An adaptive illumination correction method and system based on multi-scale feature coupling

Abstract

The application provides a kind of adaptive illumination correction method and system based on multiscale feature coupling.The method comprises: extracting the luminance component and chrominance component of the low-illumination image to be corrected;Using a preset region-aware sparse sampling strategy to generate the importance score of each image block to determine the key area in the image, so as to perform synchronous sampling at the corresponding positions of the luminance component and the chrominance component, generate the luminance feature and the chrominance feature of the key area;Using a cross-channel feature fusion mechanism to fuse the luminance feature and the chrominance feature, generate an enhanced luminance feature containing chrominance prior, and input it into a preset neural network together with the coordinate set of the key area to predict a global illumination gain map;Using the global illumination gain map to enhance the original luminance component to obtain an enhanced luminance component;Merging the enhanced luminance component with the original chrominance component and converting back to RGB space to obtain the final enhanced image.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and digital image processing technology, and in particular to an adaptive illumination correction method and system based on multi-scale feature coupling. Background Technology

[0002] In actual image acquisition, due to insufficient ambient light or uneven lighting distribution (such as backlighting or shadows), the acquired images often suffer from low brightness, poor contrast, and noise interference. This not only reduces the visual quality of the image but also affects the accuracy of subsequent computer vision tasks (such as object detection and object tracking).

[0003] To improve the quality of low-light images, researchers have proposed various image enhancement methods. Early traditional methods mainly include histogram equalization (HE) and its variants, as well as methods based on Retinex theory. These methods mainly rely on statistical models or artificially designed physical priors. Although computationally simple, they are often prone to noise amplification, color deviation, or halo artifacts when dealing with complex scenes.

[0004] In recent years, with the rise of deep learning technology, data-driven methods based on convolutional neural networks (CNNs) and generative adversarial networks (GANs) have made significant progress in the field of image enhancement. These methods improve the enhancement effect to a certain extent by learning the mapping relationship between low-light and normal-light images on large-scale datasets.

[0005] While existing deep learning methods have shown great potential in image enhancement, they still face significant challenges when dealing with high-resolution images and complex lighting environments. The main technical bottlenecks are as follows:

[0006] 1. Computational Redundancy and Inefficiency in High-Resolution Image Processing: With the development of imaging technology, 4K and even 8K ultra-high-definition images are becoming increasingly common. Existing methods based on convolutional neural networks (CNNs) typically perform convolution operations on each pixel in an image with equal weight. However, low-light images often contain a large number of flat areas with low information content or invalid backgrounds. This pixel-by-pixel processing method not only leads to huge computational overhead and memory usage, resulting in a serious waste of computing power, but also makes it difficult to meet the needs of real-time processing or edge deployment.

[0007] 2. Distortion issues caused by decoupling brightness and color features: Traditional methods and some deep learning solutions often treat brightness adjustment and color restoration as two independent tasks when processing low-light images. This processing approach, lacking multi-scale feature coupling and cross-channel interaction, ignores the complex impact of non-linear brightness changes on color saturation and hue. Therefore, while significantly increasing image brightness, it easily disrupts the color balance of the original image, leading to color casts, loss of saturation, or a "washed-out" appearance in the enhanced image, failing to maintain its naturalness.

[0008] 3. Discrete modeling struggles to fit continuous illumination distributions: Illumination changes in the physical world are typically continuous and smooth. However, existing neural network methods based on fixed convolutional kernels essentially operate on a discrete pixel grid, making it difficult to accurately simulate the continuous spatial distribution of illumination. This discrete modeling approach is prone to producing block artifacts, halo artifacts, or unnatural illumination transitions when dealing with boundary regions where illumination changes drastically. Summary of the Invention

[0009] To address the problems of inaccurate illumination estimation and color distortion in existing methods, this invention proposes an adaptive illumination correction method and system based on multi-scale feature coupling.

[0010] In a first aspect, the present invention provides an adaptive illumination correction method based on multi-scale feature coupling, comprising:

[0011] The low-light image to be corrected is converted to HSV space to separate the luminance and chrominance components;

[0012] The brightness components are downsampled using an image pyramid construction method to generate a low-resolution brightness feature map;

[0013] A preset region-aware sparse sampling strategy is used to divide the low-resolution brightness feature map into multiple image blocks and generate an importance score for each image block.

