A Low-Light Image Enhancement Method and System Based on Illumination Decomposition

Abstract

This invention discloses a low-light image enhancement method and system based on illumination decomposition, belonging to the fields of artificial intelligence and computer vision technology. The method constructs an illumination-decomposable low-light enhancement model based on coordinate attention and an unsupervised denoising enhancement model. The illumination-decomposable low-light enhancement model based on coordinate attention performs initial enhancement on the low-light image. Then, the enhanced image and the original low-light image are simultaneously input into the unsupervised denoising enhancement model for further enhancement and denoising. The unsupervised denoising enhancement model improves the network's generalization and noise resistance, better handling noise from various environments and enhancing the network's performance under complex conditions, thereby further improving the performance of this invention in low-light image enhancement tasks.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and computer vision technology, specifically relating to a low-light image enhancement method and system based on illumination decomposition. Background Technology

[0002] Low-light images are typically captured or photographed under conditions of insufficient illumination. Insufficient lighting significantly degrades the visual quality of images, resulting in loss of detail, low contrast, and increased noise. This not only affects the subjective experience of viewing the image, but also, since existing vision systems are often designed for images captured under normal lighting, this can negatively impact the performance of many computer vision tasks. Factors contributing to the low quality of low-light images include insufficient ambient light, poor light sensitivity of low-end imaging equipment, and inappropriate exposure settings by the photographing device. To address the issues of restoring image details and improving image quality, researchers have proposed various low-light image enhancement and denoising methods.

[0003] In recent years, low-light enhancement techniques based on deep learning have attracted much attention. Deep learning-based methods have better performance than traditional low-light image enhancement methods, and have good generalization capabilities, adapting to various scenarios. The RetinexNet model consists of a decomposition network (Decom) for decomposition and an enhancement network for illumination adjustment (Relight). First, the decomposition network decomposes the low-light image into a reflectance map and an illuminance map; then, the enhancement network enhances and adjusts the image. However, the enhanced image lacks detail, and the network model is large, resulting in significant computational time consumption. The KinD model is similar to Retinex-Net, decomposing the image into illumination and reflectance components, and using convolutional neural networks to process the illumination and reflectance components respectively, improving the problems encountered in RetinexNet, but its model size is much larger than RetinexNet. The SSIENet model decomposes the image based on Retinex theory, using the result of histogram equalization as a reference, assuming that the maximum reflectance channel is consistent with the maximum entropy of the low-light image. Based on this, a simple network can achieve good results, but the enhanced image suffers from color difference and loss of detail. Inspired by Retinex, Liu et al. first established the enhancement network by expanding the optimization process of the Retinex-inspired model; then, they introduced a distillation-based prior module search space; finally, they proposed a cooperative two-layer search strategy to discover the required illumination estimation and noise removal architecture. This network requires a very small model size, floating-point computations, and time. Although image details are enhanced, visible noise is introduced, degrading image quality. The large number of parameters and the presence of noise in the enhanced image in these low-light image enhancement models make them difficult to directly apply to mobile devices or provide assurance for backend tasks. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a low-light image enhancement method and system based on illumination decomposition, which solves the problem that existing low-light image enhancement methods have large models and many parameters, and cannot be directly used for mobile devices or backend tasks.

[0005] This invention is achieved through the following technical solution:

[0006] A low-light image enhancement method based on illumination decomposition includes the following steps:

[0007] Step 1: Construct a weak light enhancement model based on coordinate attention, which includes an illumination decomposition and learning module and a reflection component self-calibration module;

[0008] The illumination decomposition learning module embeds a coordinate attention mechanism in the feature extraction layer. The feature extraction layer obtains the feature map of the original low-light image, and the coordinate attention mechanism is used to embed the position information of the original image into the feature map to obtain the illumination map of the original low-light image. The original low-light image is divided by the illumination map to obtain the reflection map of the original low-light image.

[0009] The reflection component self-calibration module acquires the reflection component of the reflection map, combines the reflection component with the original low-light image, and obtains an enhanced low-light image.

[0010] The enhanced low-light image is input into the illumination decomposition learning module, and the reflection component self-calibration module is repeatedly executed until the number of iterations is reached to obtain the reflection map of the preliminary enhanced image.

