A method and system for super-resolution reconstruction of low-light images based on illumination priors

Abstract

A method and system for super-resolution reconstruction of low-light images incorporating illumination priors are disclosed. The method includes the following steps: reading and organizing multiple original images, constructing multiple training sample pairs according to their correspondence; constructing a low-light image super-resolution reconstruction network, including a dual-priority guided network, a super-resolution reconstruction module, and a color correction module connected in sequence; training the low-light image super-resolution reconstruction network using the training sample pairs to obtain the trained network; deploying the trained network to a device and performing image reconstruction to obtain the reconstruction result. This invention can recover clear, low-noise image details under extremely low illumination conditions. Simultaneously, it significantly improves the color fidelity of the image, making the reconstructed image more consistent with human visual perception.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and system for super-resolution reconstruction of low-light images by incorporating illumination priors. Background Technology

[0002] Images acquired at night or in extremely low-light environments often suffer from severe visibility degradation, characterized by low brightness, poor contrast, color distortion, and increased noise. This degradation not only affects direct human perception but also weakens the performance of downstream visual tasks such as object detection and semantic segmentation. To address these problems, low-light image enhancement has become an important research direction.

[0003] Traditional low-light enhancement methods include histogram equalization, gamma correction, and Retinex theory. Retinex theory, in particular, decomposes the image into illumination and reflection components to explain color constancy, enhancing image brightness by estimating the illumination map. However, these traditional methods often have the following shortcomings: Histogram equalization can easily lead to loss of detail, and Retinex enhancement is sensitive to parameters; improper handling can result in color distortion or excessive detail enhancement. Simple gamma correction is ill-suited to areas with uneven lighting. Overall, traditional methods fall short when dealing with complex and varied low-light scenes.

[0004] The rise of deep learning technology has greatly promoted the development of low-light enhancement. Early convolutional networks such as LLNet and MBLLEN demonstrated the potential for end-to-end enhancement of low-light images. Meanwhile, some studies have incorporated physical priors into deep models; for example, RetinexNet estimates illumination and reflection components through the network and then enhances the image based on illumination. These Retinex-inspired networks utilize illumination priors to improve interpretability and stability. Nevertheless, even advanced deep models still face challenges in restoring details and maintaining color consistency under diverse low-light conditions. Structural information such as edges and textures in low-light images is very fragile, and ordinary single-scale models struggle to effectively preserve these details. Furthermore, the complexity of imaging devices and environments limits the generalization ability of models in real-world scenarios.

[0005] For the combined task of low-light super-resolution, the traditional approach is to process "low-light enhancement" and "image super-resolution" in sequence. However, feeding the output of the brightness enhancement model into the super-resolution model can lead to error accumulation: for example, brightening the image with the enhancement model before super-resolution can easily amplify artifacts or noise generated during the enhancement process, resulting in significant color distortion and difficulty in restoring details.

[0006] In summary, existing technologies for super-resolution of low-light images face two major challenges: first, how to effectively enhance modulation brightness and suppress noise while improving resolution, avoiding the accumulation of errors caused by simple concatenation processing; and second, how to ensure that the enhanced image has realistic and natural colors and reduce color distortion. Therefore, it is necessary to propose a new technical solution to address the issues of sharpness, brightness, and color in low-light images. Summary of the Invention

[0007] This invention provides a method and system for super-resolution reconstruction of low-light images by incorporating illumination priors, in order to solve the technical problems mentioned in the background art.

[0008] To achieve the above objectives, the technical solution of the present invention is implemented as follows: This invention provides a method for super-resolution reconstruction of low-light images by incorporating illumination priors, comprising the following steps: S1. Read and organize multiple original images, and construct multiple training sample pairs; S2. Construct a low-light image super-resolution reconstruction network, which includes a dual priori guided network DPGN, a super-resolution reconstruction module, and a color correction module SCCM connected in sequence. S3. Upsample the training samples using the upsampling operator to obtain the upsampled image. Then, input the upsampled image into the dual prior guidance network DPGN to obtain multi-channel illumination guidance prior features, illumination prior map, and noise prior map. S4. Input the upsampled image and multi-channel illumination-guided prior features into the super-resolution reconstruction module to obtain the super-resolution reconstruction result; S5. Input the super-resolution reconstruction results into the color correction module SCCM to obtain the final result; S6. Construct the total loss function based on the final result, the illumination prior map, and the noise prior map, and adjust the low-light image super-resolution reconstruction network according to the total loss function; determine whether the set iteration stopping condition has been met. If so, output the trained low-light image super-resolution reconstruction network; otherwise, return to S3. S7. Deploy the trained low-light image super-resolution reconstruction network to the device and perform image reconstruction to obtain the reconstruction results.

[0009] Furthermore, step S1 specifically includes the following steps: S11. Acquire or select multiple raw images, including multiple low-light, low-resolution images. and multiple corresponding high-resolution images of normal light ; Organize low-light, low-resolution images according to filename correspondence High-resolution images with normal light The images are then paired, converted into array and tensor formats, and pixel values ​​are limited to a preset range to obtain multiple preprocessed original images. S12. Construct multiple training sample pairs using the preprocessed original images. Each training sample pair includes a low-light, low-resolution image. and corresponding high-resolution images of normal light .

[0010] Furthermore, the dual prior guided network DPGN includes a color space transformation operator R1, a first illumination feature extraction U-shaped network IFU, a second illumination feature extraction U-shaped network IFU, and two convolutional layers C. 11 Activation function S 11 And a projection module; wherein, the output of the color space conversion operator R1 is connected to the input of the first illumination feature extraction U-shaped network (IFU) and the second illumination feature extraction U-shaped network (IFU), and the outputs of the first illumination feature extraction U-shaped network (IFU) and the second illumination feature extraction U-shaped network (IFU) are connected to the corresponding convolutional layers C. 11 The connections are made to obtain the luminance prior activation pre-mapped value Ly and the chrominance perturbation activation pre-mapped value Lc, respectively; the luminance prior activation pre-mapped value Ly and the chrominance perturbation activation pre-mapped value Lc are then corrected and input into the activation function S. 11 To obtain the illumination prior map and noise prior maps The illumination prior map The projection module is input, and the projection module includes convolutional layers C connected in sequence. 12 Activation function R 11 Convolutional layer C 13 and activation function S 12 The activation function S 12 The output has three branches.

