An edge information guidance-based pseudo-supervised learning traffic image quality enhancement method

Abstract

The application discloses a pseudo-supervised learning traffic image quality enhancement method based on edge information guidance, aims to solve the problem of target detection precision decline caused by the "underexposure" and "overexposure" phenomenon, and the method is as follows: first, a visual content generator based on a visible light image-SAM mask is designed, image and mask features are extracted respectively and are deeply fused to generate a corresponding enhancement matrix gamma; second, a pseudo-label generator is designed, 2N reference images are randomly generated by using input images, aiming to provide rich reference data for the model and enhance the generalization and robustness of the model; third, a global-local image quality evaluator is constructed to evaluate the image quality of the reference images, and the best quality image is selected as the pseudo-label; finally, L1-SSIM is introduced as the structure reconstruction loss of the model, which effectively captures the details and structure information of the image, improves the image quality and ensures the visual coherence of the image.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and deep learning technology, and in particular to a pseudo-supervised learning traffic image quality enhancement method based on edge information guidance, which aims to solve the problem of decreased target detection accuracy in visible light images caused by underexposure and overexposure in the field of intelligent transportation. Background Technology

[0002] The field of intelligent transportation has developed rapidly in recent years, and computer vision technology has played a crucial role in this development. Real-time image acquisition via cameras and processing and analysis using computer vision technology helps optimize traffic management, improve traffic safety, and increase traffic efficiency. However, due to the complex and ever-changing real-world traffic environment, factors such as weather and lighting variations lead to serious quality problems in the visible light images acquired by cameras, with overexposure and underexposure being the most common issues.

[0003] Overexposure causes images to be too bright, resulting in the loss of detail in some areas and the appearance of white areas or bright spots, making information unreadable and unrecognizable, thus affecting subsequent visual processing and decision-making. Underexposure, on the other hand, causes visible light images to appear dim, lacking detail, with blurred object outlines and distorted colors, making the image difficult to identify and analyze. These phenomena severely impact the detection accuracy of object detection algorithms. Therefore, it is necessary to enhance the quality of visible light images to suppress overexposure and underexposure, providing a good data foundation for downstream tasks such as object detection.

[0004] Visible light image quality enhancement techniques are mainly divided into traditional visible light image enhancement algorithms and deep learning-based image enhancement algorithms. Traditional image enhancement algorithms improve the brightness and contrast of an image by directly adjusting its pixel values, thereby achieving better visual effects. Deep learning-based methods, primarily using convolutional neural networks and generative adversarial networks, effectively enhance low-quality visible light images by constructing deep learning network models and training them with large amounts of data. Although some progress has been made in related research, traditional methods are sensitive to parameter settings, lack automation and adaptability, and struggle to adapt to varying lighting environments and different application scenarios. Deep learning methods require large amounts of paired training data, which is extremely difficult to obtain in the current intelligent transportation field. This makes it difficult for deep learning methods to fit the data, resulting in unclear edges in the enhanced image and difficulty in effectively distinguishing the foreground from the background. These limitations make it difficult for mainstream image enhancement methods to guarantee the visual quality of the generated images in intelligent transportation applications. Therefore, achieving efficient and stable image quality enhancement for the intelligent transportation field remains a significant challenge.

[0005] To address the aforementioned problems, this invention proposes an edge-guided traffic image enhancement method based on pseudo-supervised learning. This method generates visually excellent image enhancement results under varying input lighting conditions without requiring paired training data. Specifically, it cleverly combines the advantages of deep learning techniques with traditional image processing methods. A deep learning model generates corresponding enhancement matrices for different input images, effectively handling various lighting conditions and achieving significant image quality improvement. This technology provides strong data support for downstream tasks such as object detection in the field of intelligent transportation.

[0006] The technical solution of this invention not only includes the design and optimization of the algorithm model, but also covers the testing and verification of the algorithm, forming a complete technical solution. It effectively improves the generalization of image enhancement algorithms in the field of intelligent transportation, suppresses the "overexposure" and "underexposure" phenomena in traffic scenes, provides a good data foundation for downstream tasks such as target detection, and improves the detection accuracy of detection algorithms. It has important practical application value and broad market prospects. Summary of the Invention

[0007] To address the problems of current methods in the field of intelligent transportation being difficult to train and having unclear differences between the foreground and background in the generated results, this invention proposes a pseudo-supervised learning method for traffic image quality enhancement based on edge information guidance.

