Low-illumination image enhancement algorithm based on deep convolution

Abstract

The invention discloses a low-illumination image enhancement algorithm based on deep convolution, which is suitable for the field of image processing, and the core of the algorithm lies in an efficient noise reduction process and accurate decomposition of an illumination image and a reflection image. Comprising the following steps that firstly, a channel noise reduction module works cooperatively through multiple noise reduction strategies, noise in an image is effectively removed, and a foundation is laid for follow-up processing; then, the illumination image decomposition network and the reflection image decomposition network accurately separate the brightness information, texture and color information of the image based on a deep learning structure. And finally, by ingeniously fusing the illumination image and the reflection image, reconstructing a high-quality enhanced image, remarkably improving the brightness, contrast and details of the image, and meanwhile, retaining the natural color of the image. The algorithm has the advantages of being high in noise reduction capacity, accurate in illumination image decomposition, efficient in reflection image decomposition, good in enhancement effect and the like, and an effective and practical solution is provided for image enhancement under the low-illumination condition.

Description

Technical Field

[0001] The present invention relates to a low-light image enhancement algorithm based on deep convolution, which is based on cutting-edge technology in the field of image processing, and in particular focuses on image enhancement algorithms under low-light conditions, belonging to the field of image processing. Background Art

[0002] Low-light image enhancement technology is a technology that aims to improve image quality and make it more suitable for human vision or machine vision processing. In low-light conditions, due to insufficient light, images often show low brightness, poor contrast, missing details and increased noise, which greatly affects the usability and visual effects of the image. The goal of image enhancement is to improve the visibility and analyzability of images through algorithms, so that images can still present clear, bright and detailed pictures in low-light environments.

[0003] Before the popularization of deep learning technology, image enhancement mainly relied on traditional image processing methods, including but not limited to: Histogram equalization: By stretching the grayscale distribution of the image, the contrast of the image is enhanced to make the image look brighter and more vivid. Retinex theory: Based on the fact that the image can be decomposed into the product of reflectance and light intensity, the Retinex method attempts to restore the true reflectance of the image to overcome the effects of uneven lighting or poor lighting conditions. Multiscale morphological operations: By opening and closing the image through structural elements of different scales, noise can be removed, edges and textures can be highlighted, and the details of the image can be enhanced. Frequency domain filtering: Filtering the image in the frequency domain can suppress low-frequency noise and enhance high-frequency details, thereby improving image quality. However, traditional methods are often unable to cope with images with complex lighting conditions and high noise, and are prone to problems such as over-enhancement, color distortion, or loss of details.

[0004] With the development of deep learning technology, especially the emergence of deep convolutional neural networks (DCNNs), the field of image processing has ushered in a revolutionary change. DCNNs can automatically learn the multi-level features of images, from low-level textures and edges to high-level semantic information. This ability makes DCNNs perform well in tasks such as image classification, object detection, and semantic segmentation. Similarly, DCNNs have also been applied to image enhancement, especially low-light image enhancement, showing advantages over traditional methods. The following are some deep learning networks that have been widely studied and applied in this field: Generative Adversarial Network (GAN)-based methods use the generative adversarial network framework, in which the generator is responsible for converting low-light images into images under normal lighting conditions, while the discriminator distinguishes whether the generated image is real. This method can generate more natural and realistic enhanced images. Methods based on the encoder-decoder architecture, Zero-DCE (Zero-reference Deep Curve Estimation): No reference image is required, and the curve parameters are directly estimated from the input image to achieve image enhancement. This method combines the encoder-decoder structure and curve estimation, which can effectively improve image brightness and contrast. Residual-to-Space Image (RESIDE) is a residual learning-based method that contains multiple residual learning modules that learn the residual between the input image and the ideal image, and then add the residual back to the input image to achieve enhancement.