[0011] The super-resolution reconstruction module consists of C convolutional layers connected in sequence. 21 Encoder illumination fusion module EIFB1, downsampling operator D 21 Residual channel attention network RCAN1, encoder illumination fusion module EIFB2, downsampling operator D 22 1. Residual channel attention network RCAN2, encoder illumination fusion module EIFB3, upsampling layer U 21 Residual Channel Attention Network RCAN3, Upsampling Layer U 22 Residual Channel Attention Network RCAN4, Convolutional Layer C 22 ; where the activation function S 12The three branches of the output end are connected to encoder illumination fusion modules EIFB1, EIFB2, and EIFB3, respectively; the output end of encoder illumination fusion module EIFB1 is split into two branches, and one branch of encoder illumination fusion module EIFB1 is connected to the downsampling operator D. 21 Another branch uses element-wise addition operations with the convolutional layer C. 22 The input terminal is connected; the downsampling operator D 21 The output is split into two branches. The first branch is connected to the residual channel attention network RCAN1, and the other branch is connected to the upsampling layer U through element-wise addition. 22 The input terminal is connected; the downsampling operator D 22 The output is split into two branches. The first branch is connected to the residual channel attention network RCAN2, and the other branch is connected to the upsampling layer U through element-wise addition. 22 connect; The color space correction module SCCM includes a color space conversion operator R2, a color residual prediction network, and a clipping operator Clip. The output of the color space conversion operator R2 is split into three branches, namely the first branch to the third branch. The first branch is connected in series with the color residual prediction network. The second branch is connected to the output of the color residual prediction network through an element-wise addition operation to form the fourth branch. The third branch and the fourth branch are connected through a channel-dimensional splicing operation.

[0012] Furthermore, both the first illumination feature extraction U-shaped network (IFU) and the second illumination feature extraction U-shaped network (IFU) include a convolutional block CB1 and a downsampling operator D connected in sequence. 11 Convolutional block CB2, downsampling operator D 12 Convolutional block CB3, upsampling layer U 11 Convolutional block CB4, upsampling layer U 12 Convolutional block CB5, convolutional layer C 14 ; The output of convolutional block CB1 is split into two branches. The first branch is connected to the downsampling operator D. 11 Connect, another branch with the upsampling layer U 12 The output features are concatenated in the channel dimension; the output of convolutional block CB2 is split into two branches, the first branch being connected to the downsampling operator D. 12 Connect, another branch with the upsampling layer U 11 The output features are concatenated in the channel dimension to form skip connections.

[0013] Furthermore, step S3 specifically includes the following steps: S31. Upsample the training samples using the upsampling operator to obtain the upsampled image. ; S32, Upsample the image The input to the dual-prior guided network DPGN first passes through the color space conversion operator R1 to upsample the image. Converting from RGB color space to YCbCr space yields three-channel YCbCr features. Then, the three-channel YCbCr features were analyzed. The data is split to obtain the brightness features of one channel. and two-channel chromaticity characteristics ; S33, convert one channel brightness component and two-channel chromaticity components The inputs are respectively fed into the first illumination feature extraction U-shaped network IFU and the second illumination feature extraction U-shaped network IFU, and illumination-related features are extracted respectively. and noise / chromaticity perturbation related features ; S34. Illumination-related features and noise / chromaticity perturbation related features The inputs are fed into two convolutional layers C respectively. 11 In the process, the features are compressed into single-channel features, and the luminance prior activation pre-mapped value Ly and the chrominance perturbation activation pre-mapped value Lc are obtained respectively. S35. The luminance prior activation pre-mapped value Ly and the chromaticity perturbation activation pre-mapped value Lc are subtracted element-wise to obtain the corrected illumination activation pre-mapped value Lr. S36. Input the pre-activation mapping value Lr of the corrected illumination and the pre-activation mapping value Lc of the chromaticity perturbation into the activation function S respectively. 11 In the process, the illumination prior map is obtained. Noise Prior Diagram ; S37. Illumination prior diagram The image is input into the projection module, which then projects the illumination prior image. Mapping back to multi-channel illumination-guided prior features .

[0014] Furthermore, the upsampled image in S31 The expression is as follows: ; in, This indicates an upsampling operator using bicubic interpolation; This indicates that the upsampling factor is n; The three-channel YCbCr feature in S32 The expression is as follows: ; in, Let be the linear transformation matrix from RGB to YCbCr. It is the bias vector; This represents the convolutional layer within the color space conversion operator R1; The illumination-related features in S33 and noise / chromaticity perturbation related features The expressions are as follows: ; ; in, This represents the brightness branch, i.e., the first illumination feature extraction U-shaped network (IFU). This represents the chromaticity branch, i.e., the second illumination feature extraction U-shaped network (IFU). The expressions for the luminance prior activation pre-mapped value Ly and the chrominance perturbation activation pre-mapped value Lc in S34 are as follows: ; ; in, Represents the convolutional layer C 11 ; The mapping value before illumination activation in S35 is corrected. The expression is as follows: ; in, Represents the learnable coefficient; The illumination prior map in S36 Noise Prior Diagram The expressions are as follows: ; ; in, S represents the activation function. 11 Specifically, the Sigmoid function; The multi-channel illumination-guided prior features in S37 The expression is as follows: ; in, Represents the convolutional layer C 12 ; Represents the activation function R 11 ; Represents the convolutional layer C 13 ; S represents the activation function. 12Specifically, this refers to the Sigmoid function.