[0008] To ensure effective traffic image quality enhancement even in the absence of paired datasets, the following steps are taken: First, the Segment Anything Model (SAM) is used to obtain the SAM mask corresponding to the image. An RGB-SAMBased Visual Content Generator (R-SVCG) is designed, based on the visible light image and SAM mask, to extract image and mask features using a dual-ended encoder structure. Then, a HybridAttention Fusion Module (HAFM) is used, guided by edge information provided by the SAM mask, to achieve deep fusion of image and mask features at the channel and spatial levels, generating the corresponding enhancement matrix γ. Second, a pseudo-label generator is designed to randomly generate 2N pseudo-labels based on the visible light image, corresponding to N brighter and N darker images, respectively. This design aims to provide rich pseudo-label reference data for the model when facing unknown input lighting conditions, enhancing the model's generalization and robustness. Then, a global-local image quality evaluator is constructed to comprehensively evaluate the visual effect of the image from three aspects: global image spatial quality, local pixel contrast, and local color saturation. The image with the best visual effect in the reference image set is selected as the pseudo-label for training. Finally, L1-SSIM is introduced as the structural reconstruction loss of the model, which considers both image structural similarity and global similarity to minimize the difference between the enhancement result and the pseudo-label. This loss function can effectively capture the details and structural information of the image, improving the image quality enhancement effect while ensuring the visual coherence of the image.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] Step 1: Dataset Construction and Expansion. For existing common datasets in the field of intelligent transportation, data from multiple shooting locations are selected to enrich the scenarios in the dataset for training deep learning models. Since the images in the current dataset are too monotonous and lack sufficient positive and negative samples, the Retinex method is used to process the original image dataset to simulate visible light images under different lighting conditions, thus expanding the original dataset and adding positive and negative samples. After expanding the dataset, the original data is divided into training and validation samples in a 7:3 ratio.

[0011] Step 2: Generate an enhancement matrix for image enhancement using a visual content generator based on visible light image-SAM mask image pairs. First, feature information from the SAM mask image and the original image is acquired using a dual-ended encoder. Second, considering both channel and spatial levels, edge information contained in the SAM features is used as a guide for deep fusion with the original image features. Finally, the fused features are input into the decoder to generate the enhancement matrix, which is then used for subsequent gamma transform to achieve image enhancement.

[0012] Step 3: Obtain a set of pseudo-label reference images using a pseudo-label generator. The pseudo-label generator performs a gamma transform on the original image with random parameters. The gamma value γ ranges from [0.1, 1) to [1, 4), generating N reference labels. If the image is underexposed, γ should be less than 1 to enhance the brightness of the dark areas; if the image is overexposed, γ should be greater than 1 to reduce the brightness of the bright areas. This ultimately generates 2N reference images, i.e., N images that are either brighter or darker than the original image.

[0013] Step 4: Evaluate the image quality in the pseudo-label reference image set using a global-local image quality evaluator to obtain pseudo-labels. The Blind / Referenceless Image Spatial Quality Evaluator (BRISQUE) is used to calculate image quality from a global perspective. A Support Vector Machine (SVM) is used to learn the mapping relationship between image quality and image features, and a global quality score is calculated. Furthermore, the visual effect of the pseudo-label reference images is evaluated from two levels: local pixel contrast and local color saturation.

[0014] Step 5: Calculate the reconstruction loss function between the pseudo-labels and the output, and update the model parameters through backpropagation to the visual content generator. Train the model using an iterative enhancement training strategy until it fits perfectly.

[0015] Step 6: Input the test samples into the trained model to obtain the image quality enhancement results, and verify the effect of image quality enhancement by combining human preference evaluation and object detection accuracy.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] (1) To address the current lack of paired datasets in the field of intelligent transportation, a pseudo-supervised learning traffic image quality enhancement method based on edge information guidance is proposed. By combining the advantages of deep learning and traditional methods, the enhancement matrix of the image is generated by the deep learning model and adaptively adjusted for different input images, which effectively improves the generalization performance of the model.