[0014] The top K image patches with the highest importance scores are selected as key regions, and the coordinate set of the key regions in the low-light image to be corrected is obtained so that synchronous sampling is performed at the corresponding positions of the luminance component and the chrominance component to generate the luminance features and chrominance features of the key regions.

[0015] A cross-channel feature fusion mechanism is used to fuse the luminance features and chrominance features to generate enhanced luminance features that include chrominance priors;

[0016] The enhanced brightness features, which include chromaticity priors, and the coordinate set of the key region are input into a preset neural network to predict the global illumination gain map; the neural network has a pre-established continuous mapping relationship between the enhanced brightness features and spatial coordinates to the illumination gain coefficient;

[0017] The original luminance component is enhanced using the global illumination gain map to obtain the enhanced luminance component; the enhanced luminance component is then merged with the original chrominance component and converted back to RGB space to obtain the final enhanced image.

[0018] Furthermore, the importance score of each image block in the low-resolution brightness feature map is calculated using the following formula;

[0019]

[0020] in, Low-resolution brightness feature map Indicates the edge extraction kernel. This represents the convolution operation. This represents the local spatial attention extraction function. This represents the global context extraction function. These are learnable weight coefficients. This is a normalization operation.

[0021] Furthermore, the cross-channel feature fusion mechanism includes:

[0022] A channel attention mechanism is introduced to reweight the luminance and chrominance features respectively according to the following formula, generating their respective channel weights:

[0023]

[0024]

[0025] in, Brightness characteristics For chromaticity characteristics, It is a multilayer perceptron. Use the Sigmoid activation function; and For each corresponding channel weight;

[0026] Based on their respective channel weights, the bidirectional feature interaction paths constructed according to the following formula generate their respective modulation weights for feature coupling:

[0027]

[0028]

[0029]

[0030] in, and Indicates a gating network. This indicates element-wise multiplication. This indicates a splicing operation. and For their respective modulation weights, Indicates the characteristics after coupling;

[0031] The coupled features are fused with the original luminance features according to the following formula to obtain enhanced luminance features that include chromaticity prior. :

[0032]

[0033] in, A learnable residual scaling factor. It is a nonlinear transformation layer.

[0034] Furthermore, the neural network layers use a sinusoidal periodic function as the activation function, and the output of the i-th network layer... The calculation formula is:

[0035]

[0036] in, The weights and biases of the learnable i-th network layer. This is the preset frequency control factor.

[0037] Secondly, the present invention provides an adaptive illumination correction system based on multi-scale feature coupling, comprising:

[0038] The image component extraction module is used to convert the low-light image to be corrected to the HSV space to separate the luminance and chrominance components.

[0039] The pyramid module is used to downsample the brightness components using an image pyramid construction method to generate a low-resolution brightness feature map;

[0040] The region awareness module is used to divide the low-resolution brightness feature map into multiple image blocks using a preset region awareness sparse sampling strategy and generate an importance score for each image block; select the top K image blocks with the highest importance scores as key regions, and obtain the coordinate set of the key regions in the low-light image to be corrected, so as to perform synchronous sampling at the corresponding positions of the brightness component and the chromaticity component to generate the brightness features and chromaticity features of the key regions.

[0041] The feature fusion module is used to fuse the luminance features and chrominance features using a cross-channel feature fusion mechanism to generate enhanced luminance features that include chrominance priors;

[0042] The illumination field modeling module is used to input the enhanced brightness features including chromaticity priors and the coordinate set of the key regions into a preset neural network to predict the global illumination gain map; the neural network has a pre-established continuous mapping relationship between spatial coordinates and illumination gain coefficients.

[0043] The image reconstruction module is used to enhance the original luminance component using the global illumination gain map to obtain the enhanced luminance component; the enhanced luminance component is then merged with the original chrominance component and converted back to RGB space to obtain the final enhanced image.

[0044] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in the first aspect.

