Low-light image enhancement method and system based on edge extraction and feature fusion

Abstract

The invention discloses a low-light image enhancement method and system based on edge extraction and feature fusion, and aims to improve the image quality under a low-light condition, and the method comprises the steps: differential convolution edge extraction, feature fusion and image enhancement, effective extraction of image edge information through a differential convolution kernel, and combination of global and local features through a feature fusion module. The image enhancement module utilizes a deep learning network to improve image brightness and suppress noise, the method is suitable for embedded equipment with a high-performance computing environment and limited resources, the performance is excellent on a plurality of data sets through experimental verification, the image definition and details are remarkably improved, and the method is suitable for the fields of security monitoring, automatic driving and the like and has a wide application prospect.

Description

TECHNICAL FIELD

[0001] The present application relates to a low-light image enhancement method and system based on edge extraction and feature fusion, more specifically, to a low-light image enhancement method and system based on an edge extraction module, a feature fusion module and an image enhancement module, belonging to the field of low-light image enhancement. BACKGROUND

[0002] In low-light environments, image quality will be greatly reduced due to insufficient illumination, which is manifested as loss of details, increase of noise and color distortion. This quality degradation poses a serious challenge to computer vision tasks such as object detection, image segmentation and scene understanding, especially in critical fields such as autonomous driving, security monitoring and medical imaging, where high-quality image input is crucial to system performance. Therefore, there is an urgent need to develop a low-light image enhancement technology that can effectively restore image details, improve contrast and suppress noise.

[0003] Traditional methods, such as histogram equalization and gamma correction, are easy to implement, but they are not satisfactory in detail restoration and may cause over-enhancement or noise amplification problems, and their generalization ability is limited, making it difficult to adapt to diverse low-light scenes.

[0004] With the rise of deep learning technology, low-light image enhancement methods based on convolutional neural networks (CNN) have gradually become mainstream. These methods can automatically learn the mapping from low-light to normal-light images from massive data through end-to-end learning mechanism, significantly improving the enhancement effect. However, current deep learning methods often increase network size and complexity to improve performance, resulting in a sharp increase in model parameter quantity and computational complexity, making it difficult to deploy and process in real time on resource-constrained edge devices. SUMMARY

[0005] To solve the above technical problems, the present application proposes a low-light image enhancement method based on an edge extraction module, a feature fusion module and an image enhancement module. By introducing the edge extraction module, the edge and texture information of the image is effectively captured, providing important structural features for subsequent enhancement. The feature fusion module combines low-level details and high-level semantic information through a multi-scale feature fusion strategy, further improving the feature expression capability. The image enhancement module uses the fused features for nonlinear mapping to generate high-quality enhanced images. The present application has the technical features of significantly reducing the computational complexity of the model while ensuring the enhancement effect, suitable for efficient operation on cloud servers and embedded edge devices, and meeting the needs of actual applications in terms of computational efficiency and deployment flexibility while improving image quality.

[0006] To achieve the above purpose, the present application is realized by the following technical scheme: The application discloses a low-light image enhancement method based on edge extraction and feature fusion, and the method comprises the following steps: Step 1: edge and gradient information of an image is extracted through a multi-direction differential convolution kernel to generate a multi-dimensional edge feature map; Step 2: the edge feature map is fused with original low-light image features in multiple levels to combine global semantics and local details, so that more abundant feature representation is provided for image enhancement; Step 3: a deep learning network is used to perform nonlinear mapping on the image to output an enhanced image; Step 4: model quantization and hardware acceleration are used to realize real-time processing on an embedded device, and step 4 comprises the following steps: a PyTorch model is converted into an ONNX format, and then the model is optimized into an FP16 precision execution engine through TensorRT on an NVIDIJetsonNano edge device; a real-time processing flow is as follows: a camera is used for input; the NVIDIJetsonNano is used for enhancement calculation; and a USB interface is used for output to an LED screen, and a time delay is less than 50ms@720p; a hardware acceleration method comprises the following steps: a differential convolution calculation is parallelized by using a CUDA kernel function; and a sparse attention matrix operation is accelerated by using a TensorCore.

