Low-light image enhancement method based on pixel point RGB value imbalance feature guidance

Abstract

The application discloses a low-light image enhancement method based on pixel point RGB value imbalance feature guidance, and belongs to the low-light image enhancement field.The method comprises the following steps: acquiring a low-light image; calculating an RGB value imbalance point graph of the low-light image; constructing a Transform-Unet low-light image enhancement network based on RGB value imbalance point guidance; inputting the low-light image and the RGB value imbalance point graph into the Transform-Unet low-light image enhancement network based on RGB value imbalance point guidance; and outputting an enhanced image by the Transform-Unet low-light image enhancement network based on RGB value imbalance point guidance.Under the principle of traditional self-attention, the RGB value imbalance feature is used to construct an imbalance point guided Transform.The application guides the learning of attention in the Transform by using the pixel point RGB value imbalance characteristics in the image, improves the poor learning condition of the traditional model on the information distortion serious area in the low-light image, and can significantly improve the visual effect of the enhanced image.

Description

Technical Field

[0001] This invention relates to the field of low-light image enhancement, specifically a low-light image enhancement method guided by the RGB value imbalance feature of pixels. Background Technology

[0002] Images captured in low-light environments often exhibit uncertain degradation factors, such as low visibility, color degradation, and low contrast, due to insufficient light, short exposure times, or hardware limitations of the camera's CMOS sensor. These low-quality images not only make it difficult for humans to extract information but also pose significant challenges to subsequent advanced computer vision tasks such as object detection, object recognition, and scene understanding. Therefore, improving the perceptual quality of low-light images is of great importance.

[0003] Current low-light image enhancement methods often focus on the visual characteristics of low-light images, such as increasing brightness or suppressing noise. However, few methods consider designing enhancement methods based on the numerical characteristics of low-light images. For RGB images, each pixel has values ​​in three channels: Red (R), Green (G), and Blue (B). In a properly exposed image, the R, G, and B channel values ​​should be balanced and within a suitable range. Low-light conditions often cause image degradation due to uncertainty, leading to significant deviations in RGB values. For example, in extremely dark areas, the R, G, or B channel values ​​may become 0, resulting in information loss, or large differences in R, G, or B channel values ​​may cause information or color distortion. Therefore, we define pixels with significant differences in R, G, and B channel values ​​as imbalance points. We mark these RGB imbalanced pixels in low-light images to form a binary imbalance point map, which reflects the distortion of information at each pixel in the image.

[0004] The differences compared to existing technologies are as follows:

[0005] In comparison with the technology of patent CN202110142918.X, "A method and system for enhancing low-light images based on attention mechanism",

[0006] Compared to patent CN202110142918.X, which employs retinal decomposition for low-light image enhancement, this patent designs a decomposition network and a fusion network. Specifically, the decomposition network decomposes the low-light image into illumination and reflection components, and then trains the fusion network to combine the illumination and reflection components to obtain the enhanced image. This patent, however, does not use retinal decomposition but instead employs an end-to-end training method. The network's input and output are both complete images, avoiding information loss caused by image decomposition and component fusion. Furthermore, this patent's design, based on the RGB value imbalance map of image pixels, can accurately locate points in the image that verify information distortion, thereby enabling accurate information repair and enhancement.

[0007] In comparison with the technology of patent CN2022105102025 "A method and system for enhancing low-light images based on attention mechanism"

[0008] Compared to patent CN2022105102025, which uses cross-convolution to extract local detail information and employs gating and self-attention mechanisms to extract distant pixel information in the image to obtain global information, this method enhances distorted areas entirely through network learning, thus failing to accurately locate and enhance these distorted areas. This patent uses a Transformer structure for local feature extraction. Furthermore, the patent's design based on the RGB value imbalance map of image pixels can accurately locate points in the image that verify information distortion, thereby accurately guiding the Transformer structure to effectively repair and enhance these distorted areas.