[0045] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in the first aspect.

[0046] The beneficial effects of this invention are as follows:

[0047] (1) Region-aware sparse sampling strategy based on multidimensional saliency assessment: To address the computational redundancy caused by a large number of flat background regions in high-resolution images, this invention proposes an attention-driven sparse computation paradigm. Unlike traditional methods that perform indiscriminate convolution operations on all pixels in the image, this invention constructs a multidimensional saliency scoring model that integrates high-frequency edge features, local spatial attention, and global brightness context. This model can accurately identify key regions with high information density in the image and perform depth feature extraction and illumination estimation only on these sparse key regions. This improvement significantly reduces the algorithm's memory usage and inference latency while ensuring illumination correction accuracy, achieving efficient processing of ultra-high-definition images.

[0048] (2) Feature Coupling and Bidirectional Interaction Mechanism Based on Cross-Channel Collaboration: To address the color distortion and saturation loss issues caused by traditional methods, this invention designs a cross-channel feature coupling mechanism. This mechanism utilizes the structural stability of chroma components as prior knowledge to construct a bidirectional interaction path between luminance and chroma features. Through channel attention reweighting and bidirectional gating modulation, the model can dynamically suppress artifacts during luminance enhancement using chroma features, while simultaneously enhancing chroma expression using luminance features. This effectively ensures that the enhanced image maintains the color balance and naturalness of the original scene while significantly increasing brightness.

[0049] (3) Implicit modeling method for continuous illumination field based on periodic activation function: This method abandons the traditional discrete convolutional network method for generating illumination maps and adopts an implicit neural network based on coordinate mapping to fit the illumination distribution. In particular, a sinusoidal periodic function is introduced as the activation function. By utilizing its excellent high-frequency fitting ability, it achieves accurate simulation of continuous illumination changes and shadow boundaries in the physical world, and completely eliminates the block effect and halo artifacts caused by block processing.

[0050] In summary, this invention achieves end-to-end correction of low-light images by constructing a processing flow that includes sparse sampling, cross-channel coupling, and implicit illumination modeling, significantly improving the quality and efficiency of low-light image enhancement. Attached Figure Description

[0051] Figure 1 A flowchart illustrating an adaptive illumination correction method based on multi-scale feature coupling, provided for an embodiment of the present invention;

[0052] Figure 2 A schematic diagram of the region-aware sparse sampling strategy provided in an embodiment of the present invention;

[0053] Figure 3 Comparison of the effects of the method of the present invention and the prior art in image detail restoration provided in the embodiments of the present invention: (a) full image of the method of the present invention, (b) original input image, (c) restoration result of LIME, (d) restoration result of Zero-DCE, (e) restoration result of SCI, (f) restoration result of RUAS, (g) restoration result of the method of the present invention;

[0054] Figure 4 The overall visual effect of the method of the present invention provided in the embodiments of the present invention and various mainstream low-light enhancement algorithms in real complex scenes is compared: (a) the original low-light image, (b) the enhancement result of LIME, (c) the enhancement result of Zero-DCE, (d) the enhancement result of SCI, (e) the enhancement result of RUAS, and (f) the enhancement result of the method of the present invention.

[0055] Figure 5A schematic diagram of an adaptive illumination correction system based on multi-scale feature coupling provided in an embodiment of the present invention;

[0056] Figure 6 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0058] This invention effectively solves the problems of inaccurate illumination estimation and color distortion in traditional methods, and can be applied to scenarios such as security monitoring, autonomous driving visual perception, and consumer electronics imaging.

[0059] like Figure 1 As shown, this invention provides an adaptive illumination correction method based on multi-scale feature coupling, comprising the following steps:

[0060] S101: Convert the low-light image to be corrected to HSV space to separate the luminance and chrominance components;

[0061] S102: The brightness components are downsampled using an image pyramid construction method to generate a low-resolution brightness feature map;

[0062] Specifically, in order to capture contextual information at different scales and reduce the computational complexity of subsequent processing, this embodiment of the invention employs existing image pyramid construction techniques (such as Gaussian pyramid sampling or bilinear interpolation downsampling) to downsample the original luminance components. Specifically, the original high-resolution luminance components are used as the bottom layer of the pyramid, and a low-resolution luminance feature map is generated through a downsampling algorithm. This low-resolution luminance feature map preserves the global illumination distribution structure of the image and will serve as input data for subsequent region-aware sparse sampling modules and global illumination analysis.