[0007] Preferably, the step 3) is nonlinear mapping of the fused features by using a lightweight Transformer network (a deep learning model based on a self-attention mechanism).

[0008] Preferably, the step 1) specifically comprises the following steps: Ring, radial and cross-shaped differential convolution kernels are used, wherein a calculation formula of the ring-shaped differential convolution kernel is as follows: ; In the formula, W and X respectively represent learnable parameters of the convolution kernel and a feature map to be convolved, y represents an output of the differential convolution, and the subscripts represent rows and columns in a 3*3 region, wherein x 11 represents the first row and the first column in the 3*3 region; multi-scale enhancement processing is performed on the extracted gradient information, and the multi-scale enhancement processing comprises a compact dilated convolution module and a compact spatial attention module; the compact dilated convolution module uses convolution kernels with a dilated rate of 1, 2 and 3; and the compact spatial attention module generates a dynamic weight mask through a Sigmoid activation function.

[0009] Preferably, the step 2) comprises the following steps: a double-branch structure is used in feature fusion; a first branch is used to extract a bottom layer feature F1 through 3*3 convolution on an original image; and a second branch is used to extract a high layer feature F2 through 2-layer convolution on an edge map; A fusion formula is as follows: F fuse= alpha * F1 + (1-alpha) * F2, where alpha is dynamically calculated from the variance of F1 and F2, and the calculation formula is: alpha = Var(F1) / (Var(F1) + Var(F2)) By spatially adaptive normalization, the feature distribution is dynamically adjusted, and the normalized parameters mu and sigma are generated by inputting the semantic segmentation graph.

[0010] Preferably, step 3) is specifically: using a four-level UNet structure, the encoder uses step 2 convolution downsampling to 1 / 16 resolution; Each level contains a Transformer module based on sparse attention mechanism, which divides the feature map into 8x8 windows for local self-attention calculation; The decoder gradually recovers the resolution through transposed convolution, and the edge feature map is injected through the jump connection.

[0011] The system of the low-light image enhancement method based on edge extraction and feature fusion includes a differential convolution-based edge extraction module, a feature fusion module and an image enhancement module connected by data; The differential convolution-based edge extraction module in step 1) can effectively extract edge information in the image by using a plurality of differential convolution kernels in different directions; The feature fusion module in step 2) fuses the edge features output by the edge extraction module with the features of the original image, and through a multi-level fusion strategy, the global information and local details of the image can be effectively combined to provide more rich feature representation for image enhancement; The image enhancement module in step 3) uses the fused features for image enhancement processing, and through a deep learning network, the image is nonlinearly mapped, which can effectively suppress noise and artifacts while enhancing image brightness, and improve the visual quality of the image.

[0012] Preferably, the operation mode of the differential convolution-based edge extraction module includes the following: S1) Multi-dimensional edge information extraction branch: differential convolution is used to extract edge information of different dimensional features, and differential convolution encodes prior information contained in traditional edge detection operators into convolution layers through careful design of pixel gradient calculation strategy, which can more flexibly extract complex image gradient information; Differential convolution first calculates the gradient corresponding to each pixel according to the gradient calculation strategy designed by hand, and then convolves the calculated gradient using a learnable convolution kernel; According to the corresponding pixel gradient calculation strategy, the gradient information extracted by a single differential convolution is limited, in order to pay attention to local edge information and global edge information at the same time, the present application uses differential convolution to extract edge information of different dimensions and designs different differential convolution to extract multiple gradient information. Further, taking a 3x3 ring differential convolution as an example, it first calculates the gradient between adjacent pixels in the clockwise direction, and then convolves the calculated gradient information through a learnable convolution kernel, and the corresponding calculation formula is as follows: ; In the above formula, W and X represent the learnable parameters of the convolution kernel and the feature map to be convolved, respectively, y represents the output of the differential convolution, and the subscripts represent the row and column in the 3x3 region, x 11 represents the first row and first column in the 3x3 region; S2) Edge information processing branch: further process the obtained edge information to better enhance the texture and detail features of the image. In the present application, the edge information processing branch further processes the extracted edge information by introducing a compact dilated convolution module and a compact spatial attention module. The compact dilated convolution module extracts multi-scale edge features and fuses them to enhance the edge expression ability. The compact spatial attention module dynamically adjusts the feature weights through the spatial attention mechanism to highlight the key edge regions. The combination of the two significantly improves the texture and detail features of the image, providing high-quality edge information support for subsequent image enhancement; S3) Multi-dimensional edge information fusion: process and fuse the multi-dimensional edge information extracted by the edge information processing branch. The fused feature map is output as an image gradient map after Sigmoid activation function.