[0009] The low-light image enhancement method designed in this invention, guided by the RGB value imbalance feature of pixels, considers using the low-light image and its RGB value imbalance point map as input to the enhancement network to guide the network to learn and train, thereby improving the enhancement network's ability to repair information distortion caused by low-light environments. Summary of the Invention

[0010] To address the aforementioned technical problems, this invention proposes a low-light image enhancement method guided by the RGB value imbalance feature of pixels, which enhances the ability of neural networks to recover image information distortion caused by low-light conditions, while improving the visual quality of the enhanced image.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0012] A low-light image enhancement method guided by pixel RGB value imbalance features includes the following steps:

[0013] Step S1: Data preprocessing;

[0014] Acquire a low-light image and obtain its RGB value imbalance point map;

[0015] Step S2: Construct a Transformer-Unet low-light image enhancement network guided by RGB value imbalance points, and complete network training;

[0016] The Transformer-Unet low-light image enhancement network based on RGB value imbalance points includes: 4 local-global feature encoding modules, 4 local-global feature decoding modules, a downsampling module, and an upsampling module;

[0017] Step S3: Identify the imbalance points between the image to be enhanced in step S1 and its RGB values. Figure 1 In the input step 2, the Transformer-Unet low-light image enhancement network guided by RGB value imbalance points outputs the enhanced image.

[0018] Preferably, in step S2, the four local-global feature encoding modules and the four local-global feature decoding modules have the same local-global feature extraction structure, which is composed of a local Transformer block guided by RGB value imbalance points and a global Transformer block based on self-attention connected in parallel.

[0019] Preferably, the four local-global feature encoding modules include: local-global feature encoding module DF1, local-global feature encoding module DF2, local-global feature encoding module DF3 and local-global feature encoding module DF4;

[0020] The four local-global feature decoding modules include: local-global feature decoding module UF1, local-global feature decoding module UF2, local-global feature decoding module UF3 and local-global feature decoding module UF4;

[0021] The downsampling module includes downsampling module D1, downsampling module D2 and downsampling module D3;

[0022] The upsampling modules include upsampling module U1, upsampling module U2, and upsampling module U3.

[0023] Preferably, the local Transformer block guided by RGB value imbalance points is composed of an attention network A1 and a feedforward network N1 connected in series; the global Transformer block based on self-attention is composed of a global self-attention network A2 and a feedforward network N2 connected in series.

[0024] The attention network A1, guided by RGB value imbalance points, comprises three branches, whose inputs are q. sp k sp and v sp , where q sp and k sp Derived from the RGB value imbalance point map features, v sp The feature map is derived from the image, and the input is q. sp The branch consists of a depthwise convolution DW1 with a kernel size of 3×3, and the input is k. sp The branch consists of a depthwise convolution (DW2) with a kernel size of 3×3, and the input is v. sp The branch consists of a normalized layer LN1 and a cascaded depthwise convolution DW3 with a kernel size of 3×3, q sp k sp and v sp After the three branches perform attention calculations guided by the imbalance point of RGB values, their outputs will be used as the input of the linear layer Linear1. The feedforward network N1 consists of a normalization layer LN2, a linear layer Linear2, an activation function R1, and a linear layer Linear3.

[0025] The global self-attention network A2 includes three branches, whose inputs are q respectively. sl k sl and v sl Both are derived from the feature maps of the image, with q as the input. sl The branch consists of a normalized layer LN3 and a depthwise convolution DW4 with a kernel size of 3×3, with input k. sl The branch consists of a normalized layer LN4 and a depthwise convolution DW5 with a kernel size of 3×3, connected in series, with the input being v. sl The branch consists of a normalized layer LN5 and a depthwise convolution DW6 with a kernel size of 3×3 connected in series. The three branches perform attention calculations, and their outputs will be used as the input of the linear layer Linear4. The feedforward network N2 consists of a normalized layer LN6, a linear layer Linear5, an activation function R2, and a linear layer Linear6.

[0026] Preferably, the downsampling sub-block D1 is composed of a 3×3 convolution kernel DConv1 and a bilinear interpolation DB1 with a scaling factor of 2 connected in series; the downsampling sub-block D2 is composed of a 3×3 convolution kernel DConv2 and a bilinear interpolation DB2 with a scaling factor of 2 connected in series; and the downsampling sub-block D3 is composed of a 3×3 convolution kernel DConv3 and a bilinear interpolation DB3 with a scaling factor of 2 connected in series.

[0027] The upsampling sub-block U1 is composed of a 3×3 convolution kernel UConv1 and a bilinear interpolation UB1 with a magnification factor of 2 connected in series. The upsampling sub-block U21 is composed of a 3×3 convolution kernel UConv2 and a bilinear interpolation UB2 with a magnification factor of 2 connected in series. The upsampling sub-block U3 is composed of a 3×3 convolution kernel UConv3 and a bilinear interpolation UB3 with a magnification factor of 2 connected in series.