[0011] Step 2: Construct an unsupervised denoising and enhancement model;

[0012] The unsupervised denoising enhancement model uses upsampling and downsampling to denoise the final reflection map output by the illumination decomposition learning module, and performs skip connections between feature maps obtained from different sampling layers to fuse the feature maps from different sampling layers to obtain the final enhanced image.

[0013] Preferably, the method for extracting the illumination map of the original low-light image in step 1 is as follows:

[0014] The illumination decomposition learning module comprises at least four convolutional layers;

[0015] The first convolutional layer contains at least one 2D convolution, a ReLU activation function, and coordinate attention;

[0016] The second and third convolutional layers have the same structure, including at least one two-dimensional convolution, normalization, and ReLU activation function;

[0017] The fourth convolutional layer contains at least one two-dimensional convolution and a sigmoid activation function.

[0018] Preferably, the method for extracting the illumination map of the original low-light image in step 1 is as follows:

[0019] Each channel of the feature sequence of the low-light image is encoded. The obtained positional encoding is then subjected to concat and transform operations to obtain a feature map. The feature map is normalized and activated by the ReLU function. It is then decomposed into two separate tensors along the spatial dimension. The two tensors are transformed into tensors with the same number of channels while keeping the dimension of the input tensor unchanged. The two tensors are activated by the sigmoid function to obtain two attention weight tensors, which are used to weight the input tensor. The two weighted input tensors are multiplied by the original input to obtain the illumination map.

[0020] Preferably, the number of channels in the tensor transform is reduced using a reduction ratio.

[0021] Preferably, the reflection component self-calibration module comprises at least six convolutional layers;

[0022] The first convolutional layer consists of a two-dimensional convolution, normalization, ReLU activation function, and coordinate attention;

[0023] The last convolutional layer consists of a two-dimensional convolution and a sigmoid activation function;

[0024] The structure of the multiple convolutional layers between the first and last convolutional layers is the same, including two sets of two-dimensional convolutions, normalization, and ReLU activation functions.

[0025] Preferably, a loss function is constructed and the backpropagation algorithm is used to train the illumination decomposable low-light enhancement model and the unsupervised denoising enhancement model until the model converges to obtain the trained lightweight low-light image enhancement model.

[0026] The loss function includes fidelity loss, smoothing loss, colorfastness loss, and reflection loss.

[0027] Preferably, the expression for the loss function is as follows:

[0028] L total =1.5·L f +L s +2·L color +L R

[0029] Among them, L f To preserve fidelity, L s To smooth out the loss, L color For color constancy loss, L R This is the reflection loss.

[0030] Preferably, the expression for the fidelity loss is as follows:

[0031]

[0032] Where T is the total number of stages, y+s t-1 Input lighting, x t Output lighting;

[0033] The expression for the smoothing loss is as follows:

[0034]

[0035] Where N is the total number of pixels, i is the i-th pixel, N(i) represents the neighboring pixels of i in the window, and ω i,j Indicates the weight;

[0036]

[0037] Where c is the image channel in the YUV color space, and σ is the standard deviation of the Gaussian kernel;

[0038] The expression for the colorfastness loss is as follows:

[0039]

[0040] The expression for the reflection loss is as follows:

[0041]

[0042] In the formula, F(x) is the histogram equalization operator for image x, and λ is the weight parameter. This is the gradient operator.

[0043] A system for low-light image enhancement based on illumination decomposition, comprising:

[0044] The module for constructing a weak light enhancement model with decomposable illumination is used to build a weak light enhancement model with decomposable illumination based on coordinate attention. It includes an illumination decomposition and learning module and a reflection component self-calibration module.

[0045] The illumination decomposition learning module embeds a coordinate attention mechanism in the feature extraction layer. The feature extraction layer obtains the feature map of the original low-light image, and the coordinate attention mechanism is used to embed the position information of the original image into the feature map to obtain the illumination map of the original low-light image. The original low-light image is divided by the illumination map to obtain the reflection map of the original low-light image.

[0046] The reflection component self-calibration module acquires the reflection component of the reflection map, combines the reflection component with the original low-light image, and obtains an enhanced low-light image.

[0047] The enhanced low-light image is input into the illumination decomposition learning module, and the reflection component self-calibration module is repeatedly executed until the number of iterations is reached to obtain the reflection map of the preliminary enhanced image.

[0048] The Unsupervised Denoising and Enhancement Model Building Module is used to build unsupervised denoising and enhancement models.