[0018] (2) To address the problem that current deep learning methods are difficult to fit, resulting in difficulty in ensuring the difference between the foreground and background after enhancement, i.e., the edges of the image enhancement results after existing methods are not clear, SAM mask is introduced as edge guidance information. The features of SAM mask and visible light image are extracted by dual-ended encoder, and the two features are fully fused to effectively enhance the model's ability to distinguish between the foreground and background of the image. Attached Figure Description

[0019] Figure 1 This is the overall process of a pseudo-supervised learning traffic image quality enhancement method based on edge information guidance;

[0020] Figure 2 This is a schematic diagram of the architecture of a pseudo-supervised traffic image quality enhancement method guided by edge information based on pseudo-supervised learning;

[0021] Figure 3 This is a schematic diagram of the structure of a visual content generator (R-SVCG) based on visible light image-SAM mask;

[0022] Figure 4 This is a schematic diagram of the Hybrid Attention Fusion Module (HAFM); Detailed Implementation

[0023] The invention will now be described in detail with reference to a specific implementation example, titled "Edge-Guided Image Enhancement Based on Pseudo-Supervised Learning on the FLIR and LLVIP Hybrid Datasets." This implementation example is provided to enable those skilled in the art to understand the specific process of the invention, but it does not limit the invention in any way. Furthermore, variations and improvements can be made based on this invention without departing from its conceptual framework.

[0024] In related research, deep learning-based image enhancement algorithms aim to enhance images by directly inputting high-quality images or learning specific parameters of a parametric model from paired datasets. However, collecting paired datasets is extremely time-consuming and difficult to implement in the field of intelligent transportation. Therefore, this method proposes an image enhancement framework based on pseudo-supervised learning to address the current lack of paired training data in the field of intelligent transportation. Simultaneously, a SAM mask is introduced to guide the network model in distinguishing foreground and background information of the image, thereby generating more accurate image enhancement results.

[0025] The overall process of an edge-guided image enhancement method based on pseudo-supervised learning in this implementation case is as follows: Figure 1 The method includes the following steps:

[0026] Step 1: Construct a hybrid dataset of FLIR and LLVIP.

[0027] In this implementation case, the mainstream datasets FLIR and LLVIP in the field of intelligent transportation are selected as the main datasets. Images are extracted from the original datasets at a 2:1 ratio to construct the dataset for this implementation case, totaling 12,384 images. The obtained data are uniformly scaled to 640×640 pixels, and the Retinex method is used to obtain images under different lighting conditions. The specific representation is shown in the formula:

[0028] L(x,y)=I(x,y)*G(k,σ) (1)

[0029] E(x,y)=Log(I(x,y))-Log(L(x,y)) (2)

[0030] Where I(x,y) represents the input image data, G represents Gaussian blur, k is the kernel size corresponding to the Gaussian blur, σ is the standard deviation of the Gaussian blur, L(x,y) represents the result of the input image after Gaussian blur, which is used to approximate the incident component in the single-scale Retinex algorithm, and E(x,y) represents the result after image enhancement. By setting different Gaussian kernel sizes and adjusting code parameters, the phenomena of "underexposure" and "overexposure" are simulated. In this implementation case, 3×3 and 7×7 Gaussian convolution kernels are used to process the original image, with a standard deviation of 0.

[0031] The expanded dataset was divided in a 7:3 ratio, with 26,007 training samples and 11,145 validation samples.

[0032] Step 2: Construct a pseudo-supervised learning traffic image quality enhancement model framework based on edge information and complete the training.

[0033] The framework structure of the edge-guided traffic image quality enhancement model based on pseudo-supervised learning is as follows: Figure 2 As shown, it mainly includes a visible light-masked image-based visual content generator (R-SVCG), a pseudo-label generator, and an image quality evaluator.

[0034] The visual content generator based on visible light-mask images (R-SVCG) serves as the core of the overall model, and its specific structure is as follows: Figure 3 As shown in the diagram, firstly, the Segment Anything Model (SAM) is used to obtain the SAM mask corresponding to the image, and then a dual-encoder is used to extract the feature information of the visible light image and the SAM mask respectively. A Hybrid Attention Fusion (HAFM) module is then used to fuse the SAM feature information with the image features, fully integrating their complementary information. Finally, after obtaining the fused features, a decoder generates the enhancement matrix γ of the image and performs a gamma transform on the visible light image. The specific representation is shown in the formula:

[0035] F r ,F s =Dual-Encoder(I r ,I s (3)

[0036] F fus =HAFM(F r ,F s (4)

[0037]

[0038] Where I r ,I s F represents the input image and the SAM mask result, respectively. r ,F s The image features and mask features obtained by the dual-encoder are used to obtain the fused feature F through a hybrid attention fusion module (HAFM). fus The enhancement matrix γ is predicted stepwise from the features using a decoder, and the enhanced image I is obtained by performing a gamma transform on the original visible light image based on the enhancement matrix γ. e .