[0028] Preferably, the training process of the Transformer-Unet low-light image enhancement network guided by RGB value imbalance points in step S2 is as follows:

[0029] Using an existing image dataset as the training set, during the data loading phase, the low-light images to be enhanced and the normally exposed images used as labels are subjected to the same random cropping, flipping, and rotation to obtain N×N image patches for network training. During the model training phase, the RGB value imbalance map of the low-light images is obtained, and the low-light images and their RGB value imbalance maps are used as inputs to the Transformer-Unet low-light image enhancement network guided by RGB value imbalance to form training samples. The loss function and training parameters are set, and the Transformer-Unet low-light image enhancement network guided by RGB value imbalance is trained using a supervised learning method.

[0030] Preferably, the loss function is:

[0031]

[0032] Where L(x, y) represents the loss function, GT(x, y) represents the label of the normally exposed image, E(x, y) represents the output image generated by the Transformer-Unet low-light image enhancement network guided by the RGB value imbalance point, (x, y) represents the pixel coordinates, and N is the cropped image size.

[0033] Compared with the prior art, the present invention has the following improvements and advantages:

[0034] 1. This invention defines RGB value imbalance points based on the numerical characteristics exhibited by low-light images. These imbalance points reflect the information distortion of different pixels in low-light images. Replacing the self-attention mechanism in the original Transformer with these imbalance points enables the network to accurately recover information from image distortion points.

[0035] 2. This invention constructs a parallel block of a local Transformer guided by RGB value imbalance points and a global Transformer based on self-attention. The local Transformer block guided by RGB value imbalance points can accurately restore distorted information points, while the global Transformer block based on self-attention can compensate for global information, achieving overall image enhancement. By combining local and global features, the global consistency of the image and the clarity of local details can be enhanced. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the low-light image enhancement method based on pixel RGB value imbalance in this invention.

[0037] Figure 2 This is a schematic diagram of the Transformer-Unet low-light image enhancement network guided by RGB value imbalance points in this invention.

[0038] Figure 3 This is a schematic diagram of the local-global feature encoding / decoding module in this invention;

[0039] Figure 4 This is a schematic diagram of the downsampling module in this invention;

[0040] Figure 5 This is a schematic diagram of the modules used in this invention;

[0041] Figure 6 This is a rendering of the output image in this invention. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0043] like Figure 1 As shown, the low-light image enhancement method based on the RGB value imbalance feature of pixels specifically includes the following steps:

[0044] Step S1: Data preprocessing;

[0045] Obtain a low-light image and extract its RGB value imbalance map. The process of obtaining the imbalance map can be represented by the following formula:

[0046]

[0047] Where (x,y) represents the pixel at position (x,y), T represents the threshold for detecting whether P(x,y) is an imbalance point, M(x,y) represents the value assigned to the imbalance point at position (x,y), which is 0 or 1, and P(x,y) is the minimum ratio of the R, G, and B channels of the pixel at position (x,y) to the R+B+G channels, respectively. This can be expressed by the following formula:

[0048]

[0049] Where min represents the smallest ratio among these three ratios, and sum(x,y) represents the sum of the pixel values ​​of the R, G, and B channels at pixel (x,y), which can be expressed by the following formula:

[0050] sum(x,y)=R(x,y)+G(x,y)+B(x,y) (3)

[0051] Where R(x,y), G(x,y) and B(x,y) represent the pixel values ​​of the R, G, and B channels at pixel (x,y), respectively.

[0052] In formula (2), we calculated the ratio of each of the R, G, and B channels to the total three channels, and obtained the smallest ratio among them. The smaller the ratio, the greater the difference in R, G, and B values ​​at that point, indicating more severe information distortion at that point. We used the threshold T in formula (1) to detect whether the minimum ratio in formula (2) satisfies the condition. If it does, it is marked as an imbalance point, such that M(x,y) = 1; otherwise, it is 0. Each pixel in the low-light image was marked using formula (1), resulting in an imbalance point map, which reflects which pixels in the image have information distortion. The low-light image and the imbalance point map are used as inputs to the augmentation network to guide the network in learning and restoring information-distorted regions.

[0053] In this embodiment, the size of the randomly cropped low-light image is (3, 256, 256), that is, the number of channels is 3, the width and height are 256 and 256 respectively. The obtained imbalance point map has 1 channel, and the width and height are the same as the low-light image, that is, the size is (1, 256, 256).

[0054] Step S2: Construct a Transformer-Unet low-light image enhancement network guided by RGB value imbalance points, and complete network training;

[0055] The Transformer-Unet low-light image enhancement network guided by RGB value imbalance points, such as... Figure 2As shown, it includes: 4 local-global feature encoding modules, 4 local-global feature decoding modules, 3 downsampling modules and 3 upsampling modules.