[0013] Preferably, the operation mode of the feature fusion module includes the following: S1) Feature preprocessing branch: feature extraction and processing of the edge map and input image through the combination of conventional convolution layers and Sigmoid activation functions; S2) Feature fusion attention mechanism: the feature fusion attention mechanism effectively fuses the gradient image (such as edge information) and input features through convolution and standardization operations, thereby enhancing the feature expression ability of the image. Specifically, this module first performs convolution operations on the gradient image and input features to extract their respective multi-level features. Then, the features are normalized through standardization to ensure the stability of the feature distribution. Finally, the processed gradient features and input features are weighted and fused to generate a fused feature map. This fusion method not only preserves the global information of the input image, but also introduces the details and texture features from the gradient image, thereby providing more rich feature representations for subsequent image enhancement.

[0014] S3) Feature normalization: normalize image features through spatial adaptive normalization. By introducing a semantic layout map as input, the parameters (such as mean and variance) of the normalization layer are dynamically adjusted to better preserve the input semantic information when generating images. Unlike traditional normalization methods, spatial adaptive normalization can adaptively adjust the normalization parameters of the feature map according to the spatial position of the input semantic map, thereby generating higher quality images.

[0015] Preferably, the network structure in the image enhancement module adopts a UNet-style network structure, which is divided into three parts: encoder, feature conversion module and decoder; the network has a total of four levels, each level represents a different resolution, and each level is composed of a stack of designed basic modules; The sparsity-improved Transformer (a deep learning model based on the self-attention mechanism) will be introduced to reduce computational complexity while maintaining the model's efficient performance. The sparse attention mechanism only calculates the association between local areas or key positions, significantly reducing the amount of computation. The attention range is limited through fixed patterns (such as local windows, strip areas) or dynamic learning (such as content-based sparsification), thereby reducing computational overhead while maintaining performance.

[0016] The present invention provides an edge computing device for a low-light image enhancement method based on edge extraction and feature fusion, comprising an edge computing module, a video input module, and a result output module connected in communication; Edge computing module: uses NVIDIA Jetson Nano for real-time low-light image enhancement calculations. Specifically, Jetson Nano uses the NVIDIA Maxwell architecture with a quad-core Cortex A57 MPCore processor and 128 CUDA cores, capable of completing complex neural network calculations while using only 5 watts of power. Video input module: uses a camera to capture images as input for low-light images; Result output module: A portable LED screen powered by USB is used to output and display the results; Use the following steps to deploy the deep learning model trained on the server cluster on the edge computing device: Convert neural network models trained using the deep learning framework PyTorch to ONNX, a common deep learning format. Convert neural network models in ONNX format to TensorRT format. TensorRT is a C++ inference framework for NVIDIA GPUs that facilitates high-performance inference on GPUs. TensorRT is used to optimize the pre-trained model into an execution engine through quantization, operator fusion, kernel automatic adjustment, dynamic tensor memory, and multi-stream execution. The execution engine TensorRT is then called using the Python interface to perform low-light image enhancement calculations on the neural network.