[0039] The Dual-Encoder consists of five sequentially iterative deep learning modules. The encoder is constructed in the following order: a 1×1 convolutional layer (1×1Conv), a depthwise separable convolutional layer (DWConv), an activation function (Hswish), a 3×3 convolutional kernel (3×3Conv), and a compression-activation module (SE). The 1×1Conv and 3×3Conv layers work together to extract features of different dimensions, effectively enhancing the network's non-linear expressive power. The DWConv separable convolution preserves feature extraction capabilities through channel-wise and pointwise convolutions of the feature map, while improving computational efficiency. The SE module enhances the expressive power of channel features by introducing global information. The number of output channels in each of the five layers is set to (16, 16, 32, 32, 16), effectively extracting and representing image features through the progressively varying number of channels.

[0040] After feature extraction is completed, the two features are fused using a Hybrid Attention Fusion Module (HAFM). The structure of the Hybrid Attention Fusion Module (HAFM) is as follows: Figure 4As shown, this module first performs max pooling and average pooling on the channel dimension to obtain the feature components of image features and mask features, respectively, and then learns the feature components through a linear layer. This learning allows the model to adaptively determine which channels in the SAM features are more important for the current task, thus providing weights for subsequent fusion. Then, the SAM features and the fused features are input into a weight-sharing convolutional layer to complete feature fusion at the spatial level, obtaining the final fused feature F. fus The specific representation is shown in the formula:

[0041] F fc =Tanh(Linear(Linear(Adapt-avgpool c (F s +Adapt-maxpool c (F s (6)

[0042] +Linear(Adapt-avgpool c (F r +Adapt-maxpool c (F r ))))·F r

[0043] F fus =Tanh(Concat(7×7Conv(F)) s ),7×7Conv(F fc )))·F fc (7)

[0044] Where F r ,F s For image features and mask features obtained by a dual-encoder, F fc The features represented by channel attention fusion, Adapt-avgpool c Adapt-maxpool c This is an adaptive pooling layer in the PyTorch framework. It performs max pooling and average pooling on the feature maps at the channel level while ensuring the feature map size remains consistent before and after output. It calculates the maximum and average values ​​for each fixed-size grid along the feature channel dimension and adds them together to obtain the corresponding feature components. A linear layer (Linear) is used to perform a linear transformation on the original data, enhancing the model's generalization ability through learning. The Tanh activation function is applied to obtain the final fusion weights and complete channel fusion to obtain the F-value. fc The Tanh activation function formula is:

[0045]

[0046] After completing channel fusion, F fc and the original SAM feature F s In a 7×7 Conv with shared input weights, the convolutional kernel has a stride of 1 and padding of 3. It acquires spatial feature components from visible light and the SAM mask, respectively, and fuses these components using the Concat dimension concatenation method. The Tanh activation function is then used to obtain spatial fusion weights to complete feature-level fusion, resulting in the fused feature F. fus .

[0047] The decoder structure is similar to the encoder structure, generating an enhancement matrix γ and performing a gamma transform on the original image, as shown in the formula:

[0048] γ=MLP(Up(Concat(F fus ,3×3Conv(Up(F fus ))))) (9)

[0049] The decoder consists of a 3×3 convolutional kernel (3×3Conv) with a stride of 1, upsampling (Up), and a multilayer perceptron (MLP). Upsampling (Up) employs bilinear interpolation, generating new pixels by weighting existing pixels to achieve the fusion of features F. fus The dimensions of the feature F are expanded, and the fused features are input into a 3×3 Conv for information extraction. The Concat channel dimension concatenation method is used to combine the features F. fus The features are concatenated with the initially decoded features to effectively preserve information. The concatenated information is then input into a multilayer perceptron (MLP) to construct a complex function mapping to complete feature decoding and obtain the enhancement matrix γ. Based on the enhancement matrix, a gamma transform is performed on the original visible light image to obtain the single training result.