[0015] Furthermore, step S4 specifically includes the following steps: S41. Guiding prior features with multi-channel illumination The input is fed into the super-resolution reconstruction module, where it is first constructed as a three-scale pyramid feature using bilinear interpolation. These correspond to the backbone network. , Compared to the original scale; features of the three-scale pyramid The expressions are as follows: ; ; ; in, This indicates that bilinear interpolation is used to scale the feature map. Scaling is applied to the feature map by a ratio of s=0.25 or 0.5; , , They are respectively , Prior lighting features at the original scale; S42, Upsample the image The input is first processed by the mean-variance normalization operator and then by the convolutional layer C in the super-resolution reconstruction module. 21 Obtain input features Then, at the original scale, the input features are... Prior features of illumination The original scale-corrected features are obtained by fusion. Input features and original scale correction features The expressions are as follows: ; in, This represents the mean-variance normalization operator. Represents the convolutional layer C 21 ; ; S43. Correct the original scale features Input to downsampling operator D 21 Perform downsampling to obtain encoded features Next, the encoded features will be... The input is fed into the residual channel attention network RCAN1 to obtain the encoded features. Then At scale, the encoded features will be and prior features of illumination The input is fed into the encoder illumination fusion module EIFB2 for fusion to obtain 1 / 2 scale correction features. The expressions are as follows: ; ; ; in, The downsampling operator D is represented. 21 ; S44, Correct the 1 / 2 scale feature Input to downsampling operator D 22 Downsampling is performed to obtain encoded features. Encode features The input is fed into the residual channel attention network RCAN2 to obtain the encoded features. Then At scale, the encoded features will be Prior features of illumination The input is fed into the encoder illumination fusion module EIFB3 for fusion to obtain 1 / 4 scale correction features. The expressions are as follows: ; ; ; in, The downsampling operator D is represented. 22 ; S45. Encode features and 1 / 4 scale correction features The fused input is then fed into the upsampling layer U. 21 Upsampling is performed to obtain the first decoded feature. Then the first decoded feature The input is fed into the residual channel attention network RCAN3 for feature enhancement, resulting in the first enhanced decoding feature. Next, the first enhanced decoding feature will be... With coding features The fused input is then fed into the upsampling layer U. 22 In the process, spatial resolution is restored to obtain the second decoding feature. Then the second decoding feature The input is fed into the residual channel attention network RCAN4 for feature enhancement, resulting in the second enhanced decoding feature. Then the second enhanced decoding feature and original scale correction features Input to convolutional layer C22 In the process, the super-resolution reconstruction results are obtained. The expressions are as follows: ; ; ; ; ; in, Indicates the upsampling layer U 21 ; Indicates upsampling layer U 22 ; Represents the convolutional layer C 22 .

[0016] Furthermore, step S5 specifically includes the following steps: S51. The super-resolution reconstruction results... The input is fed into the color correction module SCCM, where it first undergoes a color space conversion operator R2 to convert the super-resolution reconstruction result. Transform to YCbCr space to obtain reconstructed YCbCr features Then, the YCbCr features will be reconstructed. Decomposition yields reconstructed chromaticity features and reconstructed brightness features Reconstructing YCbCr characteristics The expression is as follows: ; S52, Reconstructing chromaticity features The input is fed into the chromaticity residual prediction network to obtain the chromaticity residual. Next, the chromaticity residuals Reconstructing chromaticity features Input to truncation operator In the process, the corrected reconstructed chromaticity features are obtained. The expressions are as follows: ; ; in, This represents a chromaticity residual prediction network; S53, Reconstruct brightness features and the corrected reconstructed chromaticity features Perform an inverse transformation from the YCbCr color space to the RGB color space to obtain the final result. The expression is as follows: ; in, This represents the inverse transformation from the YCbCr color space to the RGB color space. This indicates feature splicing along the channel dimension.

[0017] Furthermore, the expression for the total loss function in S6 is as follows: ; in, Total loss; This represents the loss in super-resolution primary reconstruction. To estimate the loss weighting coefficients a priori; To estimate the loss using two priors; For chromaticity correction constraint loss; Super-resolution primary reconstruction loss The expression is as follows: ; in, For L1 loss, In order to perceive loss, For frequency domain loss; These are the corresponding weighting coefficients; Two prior estimation loss The expression is as follows: ;

[0018] in, and For the corresponding monitoring signal; The total variation regularization weight, Weights for noise priors; Represents the L1 norm; This represents the regularization term for the total variation; Colorimetric correction constraint loss The expression is as follows: ; in, This indicates that the chroma channels are extracted after the image is converted to the YCbCr color space; Weights for chromaticity correction loss; The weights are for pixel auxiliary terms.

[0019] In another aspect, the present invention provides a low-light image super-resolution reconstruction system incorporating illumination priors, configured to perform the aforementioned low-light image super-resolution reconstruction method incorporating illumination priors, comprising: The building unit is configured to build a low-light image super-resolution reconstruction network, which includes a dual priori guided network DPGN, a super-resolution reconstruction module, and a color correction module SCCM. The prior extraction unit is configured to use the dual prior guidance network DPGN to obtain multi-channel illumination guidance prior features, illumination prior map, and noise prior map. The super-resolution reconstruction unit is configured to use the super-resolution reconstruction module to obtain super-resolution reconstruction results. The colorimetric correction unit is configured to obtain the final result using the colorimetric correction module SCCM.

[0020] The beneficial effects of this invention are: 1. It breaks the limitation of error accumulation; This invention constructs a super-resolution reconstruction network for low-light images, abandoning the traditional serial processing mode of "enhancing first and then super-resolution". Instead, it inputs multi-channel illumination-guided prior features into the super-resolution reconstruction module for multi-scale fusion, effectively avoiding the problems of amplified enhancement artifacts and error accumulation in the traditional two-stage method.

[0021] 2. This invention enhances robustness through dual physical prior guidance; This invention extracts illumination prior maps simultaneously using a dual prior guidance network (DPGN). Noise Prior Diagram Illumination prior diagram Noise prior map used to guide low-light image super-resolution reconstruction networks in recovering textures in dark areas. This invention helps super-resolution reconstruction networks for low-light images distinguish high-frequency details from noise by utilizing illumination prior maps. Noise Prior Diagram The original scale correction features, 1 / 2 scale correction features, and 1 / 4 scale correction features in the super-resolution reconstruction module are guided and constrained, so that clear and low-noise image details can be recovered even under extremely low illumination conditions.

[0022] 3. This invention ensures structural consistency through multi-scale fusion; This invention introduces an encoder illumination fusion block (EIFB) at the original scale, 1 / 2 scale, and 1 / 4 scale, which enables deep interaction between illumination information and content features. This ensures that while improving image resolution, it also guarantees natural brightness adjustment and consistency of structural information (such as edges and contours).