[0049] The unsupervised denoising enhancement model uses upsampling and downsampling to denoise the final reflection map output by the illumination decomposition learning module, and performs skip connections between feature maps obtained from different sampling layers to fuse the feature maps from different sampling layers to obtain the final enhanced image.

[0050] Compared with the prior art, the present invention has the following beneficial technical effects:

[0051] This invention provides a low-light image enhancement method based on illumination decomposition, employing a staged reflectance component self-calibration learning technique to ensure the effectiveness of the low-light image enhancement network. This technique ensures that the network's outputs at different stages converge to the same state during training, thereby improving the network's convergence. Simultaneously, separating the self-calibration module from the prediction stage significantly reduces the number of parameters required by the network model, making the network more lightweight. Furthermore, to further improve the network's generalization and noise resistance, this invention designs an unsupervised noise-resistant enhancement module. Through this module, the network can learn to suppress noise and enhance low-light images without explicit labels, further improving the low-light image enhancement performance of this invention. Finally, the loss function of this invention introduces a calibrable color deviation loss term, making the enhanced image colors more realistic and accurate. Attached Figure Description

[0052] Figure 1 This is a flowchart of the low-light image enhancement method based on illumination decomposition according to the present invention;

[0053] Figure 2 This is a flowchart of the lightweight learning network of the present invention;

[0054] Figure 3 This is a structural diagram of the coordinate attention module of the present invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings. These descriptions are intended to explain the invention and not to limit it.

[0056] See Figure 1 A low-light image enhancement method based on illumination decomposition includes the following steps:

[0057] Step 1: Construct a weak light enhancement model based on coordinate attention, which includes an illumination decomposition and learning module and a reflection component self-calibration module;

[0058] The illumination decomposition learning module embeds a coordinate attention mechanism in the feature extraction layer. The feature extraction layer obtains the feature map of the original low-light image, and the coordinate attention mechanism is used to embed the position information of the original image into the feature map to obtain the illumination map of the original low-light image. The original low-light image is divided by the illumination map to obtain the reflection map of the original low-light image.

[0059] The reflection component self-calibration module acquires the reflection component of the reflection map, combines the reflection component with the original low-light image, and obtains an enhanced low-light image.

[0060] The enhanced low-light image is input into the illumination decomposition learning module, and the reflection component self-calibration module is repeatedly executed until the number of iterations is reached to obtain the reflection map of the preliminary enhanced image.

[0061] The illumination decomposition learning module is used to learn the illumination residuals in stages, greatly reducing the computational difficulty. The reflection component self-calibration module is used to learn the reflection components, bridging the input of each stage with weak light observations, indirectly exploring the convergence behavior between each stage, and accelerating the convergence of the entire network. The entire network is implemented in stages. This network incorporates coordinate attention, which, by embedding positional information into channel attention, improves the model's representational ability while maintaining low computational cost, making its application in lightweight networks more feasible and effective. In traditional attention mechanisms, attention weights are usually applied to the channel dimension of the input features to enhance the model's attention to different channels. This allows the model to be more flexible in processing features, selectively focusing on the channel information most important to the current task. Unlike traditional channel attention, coordinate attention embeds positional information into the channel attention mechanism. Specifically, coordinate attention associates each position in the feature map with its corresponding coordinate information, combining the positional information with the channel attention weights, so that the model can consider the importance of each position while focusing on the channel. Coordinate attention achieves positional attention weighting by considering positional information when calculating channel attention weights. This means that when calculating the attention weights for each channel, the model considers not only the relationships between channels but also the relationships between different positions. This allows the model to finely adjust the feature representations for each position, thereby improving the model's ability to process spatial information.

[0062] Step 2: Construct an unsupervised denoising and enhancement model;

[0063] The unsupervised denoising enhancement model uses upsampling and downsampling to denoise the final reflection map output by the illumination decomposition learning module, and performs skip connections between feature maps obtained from different sampling layers to fuse the feature maps from different sampling layers to obtain the final enhanced image.