[0050] The pseudo-label generator is used to generate photos of the original image under different lighting conditions. It uses gamma transformation, with the gamma value γ′ ranging from [0.1, 1) to [1, 4). Assuming the image is underexposed, γ′ should be less than 1 to enhance the brightness of the dark areas; conversely, assuming the image is overexposed, γ′ should be greater than 1 to reduce the brightness of the bright areas. In this embodiment, two γ′ values ​​are randomly selected from each of the two ranges to obtain the pseudo-label images. In this example, 0.3, 0.7, 1.5, and 3.5 are used for pseudo-label generation. The specific representation is shown in the formula:

[0051]

[0052] in, This represents the generated set of pseudo-label reference images, which contains a total of 2N reference pseudo-labels, I rThis represents the input visible light image. The nth processed pseudo-label reference image is represented; rand(i,(a,b)) represents the i-th gamma value γ′ obtained from the range (a,b).

[0053] The global-local image quality evaluator is used to calculate the quality scores of images in the pseudo-label reference image set and selects the image with the highest quality score as the pseudo-label for the model to complete training. The image quality evaluator mainly consists of two parts: global image quality evaluation is implemented using a no-reference spatial image quality evaluator (BRISQUE), which extracts local contrast and brightness features of the image, uses a Natural Scene Statistics (NSS) model to capture the natural visual characteristics of the image, and then calculates the image quality score using a trained Support Vector Regression (SVM) model. Furthermore, local image quality evaluation measures the vividness and visual effect of the image by evaluating local pixel contrast and local color saturation, as shown in the formula:

[0054]

[0055] Among them, P x,y Let L represent a patch centered at pixel (x,y) with a size of K×K; I(i,j) represents the pixel value at position i,j, and μ is the average pixel value of the patch; calculate L c To represent the local pixel contrast of an image. Furthermore, C s It represents the local color saturation of an image, using the saturation channels of the HSV color space to measure the color saturation of a pixel, where R, G, and B correspond to the pixel values ​​at the (x, y) positions of the three channels of the visible light image, respectively.

[0056] In this implementation case, after the deep learning framework is built, the specific training steps of the pseudo-supervised edge-guided image enhancement model can be divided into the following steps:

[0057] 2.1 Obtain the SAM mask corresponding to the visible light image using the SAM model.

[0058] 2.2 The visible light image and the SAM mask image are input into the visual content generator (R-SVCG) based on the visible light image-SAM mask image pair. After extracting features from each, the feature information of the two images is fused using a hybrid attention fusion module (HAFM). The fused feature map is then input into the decoder to generate an image enhancement matrix, and a gamma transform is performed on the visible light image based on this matrix.

[0059] 2.3 Input the visible light image into the pseudo-label generator to obtain the pseudo-label reference image set, and add the training result of the (n-1)th iteration to the pseudo-label reference image set. Calculate the quality score of the images in the reference set using a global-local image quality evaluator, and select the image with the highest score as the pseudo-label for model training.

[0060] 2.4 Calculate the loss function between the pseudo-label and the output result, and backpropagate the overall model. Use the iterative enhancement training strategy to continuously train the model, enhance the image enhancement effect of the model, and retain the corresponding model weights.

[0061] In this embodiment, the input image size in the leakage condition identification model is 640×640, the initial learning rate during model training is set to 0.01, the momentum parameter is 0.937, the decay rate is set to 0.0005, and the stochastic gradient descent algorithm is used to train the identification model for 100 rounds.

[0062] The model uses L1-SSIM as the structure reconstruction loss. The L1-SSIM loss function calculates the structural difference between the output image and the pseudo-label, as shown in the formula:

[0063]

[0064] L c =α·SSIM(I o ,I p )+(1-α)·L1(I o ,I p (14)

[0065] SSIM operates on a fixed-size pixel window, where x and y represent two different input images, and μ x ,μ y This represents the average pixel value within a window at the same x, y position. σ represents the standard deviation of the corresponding window. xy The covariance of the windows at the same x and y positions is defined by c1 and c2, which are set as fixed proportions of the input pixel value range. These parameters are used to maintain the stability of the calculation results and are set to 6.5025 and 58.5025, respectively. The L1 loss function calculates the absolute difference between the pseudo-label and the output image, where I... o The image output of the model, I p The pseudo-labeled image is represented by α, which is a learnable parameter obtained through model iteration and used to balance the loss function. The initial value is set to 0.5, and the value tends to stabilize at about 0.275 when the training epochs reach 20.