[0023] 4. This invention solves the color cast problem through the color correction module SCCM; To address the color distortion problem often associated with enhanced low-light performance, a colorimetric correction module (SCCM) was designed. The SCCM performs residual learning independently on the chroma channels in the YCbCr color space without interfering with the structural reconstruction of the luminance channels, thereby significantly improving the color fidelity of the image and making the reconstructed image more consistent with human visual perception. Attached Figure Description

[0024] Figure 1 This is a flowchart of the low-light image super-resolution reconstruction method in this invention; Figure 2 A block diagram of a network for super-resolution reconstruction of low-light images; Figure 3 This is a structural block diagram of the Dual Prior Guidance Network (DPGN) in this invention; Figure 4 This is a block diagram of the Residual Channel Attention Network (RCAN) in this invention. Figure 5 This is a structural block diagram of the encoder illumination fusion module EIFB in this invention; Figure 6 This is a structural block diagram of the colorimetric correction module SCCM in this invention; Figure 7 This is a schematic diagram of the reconstruction result in an embodiment of the present invention. Detailed Implementation

[0025] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many other different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0026] The purpose of this invention is to overcome the problems of error accumulation, noise amplification, and color distortion caused by simple concatenation enhancement and super-resolution models in the super-resolution reconstruction of low-light images in existing technologies. This invention aims to better restore natural lighting and color while improving image resolution, and provides a method for super-resolution reconstruction of low-light images that integrates lighting priors, as detailed below: Reference Figure 1 and Figure 2 This application provides a method for super-resolution reconstruction of low-light images by fusing illumination priors, including the following steps: S1. Read and organize multiple original images, and construct multiple training sample pairs; S2. Construct a super-resolution reconstruction network for low-light images, including a dual-prior guidance network (DPGN), a super-resolution reconstruction module, and a spatial chroma correction module (SCCM) connected in sequence. S3. Upsample the training sample pairs using the upsampling operator to obtain upsampled images. Then, input the upsampled images into the dual prior guidance network DPGN to obtain multi-channel illumination guidance prior features, illumination prior maps, and noise prior maps. S4. Input the upsampled image and multi-channel illumination-guided prior features into the super-resolution reconstruction module to obtain the super-resolution reconstruction result; S5. Input the super-resolution reconstruction results into the color correction module SCCM to obtain the final result; S6. Construct the total loss function based on the final result, the illumination prior map, and the noise prior map, and adjust the low-light image super-resolution reconstruction network according to the total loss function; determine whether the set iteration stopping condition has been met. If so, output the trained low-light image super-resolution reconstruction network; otherwise, return to S3. S7. Deploy the trained low-light image super-resolution reconstruction network to the device and perform image reconstruction to obtain the reconstruction results. The reconstruction results are shown below. Figure 7 As shown, the left and right images constitute a training sample pair, and the middle image is the reconstructed image of the left image.

[0027] In some embodiments, S1 specifically includes the following steps: S11. Acquire or select multiple raw images, including multiple low-light, low-resolution images. and multiple corresponding high-resolution images of normal light ; Organize low-light, low-resolution images according to filename correspondence High-resolution images with normal light The images are then paired, converted into array and tensor formats, and pixel values ​​are limited to a preset range to obtain multiple preprocessed original images. The low-light, low-resolution image This refers to an input image that simultaneously meets the conditions of low illumination and low spatial resolution; the normal light high-resolution image This refers to the low-light, low-resolution image. A reference image is defined as having a one-to-one correspondence in scene content and possessing a set brightness level (e.g., a high brightness level) and a set spatial resolution (e.g., a high spatial resolution) or higher. In some embodiments, low-light conditions are determined by brightness statistical features, which include at least one of the following: average grayscale value, average brightness channel value, median brightness value, or exposure level. When the brightness statistical features are lower than a preset brightness threshold, the image is determined to be a low-light image. In some embodiments, low spatial resolution conditions are determined by image pixel size or scaling factor. When the width and / or height of the input image is lower than the corresponding reference image, or when the input image is obtained by downsampling the corresponding reference image at a preset scaling factor, the input image is determined to be a low-resolution image. The preset scaling factor can be ×2, ×4, or other scaling factors.

[0028] S12. Construct multiple training sample pairs using the preprocessed original images. Each training sample pair includes a low-light, low-resolution image. and corresponding high-resolution images of normal light .

[0029] In some embodiments, refer to Figure 3 The dual prior guided network DPGN includes a color space transformation operator R1, a first illumination feature extraction U-shaped network IFU, a second illumination feature extraction U-shaped network IFU, and two convolutional layers C. 11 Activation function S 11 And a projection module; wherein, the output of the color space conversion operator R1 is connected to the input of the first illumination feature extraction U-shaped network (IFU) and the second illumination feature extraction U-shaped network (IFU), and the outputs of the first illumination feature extraction U-shaped network (IFU) and the second illumination feature extraction U-shaped network (IFU) are connected to the corresponding convolutional layers C. 11 The connections are made to obtain the luminance prior activation pre-mapped value Ly and the chrominance perturbation activation pre-mapped value Lc, respectively; the luminance prior activation pre-mapped value Ly and the chrominance perturbation activation pre-mapped value Lc are then corrected and input into the activation function S. 11 To obtain the illumination prior map and noise prior maps The illumination prior map The projection module is input, and the projection module includes convolutional layers C connected in sequence. 12 Activation function R 11 Convolutional layer C 13 and activation function S 12 The activation function S 12 The output terminal has three branches; The super-resolution reconstruction module consists of C convolutional layers connected in sequence. 21Encoder illumination fusion module EIFB1, downsampling operator D 21 Residual channel attention network RCAN1, encoder illumination fusion module EIFB2, downsampling operator D 22 1. Residual channel attention network RCAN2, encoder illumination fusion module EIFB3, upsampling layer U 21 Residual Channel Attention Network RCAN3, Upsampling Layer U 22 Residual Channel Attention Network RCAN4, Convolutional Layer C 22 ; where the activation function S 12 The three branches of the output end are connected to encoder illumination fusion modules EIFB1, EIFB2, and EIFB3, respectively; the output end of encoder illumination fusion module EIFB1 is split into two branches, and one branch of encoder illumination fusion module EIFB1 is connected to the downsampling operator D. 21 Another branch uses element-wise addition operations with the convolutional layer C. 22 The input terminal is connected; the downsampling operator D 21 The output is split into two branches. The first branch is connected to the residual channel attention network RCAN1, and the other branch is connected to the upsampling layer U through element-wise addition. 22 The input terminal is connected; the downsampling operator D 22 The output is split into two branches. The first branch is connected to the residual channel attention network RCAN2, and the other branch is connected to the upsampling layer U through element-wise addition. 22 Connections; the residual channel attention network RCAN1 and residual channel attention network RCAN2 have the same structure, see details. Figure 4 As shown; Reference Figure 5 The encoder illumination fusion modules EIFB1, EIFB2, and EIFB3 have the same structure, all consisting of channel-dimensional splicing. Convolution dimensionality reduction, ReLU activation function and The input features of the encoder's illumination fusion module are enhanced using convolutional methods and a scaling residual stacking technique. Mathematically, this can be represented as: ; ; ; in, This represents the input characteristics of the encoder's illumination fusion module; These are prior features of illumination at the same scale; Indicates fusion characteristics; Represents the characteristics of residual mapping; express Convolution dimensionality reduction; Represents the ReLU activation function; This represents the residual scaling factor, which is the residual scaling factor in this invention. Take 0.1; This indicates the output characteristics of the encoder's illumination fusion module; Reference Figure 6 The color space correction module (SCCM) includes a color space conversion operator R2, a color residual prediction network, and a clipping operator Clip. The output of the color space conversion operator R2 is split into three branches, namely the first branch to the third branch. The first branch is connected in series with the color residual prediction network. The second branch is connected to the output of the color residual prediction network through an element-wise addition operation to form the fourth branch. The third branch and the fourth branch are connected through a channel-dimensional splicing operation.