[0064] The enhanced image and the original low-light image are input into an unsupervised denoising and enhancement network model for further enhancement and denoising. By utilizing the contrast information of the enhanced image, the contrast and brightness of the low-light image can be adjusted more effectively, thereby improving image clarity and visual quality. Simultaneously, noise present in the image can be denoised to further improve image quality. It is important to note that this invention chooses to perform re-enhancement and denoising on the original low-light image, rather than only operating on the already enhanced image. This is because image processing typically leads to information loss, especially during enhancement operations. Therefore, to preserve image details and information to the greatest extent possible, this invention chooses to perform subsequent processing on the original low-light image. Through this re-enhancement strategy, this invention can more effectively remove noise from the image and further improve image clarity and quality.

[0065] Step 3: Construct a loss function and combine it with the backpropagation algorithm to train the low-light enhancement model and the unsupervised denoising enhancement model that can be decomposed by illumination until the model converges to obtain the lightweight low-light image enhancement model after training. Use the lightweight low-light image enhancement model to enhance the low-light image.

[0066] Two models were trained using a dataset. The network models were trained with a large amount of training data, and the model parameters were optimized by minimizing the loss function. Finally, the model performance was evaluated using test set data. By comparing the model with the test set data, the effectiveness of the model in denoising and enhancement can be verified, and necessary adjustments and optimizations can be made.

[0067] Using a pre-trained network model for low-light image enhancement, and employing only a single illumination learning network module and an unsupervised denoising and enhancement network module for low-light image enhancement and denoising, not only ensures the enhancement effect but also reduces the computational load.

[0068] In some embodiments, the illumination decomposition and learning module and the reflection component self-calibration module are used to learn the illumination components and reflection components of the image, respectively. The illumination decomposition and learning module and the reflection component self-calibration module each contain at least four convolutional layers, which consist of two-dimensional convolution, normalization, ReLU activation function, and sigmoid activation function. First, the low-light image is input into the illumination decomposition and learning network, which learns the illumination components of the image and outputs an illumination map. Then, the original low-light image is divided by the illumination map to obtain the reflection map for this stage. The reflection map is then input into the reflection component self-calibration network, which learns the reflection components of the image. The output reflection component self-calibration map is added to the original low-light image as the input for the next stage.

[0069] The illumination decomposition learning module does not employ a single-step decomposition method to calculate the illumination components of an image. Instead, it uses an iterative, multi-stage illumination component learning module. This module uses a shared parameter approach, meaning that each stage uses the same network architecture and parameters to progressively calculate accurate illumination components in a multi-stage manner. Furthermore, this module can adjust the number of iterations in the multi-stage illumination decomposition to suppress overexposure issues caused by image enhancement. The module consists of at least four convolutional layers. The first convolutional layer contains at least one 2D convolution, a ReLU activation function, and coordinate attention. The second and third convolutional layers are identical, containing at least one 2D convolution, normalization, and a ReLU activation function. The fourth convolutional layer contains at least one 2D convolution and a sigmoid activation function.

[0070] Coordinate attention is incorporated into the feature extraction layer of the illumination decomposition learning module. This helps to accurately capture the spatial structure information in the image, enabling the network module to perform efficient illumination decomposition and learning. To implement coordinate attention, firstly, to introduce positional information, each position in the input sequence needs to be encoded. The positional encoding is generated using a sine or cosine function, with the same dimension as the input feature dimension. For a given input sequence, the positional encoding is added to the input to obtain the key (k) and query (q), ensuring that each position has its own positional information. The value (v) typically uses the feature representation from the input sequence directly. Next, the dot product between the key (k) and query (q) is calculated and then normalized using a softmax function to obtain the attention weights. Finally, the calculated attention weights are used to perform a weighted summation of the values ​​(v) to obtain the output representation for each position.

[0071] The reflection component self-calibration module consists of at least six convolutional layers. The first convolutional layer contains a 2D convolution, normalization, ReLU activation, and coordinate attention. The second convolutional layer contains two sets of 2D convolutions, normalization, and ReLU activation. The next three convolutional layers have the same structure as the second layer. The sixth convolutional layer contains a 2D convolution and a sigmoid activation. During training, this module assists the illumination decomposition and learning module in converging to the same state at each stage, accelerating the convergence speed of the entire network and reducing the number of iterations in the illumination decomposition and learning module, thus reducing computational cost.