[0017] Beneficial effects: the present application has the technical characteristics of being able to significantly reduce the computational complexity of the model while ensuring the enhancement effect, being suitable for efficient operation on cloud servers and embedded edge devices, being able to improve the image quality while meeting the demand for computing efficiency and deployment flexibility in actual application, etc.; the present application has a significant performance advantage in low-light image enhancement tasks, can effectively improve the clarity and detail performance of the image, and is suitable for multiple fields such as security monitoring, automatic driving, etc. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a global schematic diagram of the present application.

[0019] Figure 2 is a different type of hole convolution mode diagram of the present application.

[0020] Figure 3 is a hole convolution module diagram of the present application.

[0021] Figure 4 is a channel and spatial attention module diagram of the present application.

[0022] Figure 5 is a multi-hole rate convolution module diagram of the present application.

[0023] Figure 6 is one of the specific detail diagrams of the image enhancement module of the present application.

[0024] Figure 7 is the second specific detail diagram of the image enhancement module of the present application.

[0025] Figure 8 is the third specific detail diagram of the image enhancement module of the present application. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application Figures 1 to 8 , and it is obvious that the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0027] Figure 1Comprise a, b, c three parts, a part: present the overall flow framework of low light image enhancement method. After inputting "image", it is processed by "residual module", "transform block + structure perception fusion block" (multiple groups in series), and "residual module" again, and finally output the enhanced "image", which embodies the multi-module series processing flow from the original image to the enhanced image. The structure perception fusion block and the transform block cooperate to gradually optimize the image features. b part: edge extraction and preliminary fusion sub-process. After the "pyramid dilated convolution module" (multiple levels) extracts the features, it is processed by "channel dimension convolution module" and "channel space attention module" respectively. After the outputs of the two are spliced, the final output is processed "image" through "Sigmoid activation function". It mainly realizes feature extraction and preliminary fusion based on pyramid dilated convolution and attention mechanism. c part: feature fusion and transformation sub-process. After the "feature map" is input, it is processed by "convolution, residual block, convolution", and then output a new "feature map" through "convolution" after being processed by "fusion block", "ReLU + convolution", "spatial adaptive normalization + ReLU", etc. It focuses on the depth fusion and transformation of features to optimize the image feature representation.

[0028] Figure 2 Display the operation logic of three difference convolution kernels, each group contains the original pixel matrix and the result matrix after operation: the first group: 3x3 original pixel matrix (x1-x9), after operation, a new 3x3 matrix is obtained, the center position is 0, and the rest positions are the difference values of corresponding original pixels (such as x1-x5, etc.), which embodies a difference calculation method for extracting image gradient information. The second group: 3x3 original pixel matrix (x1-x9, center x5 is blank), after operation, the center is 0, and the rest positions are the difference values between original pixels (such as x1-x2, etc.), which presents another pixel difference logic for capturing edge gradients in different directions. The third group: 5x5 original pixel matrix (x1-x25, containing blank pixels), after operation, the center is 0, and the rest positions are the difference values of corresponding original pixels (such as x1-x7, etc.), which is used to extract image gradients in more complex scenes and adapt to image edge extraction of different sizes and structures. Figure 3 The processing flow of the sub-blocks / function blocks: input is processed by "sub-block / function block_1", "sub-block / function block_2", then "3x3 convolution kernel dilated convolution operation", then "ReLU", "1x1" convolution, and the branch signal from "sub-block / function block_2" is added, and then processed by "sub-block / function block_3", "sub-block / function block_4", which presents the feature processing and residual connection logic within the local module, which is used to optimize feature extraction and transmission and enhance the processing capability of the module for image features.

[0029] Figure 4is the attention mechanism and convolution fusion process: the input is output after being multiplied by the original input branch after "rectified linear unit", "convolution 1x1", "convolution 3x3", and "sigmoid activation function". It embodies the dimensionality reduction of input features, the extraction of local features, the generation of attention weights through activation functions, the fusion with original features, the highlighting of key features, the suppression of irrelevant information, and the improvement of the relevance of feature expression.