[0056] The input size of the local-global feature encoding module DF1 is (3, 256, 256), and the output is (48, 256, 256); the input size of the downsampling module D1 is (48, 256, 256), and the output is (96, 128, 128); the input size of the local-global feature encoding module DF2 is (96, 128, 128), and the output is (96, 128, 128); the input size of the downsampling module D2 is (96, 128, 128). The input size of the local-global feature encoding module DF3 is (192, 64, 64), and the output is (192, 64, 64). The input size of the downsampling module D3 is (96, 128, 128), and the output is (384, 32, 32). The input size of the local-global feature encoding module DF4 is (384, 32, 32), and the output is (384, 32, 32). The input dimensions of the local-global feature decoding module UF1 are (384, 32, 32), and its output is (384, 32, 32); the input dimensions of the upsampling module U1 are (384, 32, 32), and its output is (192, 64, 64); the input dimensions of the local-global feature decoding module UF2 are (384, 64, 64), and its output is (192, 64, 64); the input dimensions of the upsampling module U2 are (192, 64, 64). The output is (96, 128, 128); the input size of the local-global feature decoding module UF3 is (192, 128, 128), and the output is (192, 128, 128); the input size of the upsampling module U3 is (192, 128, 128), and the output is (96, 256, 256); the input size of the local-global feature decoding module UF4 is (144, 256, 256), and the output is (3, 256, 256).

[0057] The four local-global feature encoding modules and four local-global feature decoding modules have the same local-global feature extraction structure, as shown in the example below. Figure 3As shown, this structure consists of a local Transformer block guided by RGB value imbalance points and a global Transformer block based on self-attention, connected in parallel. First, the local Transformer block on the left, guided by RGB value imbalance points, is part of the attention network A1. The depthwise convolutions DW1 and DW2 have kernel sizes of 3×3 and a stride of 1. Their outputs are multiplied to obtain different weights for different pixels. This weight matrix is ​​then normalized to the range of 0 to 1 after Softmax processing. Pixels with more severe information distortion have weight values ​​closer to 0, while pixels with more intact information have weight values ​​closer to 1. Image features in the attention network A1 first pass through a normalization layer LN1, followed by a depthwise convolution DW3 with a kernel size of 3×3 and a stride of 1. The output of DW3 and the weight matrix from the Softmax output are multiplied to assign different weights to different pixels in the image. The linear layer Linear1 is responsible for transforming the dimensionality of the output features. Next is the self-attention-based global Transformer block on the right. In the attention network A2, image features pass through three normalization layers LN3, LN4, and LN5, followed by depthwise convolutions DW4, DW5, and DW6 with a kernel size of 3×3 and a stride of 1. The outputs of DW4 and DW5 are multiplied to obtain different weights based on the different global features of the image. This weight matrix is ​​then normalized to the range of 0 to 1 after Softmax processing. This weight matrix is ​​multiplied with the output of DW6 to adaptively assign different weights to different pixels based on different image features, thus achieving global feature compensation. The linear layer Linear4 is responsible for transforming the dimensionality of the output features. Finally, there are feedforward networks N1 and N2 connecting the attention networks A1 and A2. Feedforward networks N1 and N2 have the same network structure, consisting of a normalization layer, a linear layer, an activation function, and another linear layer connected in sequence. The feedforward network aims to perform fully connected transformations to enhance feature extraction. The outputs of the two feedforward networks are eventually connected in parallel through element-wise addition of matrices. In this way, the local-global feature encoding / decoding module can simultaneously combine local features guided by imbalance points and global features based on self-attention, achieving a combination of local details and global consistency.

[0058] downsampling module such as Figure 4As shown, the downsampling module consists of a convolutional layer and a bilinear interpolator connected sequentially. The convolutional layer has a kernel size of 3×3 and a stride of 1, which is mainly used to change the number of feature channels before and after the input and output. The bilinear interpolation has a reduction factor of 2, which can halve the image size on the scale H×W. That is, a (C, H, W) feature map will become (2×C, H / 2, W / 2) after passing through a downsampling module.

[0059] Upsampling module, such as Figure 5 As shown, the upsampling module consists of a convolutional layer and a bilinear interpolator connected in sequence. The convolutional layer has a kernel size of 3×3 and a stride of 1, which is mainly used to change the number of feature channels before and after the input and output. The bilinear interpolation has an amplification factor of 2, which can realize the expansion on the image size scale H×W. That is, a feature map of (C, H, W) will become (C / 2, 2×H, 2×W) after passing through an upsampling module.