[0072] In some embodiments, the unsupervised enhancement and denoising module takes the output of an illumination-decomposable low-light enhancement network and the original low-light image as input. It utilizes the contrast information of the enhanced image to adjust and denoise the contrast and brightness information of the enhanced image. This unsupervised enhancement and denoising module contains at least nine convolutional layers, each consisting of a two-dimensional convolution, a ReLU activation function, and a sigmoid activation function. The invention also employs upsampling and downsampling operations to reduce noise. However, downsampling can blur the image; therefore, the invention uses a concat operation in the fifth and sixth layers to fuse the downsampling results with the original image features, thereby increasing the detail information of the enhanced image. The invention sets a reflection loss function. Unlike existing methods that only use the gradient norm, this invention proposes a novel distance measurement method that considers the information preservation and consistency required for image enhancement tasks and leverages the properties of histogram equalization to improve the image's information entropy. Therefore, unsupervised low-light image enhancement and denoising is achieved without the need for paired datasets for training.

[0073] Example 1

[0074] See Figure 2 and 3 A low-light image enhancement method based on illumination decomposition includes the following steps:

[0075] Step 1: Build a weak light enhancement network model based on coordinate attention that can decompose illumination. The network consists of an illumination decomposition and learning module and a reflection component self-calibration module.

[0076] S1.1 Illumination Decomposition and Learning Module: This module mainly consists of multiple sets of two-dimensional convolutional layers, ReLU activation function, normalization, and sigmoid activation function. Assuming the input to this module is a 3×600×400 image, its output is a 3×600×400 feature map. The output shape of this illumination decomposition and learning module should be the same as the input shape of the reflection component self-calibration module. The illumination decomposition and learning module learns the illumination residuals in stages, greatly reducing the computational difficulty.

[0077] 1.2 Reflectance Component Self-calibration Module: This module consists of multiple sets of two-dimensional convolutional layers, ReLU activation functions, normalization, and sigmoid activation functions. Assuming the input to this module is a 3×600×400 image, it will produce a 16×600×400 feature map, and the final output will be a 3×600×400 feature map. The reflection component self-calibration module bridges the input of each stage (except the first stage) with the low-light observation (i.e., the input of the first stage), indirectly exploring the convergence behavior between each stage.

[0078] The construction process of the illumination decomposition learning network is as follows:

[0079] The reflection component self-calibration module's mapping is added to the original low-light input, resulting in a 3-channel image matrix, which serves as the input for the next stage of illumination estimation. The first convolutional layer contains a 2D convolution with a ReLU activation function, a 3×3 kernel size, 16 output channels, and a stride of 1, denoted as 3×3×16×1. The feature map output from the first convolution is input to the coordinate attention layer. First, to incorporate positional information, the input feature sequence is encoded along the horizontal and vertical coordinates using pooling kernels of size (H,1) and (1,W), respectively. Then, the resulting positional encodings are concatenated and transformed using a 16-kernel convolutional transformation function, ensuring each position has its own location information. Next, the input feature map is normalized and activated with a ReLU function, then decomposed into two separate tensors along the spatial dimension. Finally, two additional convolutional transformations are used to transform these two tensors into tensors with the same number of channels, maintaining the dimensionality of the input tensors. In this step, an appropriate reduction ratio (e.g., 16) is used to reduce the number of channels in the tensors. This helps reduce the complexity and computational cost of the model while preserving key feature information. Then, the resulting two tensors are subjected to a sigmoid activation function to obtain two attention weight tensors. These two attention weights are used to weight the input tensors to enhance the model's attention to different locations. Finally, the two weighted input tensors are multiplied by the original input to obtain the coordinate attention output, with 16 output channels. The second and third convolutional layers are similar, containing a 2D convolution, normalization, and ReLU activation function, with a 3×3 kernel size, 16 output channels, and a stride of 1, denoted as 3×3×16×1. The fourth convolutional layer contains a 2D convolution and a sigmoid activation function, with a 3×3 kernel size, 3 output channels, and a stride of 1, denoted as 3×3×3×1.