[0030] In some embodiments, the first illumination feature extraction U-Net (Illumination Feature U-Net) and the second illumination feature extraction U-Net both include two levels of encoders, a bottleneck layer, and two levels of decoders connected in sequence. Each level of encoder is connected to the decoder of the same level through a skip connection. The encoder uses two levels of max pooling downsampling, and the decoder uses bilinear upsampling and is concatenated with the corresponding coding layer features to form a skip connection. Specifically, both the first illumination feature extraction U-shaped network IFU and the second illumination feature extraction U-shaped network IFU include a convolutional block CB1 and a downsampling operator D connected in sequence. 11 Convolutional block CB2, downsampling operator D 12 Convolutional block CB3, upsampling layer U 11 Convolutional block CB4, upsampling layer U 12 Convolutional block CB5, convolutional layer C 14 Among them, convolutional blocks CB1 and CB2 correspond to encoders, which are used to encode the input features step by step; convolutional block CB3 serves as the bottleneck layer between the two-stage encoder and the two-stage decoder, and is used to extract the deep features after downsampling; convolutional blocks CB4 and CB5 correspond to decoders, which are used to decode and restore the upsampling features step by step. In addition, the output of the convolutional block CB1 is split into two branches, the first branch being connected to the downsampling operator D. 11 Connect, another branch with the upsampling layer U 12 The output features are concatenated in the channel dimension; the output of convolutional block CB2 is split into two branches, the first branch being connected to the downsampling operator D. 12 Connect, another branch with the upsampling layer U 11 The output features are concatenated in the channel dimension to form skip connections.

[0031] In some embodiments, S3 specifically includes the following steps: S31. Upsample the training samples using the upsampling operator to obtain the upsampled image. ; S32, Upsample the image The input to the dual-prior guided network DPGN first passes through the color space conversion operator R1 to upsample the image. Converting from RGB color space to YCbCr space yields three-channel YCbCr features. Then, the three-channel YCbCr features were analyzed. The data is split to obtain the brightness features of one channel. and two-channel chromaticity characteristics ; S33, convert one channel brightness component and two-channel chromaticity components The inputs are respectively fed into the first illumination feature extraction U-shaped network IFU and the second illumination feature extraction U-shaped network IFU, and illumination-related features are extracted respectively. and noise / chromaticity perturbation related features ; S34. Illumination-related features and noise / chromaticity perturbation related features The inputs are fed into two convolutional layers C respectively. 11 In the process, the features are compressed into single-channel features, and the luminance prior activation pre-mapped value Ly and the chrominance perturbation activation pre-mapped value Lc are obtained respectively. S35. The luminance prior activation pre-mapped value Ly and the chromaticity perturbation activation pre-mapped value Lc are subtracted element-wise to obtain the corrected illumination activation pre-mapped value Lr. S36. Before activating the corrected illumination, adjust the mapping value Lr. The chromaticity perturbation before activation mapping value Lc is input into the activation function S. 11 In the process, the illumination prior map is obtained. Noise Prior Diagram ; S37. Illumination prior diagram The image is input into the projection module, which then projects the illumination prior image. Mapping back to multi-channel illumination-guided prior features .

[0032] In some embodiments, the upsampled image in S31 The expression is as follows: ; in, This indicates an upsampling operator using bicubic interpolation; This indicates that the upsampling factor is n; This represents the image after n times upsampling (compared to a normal high-resolution image). (equal in size) The three-channel YCbCr feature in S32 The expression is as follows: ; in, Let be the linear transformation matrix from RGB to YCbCr. It is the bias vector; This represents the convolutional layer within the color space conversion operator R1; The illumination-related features in S33 and noise / chromaticity perturbation related features The expressions are as follows: ; ; in, This represents the brightness branch, i.e., the first illumination feature extraction U-shaped network (IFU). This represents the chromaticity branch, i.e., the second illumination feature extraction U-shaped network (IFU). The expressions for the luminance prior activation pre-mapped value Ly and the chrominance perturbation activation pre-mapped value Lc in S34 are as follows: ; ; in, Represents the convolutional layer C 11 ; The mapping value before illumination activation in S35 is corrected. The expression is as follows: ; in, Represents the learnable coefficient; The illumination prior map in S36 Noise Prior Diagram The expressions are as follows: ; ; in, S represents the activation function. 11 Specifically, the Sigmoid function; The multi-channel illumination-guided prior features in S37 The expression is as follows: ; in, Represents the convolutional layer C 12 ; Represents the activation function R 11 ; Represents the convolutional layer C 13 ; S represents the activation function. 12 Specifically, this refers to the Sigmoid function.