[0060] Furthermore, the training process of the Transformer-Unet low-light image enhancement network guided by RGB value imbalance points is as follows:

[0061] Using an existing image dataset as the training set, in the data preprocessing stage, the low-light images to be enhanced and the normally exposed images used as labels are subjected to the same random cropping, flipping, and rotation to obtain N×N image patches for network training. Simultaneously, their RGB value imbalance maps are obtained, and the low-light images and their imbalance maps are then processed. Figure 1 As the input to the network, in this embodiment, the cropped image size N is 256.

[0062] The loss function and training parameters are set, and supervised batch learning is used to train the Transformer-Unet low-light image enhancement network guided by RGB value imbalance points. The loss function is:

[0063]

[0064] Where L(x,y) represents the loss function, GT(x,y) represents the label of the normally exposed image, E(x,y) represents the output image generated by the Transformer-Unet low-light image enhancement network guided by RGB value imbalance, (x,y) represents the pixel coordinates, and N is the cropped image size. This loss function represents the "gap" between the enhanced image output by the Transformer-Unet low-light image enhancement network guided by RGB value imbalance and the normally exposed reference image, and continuously reduces this "gap" during training.

[0065] Step S3: Identify the imbalance points between the image to be enhanced in step S1 and its RGB values. Figure 1 In the input step 2, the Transformer-Unet low-light image enhancement network guided by RGB value imbalance points outputs the enhanced image.

[0066] The present invention has the following advantages:

[0067] 1. This invention defines RGB value imbalance points based on the numerical characteristics exhibited by low-light images. These imbalance points reflect the information distortion of different pixels in low-light images. Replacing the self-attention mechanism in the original Transformer with these imbalance points enables the network to accurately recover information from image distortion points.

[0068] 2. This invention constructs a parallel block of a local Transformer guided by RGB value imbalance points and a global Transformer based on self-attention. The local Transformer block guided by RGB value imbalance points can accurately restore distorted information points, while the global Transformer block based on self-attention can compensate for global information, achieving overall image enhancement. By combining local and global features, the global consistency of the image and the clarity of local details can be enhanced.

[0069] Enhanced image illustration as shown Figure 6 As shown, the enhanced image has appropriate lighting intensity, normal colors, and rich details, and the severe loss of image information caused by the low-light environment has been significantly restored.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A low-light image enhancement method guided by pixel RGB value imbalance features, characterized in that, Includes the following steps: Step S1: Data preprocessing; Acquire a low-light image and obtain its RGB value imbalance point map; The process of obtaining the RGB value imbalance point map is represented by the following formula: (1) Where (x,y) represents the pixel at position (x,y), T represents the threshold for detecting whether P(x,y) is an imbalance point, M(x,y) represents the value assigned to the imbalance point at position (x,y), which is 0 or 1, and P(x,y) is the minimum ratio of the R, G, and B channels of the pixel at position (x,y) to the R+B+G channels, respectively. This can be expressed by the following formula: (2) Where min represents the smallest ratio among these three ratios, and sum(x,y) represents the sum of the pixel values ​​of the R, G, and B channels at pixel (x,y), which can be expressed by the following formula: (3) Where R(x,y), G(x,y) and B(x,y) represent the pixel values ​​of the R, G and B channels at the (x,y) pixel point, respectively; Step S2: Construct a Transformer-Unet low-light image enhancement network guided by RGB value imbalance points, and complete network training; The Transformer-Unet low-light image enhancement network based on RGB value imbalance points includes: 4 local-global feature encoding modules, 4 local-global feature decoding modules, a downsampling module, and an upsampling module; Step S3: Input the low-light image from step S1 and its RGB value imbalance point map into the Transformer-Unet low-light image enhancement network guided by RGB value imbalance points in step 2. The Transformer-Unet low-light image enhancement network guided by RGB value imbalance points outputs the enhanced image.

2. The low-light image enhancement method based on pixel RGB value imbalance features as described in claim 1, characterized in that: In step S2, the four local-global feature encoding modules and the four local-global feature decoding modules have the same local-global feature extraction structure, which is composed of a local Transformer block guided by RGB value imbalance points and a global Transformer block based on self-attention connected in parallel.