[0030] Figure 5 is a multi-scale dilated convolution process: the input is output after being added by four branches of "3x3 convolution (dilation rates are 5, 7, 9, and 11)" after "rectified linear unit" and "convolution 1x1". Through dilated convolutions with different dilation rates, multi-scale image features are extracted, and the feature representation is enriched after fusion, adapting to the feature extraction needs of different size targets or scenes.

[0031] Figure 6 The feature processing process related to Transformer is shown: the input is output after being added by the original input branch after "layer normalization", "Token kernel perception sparse self-attention", and "layer normalization", "multi-scale feature network". Using layer normalization to stabilize feature distribution, sparse self-attention to reduce computational complexity, and multi-scale feature network to extract rich features, efficient feature transformation and enhancement are achieved, adapting to the depth processing of image features.

[0032] Figure 7 is an attention process based on depth separable convolution: the input is output after being multiplied by the original input branch after "mask", "sparsification", "restoring feature space / channel layout", and "Softmax activation function" through three parallel branches ("1x1+3x3 depth separable convolution+dimension reshaping") to generate Q, K, and other features. With the help of depth separable convolution dimensionality reduction, combined with attention mechanism focusing on key features, feature extraction and fusion are optimized, and the ability to capture important image features is improved.

[0033] Figure 8 is a cross-fusion convolution process: the input is output after being multiplied by the original input branch after two parallel branches ("1x1+3x3 convolution to extract local features+rectified linear unit") cross-fusion (two "C" nodes), and two parallel "3x3 convolution to extract local features+rectified linear unit" cross-fusion ("C" node) again, and finally "1x1". Through multiple cross-fusion of local features extracted by different branches, feature interaction and integration are strengthened, feature richness and expression ability are improved, and image feature processing effect is optimized.

[0034] The technical scheme / principle of the present application: In view of the problem that the image quality under low light conditions affects computer vision tasks, a low light image enhancement method based on edge extraction, feature fusion and image enhancement module is proposed. The low light image enhancement method mainly includes four key steps: "edge extraction → feature fusion → nonlinear mapping → device deployment". Through the cooperative operation of different modules in each step, the low light image is realized from edge extraction, feature processing to final enhancement output, and real-time processing on embedded devices.

[0035] Edge extraction: Through multi-directional difference convolution kernel and multi-scale enhancement module, the image edge gradient information is extracted and optimized. Feature fusion: with double branch structure, attention mechanism and spatial normalization, the edge features and original image features are integrated. Image enhancement: based on lightweight U-Net and sparse Transformer, the feature nonlinear mapping and image quality improvement are realized. Embedded deployment: through model format conversion and optimization, it is adapted to NVIDIA JetsonNano (NVIDIA-Computer) real-time processing, and each module is connected through data flow and function, so as to achieve the goal of low light image enhancement.

[0036] A specific step implementation: (1) Step 1: Multi-dimensional edge information extraction Difference convolution kernel operation: adopt ring, radial, and cross convolution kernels to encode image gradient information. Taking ring difference convolution as an example, for a 3x3 pixel area (such as building outline and vehicle edge in night road monitoring image), the gradient of adjacent pixels is calculated according to the formula: ; Encode prior edge detection logic into convolution layer. In actual operation, three convolution kernels process low light image in parallel, such as: ring convolution kernel: capture building arc outline and vehicle ring-shaped tail light edge. Radial convolution kernel: extract road marking and light line extension direction gradient. Cross convolution kernel: strengthen the horizontal and vertical edges of power poles and traffic signal poles, and output multi-dimensional gradient feature map.