[0080] The process of building the reflection component self-calibration module is as follows:

[0081] The input to each stage of the reflectance component self-calibration network originates from the previous stage. The input to the first stage is explicitly defined as a weak-light observation with 3 channels. The first convolutional layer contains a 2D convolution, normalized, with a ReLU activation function, a 3×3 kernel size, 16 output channels, and a stride of 1, denoted as 3×3×16×1. The feature map output from the first convolution is input to the coordinate attention layer. First, to incorporate positional information, the input feature sequence is encoded along the horizontal and vertical coordinates using pooling kernels of sizes (H,1) and (1,W), respectively. Then, the resulting positional encodings are concatenated and transformed using a 16-kernel convolutional transformation function, ensuring each position has its own location information. Next, the input feature map is normalized and activated with a ReLU function, then decomposed into two separate tensors along the spatial dimension. Finally, two additional convolutional transformations are used to transform these two tensors into tensors with the same number of channels, maintaining the dimensionality of the input tensors. In this step, an appropriate reduction ratio (e.g., 16) is used to reduce the number of channels in the tensors. This helps reduce the complexity and computational cost of the model while preserving key feature information. Then, the resulting two tensors are subjected to a sigmoid activation function to obtain two attention weight tensors. These two attention weights are used to weight the input tensors to enhance the model's attention to different locations. Finally, the two weighted input tensors are multiplied by the original input to obtain the coordinate attention output, with 16 output channels. The second convolutional layer contains two sets of 2D convolutions, normalization, and ReLU activation functions, with a kernel size of 3×3, 16 output channels, and a stride of 1, denoted as 3×3×16×1. The next three convolutional layers have the same structure as the second convolutional layer, with a kernel size of 3×3, 16 output channels, and a stride of 1, denoted as 3×3×16×1. The sixth convolutional layer contains a two-dimensional convolution and a sigmoid activation function. The kernel size is 3×3, the output channels are 3, and the stride is 1, denoted as 3×3×3×1.

[0082] Step 2: Construct an unsupervised denoising and enhancement network model. This module mainly consists of multiple sets of two-dimensional convolutional layers, ReLU activation function, normalization, and sigmoid activation function. The 3×600×400 feature map obtained in Step 1.1 is concatted with the original low-light image to obtain a 6×600×400 image matrix, which is then input into the unsupervised denoising and enhancement network to obtain a 3×600×400 feature matrix.

[0083] The process of building an unsupervised denoising and enhancement network is as follows:

[0084] The network consists of nine convolutional layers. The first convolutional layer contains a 2D convolution with 6 input channels, 64 output channels, and a 3×3 kernel. The second convolutional layer contains a 2D convolution and a ReLU activation function, with 64 input and 64 output channels and a 3×3 kernel. The third convolutional layer contains a 2D convolution and a ReLU activation function, with 64 input channels, 128 output channels, and a 3×3 kernel. The fourth and sixth convolutional layers contain a 2D convolution and a ReLU activation function, with 128 input channels, 128 output channels, and a 3×3 kernel. The seventh convolutional layer contains a 2D convolution and a ReLU activation function, with 96 input channels, 64 output channels, and a kernel size of 3. The eighth convolutional layer contains a 2D convolution and a ReLU activation function, with 64 input channels, 64 output channels, and a kernel size of 3. The ninth convolutional layer contains a 2D convolution, with 64 input channels, 3 output channels, and a kernel size of 3.

[0085] Step 3: Train the network model constructed in Steps 1 and 2 using the training dataset. In this invention, a fidelity loss is employed to ensure pixel-level consistency between the estimated illumination components and the input at each stage. Simultaneously, a smoothing term with spatial variation norm L1 is introduced to constrain the smoothness of the output image. Furthermore, a calibrable color deviation loss is designed to correct potential color deviations in the enhanced image, and the relationship between the three adjustment channels is established. Finally, a reflection loss is used to improve image contrast and suppress noise. During training, the model parameters are updated by minimizing the total loss value. Specifically, the input data is passed to the model to calculate the loss, and the model parameters are updated using the backpropagation algorithm. Training was performed on the training set for 1000 epochs, with each epoch involving passing the entire training set to the model once and updating the parameters. The training loss function is shown in Equation (1).

[0086] The loss function is defined as shown in Equation 1:

[0087] L total =1.5·L f +L s +2·L color +L R (1)

[0088] Among them, L f To preserve fidelity, L s To smooth out the loss, Lcolor For color constancy loss, L R The reflection loss is shown in Equations (2), (3), (5) and (6), respectively.

[0089]

[0090] In formula (2), T represents the total number of stages. In fact, this function utilizes the redefined input y+s t-1 To constrain the output lighting x t Instead of manually generated true values ​​or ordinary low-light input, the fidelity loss is to ensure that the estimated illuminance is consistent with the input pixel-level at each stage.