[0033] In some embodiments, S4 specifically includes the following steps: S41. Guiding prior features with multi-channel illumination The input is fed into the super-resolution reconstruction module, where it is first constructed as a three-scale pyramid feature using bilinear interpolation. These correspond to the backbone network. , Compared to the original scale; features of the three-scale pyramid The expressions are as follows: ; ; ; in, This indicates that bilinear interpolation is used to scale the feature map. Scaling is applied to the feature map by a ratio of s=0.25 or 0.5; , , They are respectively , Prior lighting features at the original scale; S42, Upsample the image The input is first processed by the mean-variance normalization operator and then by the convolutional layer C in the super-resolution reconstruction module. 21 Obtain input features Then, at the original scale, the input features are... Prior features of illumination The original scale-corrected features are obtained by fusion. Input features and original scale correction features The expressions are as follows: ; in, This represents the mean-variance normalization operator (MeanShift). Represents the convolutional layer C 21 ; ; S43. Correct the original scale features Input to downsampling operator D 21 Perform downsampling to obtain encoded features Next, the encoded features will be... The input is fed into the residual channel attention network RCAN1 to obtain the encoded features. Then At scale, the encoded features will be and prior features of illumination The input is fed into the encoder illumination fusion module EIFB2 for fusion to obtain 1 / 2 scale correction features. The expressions are as follows: ; ; ; in, The downsampling operator D is represented. 21 ; S44, Correct the 1 / 2 scale feature Input to downsampling operator D 22 Downsampling is performed to obtain encoded features. Encode features The input is fed into the residual channel attention network RCAN2 to obtain the encoded features. Then At scale, the encoded features will be Prior features of illumination The input is fed into the encoder illumination fusion module EIFB3 for fusion to obtain 1 / 4 scale correction features. The expressions are as follows: ; ; ; in, The downsampling operator D is represented. 22 ; S45. Encode features and 1 / 4 scale correction features The fused input is then fed into the upsampling layer U. 21 Upsampling is performed to obtain the first decoded feature. Then the first decoded feature The input is fed into the residual channel attention network RCAN3 for feature enhancement, resulting in the first enhanced decoding feature. Next, the first enhanced decoding feature will be... With coding features The fused input is then fed into the upsampling layer U. 22 In the process, spatial resolution is restored to obtain the second decoding feature. Then the second decoding feature The input is fed into the residual channel attention network RCAN4 for feature enhancement, resulting in the second enhanced decoding feature. Then the second enhanced decoding feature and original scale correction features Input to convolutional layer C 22 In the process, the super-resolution reconstruction results are obtained. The expressions are as follows: ; ; ; ; ; in, U represents the upsampling layer with a scaling factor of 2. 21 ; U represents the upsampling layer with a scaling factor of 2. 22 ; Represents the convolutional layer C 22 .

[0034] In some embodiments, S5 specifically includes the following steps: S51. The super-resolution reconstruction results... The input is fed into the color correction module SCCM, where it first undergoes a color space conversion operator R2 to convert the super-resolution reconstruction result. Transform to YCbCr space to obtain reconstructed YCbCr features Then, the YCbCr features will be reconstructed. Decomposition yields reconstructed chromaticity features and reconstructed brightness features Reconstructing YCbCr characteristics The expression is as follows: ; S52, Reconstructing chromaticity features The input is fed into the chromaticity residual prediction network to obtain the chromaticity residual. Next, the chromaticity residuals Reconstructing chromaticity features Input to truncation operator In the process, the corrected reconstructed chromaticity features are obtained. The expressions are as follows: ; ; in, The chroma residual prediction network consists of three convolutional downsampling layers, an intermediate convolutional layer, and bilinear upsampling and output convolutional layers, thereby generating two-channel chroma residuals of the same size as the input. ; S53, Reconstruct brightness features and the corrected reconstructed chromaticity features Perform an inverse transformation from the YCbCr color space to the RGB color space to obtain the final result. The expression is as follows: ; in, This represents the inverse transformation from the YCbCr color space to the RGB color space. This indicates feature splicing along the channel dimension.

[0035] In some embodiments, the expression for the total loss function in S6 is as follows: ; in, Total loss; This represents the loss in super-resolution primary reconstruction. To estimate the loss weighting coefficients a priori; To estimate the loss using two priors; For chromaticity correction constraint loss; Super-resolution primary reconstruction loss The expression is as follows: ; in, For L1 loss, In order to perceive loss, For frequency domain loss; These are the corresponding weighting coefficients; Two prior estimation loss The expression is as follows: ;

[0036] in, and For the corresponding monitoring signal; The total variation regularization weight, Weights for noise priors; Represents the L1 norm; This represents the regularization term for the total variation; Colorimetric correction constraint loss The expression is as follows: ; in, This indicates that the chroma channels are extracted after the image is converted to the YCbCr color space; Weights for chromaticity correction loss; The weights are for pixel auxiliary terms.

[0037] A second aspect of the present invention also provides a low-light image super-resolution reconstruction system incorporating illumination priors, configured to perform the aforementioned low-light image super-resolution reconstruction method incorporating illumination priors, comprising: The building unit is configured to build a low-light image super-resolution reconstruction network, which includes a dual priori guided network DPGN, a super-resolution reconstruction module, and a color correction module SCCM. The prior extraction unit is configured to use the dual prior guidance network DPGN to obtain multi-channel illumination guidance prior features, illumination prior map, and noise prior map. The super-resolution reconstruction unit is configured to use the super-resolution reconstruction module to obtain super-resolution reconstruction results. The colorimetric correction unit is configured to obtain the final result using the colorimetric correction module SCCM.

[0038] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for super-resolution reconstruction of low-light images by incorporating illumination priors, characterized in that, Includes the following steps: S1. Read and organize multiple original images, and construct multiple training sample pairs; S2. Construct a low-light image super-resolution reconstruction network, which includes a dual priori guided network DPGN, a super-resolution reconstruction module, and a color correction module SCCM connected in sequence. S3. Upsample the training samples using the upsampling operator to obtain the upsampled image. Then, input the upsampled image into the dual prior guidance network DPGN to obtain multi-channel illumination guidance prior features, illumination prior map, and noise prior map. S4. Input the upsampled image and multi-channel illumination-guided prior features into the super-resolution reconstruction module to obtain the super-resolution reconstruction result; S5. Input the super-resolution reconstruction results into the color correction module SCCM to obtain the final result; S6. Construct the total loss function based on the final result, the illumination prior map, and the noise prior map, and adjust the low-light image super-resolution reconstruction network according to the total loss function; determine whether the set iteration stopping condition has been met. If so, output the trained low-light image super-resolution reconstruction network; otherwise, return to S3. S7. Deploy the trained low-light image super-resolution reconstruction network to the device and perform image reconstruction to obtain the reconstruction results. S5 specifically includes the following steps: S51. The super-resolution reconstruction results... The input is fed into the color correction module SCCM, where it first undergoes a color space conversion operator R2 to convert the super-resolution reconstruction result. Transform to YCbCr space to obtain reconstructed YCbCr features Then, the YCbCr features will be reconstructed. Decomposition yields reconstructed chromaticity features and reconstructed brightness features Reconstructing YCbCr characteristics The expression is as follows: ； S52, Reconstructing chromaticity features The input is fed into the chromaticity residual prediction network to obtain the chromaticity residual. Next, the chromaticity residuals Reconstructing chromaticity features Input to truncation operator In the process, the corrected reconstructed chromaticity features are obtained. The expressions are as follows: ； ； in, This represents a chromaticity residual prediction network; S53, Reconstruct brightness features and the reconstructed chromaticity features after correction Perform an inverse transformation from the YCbCr color space to the RGB color space to obtain the final result. The expression is as follows: ； in, This represents the inverse transformation from the YCbCr color space to the RGB color space. This indicates feature splicing along the channel dimension.