[0037] Multi-scale enhancement processing. Compact dilated convolution: Deploy convolution kernels with dilation rates 1, 2, 3 in parallel. When processing distant blurred vehicles in low-light images, the dilation rate 1 convolution kernel focuses on vehicle edge details (such as tire outlines). The dilation rate 2 convolution kernel relates vehicles to environmental features such as road shadows. The dilation rate 3 convolution kernel covers the global position of the vehicle (such as lane distribution) and fuses multi-scale features to enhance edge expression. Compact spatial attention: After ReLU (Rectified Linear Unit, which introduces nonlinearity into the network, improves the ability to fit complex functions, and speeds up network training) activation, 1x1 convolution, 3x3 convolution, and Sigmoid activation, dynamic weight masks are generated. In low-light road scenes, key edges such as road markings and pedestrian outlines are given high weights, and weakly related areas such as background vegetation and sky are given low weights, highlighting effective edge information.

[0038] (II) Step 2: Feature fusion - dual-branch feature processing.

[0039] Feature preprocessing branch: The first branch extracts low-level features (such as low-light road texture and vegetation outline) from the original image using a 3x3 convolution; the second branch performs 2-layer convolution on the edge map (output of Step 1) to extract high-level features (such as building outlines and vehicle edge abstract features).

[0040] Dynamic weighted fusion: According to the formula F fuse = α·F1 + (1-α)·F2 fused features, where α = Var(F1) / (Var(F1) + Var(F2)). Taking the example of autonomous driving in low-light road conditions: If the original image is blurred due to poor lighting (F1 variance is small), and the edge map has clear road markings (F2 variance is large), then α is small, and the fusion focuses on the high-level edge features of F2, highlighting the key edges of the road.

[0041] Spatially adaptive normalization: Introduce a semantic segmentation map (such as the pedestrian area, building area, and road area of a city street image) to generate normalization parameters μ and σ, and dynamically adjust the feature distribution. For building area features, the adjusted building edges are more prominent and stable in the fused features, preserving the input semantic information and providing high-quality fused features for subsequent enhancement.

[0042] (III) Step 3: Nonlinear mapping based on lightweight network Lightweight U-Net architecture. A four-level U-Net structure is adopted. The encoder down-samples to 1 / 16 resolution (e.g., the lesion area of low-light medical images) through step-2 convolution, gradually compresses the feature dimension, and extracts high-level semantics. A sparse attention Transformer module is embedded in each level: the feature map is divided into 8x8 windows (e.g., the local area of the lesion in medical images), local self-attention is calculated, and the lesion edge and internal texture association are focused. By fixing the window to limit the attention range, the computational complexity is reduced (e.g., only the pixel correlation within the window is calculated), and the embedded device is adapted.

[0043] Decoder with skip connection. The decoder restores the resolution (from 1 / 16 to gradually restore) through transposed convolution, and the skip connection injects the edge feature map of step 1. Taking medical images as an example: the lesion edge features are supplemented in the decoder restoration stage, providing details of the fracture line and tumor boundary, assisting in clearly restoring the lesion morphology, and improving the diagnostic value of the image. For example, in low-light X-ray images, the lesion outline is clearer after edge feature injection, optimizing the visual quality and diagnostic accuracy.

[0044] (4) Step 4: Embedded device deployment Edge device architecture. Computing module: NVIDIA Jetson Nano is equipped with a quad-core Cortex-A57 processor and 128 CUDA cores, running at 5W power. When processing surveillance camera low-light images, it quickly performs convolution and attention calculation (e.g., pedestrian and vehicle feature extraction in street images), meeting real-time requirements.

[0045] Input and output module: the camera captures low-light images (e.g., night park surveillance), the USB-powered LED screen outputs the enhanced results, realizing a "capture-enhance-display" closed loop.

[0046] Model deployment process. Server side: the PyTorch-trained model is converted to ONNX format (cross-framework universal), and then to TensorRT format (adapted to NVIDIA GPU). Edge side: TensorRT optimizes the model through quantization (e.g., converting 32-bit floating-point parameters to 8-bit integers) and operator fusion (combining consecutive convolution and activation operations), generating an execution engine. Inference call: through the Python interface, the execution engine is called to process low-light images (e.g., automatic driving vehicle camera images), quickly complete the enhancement calculation, and output the enhanced image in real time to assist in driving decision-making.