[0091]

[0092] In formula (3), N is the total number of pixels. i is the i-th pixel. N(i) represents the neighboring pixels of i in its 5×5 window. ω i,j The weight is represented by the formula:

[0093]

[0094] Where c represents the image channel in the YUV color space. σ = 0.1 is the standard deviation of the Gaussian kernel.

[0095]

[0096] In formula (5), J p This represents the average intensity value of the p channel in the enhanced image, where (p,q) represents a pair of channels.

[0097]

[0098] In the formula, F(x) is the histogram equalization operator for image x, and λ is the weight parameter. This is the gradient operator. The first term of the loss function is used to directly control the enhancement of image contrast. The second term is used to suppress noise.

[0099] Step 4: Perform low-light enhancement using the trained network model. The low-light image enhancement task is performed using the illumination decomposition and learning module and the unsupervised denoising and enhancement network module trained in Step 3. First, the network built in Steps 1 and 2 is initialized using the network parameters saved from Step 3 training; then, the test dataset is input into the trained network, and finally, the enhanced image result is output.

[0100] During the training phase, an assisted learning approach is adopted, that is, the reflection component self-calibration module is used to assist the training of the illumination decomposition and learning module, ensuring that the outputs at different stages of training can converge to the same state, and reducing the computational burden of the cascaded mode of illumination decomposition and learning modules. During the testing phase, the reflection component self-calibration module can be discarded after training, and only the illumination decomposition and learning module and the unsupervised denoising and enhancement module are used to enhance the low-light image.

[0101] Example 3

[0102] Another method of the present invention also provides a system for a low-light image enhancement method based on illumination decomposition, including an illumination decomposable low-light enhancement model construction module and an unsupervised denoising enhancement model construction module;

[0103] The module for constructing a weak light enhancement model with decomposable illumination is used to build a weak light enhancement model with decomposable illumination based on coordinate attention. It includes an illumination decomposition and learning module and a reflection component self-calibration module.

[0104] The illumination decomposition learning module embeds a coordinate attention mechanism in the feature extraction layer. The feature extraction layer obtains the feature map of the original low-light image, and the coordinate attention mechanism is used to embed the position information of the original image into the feature map to obtain the illumination map of the original low-light image. The original low-light image is divided by the illumination map to obtain the reflection map of the original low-light image.

[0105] The reflection component self-calibration module acquires the reflection component of the reflection map, combines the reflection component with the original low-light image, and obtains an enhanced low-light image.

[0106] The enhanced low-light image is input into the illumination decomposition learning module, and the reflection component self-calibration module is repeatedly executed until the number of iterations is reached to obtain the reflection map of the preliminary enhanced image.

[0107] The Unsupervised Denoising and Enhancement Model Building Module is used to build unsupervised denoising and enhancement models.

[0108] The unsupervised denoising enhancement model uses upsampling and downsampling to denoise the final reflection map output by the illumination decomposition learning module, and performs skip connections between feature maps obtained from different sampling layers to fuse the feature maps from different sampling layers to obtain the final enhanced image.

[0109] Example 4

[0110] An image detection system includes an execution program for the aforementioned low-light image enhancement method based on illumination decomposition. During operation, the low-light image enhancement method based on illumination decomposition enhances the acquired low-light image, and image detection is performed based on the enhanced image.

[0111] This image detection system can be used as a monitoring system for factory areas or a mobile monitoring system.

[0112] This invention discloses a low-light image enhancement method based on illumination decomposition. The method first performs preliminary enhancement on the low-light image using an illumination-decomposable low-light enhancement model based on coordinate attention. Then, the enhanced image and the original low-light image are simultaneously input into an unsupervised denoising enhancement model for further enhancement and denoising. A novel parameter-sharing mechanism is introduced, using the same network architecture and weights in the feature extraction stage, thereby significantly reducing the computational cost of learning the illumination components. Furthermore, an unsupervised denoising enhancement module is designed to further improve the network's generalization and noise resistance. Through this module, the network can better handle noise from various environments, improving its performance under complex conditions, thus further enhancing the performance of this invention in low-light image enhancement tasks. This invention also incorporates coordinate attention, which associates each position in the feature map with its corresponding coordinate information, combining positional information with channel attention weights. This allows the model to consider the importance of each position while focusing on the channel, improving the model's representational ability while maintaining low computational cost.