2. The method for super-resolution reconstruction of low-light images based on illumination priors according to claim 1, characterized in that, S1 specifically includes the following steps: S11. Acquire or select multiple raw images, including multiple low-light, low-resolution images. and multiple corresponding high-resolution images of normal light ; Organize low-light, low-resolution images according to filename correspondence High-resolution images with normal light The images are then paired, converted into array and tensor formats, and pixel values ​​are limited to a preset range to obtain multiple preprocessed original images. S12. Construct multiple training sample pairs using the preprocessed original images. Each training sample pair includes a low-light, low-resolution image. and corresponding high-resolution images of normal light .

3. The method for super-resolution reconstruction of low-light images based on illumination priors according to claim 2, characterized in that, The dual prior guided network DPGN includes a color space transformation operator R1, a first illumination feature extraction U-shaped network (IFU), a second illumination feature extraction U-shaped network (IFU), and two convolutional layers C. 11 Activation function S 11 And a projection module; wherein, the output of the color space conversion operator R1 is connected to the input of the first illumination feature extraction U-shaped network (IFU) and the second illumination feature extraction U-shaped network (IFU), and the outputs of the first illumination feature extraction U-shaped network (IFU) and the second illumination feature extraction U-shaped network (IFU) are connected to the corresponding convolutional layers C. 11 The connections are made to obtain the luminance prior activation pre-mapped value Ly and the chrominance perturbation activation pre-mapped value Lc, respectively; the luminance prior activation pre-mapped value Ly and the chrominance perturbation activation pre-mapped value Lc are then corrected and input into the activation function S. 11 To obtain the illumination prior map and noise prior maps The illumination prior map The projection module is input, and the projection module includes convolutional layers C connected in sequence. 12 Activation function R 11 Convolutional layer C 13 and activation function S 12 The activation function S 12 The output terminal has three branches; The super-resolution reconstruction module consists of C convolutional layers connected in sequence. 21 Encoder illumination fusion module EIFB1, downsampling operator D 21 Residual channel attention network RCAN1, encoder illumination fusion module EIFB2, downsampling operator D 22 1. Residual channel attention network RCAN2, encoder illumination fusion module EIFB3, upsampling layer U 21 Residual Channel Attention Network RCAN3, Upsampling Layer U 22 Residual Channel Attention Network RCAN4, Convolutional Layer C 22 ; where the activation function S 12 The three branches of the output end are connected to encoder illumination fusion modules EIFB1, EIFB2, and EIFB3, respectively; the output end of encoder illumination fusion module EIFB1 is split into two branches, and one branch of encoder illumination fusion module EIFB1 is connected to the downsampling operator D. 21 Another branch uses element-wise addition operations with the convolutional layer C. 22 The input terminal is connected; the downsampling operator D 21 The output is split into two branches. The first branch is connected to the residual channel attention network RCAN1, and the other branch is connected to the upsampling layer U through element-wise addition. 22 The input terminal is connected; the downsampling operator D 22 The output is split into two branches. The first branch is connected to the residual channel attention network RCAN2, and the other branch is connected to the upsampling layer U through element-wise addition. 22 connect; The color space correction module SCCM includes a color space conversion operator R2, a color residual prediction network, and a clipping operator Clip. The output of the color space conversion operator R2 is split into three branches, namely the first branch to the third branch. The first branch is connected in series with the color residual prediction network. The second branch is connected to the output of the color residual prediction network through an element-wise addition operation to form the fourth branch. The third branch and the fourth branch are connected through a channel-dimensional splicing operation.

4. The method for super-resolution reconstruction of low-light images based on illumination priors according to claim 3, characterized in that, Both the first illumination feature extraction U-shaped network IFU and the second illumination feature extraction U-shaped network IFU include a convolutional block CB1 and a downsampling operator D connected in sequence. 11 Convolutional block CB2, downsampling operator D 12 Convolutional block CB3, upsampling layer U 11 Convolutional block CB4, upsampling layer U 12 Convolutional block CB5, convolutional layer C 14 ; The output of convolutional block CB1 is split into two branches. The first branch is connected to the downsampling operator D. 11 Connect, another branch with the upsampling layer U 12 The output features are concatenated in the channel dimension; the output of convolutional block CB2 is split into two branches, the first branch being connected to the downsampling operator D. 12 Connect, another branch with the upsampling layer U 11 The output features are concatenated in the channel dimension to form skip connections.

5. The method for super-resolution reconstruction of low-light images based on illumination priors according to claim 4, characterized in that, S3 specifically includes the following steps: S31. Upsample the training samples using the upsampling operator to obtain the upsampled image. ; S32, Upsample the image The input to the dual-prior guided network DPGN first passes through the color space conversion operator R1 to upsample the image. Converting from RGB color space to YCbCr space yields three-channel YCbCr features. Then, the three-channel YCbCr features were analyzed. The data is split to obtain the brightness features of one channel. and two-channel chromaticity characteristics ; S33, convert one channel brightness component and two-channel chromaticity components The inputs are respectively fed into the first illumination feature extraction U-shaped network IFU and the second illumination feature extraction U-shaped network IFU, and illumination-related features are extracted respectively. and noise / chromaticity perturbation related features ; S34. Illumination-related features and noise / chromaticity perturbation related features The inputs are fed into two convolutional layers C respectively. 11 In the process, the features are compressed into single-channel features, and the luminance prior activation pre-mapped value Ly and the chrominance perturbation activation pre-mapped value Lc are obtained respectively. S35. The luminance prior activation pre-mapped value Ly and the chromaticity perturbation activation pre-mapped value Lc are subtracted element-wise to obtain the corrected illumination activation pre-mapped value Lr. S36. Input the pre-activation mapping value Lr of the corrected illumination and the pre-activation mapping value Lc of the chromaticity perturbation into the activation function S respectively. 11 In the process, the illumination prior map is obtained. Noise Prior Diagram ; S37. Illumination prior diagram The image is input into the projection module, which then projects the illumination prior image. Mapping back to multi-channel illumination-guided prior features .