[0005] Low-light image enhancement algorithms based on deep convolution usually adopt an end-to-end learning approach to directly learn the image mapping from low-light images to images under normal lighting conditions. This algorithm usually includes the following key points: Feature learning: Automatically learn the deep features of the image, including lighting distribution, noise patterns, and detail information, through a multi-layer convolutional neural network. Lighting estimation: Some algorithms will specially design modules to estimate the lighting distribution of the image, which helps guide the brightness and contrast adjustment of the image. Noise suppression: Deep networks can effectively identify and suppress noise in images to maintain the purity of the image. Detail enhancement: The algorithm usually emphasizes the restoration of image details, including edge sharpening and texture enhancement, to improve the clarity of the image. Adaptive adjustment: Advanced algorithms can adaptively adjust the enhancement parameters according to the image content and lighting conditions to obtain the best enhancement effect. Summary of the invention

[0006] The purpose of this invention is to disclose a low-light image enhancement algorithm based on deep convolution, which aims to imitate the adaptability of the human visual system under low-light conditions, improve the visibility and clarity of images, make images appear brighter and with higher contrast, while retaining or enhancing details. By improving the visual effect and usability of images, the development of computer vision and image processing technology is promoted, while expanding its application in various professional fields.

[0007] In an embodiment of the present disclosure, the deep convolution-based low-light image enhancement algorithm includes the following steps:

[0008] Step 1: Use the channel denoiser module to accurately identify and remove random noise and unnecessary features in the image, thereby obtaining a clearer and better quality denoised low-light image;

[0009] Step 2: The denoised low-light image is passed through the light map decomposition network. Through deep convolution operations, the brightness distribution of the image is finely analyzed, the light components of the image are separated, and an accurate light map is generated.

[0010] Step 3: The denoised low-light image is decomposed through a reflectance map and a feature extraction network to accurately extract the reflectance characteristics of the image and generate a detailed reflectance map to further enrich the information level of the image;

[0011] Step 4: Perform function operations on the illumination map and reflection map obtained in step 2 and step 3, and adopt a pixel multiplication strategy to organically integrate the two, and finally generate an enhanced normal illumination image to obtain an enhanced normal illumination image.

[0012] In a further technical solution, step 1 specifically comprises the following steps:

[0013] Step 1.1: In order for the subsequent downsampling and upsampling operations to proceed smoothly, ensure that the width and height of the input image are even numbers to avoid size mismatch problems. The calculation formula is as follows:

[0014] Input=pad(I,p)

[0015] Where I is the original input image, the pad function is used to add padding to the edge of the image, and p is the size of the patch to ensure that the image size is an even number.

[0016] Step 1.2: Use convolutional layers to extract preliminary features of the input image.

[0017] Step 1.3: Use downsampling and multi-scale feature extraction. The calculation formula for downsampling is as follows:

[0018] F i+1 =σ(W i*pool(F i )+b i ), i = 1, 2, 3

[0019] First, the feature map F of the previous level i Apply the pooling operation (pool), and then use the new convolution kernel W i and bias b i Perform convolution operation and obtain the next level feature map F through activation function σ i+1 .

[0020] Step 1.4: Use the multi-head self-attention mechanism to optimize the feature representation of the image. The calculation formula is as follows:

[0021] F′=MSA(Q,K,V)

[0022] Where Q, K, and V are query, key, and value matrices, respectively, which are derived from the input feature map F n Obtained through different linear transformations. The Multi-Head Self-Attention (MSA) module can capture the dependencies between different positions, thereby optimizing the feature representation F′.

[0023] Step 1.5: Use upsampling to fuse the feature map with the downsampled feature map. The calculation formula is as follows:

[0024] F′ i =σ(W′ i *upsample(F′ i+1 )+F i +b′ i ), i = n-1, n-2, n-3 First, the higher-level feature map F′ i+1 , perform upsampling operation, and then combine it with the low-level feature map F′ i Fusion, then, use the new convolution kernel W' i and bias b′ i Perform convolution operation and obtain the fused feature map F through the activation function σ i ′.