[0066] Step 3: Input the test samples into the trained model for image enhancement, obtain the enhanced images, and evaluate their image quality. In this implementation case, due to the lack of paired datasets in the current intelligent transportation field for effective evaluation of experimental results, a no-reference image quality assessment method is used to determine image quality. The Natural Image Quality Evaluation (NIQE) method based on pixel discrimination and the no-reference image quality evaluation method based on convolutional neural networks (BIECON) are selected to calculate image quality scores to quantify image quality. A lower NIQE value indicates better image quality, while a higher BIECON score indicates better image quality. Experimental data demonstrates that using the method of this invention to enhance 8948 validation samples in the mainstream LFIR and LLVIP datasets, calculating the mean of all image quality scores, and comparing them with other advanced methods, shows that in the FLIR dataset, the NIQE score is 9.968 and the BIECON score is 0.945, both outperforming other advanced image enhancement methods. These results not only demonstrate the effectiveness of our proposed method, but also show the potential of image enhancement through deep learning techniques in the field of intelligent transportation.

[0067]

[0068] Specific embodiments of the present invention have been described above. It should be noted that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the substantive content of the present invention. Unless otherwise specified, the embodiments and features of the present invention can be arbitrarily combined with each other.

Claims

1. A pseudo-supervised learning traffic image quality enhancement method based on edge information guidance, characterized in that: Step 1: Construct a visible light traffic image dataset and perform data augmentation. Randomly sample and merge images from the dataset to construct a visible light dataset and use the Retinex image augmentation method to augment the data. Step 2: Construct a visual content generator based on the visible light image-SAM mask image to obtain the enhancement matrix and image quality enhancement results. Use the Segment Anything Model (SAM) large model to obtain the corresponding SAM mask, and input the visible light image and SAM mask together into the visual content generator to generate the image enhancement matrix. And obtain the enhanced image, i.e., the single output result; Step 3: Construct a pseudo-label reference image set using a pseudo-label generator. Randomly generate multiple pseudo-label images with different exposure levels using the pseudo-label generator, and merge them with the training results from the previous round to construct the pseudo-label reference image set. Step 4: Construct a global-local image quality evaluator and evaluate the image quality in the reference set. Calculate the quality score of the images in the reference set and select the image with the highest score as the pseudo-label to complete the model training. Step 5: Calculate the reconstruction loss function between the pseudo-labels and the output results, realize backpropagation of the visual content generator and update the model parameters, and retain the model weights after model fitting; Step 6: Input the test samples into the trained model to obtain the image quality enhancement results, and verify the image quality enhancement effect using a no-reference image quality assessment method; Step 2 specifically involves a visual content generator based on visible light image-SAM mask image pairs (R-SVCG). This generator uses a dual-ended encoder to extract feature information from both the visible light image and the SAM mask image, and then fuses the SAM features with the visible light features using a hybrid attention fusion module (HAFM). After obtaining the fused features, a decoder is used to generate the image enhancement matrix. The image is then subjected to a gamma transform to enhance its quality, as shown in the formula: in These represent the input image and the SAM mask result, respectively. Image features and mask features obtained by the dual-encoder are combined to obtain fused features through a hybrid attention fusion module (HAFM). And use the decoder to progressively predict the enhancement matrix from the features. Based on the enhancement matrix Gamma transform is performed on the original visible light image to obtain the enhanced image. ; The Dual-Encoder consists of five layers of sequentially iterative deep learning modules. The construction order of the dual-encoder is as follows: Convolutional layer Depthwise separable convolution Activation function Hswish convolution kernel And the compression-excitation module SE; and Together, they are used to extract features from different dimensions, effectively enhancing the nonlinear expressive power of the network; the number of output channels in each of the 5 layers is set to [number missing]. ; After feature extraction is completed, the Hybrid Attention Fusion Module (HAFM) is used to fuse the two features. This module first performs max pooling and average pooling on the channel dimension to obtain the feature components of image features and mask features respectively, and then learns the feature components through a linear layer. Then, the SAM features and the fused features are input into a weight-shared convolutional layer to complete feature fusion at the spatial level, obtaining the final fused features. The specific representation is shown in the formula: ； in The image features and mask features are obtained by the dual-encoder. The characteristics represent the features after channel attention fusion. This is an adaptive pooling layer in the PyTorch framework. It performs max pooling and average pooling on the feature map based on the channel layer, ensuring the feature map size remains consistent before and after output. It calculates the maximum and average values ​​for each fixed-size grid along the feature channel dimension and adds them together to obtain the corresponding feature components. A linear layer (Linear) is used to perform a linear transformation on the original data to enhance the model's generalization ability. Finally, the Tanh activation function is applied to obtain the final fusion weights and complete the channel fusion. The Tanh activation function formula is: ； After channel fusion is completed, and original SAM features Input weight sharing In this process, with a step size of 1 and a padding of 3, the feature components of visible light and SAM masks are obtained respectively. Channel-dimensional concatenation is achieved through Concat, and spatial fusion weights are obtained through the Tanh activation function to complete feature-level fusion, resulting in fused features. ; The decoder performs a gamma transform on the original image, as shown in the formula: ； The decoder consists of a step size of 1. The system consists of convolutional kernels, upsampling (Up), and a multilayer perceptron (MLP). Upsampling uses bilinear interpolation, which generates new pixels by weighting existing pixels to improve the fused features. Dimension expansion and input of fused features Information extraction is performed; features are concatenated using Concat. The features are concatenated with the decoded features to effectively preserve information. This information is then input into a multilayer perceptron (MLP) to construct a complex function mapping, completing feature decoding and obtaining the enhancement matrix. ; The pseudo-label generator adjusts the gamma value of the gamma transform. Generate reference images under various lighting conditions; The range of values ​​is , Assuming the image is "underexposed", Less than 1, and assuming the image is "overexposed", Greater than 1; randomly select N values ​​from each of the two ranges. The value is used to obtain the pseudo-label image, as shown in the formula: ； in, This represents the generated pseudo-label reference image set, which contains a total of A pseudo-reference tag, This represents the input visible light image. Representing the Zhang's processed pseudo-label reference image; Representing the A random number from Gamma value obtained within the range .