[0047] A specific scenario implementation example: taking the security monitoring scenario as an example, the night monitoring image is blurred due to insufficient light, and the face and vehicle details are difficult to identify. The implementation process is as follows: Edge extraction: three kinds of differential convolution kernels extract face contour, vehicle edge gradient information; multi-scale enhancement module (hole convolution + spatial attention) strengthens the edge to generate clear edge map (such as vehicle tail light, face contour).

[0048] Feature fusion: dual-branch fusion of original image and edge map features, dynamic weighting highlights face and vehicle key information (such as facial features, vehicle brand); spatial adaptive normalization adapts to scene semantics (road, building area), optimizes feature distribution.

[0049] Nonlinear mapping: the lightweight U-Net network restores the image details, and the sparse attention focuses on the facial features and vehicle features, and outputs the enhanced image (such as clear and identifiable face, and recognizable vehicle brand).

[0050] Edge deployment: JetsonNano processes and displays the enhanced image in real time, assists security personnel in accurate identification, and embodies the application value of the method in actual scenes.

[0051] The present application effectively solves the problem of low-light image enhancement by multi-dimensional edge extraction, dynamic feature fusion, lightweight network mapping and embedded deployment, and covers the fields of security, automatic driving, medical imaging, etc. The present application provides a complete solution for the engineering application of low-light image enhancement technology, and helps the efficient execution of computer vision tasks in related fields.

[0052] Finally, it should be noted that the present application is not limited to the above embodiments, but can have many variations. All variations that can be directly derived or inferred from the disclosed content by those of ordinary skill in the art should be considered within the scope of the present application.

Claims

1. A low-light image enhancement method based on edge extraction and feature fusion, characterized in that The method comprises the following steps: Step 1) Extract the edge and gradient information of the image through multi-directional differential convolution kernel to generate a multi-dimensional edge feature map; Step 2) Multi-level fusion of edge feature maps and original low-light image features, combining global semantics with local details, thus providing richer feature representation for image enhancement; Step 3) Perform nonlinear mapping on the image through a deep learning network and output an enhanced image; Step 4) Realize real-time processing on embedded devices through model quantization and hardware acceleration.

2. The low-light image enhancement method based on edge extraction and feature fusion according to claim 1, characterized in that: In step 3), a lightweight Transformer network is used to perform nonlinear mapping on the fusion features.

3. The low-light image enhancement method based on edge extraction and feature fusion according to claim 1, characterized in that: Step 1) specifically includes: Three types of differential convolution kernels are used: circular, radial and cross. The calculation formula of circular differential convolution is: ； In the above formula, W and X represent the learnable parameters of the convolution kernel and the convolved feature map, respectively, and y represents the output of the differential convolution; the subscripts represent the rows and columns in the 3×3 area, where x 11 Represents the first row and first column in a 3×3 area; multi-scale enhancement processing of the extracted gradient information includes a compact void convolution module and a compact spatial attention module. The compact void convolution module uses convolution kernels with void rates of 1, 2, and 3, and the compact spatial attention module generates a dynamic weight mask through a Sigmoid activation function.

4. The low-light image enhancement method based on edge extraction and feature fusion according to claim 1, characterized in that: Step 2) includes: a dual-branch structure is used for feature fusion, where the first branch performs a 3×3 convolution on the original image to extract the bottom-level features F1, and the second branch performs a 2-layer convolution on the edge map to extract the high-level features F2; The fusion formula is: F fuse = α·F1 + (1-α)·F2, where α is dynamically calculated from the variance of F1 and F2, and the calculation formula is: α = Var(F1) / (Var(F1) + Var(F2)). The feature distribution is dynamically adjusted through spatial adaptive normalization, and the semantic segmentation map is input to generate the normalization parameters μ and σ.

5. The low-light image enhancement method based on edge extraction and feature fusion according to claim 1, characterized in that: Step 3) Specifically, a four-layer U-Net structure is used, and the encoder uses stride-2 convolution to downsample to 1 / 16 resolution. Each layer contains a Transformer module based on the sparse attention mechanism, which divides the feature map into 8×8 windows for local self-attention calculation. The decoder gradually restores the resolution through transposed convolution and injects skip connections into the edge feature map.