[0063] S103: Using a preset region-aware sparse sampling strategy, the low-resolution brightness feature map is divided into multiple image blocks and an importance score is generated for each image block;

[0064] Specifically, to reduce computational redundancy in high-resolution image processing, this invention employs a region-aware sparse sampling strategy. This strategy does not directly process the entire image; instead, it filters key feature regions by calculating the "information density" of each region. A detailed flowchart is shown below. Figure 2 As shown.

[0065] First, a multidimensional saliency scoring model was constructed. This model does not make a single-dimensional judgment, but rather a weighted fusion of high-frequency edge features, local spatial attention, and global brightness context to calculate the comprehensive importance score of each image patch. The calculation formula is as follows:

[0066]

[0067] in, This represents the low-resolution brightness feature map obtained after downsampling processing. This indicates a preset edge extraction kernel (such as the Laplacian operator). This means that the edge extraction kernel is used to perform convolution operation on the low-resolution brightness feature map in order to extract high-frequency texture detail features of each local region; This represents the local spatial attention extraction function. This means that the brightness variance or contrast features of each local region in the low-resolution brightness feature map are extracted by local convolutional layers, which are used to characterize the local saliency of each local region; The global context extraction function is obtained by mapping the entire low-resolution brightness feature map after global average pooling, and is used to characterize the relative position of each local region in the overall brightness distribution. This is a normalization operation used to map feature values ​​of different dimensions to the [0,1] interval; These are learnable weights that are automatically adjusted during network training to balance the importance of texture, contrast, and global information in the selection process.

[0068] S104: Select the top K image patches with the highest importance scores as key regions, and obtain the coordinate set of the key regions in the low-light image to be corrected, so as to perform synchronous sampling at the corresponding positions of the luminance component and the chrominance component to generate the luminance features of the key regions. and chromaticity characteristics ;

[0069] Specifically, based on the score All image patches are sorted in descending order, and only the top-K highest-scoring key coordinate points and their corresponding local image patches are selected as input to the subsequent network. This step effectively eliminates flat backgrounds or invalid black border regions, significantly reducing computational cost.

[0070] S105: The luminance features and chrominance features are fused using a cross-channel feature fusion mechanism to generate enhanced luminance features that include chrominance priors;

[0071] Specifically, to address the color distortion problem caused by independent brightness enhancement, this embodiment of the invention introduces a cross-channel feature coupling mechanism. The cross-channel feature fusion mechanism includes:

[0072] A channel attention mechanism is introduced to reweight the luminance and chrominance features respectively according to the following formula, generating their respective channel weights:

[0073]

[0074]

[0075] in, Brightness characteristics For chromaticity characteristics, It is a multilayer perceptron used to learn the dependencies between channels; Use the Sigmoid activation function; and Each channel has its corresponding weight, ranging from (0,1). This step uses a channel attention mechanism to reweight the feature channels to adaptively enhance the response of effective features.

[0076] Based on their respective channel weights, the bidirectional feature interaction path constructed according to the following formula generates corresponding modulation weights for deep feature coupling:

[0077]

[0078]

[0079]

[0080] in, and Indicates a gating network. This indicates element-wise multiplication. This represents the concatenation operation of feature vectors. and For their respective modulation weights, Indicates the characteristics after coupling;

[0081] The coupled features are fused with the original luminance features via residual connections according to the following formula to obtain enhanced luminance features that include chromaticity priors. :

[0082]

[0083] in, A learnable residual scaling factor. It is a nonlinear transformation layer.

[0084] S106: Input the enhanced brightness features including chromaticity prior and the coordinate set of the key region into a preset neural network to predict the global illumination gain map. The neural network has a pre-established continuous mapping relationship between spatial coordinates and illumination gain coefficients.