[0025] Step 1.6: The final output of the upsampling is residually connected to the input, and convolution is adjusted to obtain the adjusted feature map, and the Tanh activation function is used to limit the output range to generate the denoised image output. In the last stage, the preliminary feature map F1 is combined with the feature map F′1 obtained by upsampling and fusion through a residual connection, and then convolution is performed through the convolution kernel W″ and the bias b″, and finally the final output, that is, the enhanced image O, is obtained through the activation function σ.

[0026] As a further technical solution, a serialized layer structure is constructed in step 2 to realize deep learning and conversion of image features. First, the input data is filled with boundary reflections through the ReflectionPad2d layer to effectively retain edge information and avoid information loss in subsequent convolution operations. Next, the model uses the Conv2d layer for feature extraction, converts the input three-channel image into a feature map with a specified number (the default is 64), and uses the ReLU activation function to introduce nonlinear transformations to enhance the learning ability of the model. Subsequently, the model repeats the process of filling, convolution, and activation for a total of four times, keeping the same number of feature maps each time. This deep stacking design helps the model to deepen its understanding of the image step by step and capture more complex features. In the final convolution layer, the number of output channels is reduced to 1, condensing the previous multi-channel feature map into a single, dense feature response. The final output is further processed by the sigmoid function to limit the numerical range to between 0 and 1 to obtain an illumination map.

[0027] In a further technical solution, step 3 specifically comprises the following steps:

[0028] Step 3.1: Similar to the serialization layer structure constructed in step 2, the input data is filled with boundary reflections through the ReflectionPad2d layer; the Conv2d layer is used for feature extraction, and the input three-channel image is converted into a feature map with a specified number (the default is 64). At the same time, the ReLU activation function is used to introduce nonlinear transformations to enhance the learning ability of the model. Subsequently, the model repeats the process of padding, convolution, and activation three times, each time keeping the same number of feature maps. Given the input image I, the forward propagation of the serialization layer structure can be expressed as:

[0029] O=σ(N(I))

[0030] Among them, N represents the layer sequence operation in the network; σ is the Sigmoid activation function; O is the final output.

[0031] Step 3.2: Construct the DCFMBlock module, introduce deep convolutional layers and FocalModulation attention mechanism, and enhance the local and global feature capture capabilities of the model. The operation of DCFMBlock can be expressed as:

[0032] X out =X+(X1⊙X2)

[0033] Among them, X1=D(x), D is the depth convolution operation; X1=FM(X), FM is the FocalModulation attention mechanism; ⊙ represents element-by-element multiplication; X out is the output of DCFMBlock.

[0034] Step 3.3: Define the FocalModulation class, which is responsible for implementing focal modulation, that is, extracting multi-scale features through convolution kernels of different sizes, and using global average pooling and attention mechanisms to adjust the weights of the features. For the output feature X, the calculation process of FocalModulation is as follows:

[0035] Step 3.3.1: Pre-linear projection:

[0036] X′=F(X)

[0037] Step 3.3.2: Context Aggregation:

[0038] C i =C i (X′)

[0039] Among them, C i is the output of the i-th convolutional layer; C i It is a convolution operation with kernels of different sizes.

[0040] Step 3.3.3: Global gating:

[0041] G = avgpool(avgpool(X′))

[0042] Step 3.3.4: Context Fusion:

[0043]

[0044] Among them, g i is the gating coefficient of the i-th convolutional layer; L is the number of convolutional layers; G is the global gating signal; g L is the gating coefficient of the global gating.

[0045] Step 3.3.5: Focus modulation:

[0046] M=H(C all )

[0047] Step 3.3.6: Post-linear projection:

[0048] X″=P(X′⊙M)

[0049] Among them, H is the convolution operation and P is the linear projection operation.

[0050] The final output is:

[0051] X out =X″

[0052] This step uses deep convolution and focal modulation mechanisms to enable the model to efficiently extract and integrate multi-scale features from the input image, thereby improving its performance in extracting image features in image processing tasks.