6. The system for low-light image enhancement method based on edge extraction and feature fusion according to any one of claims 1 to 5, characterized in that: The method comprises a data-connected differential convolution-based edge extraction module, a feature fusion module, and an image enhancement module; in the step 1), the differential convolution-based edge extraction module can effectively extract edge information from the image by utilizing differential convolution kernels in multiple directions; in the step 2), the feature fusion module fuses the edge features output by the edge extraction module with the features of the original image, and can effectively combine the global information and local details of the image through a multi-level fusion strategy, thereby providing a richer feature representation for image enhancement; in the step 3), the image enhancement module uses the fused features to perform image enhancement processing, and performs nonlinear mapping on the image through a deep learning network, which can effectively suppress noise and artifacts while enhancing the image brightness, thereby improving the visual quality of the image.

7. The system according to claim 6, characterized in that The operation of the edge extraction module based on differential convolution includes the following: S1) Multi-dimensional edge information extraction branch: Differential convolution is used to extract edge information of features of different dimensions. Differential convolution explicitly encodes the prior information contained in traditional edge detection operators into the convolution layer, which can more flexibly extract complex image gradient information. Differential convolution first calculates the gradient corresponding to each pixel according to a manually pre-designed gradient calculation strategy, and then convolves the calculated gradient with a learnable convolution kernel. S2) Edge Information Processing Branch: The edge information processing branch further processes the extracted edge information through the compact dilated convolution module and compact spatial attention module introduced in [1]. The compact dilated convolution module uses dilated convolution to extract multi-scale edge features and fuse them to enhance edge representation. The compact spatial attention module dynamically adjusts feature weights through the spatial attention mechanism to highlight key edge areas. S3) Multi-dimensional edge information fusion: The multi-dimensional edge information extracted by the edge information processing branch is processed and fused, and the fused feature map is output as an image gradient map through an activation function.

8. The system according to claim 6, wherein: The operation of the feature fusion module includes the following: S1) Feature preprocessing branch: extracts and processes features from the edge map and input image through a combination of conventional convolutional layers and Sigmoid activation functions; S2) Feature Fusion Attention Mechanism: Through convolution and normalization operations, the gradient image is effectively fused with the input features, thereby enhancing the feature expression ability of the image; S3) Feature Normalization: Image features are normalized through spatially adaptive normalization. By introducing a semantic layout map as input, the parameters of the normalization layer are dynamically adjusted to better preserve the input semantic information when generating images.

9. The system according to claim 6, wherein: The network structure in the image enhancement module adopts the U-Net network structure, which is divided into three parts: encoder, feature conversion module and decoder; The Transformer will be improved by introducing sparsity to reduce computational complexity while maintaining the model's efficient performance. The sparse attention mechanism only calculates the association between local areas or key positions, significantly reducing the amount of computation; the attention range is limited through fixed patterns or dynamic learning, thereby reducing computational overhead while maintaining performance.

10. The edge computing device of any one of claims 1 to 5, wherein: Contains edge computing module, video input module, and result output module with communication connection; Edge computing module: Utilizes the NVIDIA Jetson Nano for real-time low-light image enhancement calculations. Specifically, the Jetson Nano utilizes the NVIDIA Maxwell architecture with a quad-core Cortex-A57 MPCore processor and 128 CUDA cores, enabling complex neural network calculations while using only 5 watts of power. Video input module: uses a camera to capture images as input for low-light images; Result output module: A portable LED screen powered by USB is used to output and display the results; Use the following steps to deploy a deep learning model trained on a server cluster on an edge computing device: Convert the neural network model trained using the deep learning framework PyTorch to the universal deep learning format ONNX; Convert the neural network model in ONNX format to a model in Tensor RT format, and then use the Python interface to call the execution engine Tensor RT to perform low-light image enhancement calculations on the neural network.

Images