[0085] Specifically, to generate continuous, smooth, and detailed lighting maps, embodiments of the present invention construct an implicit neural network based on coordinate mapping. This network establishes spatial coordinates. A continuous mapping relationship between the light gain coefficient and the illumination gain coefficient.

[0086] The network layers use sinusoidal periodic functions as activation functions to improve the model's ability to fit high-frequency variations in the lighting field. Layer output The calculation is as follows:

[0087]

[0088] in, For network weights and biases, This is the frequency control factor, a preset hyperparameter used to control the network's sensitivity to high-frequency spatial signals. A larger value indicates a higher sensitivity. The value enables the network to capture shadow boundaries that change drastically in the lighting field, while a smaller value... The value helps to fit a smooth gradient region.

[0089] S107: The original luminance component is enhanced using the global illumination gain map to obtain the enhanced luminance component; the enhanced luminance component is merged with the original chrominance component and converted back to RGB space to obtain the final enhanced image.

[0090] Specifically, based on the Retinex illumination reflection physical model, the original luminance component is divided pixel-by-pixel by the predicted global illumination gain map to remove the effects of uneven illumination, thereby obtaining the enhanced luminance component. Subsequently, the enhanced luminance component is merged with the original chrominance component separated in step S101, and an inverse conversion from HSV color space to RGB color space is performed to finally output a high-quality enhanced image.

[0091] The adaptive illumination correction method based on multi-scale feature coupling provided in this invention achieves end-to-end correction of low-light images by constructing a processing flow that includes sparse sampling, cross-channel coupling, and implicit illumination modeling, significantly improving the quality and efficiency of low-light image enhancement.

[0092] To verify the effectiveness of this invention, extensive experiments were conducted on multiple publicly available low-light image datasets, and comparisons were made with current mainstream image enhancement methods. The comparison results are as follows:

[0093] 1. This invention significantly improves the objective quality indicators of image enhancement.

[0094] Experimental results show that the present invention performs well in key objective indicators such as peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and perceptual image quality assessment index (LPIPS).

[0095] As shown in Table 1, the PSNR index of this invention achieved excellent quantitative results on both the high-resolution low-light test set A and the ordinary test set B in terms of color error-related indicators. This invention, through continuous illumination field modeling, results in enhanced images with better texture detail and edge sharpness.

[0096] Table 1. Quantitative comparison of the method of the present invention on datasets A and B.

[0097]

[0098] 2. Effectively suppresses noise amplification and halo effect in low-light enhancement.

[0099] Existing enhancement methods often amplify sensor noise in flat areas or produce halo artifacts at edges where brightness changes abruptly when improving the brightness of dark areas. This invention employs a region-aware sparse sampling strategy, focusing the model on key areas with rich texture and dramatic brightness variations, avoiding excessive attention to and amplification of flat background noise. Simultaneously, continuous illumination field modeling based on periodic activation functions ensures the smoothness and continuity of the predicted illumination gain in spatial distribution, effectively eliminating block artifacts and edge halos caused by traditional discrete processing. This allows the enhanced image to maintain image purity and naturalness while improving brightness.

[0100] Figure 3 This paper compares the effectiveness of the method of this invention with existing technologies in image detail restoration. The images show magnified regions with rich textures or drastic brightness variations. The comparison reveals that traditional techniques, while enhancing brightness, often result in blurred edges, loss of detail, or noticeable blocky artifacts. In contrast, this invention, thanks to continuous illumination field modeling, effectively brightens dark areas while maintaining the sharpness of object edges and the clarity of textures, without significant noise amplification or structural damage, thus verifying the advantages of this method in detail preservation.

[0101] 3. Improved subjective visual experience and eliminated common artifacts.

[0102] Subjective visual comparison experiments show that the images generated by this invention have natural lighting transitions and no obvious blockiness or halo artifacts. Figure 4 As shown, when dealing with scenes containing complex lighting changes (such as backlighting and shadow boundaries), the present invention can generate smooth and detailed lighting maps, resulting in a more transparent and natural visual effect that conforms to the perceptual characteristics of the human eye.