[0053] In a further technical solution, step 4 uses the luminance map and reflectance map obtained in step 2 and step 3 to reconstruct the final enhanced image. The mathematical formula can be expressed as:

[0054] I=L γ ×R

[0055] Among them, I represents the final enhanced image; L is the light intensity map, which reflects the illumination level of each pixel in the image; R is the reflectance map, which represents the inherent reflectivity of the surface of the object in the image; γ is the correction factor, which is used to adjust the light intensity and influence to adapt to different lighting environments and enhancement requirements.

[0056] In low light conditions, the original image has low brightness and unclear details. By decomposing the image into an illumination map and a reflectance map, the two parts can be processed and optimized independently. The illumination map L can be enhanced to increase the overall brightness, while the reflectance map R maintains the image's inherent color and texture information. By multiplying the enhanced illumination map with the reflectance map, an image can be generated that is bright while retaining the original color and details.

[0057] The present invention provides a low-light image enhancement algorithm based on deep convolution, which has the following beneficial effects:

[0058] 1) The present invention has precise control over the decomposition of illumination and reflection. Based on the Retinex theory, the framework uses two independent sub-networks - the illumination component estimation network and the reflection component enhancement network, which are responsible for estimating the illumination intensity distribution of the image and restoring the surface reflection characteristics of the image respectively. This separate processing method enables the model to more accurately adjust the brightness of the image while maintaining or enhancing the details and colors of the image, avoiding overexposure or color distortion that may be caused by a single enhancement method. Through independent optimization of illumination and reflection, the model can intelligently enhance the brightness and contrast of the image, especially in low-light conditions, so that the dark details can be clearly displayed, while maintaining the naturalness and layering of the image, improving the overall visual effect and viewing experience.

[0059] 2) The present invention uses multi-scale feature fusion and self-attention mechanism. The DCFMBlock module combines deep convolution and self-attention mechanism to capture local details and global structure of the image at the same time. Deep convolution is responsible for extracting spatial features of different scales, while the self-attention mechanism allows the model to redistribute attention resources globally and focus on key areas of the image, thereby achieving efficient fusion and optimization of features. This combination of multi-scale feature fusion and self-attention mechanism significantly enhances the model's ability to restore image details, whether it is the fine expression of texture or the clear definition of edges. In addition, the self-attention mechanism also improves the generalization ability of the model, enabling it to perform well in dealing with various complex scenes.

[0060] 3) The present invention adopts a channel-level denoising and structure-preserving method. The CWD (Channel-wise Denoiser) module achieves noise suppression at the image channel level through a series of downsampling, multi-head self-attention and upsampling operations. This design can not only effectively remove random noise in the image, but also avoid the loss of details caused by over-smoothing by retaining important structural information. Under low-light conditions, images are often accompanied by high noise levels. The addition of the CWD module ensures that no additional artifacts are introduced during the image enhancement process, while maintaining the original structure and texture of the image, which is crucial to improving the clarity and look and feel of the image.

[0061] 4) The present invention has efficient real-time processing capabilities. Despite the use of a multi-level deep learning structure, the framework achieves real-time or near real-time processing capabilities for complex models through a carefully designed network architecture and efficient hardware utilization. For example, by adopting lightweight convolution modules, efficient implementation of the self-attention mechanism, and parallel computing strategies, the model is able to maintain low computing latency while ensuring accuracy. This efficient processing capability enables the model to be widely used in scenarios such as real-time video stream processing, online image editing, and fast image analysis, such as real-time image enhancement in security monitoring systems, environmental perception of autonomous driving vehicles, and real-time analysis of medical images, greatly improving the system's response speed and user experience.