[0113] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A low-light image enhancement method based on illumination decomposition, characterized in that, Includes the following steps: Step 1: Construct a weak light enhancement model based on coordinate attention, which includes an illumination decomposition learning module and a reflection component self-calibration module; The illumination decomposition learning module embeds a coordinate attention mechanism in the feature extraction layer. The feature extraction layer obtains the feature map of the original low-light image, and the coordinate attention mechanism is used to embed the position information of the original low-light image into the feature map to obtain the illumination map of the original low-light image. The original low-light image is divided by the illumination map to obtain the reflection map of the original low-light image. The reflection component self-calibration module acquires the reflection component of the reflection map, combines the reflection component with the original low-light image, and obtains an enhanced low-light image. The enhanced low-light image is input into the illumination decomposition learning module, and the reflection component self-calibration module is repeatedly executed until the number of iterations is reached to obtain the final reflection map. Step 2: Construct an unsupervised denoising and enhancement model; The final reflection map output by the illumination decomposition learning module is stitched together with the original low-light image, and the resulting feature matrix is ​​input into the unsupervised denoising enhancement model. The unsupervised denoising enhancement model uses upsampling and downsampling to denoise the feature matrix, and performs skip connections between feature maps obtained from different sampling layers to fuse the feature maps from different sampling layers to obtain the final enhanced low-light image.

2. The low-light image enhancement method based on illumination decomposition according to claim 1, characterized in that, The method for extracting the illumination map of the original low-light image in step 1 is as follows: The illumination decomposition learning module includes at least four convolutional layers; The first convolutional layer contains at least one 2D convolution, a ReLU activation function, and coordinate attention; The second and third convolutional layers have the same structure, including at least one two-dimensional convolution, normalization, and ReLU activation function; The fourth convolutional layer contains at least one two-dimensional convolution and a sigmoid activation function.

3. The low-light image enhancement method based on illumination decomposition according to claim 1, characterized in that, The reflection component self-calibration module includes at least six convolutional layers; The first convolutional layer consists of a two-dimensional convolution, normalization, ReLU activation function, and coordinate attention; The last convolutional layer consists of a two-dimensional convolution and a sigmoid activation function; The structure of the multiple convolutional layers between the first and last convolutional layers is the same, including two sets of two-dimensional convolutions, normalization, and ReLU activation functions.

4. The low-light image enhancement method based on illumination decomposition according to claim 1, characterized in that, A loss function is constructed and the backpropagation algorithm is used to train the low-light enhancement model and the unsupervised denoising enhancement model with decomposable illumination until the model converges, resulting in a lightweight low-light image enhancement model after training. The loss function includes fidelity loss, smoothing loss, colorfastness loss, and reflection loss.

5. The low-light image enhancement method based on illumination decomposition according to claim 4, characterized in that, The expression for the loss function is as follows: in, To preserve fidelity, To smooth out the loss, For loss of color constancy, This is the reflection loss.

6. A system for implementing the low-light image enhancement method based on illumination decomposition as described in any one of claims 1-5, characterized in that, include: A module for constructing a weak light enhancement model with decomposable illumination is used to construct a weak light enhancement model with decomposable illumination based on coordinate attention. The weak light enhancement model with decomposable illumination includes an illumination decomposition learning module and a reflection component self-calibration module. The illumination decomposition learning module embeds a coordinate attention mechanism in the feature extraction layer. The feature extraction layer obtains the feature map of the original low-light image, and the coordinate attention mechanism is used to embed the position information of the original low-light image into the feature map to obtain the illumination map of the original low-light image. The original low-light image is divided by the illumination map to obtain the reflection map of the original low-light image. The reflection component self-calibration module acquires the reflection component of the reflection map, combines the reflection component with the original low-light image, and obtains an enhanced low-light image. The enhanced low-light image is input into the illumination decomposition learning module, and the reflection component self-calibration module is repeatedly executed until the number of iterations is reached to obtain the reflection map of the preliminary enhanced image. The Unsupervised Denoising and Enhancement Model Building Module is used to build unsupervised denoising and enhancement models. The final reflection map output by the illumination decomposition learning module is stitched together with the original low-light image, and the resulting feature matrix is ​​input into the unsupervised denoising enhancement model. The unsupervised denoising enhancement model uses upsampling and downsampling to denoise the feature matrix, and performs skip connections between feature maps obtained from different sampling layers to fuse the feature maps from different sampling layers to obtain the final enhanced low-light image.

Images