[0103] Figure 4 This paper presents a comparison of the overall visual effects of the method of this invention with several mainstream low-light enhancement algorithms in real-world complex scenes. The first column shows the original low-light input image, and the subsequent columns show the enhancement results of different algorithms. The comparison results show that some existing methods suffer from insufficient overall brightness enhancement or overexposure, and are prone to producing halo effects or color distortion at the boundary between light and dark areas. In contrast, the image generated by the method of this invention has a uniform overall brightness distribution, natural and smooth lighting transitions, high color fidelity, and a clear and natural visual experience, effectively solving the balance problem between brightness enhancement and color fidelity.

[0104] In summary, this invention achieves high-fidelity color reproduction and detail restoration while ensuring computational efficiency, effectively solving the pain points of existing technologies in complex lighting scenarios, and has significant practical value and broad application prospects.

[0105] Based on the same inventive concept, such as Figure 5 As shown, this embodiment of the invention provides an adaptive illumination correction system based on multi-scale feature coupling, including an image component extraction module, a pyramid module, a region perception module, a feature fusion module, an illumination field modeling module, and an image reconstruction module.

[0106] Specifically, the image component extraction module converts the low-light image to be corrected to the HSV space to separate the luminance and chrominance components; the pyramid module downsamples the luminance component using an image pyramid construction method to generate a low-resolution luminance feature map; the region awareness module divides the low-resolution luminance feature map into multiple image blocks using a preset region awareness sparse sampling strategy and generates an importance score for each image block; the top K image blocks with the highest importance scores are selected as key regions, and the coordinate set of the key regions in the low-light image to be corrected is obtained so as to perform [further processing] at the corresponding positions of the luminance and chrominance components. Synchronous sampling generates luminance and chromaticity features for the key region. A feature fusion module fuses the luminance and chromaticity features using a cross-channel feature fusion mechanism to generate enhanced luminance features with chromaticity priors. An illumination field modeling module inputs the enhanced luminance features with chromaticity priors and the coordinate set of the key region into a preset neural network to predict a global illumination gain map. An image reconstruction module enhances the original luminance component using the global illumination gain map to obtain an enhanced luminance component. The enhanced luminance component is then merged with the original chromaticity component and converted back to RGB space to obtain the final enhanced image.

[0107] It should be noted that the adaptive illumination correction system based on multi-scale feature coupling provided in this embodiment of the invention is for implementing the above method. Its specific functions can be referred to in the above method embodiments, and will not be repeated here.

[0108] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6As shown, the electronic device may include: a processor 601, a communication interface 602, a memory 603, and a communication bus 604. The processor 601, communication interface 602, and memory 603 communicate with each other via the communication bus 604. The processor 601 can call logical instructions in the memory 603 to execute an adaptive illumination correction method based on multi-scale feature coupling. This method includes: converting the low-illuminance image to be corrected to HSV space to separate the luminance and chrominance components; downsampling the luminance component using an image pyramid construction method to generate a low-resolution luminance feature map; dividing the low-resolution luminance feature map into multiple image blocks using a preset region-aware sparse sampling strategy and generating an importance score for each image block; selecting the top K image blocks with the highest importance scores as key regions and obtaining the coordinate set of the key regions in the low-illuminance image to be corrected. To generate the luminance and chrominance features of the key region, synchronous sampling is performed at corresponding positions of the luminance and chrominance components. A cross-channel feature fusion mechanism is used to fuse the luminance and chrominance features, generating enhanced luminance features with chrominance prior. The enhanced luminance features with chrominance prior and the coordinate set of the key region are input into a preset neural network to predict a global illumination gain map. The original luminance component is enhanced using the global illumination gain map to obtain the enhanced luminance component. The enhanced luminance component is merged with the original chrominance component and converted back to RGB space to obtain the final enhanced image.

[0109] Furthermore, when the logical instructions in the aforementioned memory 603 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0110] This invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer can execute an adaptive illumination correction method based on multi-scale feature coupling provided in the above-described method embodiments.

[0111] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements an adaptive illumination correction method based on multi-scale feature coupling provided in the above-described method embodiments.

