-attention, respectively. This indicates a constant value that prevents division by zero. and These represent max pooling and average pooling operations, respectively. The output features of HAB. The M output features of HAB The features are stitched together along the channel dimension to obtain the final output feature F of the multi-scale dynamic fusion module MDFM, as shown in the following formula.

[0040] (7) Step S12: Construct a detection network for target detection.

[0041] Reference Figure 2The detection network uses YOLOv3. Its architecture mainly consists of three parts: a backbone network, a neck network, and a detection head. The backbone network uses Darknet53 as the feature extractor. The neck structure borrows the idea of ​​FPN and adopts a top-down multi-scale feature fusion strategy. The detection head uses three prediction branches at different scales, each composed of multiple convolutions, and finally outputs the detection results.

[0042] Step S102: Prepare the ExDark dataset and use the ExDark training set to train the constructed low-light target detection network based on multi-scale dynamic fusion to obtain the low-light target detection model based on multi-scale dynamic fusion.

[0043] Specifically, the following steps are included: Step S21: Prepare the ExDark dataset.

[0044] The ExDark dataset is a widely used dataset in the field of low-light object detection, specifically designed for object detection tasks in low-light environments. This dataset contains 12 common object categories: Bicycle, Boat, Bottle, Bus, Car, Cat, Chair, Cup, Dog, Motorbike, People, and Table.

[0045] This method converts the annotation information in the ExDark dataset into the YOLO format, which is convenient for training. The annotation information of each target after conversion includes the target's category and the target's location in the image.

[0046] The annotation information for each target after conversion contains five numbers. X1 is an integer representing the category number of the target, with the number ranging from 0 to 11, representing 12 common categories respectively; Represents the coordinates of the top-left corner of the target bounding box. These five numbers represent the coordinates of the bottom right corner of the bounding box, indicating the target's category and location within the image. Finally, the dataset was divided into training and validation sets in an 8:2 ratio for experimentation.

[0047] Step S22: Train the improved low-light target detection network based on multi-scale dynamic fusion using the ExDark training set.

[0048] (a) The size of the input image is adjusted to 640×640 using the resize() function in the OpenCV library, and used as the input for model training.

[0049] (b) Determine all training parameters, load the pre-trained weights corresponding to the detection network, select SGD as the optimizer, adjust the learning rate to 0.01, set the weight decay to 0.0005, and set the batch size to 16.

[0050] (c) The category loss function uses the binary cross-entropy loss (BCE Loss), and the bounding box loss function uses the sum of squared errors (SSE) loss.

[0051] The improved low-light target detection network based on multi-scale dynamic fusion was trained using the ExDark training set.

[0052] Step S23: Record the training log, which contains the loss value and AP metric for each iteration during training.

[0053] Step S24: Complete the training and obtain a low-light target detection model based on multi-scale dynamic fusion.

[0054] Step S103: Use the ExDark validation set to evaluate the low-light target detection model based on multi-scale dynamic fusion, and adjust the model training parameters according to the training log and evaluation metrics to obtain the optimized model. Step S104: Select the best-performing low-light target detection model based on multi-scale dynamic fusion to perform target detection on the low-light image and obtain the target detection result image.

[0055] To better illustrate the effectiveness of this invention, experiments were conducted on the ExDark dataset and compared with several object detection methods.

[0056] The ExDark dataset is a widely used dataset in the field of low-light object detection, specifically designed for object detection tasks in low-light environments. This dataset contains 7363 images, covering 10 different lighting levels and complex lighting environments (including extremely low light, dim light, dusk, indoor / outdoor low light, etc.), encompassing challenges such as non-uniform lighting, object blurring, and objects hidden in darkness, making it an important benchmark for evaluating detection robustness.

[0057] The comparison results of different methods on the ExDark dataset are shown in Tables 1 and 2. Table 1 shows the AP index of different methods on the first 7 categories of the ExDark dataset, while Table 2 shows the AP results of the remaining 5 categories and the overall mAP50 value. The symbol "-" in the table indicates that the original paper did not provide detection data for the corresponding category.

[0058] Table 1 Comparative experimental results of different methods

[0059] Table 2 Comparative experimental results of different methods

[0060] The experimental results show that the method of this invention exhibits optimal detection performance on the ExDark dataset, achieving an mAP50 of 79.30%, which is the best performance. Specifically, compared to DENet, the method of this invention improves mAP50 by 2.0%; compared to PE-YOLO, it improves by 1.3%; and compared to the currently strong ILE-YOLO and FE-YOLO, the method of this invention also achieves a performance gain of 0.8%. Furthermore, compared to the baseline method YOLOv3l, the method of this invention improves mAP50 by 1.8%, further demonstrating the effectiveness of the method of this invention.