[0062] 5) The present invention can comprehensively improve image quality and wide applicability. The framework achieves comprehensive optimization of image brightness, contrast, color and details by combining advanced technologies such as illumination and reflection decomposition, multi-scale feature fusion, self-attention mechanism and channel-level denoising. This comprehensive improvement is not limited to specific image types or lighting conditions, but is applicable to a wide range of image processing scenarios. From everyday photographic image enhancement to professional image analysis fields such as security monitoring, autonomous driving, medical imaging and astronomical photography, the framework has demonstrated excellent performance and wide applicability. It can significantly improve the visual quality of images, improve the accuracy and efficiency of image analysis, and provide strong technical support for image processing in various fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A flowchart of a low-light image enhancement algorithm based on deep convolution according to the present invention;

[0064] Figure 2 This is a schematic diagram of the network structure of a deep convolution-based low-light image enhancement algorithm involved in the present invention. DETAILED DESCRIPTION

[0065] The present invention proposes a low-light image enhancement algorithm, which uses a deep convolutional neural network to achieve automatic enhancement of low-light images. In order to more clearly explain the purpose, technical solution and advantages of the present invention, the technical solution of the present invention will be described in detail with reference to the accompanying drawings. It should be pointed out that the specific embodiments described below are only used to explain the principles of the present invention and are not used to limit the scope of implementation of the present invention.

[0066] The present invention first provides a low-light image enhancement algorithm based on deep convolution. The specific flow chart is shown in Figure 1 , including the following steps:

[0067] Step 1: Use the channel denoiser module to accurately identify and remove random noise and unnecessary features in the image, thereby obtaining a clearer and better quality denoised low-light image;

[0068] Step 2: The denoised low-light image is passed through the light map decomposition network. Through deep convolution operations, the brightness distribution of the image is finely analyzed, the light components of the image are separated, and an accurate light map is generated.

[0069] Step 3: The denoised low-light image is decomposed through a reflectance map and a feature extraction network to accurately extract the reflectance characteristics of the image and generate a detailed reflectance map to further enrich the information level of the image;

[0070] Step 4: Perform function operations on the illumination map and reflection map obtained in step 2 and step 3, and adopt a pixel multiplication strategy to organically integrate the two, and finally generate an enhanced normal illumination image to obtain an enhanced normal illumination image.

[0071] As a preferred embodiment of the present invention, step 1 specifically includes the following steps:

[0072] Step 1.1: In order for the subsequent downsampling and upsampling operations to proceed smoothly, ensure that the width and height of the input image are even numbers to avoid size mismatch problems. The calculation formula is as follows:

[0073] Input=pad(I,p)

[0074] Where I is the original input image, the pad function is used to add padding to the edge of the image, and p is the size of the patch to ensure that the image size is an even number.

[0075] Step 1.2: Use convolutional layers to extract preliminary features of the input image.

[0076] Step 1.3: Use downsampling and multi-scale feature extraction. The calculation formula for downsampling is as follows:

[0077] F i+1 =σ(W i *pool(F i )+b i ), i = 1, 2, 3

[0078] First, the feature map F of the previous level i Apply the pooling operation (pool), and then use the new convolution kernel W i and bias b i Perform convolution operation and obtain the next level feature map F through activation function σ i+1 .

[0079] Step 1.4: Use the multi-head self-attention mechanism to optimize the feature representation of the image. The calculation formula is as follows:

[0080] F′=MSA(Q,K,V)

[0081] Where Q, K, and V are query, key, and value matrices, respectively, which are derived from the input feature map F n Obtained through different linear transformations. The Multi-Head Self-Attention (MSA) module can capture the dependencies between different positions, thereby optimizing the feature representation F′.

[0082] Step 1.5: Use upsampling to fuse the feature map with the downsampled feature map. The calculation formula is as follows:

[0083] F′ i =σ(W′ i *upsample(F′ i+1 )+F i +b′ i ), i = n-1, n-2, n-3

[0084] First, the higher-level feature map F′ i+1 , perform upsampling, and then combine it with the low-level feature map F i ' is fused, and then a new convolution kernel W' is used i and bias b′ i Perform convolution operation and obtain the fused feature map F′ through the activation function σ i .

[0085] Step 1.6: The final output of the upsampling is residually connected to the input, and convolution is adjusted to obtain the adjusted feature map, and the Tanh activation function is used to limit the output range to generate the denoised image output. In the last stage, the preliminary feature map F1 is combined with the feature map F′1 obtained by upsampling and fusion through a residual connection, and then convolution is performed through the convolution kernel W″ and the bias b″, and finally the final output, that is, the enhanced image O, is obtained through the activation function σ.