[0112] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive illumination correction method based on multi-scale feature coupling, characterized in that, include: The low-light image to be corrected is converted to HSV space to separate the luminance and chrominance components; The brightness components are downsampled using an image pyramid construction method to generate a low-resolution brightness feature map; A preset region-aware sparse sampling strategy is used to divide the low-resolution brightness feature map into multiple image blocks and generate an importance score for each image block. The top K image patches with the highest importance scores are selected as key regions, and the coordinate set of the key regions in the low-light image to be corrected is obtained so that synchronous sampling is performed at the corresponding positions of the luminance component and the chrominance component to generate the luminance features and chrominance features of the key regions. A cross-channel feature fusion mechanism is used to fuse the luminance features and chrominance features to generate enhanced luminance features that include chrominance priors; The enhanced brightness features, which include chromaticity priors, and the coordinate set of the key region are input into a preset neural network to predict the global illumination gain map; the neural network has a pre-established continuous mapping relationship between the enhanced brightness features and spatial coordinates to the illumination gain coefficient; The original luminance component is enhanced using the global illumination gain map to obtain the enhanced luminance component; The enhanced luminance component is merged with the original chrominance component and converted back to RGB space to obtain the final enhanced image.

2. The adaptive illumination correction method based on multi-scale feature coupling according to claim 1, characterized in that, The importance score of each image block in the low-resolution brightness feature map is calculated using the following formula; in, Low-resolution brightness feature map Indicates the edge extraction kernel. This represents the convolution operation. This represents the local spatial attention extraction function. This represents the global context extraction function. These are learnable weight coefficients. This is a normalization operation.

3. The adaptive illumination correction method based on multi-scale feature coupling according to claim 1, characterized in that, The cross-channel feature fusion mechanism includes: A channel attention mechanism is introduced to reweight the luminance and chrominance features respectively according to the following formula, generating their respective channel weights: in, Brightness characteristics For chromaticity characteristics, It is a multilayer perceptron. Use the Sigmoid activation function; and For each corresponding channel weight; Based on their respective channel weights, the bidirectional feature interaction paths constructed according to the following formula generate their respective modulation weights for feature coupling: in, and Indicates a gating network. This indicates element-wise multiplication. This indicates a splicing operation. and For their respective modulation weights, Indicates the characteristics after coupling; The coupled features are fused with the original luminance features according to the following formula to obtain enhanced luminance features that include chromaticity prior. : in, A learnable residual scaling factor. It is a nonlinear transformation layer.

4. The adaptive illumination correction method based on multi-scale feature coupling according to claim 1, characterized in that, The neural network layers use a sinusoidal periodic function as the activation function, and the output of the i-th layer is... The calculation formula is: in, The weights and biases of the learnable i-th network layer. This is the preset frequency control factor.

5. An adaptive illumination correction system based on multi-scale feature coupling, characterized in that, include: The image component extraction module is used to convert the low-light image to be corrected to the HSV space to separate the luminance and chrominance components. The pyramid module is used to downsample the brightness components using an image pyramid construction method to generate a low-resolution brightness feature map; The region awareness module is used to divide the low-resolution brightness feature map into multiple image blocks and generate an importance score for each image block using a preset region awareness sparse sampling strategy. The top K image patches with the highest importance scores are selected as key regions, and the coordinate set of the key regions in the low-light image to be corrected is obtained so that synchronous sampling is performed at the corresponding positions of the luminance component and the chrominance component to generate the luminance features and chrominance features of the key regions. The feature fusion module is used to fuse the luminance features and chrominance features using a cross-channel feature fusion mechanism to generate enhanced luminance features that include chrominance priors; The illumination field modeling module is used to input the enhanced brightness features including chromaticity priors and the coordinate set of the key regions into a preset neural network to predict the global illumination gain map; the neural network has a pre-established continuous mapping relationship between spatial coordinates and illumination gain coefficients. The image reconstruction module is used to enhance the original luminance component using the global illumination gain map to obtain the enhanced luminance component; The enhanced luminance component is merged with the original chrominance component and converted back to RGB space to obtain the final enhanced image.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 4.

Images