[0061] Figures 5(a) and 5(b) show a subjective quality comparison between the method of this invention and YOLOv3, where Figure 5(a) shows the detection results of YOLOv3 and Figure 5(b) shows the detection results of the method of this invention. The first row of Figures 5(a) and 5(b) shows the detection results for image number 2015_02882 in the ExDark dataset. This image is a nighttime outdoor scene, with a noticeable halo formed by the light from the streetlight in the upper right. The detection results show that YOLOv3 only detected the car in the center of the image and failed to detect the vehicle below the halo of the streetlight on the right side of the image; in contrast, the method of this invention successfully detected both the car in the center of the image and the car under the streetlight. The second row of Figures 5(a) and 5(b) shows the detection results for image number 2015_01991 in the ExDark dataset. This image is a nighttime street scene, with a giant bus occupying most of the image, surrounded by relatively smaller cars, people, and motorcycles. The detection results show that while YOLOv3 can detect large buses, it misses several small-scale pedestrians in the upper left corner of the image. In contrast, the method of this invention can detect not only large buses but also small targets in the image, such as motorcycles and more pedestrians. The comparison between Figures 5(a) and 5(b) fully demonstrates that this method solves, to some extent, the problems of abnormal light source interference and large target scale differences in low-light target detection.

[0062] Reference Figure 6 As shown, this embodiment of the invention also provides a low-light target detection device 100 based on multi-scale dynamic fusion, comprising: Module 110 is used to construct a low-light target detection network based on multi-scale dynamic fusion, which includes a detection network and an improved low-light enhancer. The improved low-light enhancer is constructed by using a multi-scale dynamic fusion enhancer, which includes a Laplacian pyramid, a multi-sensory detail enhancement module and a multi-scale dynamic fusion module. The multi-sensory detail enhancement module is applied to each multi-scale component obtained by decomposing the Laplacian pyramid to enhance the feature representation capability. Training module 120 is used to train the constructed low-light target detection network based on multi-scale dynamic fusion using the ExDark training set to obtain a low-light target detection model based on multi-scale dynamic fusion. Evaluation and optimization module 130 is used to evaluate the low-light target detection model based on multi-scale dynamic fusion using the ExDark validation set, and adjust the model training parameters according to the training log and evaluation metrics to obtain the optimized model. The detection module 140 is used to select the best-performing low-light target detection model based on multi-scale dynamic fusion to perform target detection on the low-light image and obtain the target detection result image.

[0063] like Figure 7 As shown, the present invention provides an electronic device 200 for implementing a method for low-light target detection based on multi-scale dynamic fusion, comprising: a data acquisition card 210, a memory 220, a communication bus 230, and a processor 240.

[0064] The computer device includes a processor 240 and a memory 220. The memory 220 stores a computer program, which includes program instructions. The processor 240 executes the program instructions stored in the computer storage medium. The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), or field-programmable gate arrays (FPGAs). The processor described in this embodiment of the invention can be used to implement a low-light target detection method based on multi-scale dynamic fusion.

[0065] This invention is not limited to the above embodiments. Based on the technical solutions disclosed in this invention, those skilled in the art can make some substitutions and modifications to some of the technical features without creative effort, and all such substitutions and modifications are within the protection scope of this invention.

Claims

1. A low-light target detection method based on multi-scale dynamic fusion, characterized in that, include: A low-light target detection network based on multi-scale dynamic fusion is constructed, which includes a detection network and an improved low-light intensifier. The improved low-light enhancer is a multi-scale dynamic fusion enhancer, which includes a Laplacian pyramid, a multi-sensory detail enhancement module, and a multi-scale dynamic fusion module. The multi-sensory detail enhancement module is applied to each multi-scale component obtained by decomposing the Laplacian pyramid to enhance the feature representation capability. The constructed low-light target detection network based on multi-scale dynamic fusion was trained using the ExDark training set to obtain a low-light target detection model based on multi-scale dynamic fusion. The low-light target detection model based on multi-scale dynamic fusion was evaluated using the ExDark validation set. The model training parameters were adjusted based on the training logs and evaluation metrics to obtain an optimized model. The best-performing low-light target detection model based on multi-scale dynamic fusion was selected to perform target detection on the low-light image, and the target detection result image was obtained.