2. The method according to claim 1, characterized in that: Step 1 involves using the Retinex algorithm to expand the original image dataset by adjusting the brightness of the original visible light images to simulate "overexposure" and "underexposure" phenomena, thus adding positive and negative samples to the dataset.

3. The method according to claim 1, characterized in that: A global-local image quality evaluator is used to calculate the quality scores of images in the pseudo-label reference image set, and selects the image with the highest quality score as the pseudo-label for the model to complete training. Global image quality evaluation is implemented using a referenceless spatial image quality evaluator (BRISQUE), which extracts local contrast and brightness of the image, uses a Natural Scene Statistics (NSS) model to capture the natural visual characteristics of the image, and then calculates the image quality score using a trained Support Vector Regression (SVM) model. Furthermore, local image quality evaluation measures the vividness and visual effect of the image by evaluating local pixel contrast and local color saturation, as shown in the formula: ;in, Representative with A pixel-centered patch, its size is ; Indicates in The pixel value corresponding to the position. The average pixel value of this patch; calculate To represent the local pixel contrast of an image; in addition, It represents the local color saturation of an image, using the saturation channels of the HSV color space to measure the color saturation of a pixel, where R, G, and B correspond to the pixel values ​​at the (x, y) positions of the three channels of the visible light image, respectively.

4. The method according to claim 1, characterized in that: L1-SSIM is proposed as the reconstruction loss function of the model to calculate the structural difference between the output image and the pseudo-label, as shown in the formula. ； in It applies to a fixed-size pixel window. Representing two different input images, Representative at The average pixel value of a window at the same location in two different input images. This represents the standard deviation of the corresponding window. for Covariance of windows at the same location Set as a fixed ratio to the range of input pixel values, and set them respectively. The L1 loss function calculates the absolute difference between the pseudo-label and the output image, where, The image output of the model is representative. Represents a pseudo-label image. As learnable parameters, they are obtained through iterative model learning and used to balance the loss function.

Images