[0086] As a preferred embodiment of the present invention, a serialized layer structure is constructed in step 2 to realize deep learning and conversion of image features. First, the input data is filled with boundary reflections through the ReflectionPad2d layer to effectively retain edge information and avoid information loss in subsequent convolution operations. Next, the model uses the Conv2d layer for feature extraction, converts the input three-channel image into a feature map with a specified number (the default is 64), and introduces nonlinear transformation using the ReLU activation function to enhance the learning ability of the model. Subsequently, the model repeats the process of filling, convolution and activation for a total of four times, each time keeping the same number of feature maps. This deep stacking design helps the model to deepen its understanding of the image step by step and capture more complex features. In the final convolution layer, the number of output channels is reduced to 1, and the previous multi-channel feature map is condensed into a single, dense feature response. The final output is further processed by the sigmoid function, and the numerical range is limited to between 0 and 1 to obtain an illumination map.

[0087] As a preferred embodiment of the present invention, step 3 specifically includes the following steps:

[0088] Step 3.1: Similar to the serialization layer structure constructed in step 2, the input data is filled with boundary reflections through the ReflectionPad2d layer; the Conv2d layer is used for feature extraction, and the input three-channel image is converted into a feature map with a specified number (the default is 64). At the same time, the ReLU activation function is used to introduce nonlinear transformations to enhance the learning ability of the model. Subsequently, the model repeats the process of padding, convolution, and activation three times, each time keeping the same number of feature maps. Given the input image I, the forward propagation of the serialization layer structure can be expressed as:

[0089] O=σ(N(I))

[0090] Among them, N represents the layer sequence operation in the network; σ is the Sigmoid activation function; O is the final output.

[0091] Step 3.2: Construct the DCFMBlock module, introduce deep convolutional layers and FocalModulation attention mechanism, and enhance the local and global feature capture capabilities of the model. The operation of DCFMBlock can be expressed as:

[0092] X out =X+(X1⊙X2)

[0093] Among them, X1=D(x), D is the depth convolution operation; X1=FM(X), FM is the FocalModulation attention mechanism; ⊙ represents element-by-element multiplication; X out is the output of DCFMBlock.

[0094] Step 3.3: Define the FocalModulation class, which is responsible for implementing focal modulation, that is, extracting multi-scale features through convolution kernels of different sizes, and using global average pooling and attention mechanisms to adjust the weights of the features. For the output feature X, the calculation process of FocalModulation is as follows:

[0095] Step 3.3.1: Pre-linear projection:

[0096] X′=F(X)

[0097] Step 3.3.2: Context Aggregation:

[0098] C i =C i (X′)

[0099] Among them, C i is the output of the i-th convolutional layer; C i It is a convolution operation with kernels of different sizes.

[0100] Step 3.3.3: Global gating:

[0101] G = avgpool(avgpool(X′))

[0102] Step 3.3.4: Context Fusion:

[0103]

[0104] Among them, g i is the gating coefficient of the i-th convolutional layer; L is the number of convolutional layers; G is the global gating signal; g L is the gating coefficient of the global gating.

[0105] Step 3.3.5: Focus modulation:

[0106] M=H(C all )

[0107] Step 3.3.6: Post-linear projection:

[0108] X″=P(X′⊙M)

[0109] Among them, H is the convolution operation and P is the linear projection operation.

[0110] The final output is:

[0111] X out =X″

[0112] This step uses deep convolution and focal modulation mechanisms to enable the model to efficiently extract and integrate multi-scale features from the input image, thereby improving its performance in extracting image features in image processing tasks.

[0113] In a further technical solution, step 4 uses the luminance map and reflectance map obtained in step 2 and step 3 to reconstruct the final enhanced image. The mathematical formula can be expressed as:

[0114] I=L γ ×R

[0115] Among them, I represents the final enhanced image; L is the light intensity map, which reflects the illumination level of each pixel in the image; R is the reflectance map, which represents the inherent reflectivity of the surface of the object in the image; γ is the correction factor, which is used to adjust the light intensity and influence to adapt to different lighting environments and enhancement requirements.