2. The low-light target detection method based on multi-scale dynamic fusion according to claim 1, characterized in that, An improved low-light enhancer is constructed to improve feature representation capabilities, including: The Laplacian pyramid is used for multi-scale feature decomposition of input images. The multi-sensory detail enhancement module is used to enhance feature components at different resolutions; The multi-scale dynamic fusion module is used to adaptively weight and fuse the enhanced features, and reconstruct the enhanced feature map.

3. The low-light target detection method based on multi-scale dynamic fusion according to claim 2, characterized in that, The Laplacian pyramid performs multi-scale feature decomposition on the input image, including: Gaussian pyramids are used to perform multi-scale feature decomposition on the input image to obtain feature components at different resolutions. The Laplace pyramid was constructed based on the Gaussian pyramid, and the decomposed multi-scale features were obtained.

4. The low-light target detection method based on multi-scale dynamic fusion according to claim 2, characterized in that, The multi-sensory detail enhancement module includes a local edge enhancement branch and a global feature enhancement branch; The local edge enhancement branch processing flow is as follows: Local edge enhancement branch utilization and The gradient information of the input features in the horizontal and vertical directions is calculated separately. The receptive field is further expanded by dilated convolution, and residual connections are introduced to reduce information loss. The global feature enhancement branch processing flow is as follows: The global feature enhancement branch consists of six stacked depthwise separable convolutional layers, each followed by a ReLU activation function and a skip connection, with the channel dimension aligned through a depthwise separable convolution.

5. The low-light target detection method based on multi-scale dynamic fusion according to claim 2, characterized in that, The multi-scale dynamic fusion module adaptively weights and fuses the enhanced features, reconstructing the enhanced feature map, including: The high-resolution and low-resolution feature maps of adjacent scales, which are enhanced by the multi-sensory detail enhancement module, are processed by two convolution operations and a sigmoid activation function in sequence. The fused features are obtained by adding adjacent resolution features element by element. The fused features are divided into M blocks along the channel dimension and input into the hybrid attention block for dynamic fusion. The dynamically fused features are then stitched together along the channel dimension to obtain the final fused features.

6. The low-light target detection method based on multi-scale dynamic fusion according to claim 5, characterized in that, The processing flow for mixed attention blocks is as follows: Using a self-attention mechanism, attention weights are dynamically generated through queries, keys, and values; A channel attention mechanism is introduced to adaptively adjust the weights of different channels; A spatial attention mechanism is introduced to generate a two-dimensional attention map through convolution operations, enabling the network to focus on the target contour and edge texture regions.

7. The low-light target detection method based on multi-scale dynamic fusion according to claim 1, characterized in that, The detection network uses YOLOv3 for target detection.

8. The low-light target detection method based on multi-scale dynamic fusion according to claim 1, characterized in that, The constructed low-light target detection network based on multi-scale dynamic fusion was trained using the ExDark training set, including: The annotation information in the ExDark dataset is converted into the YOLO format, which is convenient for training; the annotation information of each target after conversion includes the target's category and the target's location in the image; A low-light target detection network based on multi-scale dynamic fusion was trained using the ExDark dataset; Record the training log, which contains the loss value and AP metric for each iteration during training, to obtain a low-light target detection model based on improved multi-scale dynamic fusion.

9. A low-light target detection device based on multi-scale dynamic fusion according to any one of claims 1-8, characterized in that, include: A building block is used to construct a low-light target detection network that incorporates a detection network and an improved low-light intensifier, through multi-scale dynamic fusion. By improving the Laplacian pyramid, multi-sensory detail enhancement module, and multi-scale dynamic fusion module in the low-light enhancer, the multi-sensory detail enhancement module is applied to each multi-scale component obtained by Laplacian pyramid decomposition to enhance the feature representation capability. The training module is used to train the constructed low-light target detection network based on multi-scale dynamic fusion using the ExDark training set, so as to obtain the low-light target detection model based on multi-scale dynamic fusion. The evaluation and optimization module is used to evaluate the low-light target detection model based on multi-scale dynamic fusion using the ExDark validation set, and adjust the model training parameters according to the training logs and evaluation metrics to obtain the optimized model. The detection module is used to select the best-performing low-light target detection model based on multi-scale dynamic fusion to perform target detection on low-light images and obtain the target detection result image.

10. An electronic device, characterized in that, Including the processor and memory; The memory is used to store computer programs, the computer programs including program instructions; The processor is used to invoke a program stored in the memory to execute the target detection method as described in any one of claims 1-7.

Images