3. The low-light image enhancement method based on pixel RGB value imbalance features as described in claim 2, characterized in that: The four local-global feature encoding modules include: local-global feature encoding module DF1, local-global feature encoding module DF2, local-global feature encoding module DF3, and local-global feature encoding module DF4; The four local-global feature decoding modules include: local-global feature decoding module UF1, local-global feature decoding module UF2, local-global feature decoding module UF3 and local-global feature decoding module UF4; The downsampling module includes downsampling module D1, downsampling module D2 and downsampling module D3; The upsampling module includes upsampling module U1, upsampling module U2 and upsampling module U3.

4. The low-light image enhancement method based on pixel RGB value imbalance features as described in claim 2, characterized in that: The local Transformer block guided by RGB value imbalance points is composed of an attention network A1 and a feedforward network N1 connected in series; the global Transformer block based on self-attention is composed of a global self-attention network A2 and a feedforward network N2 connected in series. The attention network A1, guided by RGB value imbalance points, includes three branches, whose inputs are respectively... , and ,in and Derived from the characteristics of the RGB value imbalance point map, Feature maps derived from images, with input as The branch consists of a depthwise convolution DW1 with a kernel size of 3×3, and the input is... The branch consists of a depthwise convolution DW2 with a kernel size of 3×3, and the input is... The branch consists of a normalized layer LN1 and a cascaded depthwise convolution DW3 with a kernel size of 3×3. , and After the three branches perform attention calculations guided by the imbalance point of RGB values, their outputs will be used as the input of the linear layer Linear1. The feedforward network N1 consists of a normalization layer LN2, a linear layer Linear2, an activation function R1, and a linear layer Linear3. The global self-attention network A2 includes three branches, whose inputs are respectively , and Both are derived from feature maps of an image, and the input is... The branch consists of a normalized layer LN3 and a depthwise convolution DW4 with a kernel size of 3×3, connected in series. The input is... The branch consists of a normalized layer LN4 and a depthwise convolution DW5 with a kernel size of 3×3, connected in series. The input is... The branch consists of a normalized layer LN5 and a depthwise convolution DW6 with a kernel size of 3×3 connected in series. The three branches perform attention calculations, and their outputs will be used as the input of the linear layer Linear4. The feedforward network N2 consists of a normalized layer LN6, a linear layer Linear5, an activation function R2, and a linear layer Linear6.

5. The low-light image enhancement method based on pixel RGB value imbalance features as described in claim 3, characterized in that: The downsampling module D1 is composed of a 3×3 convolution kernel DConv1 and a bilinear interpolation DB1 with a scaling factor of 2 connected in series. The downsampling module D2 is composed of a 3×3 convolution kernel DConv2 and a bilinear interpolation DB2 with a scaling factor of 2 connected in series. The downsampling module D3 is composed of a 3×3 convolution kernel DConv3 and a bilinear interpolation DB3 with a scaling factor of 2 connected in series. The upsampling module U1 is composed of a 3×3 convolution kernel UConv1 and a bilinear interpolation module UB1 with a magnification factor of 2 connected in series. The upsampling module U2 is composed of a 3×3 convolution kernel UConv2 and a bilinear interpolation module UB2 with a magnification factor of 2 connected in series. The upsampling module U3 is composed of a 3×3 convolution kernel UConv3 and a bilinear interpolation module UB3 with a magnification factor of 2 connected in series.

6. The low-light image enhancement method based on pixel RGB value imbalance features as described in claim 1, characterized in that: The training process of the Transformer-Unet low-light image enhancement network guided by RGB value imbalance points in step S2 is as follows: Using an existing image dataset as the training set, during the data loading phase, the low-light images to be enhanced and the normally exposed images used as labels are subjected to the same random cropping, flipping, and rotation to obtain N×N image patches for network training. During the model training phase, the RGB value imbalance map of the low-light images is obtained, and the low-light images and their RGB value imbalance maps are used as inputs to the Transformer-Unet low-light image enhancement network guided by RGB value imbalance to form training samples. The loss function and training parameters are set, and the Transformer-Unet low-light image enhancement network guided by RGB value imbalance is trained using a supervised learning method.

7. The low-light image enhancement method based on pixel RGB value imbalance features as described in claim 6, characterized in that: The loss function is: ； Where L(x,y) represents the loss function, GT(x,y) represents the label of the normally exposed image, E(x,y) represents the output image generated by the Transformer-Unet low-light image enhancement network guided by the RGB value imbalance point, (x,y) represents the pixel coordinates, and N is the cropped image size.

Images