[0116] In low light conditions, the original image has low brightness and unclear details. By decomposing the image into an illumination map and a reflectance map, the two parts can be processed and optimized independently. The illumination map L can be enhanced to increase the overall brightness, while the reflectance map R maintains the image's inherent color and texture information. By multiplying the enhanced illumination map with the reflectance map, an image can be generated that is bright while retaining the original color and details.

[0117] In the above description, various combinations of technical features are mentioned, but not all possible combinations are listed in detail. However, it should be emphasized that as long as the combination of these technical features is reasonable in implementation and does not conflict with each other, it should be considered to be within the scope of this specification.

Claims

1. A low-light image enhancement algorithm based on deep convolution, characterized in that: The method comprises the following steps: Step 1: Use the channel denoiser module to accurately identify and remove random noise and unnecessary features in the image, thereby obtaining a clearer and better quality denoised low-light image; Step 2: The denoised low-light image is passed through the light map decomposition network. Through deep convolution operations, the brightness distribution of the image is finely analyzed, the light components of the image are separated, and an accurate light map is generated. Step 3: The denoised low-light image is decomposed through a reflectance map and a feature extraction network to accurately extract the reflectance characteristics of the image and generate a detailed reflectance map to further enrich the information level of the image; Step 4: Perform function operations on the illumination map and reflection map obtained in step 2 and step 3, and adopt a pixel multiplication strategy to organically integrate the two, and finally generate an enhanced normal illumination image to obtain an enhanced normal illumination image.

2. The deep convolution low-light image enhancement algorithm according to claim 1, characterized in that: The step 1 specifically comprises the following steps: Step 1.1: In order for the subsequent downsampling and upsampling operations to proceed smoothly, ensure that the width and height of the input image are even numbers to avoid size mismatch problems. The calculation formula is as follows: Input=pad(I,p) Where I is the original input image, the pad function is used to add padding to the edge of the image, and p is the size of the patch to ensure that the image size is an even number. Step 1.2: Use convolutional layers to extract preliminary features of the input image. Step 1.3: Use downsampling and multi-scale feature extraction. The calculation formula for downsampling is as follows: F i+1 =σ(W i *pool(F i )+b i ),i=1,2,3 First, the feature map F of the previous level i Apply the pooling operation (pool), and then use the new convolution kernel W i and bias b i Perform convolution operation and obtain the next level feature map F through activation function σ i+1 . Step 1.4: Use the multi-head self-attention mechanism to optimize the feature representation of the image. The calculation formula is as follows: F′=MSA(Q,K,V) Where Q, K, and V are query, key, and value matrices, respectively, which are derived from the input feature map F n Obtained through different linear transformations. The Multi-Head Self-Attention (MSA) module can capture the dependencies between different positions, thereby optimizing the feature representation F′. Step 1.5: Use upsampling to fuse the feature map with the downsampled feature map. The calculation formula is as follows: F′ i =σ(W′ i *upsample(F′ i+1 )+F i +b′ i ),i=n-1,n-2,n-3 First, the higher-level feature map F′ i+1 , perform upsampling operation, and then combine it with the low-level feature map F′ i Fusion, then, use the new convolution kernel W' i and bias b′ i Perform convolution operation and obtain the fused feature map F′ through the activation function σ i . Step 1.6: The final output of the upsampling is residually connected to the input, and convolution is adjusted to obtain the adjusted feature map, and the Tanh activation function is used to limit the output range to generate the denoised image output. In the last stage, the preliminary feature map F1 is combined with the feature map F′1 obtained by upsampling and fusion through a residual connection, and then convolution is performed through the convolution kernel W″ and the bias b″, and finally the final output, that is, the enhanced image O, is obtained through the activation function σ.