6. The method for super-resolution reconstruction of low-light images based on illumination priors according to claim 5, characterized in that, The upsampled image in S31 The expression is as follows: ； in, This indicates an upsampling operator using bicubic interpolation; This indicates that the upsampling factor is n; The three-channel YCbCr feature in S32 The expression is as follows: ； in, Let be the linear transformation matrix from RGB to YCbCr. It is the bias vector; This represents the convolutional layer within the color space conversion operator R1; The illumination-related features in S33 and noise / chromaticity perturbation related features The expressions are as follows: ； ； in, This represents the brightness branch, i.e., the first illumination feature extraction U-shaped network (IFU). This represents the chromaticity branch, i.e., the second illumination feature extraction U-shaped network (IFU). The expressions for the luminance prior activation pre-mapped value Ly and the chrominance perturbation activation pre-mapped value Lc in S34 are as follows: ； ； in, Represents the convolutional layer C 11 ; The mapping value before illumination activation in S35 is corrected. The expression is as follows: ； in, Represents the learnable coefficient; The illumination prior map in S36 Noise Prior Diagram The expressions are as follows: ； ； in, S represents the activation function. 11 Specifically, the Sigmoid function; The multi-channel illumination-guided prior features in S37 The expression is as follows: ； in, Represents the convolutional layer C 12 ; Represents the activation function R 11 ; Represents the convolutional layer C 13 ; S represents the activation function. 12 Specifically, this refers to the Sigmoid function.

7. The method for super-resolution reconstruction of low-light images based on illumination priors according to claim 6, characterized in that, S4 specifically includes the following steps: S41. Guiding prior features with multi-channel illumination The input is first fed into the super-resolution reconstruction module, where it is then used to construct a three-scale pyramid feature using bilinear interpolation. These correspond to the backbone network. , Compared to the original scale; features of the three-scale pyramid The expressions are as follows: ； ； ； in, This indicates that bilinear interpolation is used to scale the feature map. Scaling is applied to the feature map by a ratio of s=0.25 or 0.5; , , They are respectively , Prior lighting features at the original scale; S42, Upsample the image The input is first processed by the mean-variance normalization operator and then by the convolutional layer C in the super-resolution reconstruction module. 21 Obtain input features Then, at the original scale, the input features are... Prior features of illumination The original scale-corrected features are obtained by fusion. Input features and original scale correction features The expressions are as follows: ； in, This represents the mean-variance normalization operator. Represents the convolutional layer C 21 ; ； S43. Correct the original scale features Input to downsampling operator D 21 Perform downsampling to obtain encoded features Next, the encoded features will be... The input is fed into the residual channel attention network RCAN1 to obtain the encoded features. Then At scale, the encoded features will be and prior features of illumination The input is fed into the encoder illumination fusion module EIFB2 for fusion to obtain 1 / 2 scale correction features. The expressions are as follows: ； ； ； in, The downsampling operator D is represented. 21 ; S44, Correct the 1 / 2 scale feature Input to downsampling operator D 22 Downsampling is performed to obtain encoded features. Encode features The input is fed into the residual channel attention network RCAN2 to obtain the encoded features. Then At scale, the encoded features will be Prior features of illumination The input is fed into the encoder illumination fusion module EIFB3 for fusion to obtain 1 / 4 scale correction features. The expressions are as follows: ； ； ； in, The downsampling operator D is represented. 22 ; S45. Encode features and 1 / 4 scale correction features The fused input is then fed into the upsampling layer U. 21 Upsampling is performed to obtain the first decoded feature. Then the first decoded feature The input is fed into the residual channel attention network RCAN3 for feature enhancement, resulting in the first enhanced decoding feature. Next, the first enhanced decoding feature will be... With coding features The fused input is then fed into the upsampling layer U. 22 In the process, spatial resolution is restored to obtain the second decoding feature. Then the second decoding feature The input is fed into the residual channel attention network RCAN4 for feature enhancement, resulting in the second enhanced decoding feature. Then the second enhanced decoding feature and original scale correction features Input to convolutional layer C 22 In the process, the super-resolution reconstruction results are obtained. The expressions are as follows: ； ； ； ； ； in, Indicates upsampling layer U 21 ; Indicates upsampling layer U 22 ; Represents the convolutional layer C 22 .

8. The method for super-resolution reconstruction of low-light images based on illumination priors according to claim 7, characterized in that, The expression for the total loss function in S6 is as follows: ； in, Total loss; This represents the loss in super-resolution primary reconstruction. To estimate the loss weighting coefficients a priori; To estimate the loss using two priors; For chromaticity correction constraint loss; Super-resolution primary reconstruction loss The expression is as follows: ； in, For L1 loss, In order to perceive loss, For frequency domain loss; These are the corresponding weighting coefficients; Two prior estimation loss The expression is as follows: in, and For the corresponding monitoring signal; The total variation regularization weight, Weights for noise priors; Represents the L1 norm; This represents the regularization term for the total variation; Colorimetric correction constraint loss The expression is as follows: ； in, This indicates that the chroma channels are extracted after the image is converted to the YCbCr color space; Weights for chromaticity correction loss; The weights are for pixel auxiliary terms.

9. A low-light image super-resolution reconstruction system incorporating illumination priors, characterized in that, Configured to perform a low-light image super-resolution reconstruction method based on illumination priors as described in any one of claims 1 to 8, comprising: The building unit is configured to build a low-light image super-resolution reconstruction network, which includes a dual priori guided network DPGN, a super-resolution reconstruction module, and a color correction module SCCM. The prior extraction unit is configured to use the dual prior guidance network DPGN to obtain multi-channel illumination guidance prior features, illumination prior map, and noise prior map. The super-resolution reconstruction unit is configured to use the super-resolution reconstruction module to obtain super-resolution reconstruction results. The colorimetric correction unit is configured to obtain the final result using the colorimetric correction module SCCM.

Images