3. The deep convolution low-light image enhancement algorithm according to claim 1, characterized in that: The step 2 constructs a serialized layer structure for deep learning and conversion of image features. First, the input data is filled with boundary reflections through the ReflectionPad2d layer to effectively retain edge information and avoid information loss in subsequent convolution operations. Next, the model uses the Conv2d layer for feature extraction, converts the input three-channel image into a feature map with a specified number (the default is 64), and introduces nonlinear transformation using the ReLU activation function to enhance the learning ability of the model. Subsequently, the model repeats the process of filling, convolution, and activation for a total of four times, each time keeping the same number of feature maps. This deep stacking design helps the model to deepen its understanding of the image step by step and capture more complex features. In the final convolution layer, the number of output channels is reduced to 1, condensing the previous multi-channel feature map into a single, dense feature response. The final output is further processed by the sigmoid function to limit the numerical range to between 0 and 1 to obtain an illumination map.

4. The deep convolution low-light image enhancement algorithm according to claim 1, characterized in that: The step 3 specifically comprises the following steps: Step 3.1: Similar to the serialization layer structure constructed in step 2, the input data is filled with boundary reflections through the ReflectionPad2d layer; the Conv2d layer is used for feature extraction, and the input three-channel image is converted into a feature map with a specified number (the default is 64). At the same time, the ReLU activation function is used to introduce nonlinear transformations to enhance the learning ability of the model. Subsequently, the model repeats the process of padding, convolution, and activation three times, each time keeping the same number of feature maps. Given the input image I, the forward propagation of the serialization layer structure can be expressed as: O=σ(N(I)) Among them, N represents the layer sequence operation in the network; σ is the Sigmoid activation function; O is the final output. Step 3.2: Build the DCFMBlock module, introduce deep convolutional layers and FocalModulation attention mechanism, and enhance the local and global feature capture capabilities of the model. The operation of DCFMBIock can be expressed as: X out =X+(X1⊙X2) Among them, X1=D(x), D is the depth convolution operation; X1=FM(X), FM is the FocalModulation attention mechanism; ⊙ represents element-by-element multiplication; X out is the output of DCFMBlock. Step 3.3: Define the FocalModulation class, which is responsible for implementing focal modulation, that is, extracting multi-scale features through convolution kernels of different sizes, and using global average pooling and attention mechanisms to adjust the weights of the features. For the output feature X, the calculation process of FocalModulation is as follows: Step 3.3.1: Pre-linear projection: X′=F(X) Step 3.3.2: Context Aggregation: C i =C i (X′) Among them, C i is the output of the i-th convolutional layer; C i It is a convolution operation with kernels of different sizes. Step 3.3.3: Global gating: G = avgpool(avgpool(X′)) Step 3.3.4: Context Fusion: Among them, g i is the gating coefficient of the i-th convolutional layer; L is the number of convolutional layers; G is the global gating signal; g L is the gating coefficient of the global gating. Step 3.3.5: Focus modulation: M=H(C all ) Step 3.3.6: Post-linear projection: X″=P(X′⊙M) Among them, H is the convolution operation and P is the linear projection operation. The final output is: X out =X″ This step uses deep convolution and focal modulation mechanisms to enable the model to efficiently extract and integrate multi-scale features from the input image, thereby improving its performance in extracting image features in image processing tasks.

5. Reconstruct the luminance map and reflectance map obtained in steps 2 and 3 to obtain the final enhanced image. The mathematical formula can be expressed as: I=L γ ×R in, I represents the final enhanced image; L is the light intensity map, which reflects the light intensity of each pixel in the image; R is a reflectance map, which indicates the intrinsic reflectivity of the surface of the object in the image; γ is a correction factor, which is used to adjust the light intensity and impact to adapt to different lighting environments and enhancement requirements. In low light conditions, the original image has low brightness and unclear details. By decomposing the image into an illumination map and a reflectance map, the two parts can be processed and optimized independently. The illumination map L can be enhanced to increase the overall brightness, while the reflectance map R maintains the image's inherent color and texture information. By multiplying the enhanced illumination map with the reflectance map, an image can be generated that is bright while retaining the original color and details.

Images