Gated image low-illumination enhancement method based on improved RRDnet network

Abstract

The invention discloses a gating image low-illumination enhancement method based on an improved RRDnet network, and the method comprises the steps: carrying out the processing of a collected low-illumination gating image through employing an RGRRDnet model, and obtaining an enhanced image. In RGRRDnet, firstly, an image is divided into a reflection component, a noise component and an illumination component by using a residual decomposition network; secondly, locally adjusting an illumination component by adopting a two-dimensional adaptive Gamma correction method, and improving the problem of uneven illumination of a gated image; meanwhile, the reflection component generated by the decomposition network and the reflection component obtained through calculation are fused through an adaptive weighted fusion method, and the influence of noise estimation errors is effectively reduced; and finally, introducing a detail enhancement module, and enhancing the edge and texture information of the image by using DEConv convolution operation. Experimental results show that the method is superior to a current mainstream low-illumination enhancement algorithm in the aspects of objective indexes and subjective vision on a gated image data set.

Description

Technical Field

[0001] The present invention relates to the application of artificial intelligence technology in image processing, and in particular to a low-light enhancement method for gated images based on an improved RRDnet network. Background Art

[0002] In existing technology, gated cameras use filters or beam splitters to select specific wavelengths of light for imaging. While these cameras can capture images efficiently, these images lack color information and have low contrast, making them difficult to detect. In particular, in applications with poor lighting conditions, gated images suffer from low brightness, uneven illumination, and high noise, all of which greatly hinder their further use. Summary of the Invention

[0003] In order to solve the problems of low brightness, uneven illumination and high noise in the existing technology, a low-light enhancement method for strobed images based on an improved RRDnet network is proposed. The specific content is as follows:

[0004] A low-light enhancement method for gated images based on an improved RRDnet network is proposed. The low-light enhancement network RG_RRDnet model is used to process the collected low-light gated images to obtain enhanced images.

[0005] The RG_RRDnet is a modification of the RRDnet. Its structure includes a decomposition network and an illumination and detail enhancement network. The decomposition network includes three branches, each of which is composed of multiple residual blocks connected in series. The illumination and detail enhancement network includes an adaptive illumination correction module and a detail enhancement module.

[0006] In the RG_RRDnet network model, the steps for input strobe image processing include:

[0007] 1) Strobe image After being processed by the three branches of the decomposition network, the reflection component is obtained , illumination component and noise , ;

[0008] 2) Yes Remove the noise to get the corrected reflection component ;

[0009] 3) The corrected illumination component is obtained after processing by the adaptive illumination correction module ;

[0010] 4) and Multiply to get the corrected image, use the detail enhancement module to enhance the edge details of the image, and get the output image ;

[0011] In step 2): 2.1) First, the reflection component obtained in step 1) Further processing yields :By minus Divide by The resulting reflection component ;2.2) Then, and Perform weighted fusion to obtain ;

[0012] In step 3), an adaptive Gamma correction method is used to first calculate the local mean of the image, and then the Gamma parameter is adaptively adjusted based on the mean so that the local brightness of the image is appropriately corrected.

[0013] In the detail enhancement module of step 4),

[0014] 4.1) First, the features of the input image are extracted by the ordinary convolutional layer;

[0015] 4.2) The features extracted in step 4.1) are processed by detail enhancement convolution DEConv;

[0016] DEConv includes parallel center differential convolution CDC, angular differential convolution ADC, horizontal differential convolution HDC and vertical differential convolution VDC, as well as a normal convolution; among them, CDC convolution calculates the difference between a pixel in an image and its surrounding pixels; ADC convolution calculates the angular difference between a pixel in an image and its adjacent pixels; HDC convolution calculates the difference between a pixel in an image and its horizontally adjacent pixels; VDC convolution calculates the difference between a pixel and its vertically adjacent pixels;

[0017] In DEConv, details are captured by calculating the difference between the input feature map and the feature map after the input features are processed by different convolution kernels in parallel;

[0018] 4.3) The output of step 4.2) is activated and passed through a convolutional layer to further process and refine the extracted features.

[0019] The technical contributions of this invention mainly include:

[0020] (1) A residual decomposition network is proposed to deeply extract the reflection component, noise and illumination component in the image;

[0021] (2) An adaptive illumination correction module is proposed, which uses a two-dimensional adaptive gamma correction method to locally adjust the illumination component and improve the uneven illumination problem of the strobed image;

[0022] (3) An improved reflection component calculation method is proposed. The reflection component generated by the solution network is fused with the calculated reflection component through an adaptive weighted fusion method, effectively reducing the noise of the gated image.

[0023] (4) A detail enhancement module is proposed, which uses DEConv convolution operation to enhance the edge and texture information of the image. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is the overall structure diagram of the RG_RRDnet network (the red dotted line shows the improvements of the present invention);

[0025] Figure 2 This is the decomposed network diagram of RG_RRDnet;

[0026] Figure 3 It is the residual block structure diagram;

[0027] Figure 4 It is the structural diagram of the detail enhancement module;

[0028] Figure 5 It is a detail enhancement convolution structure diagram;

[0029] Figure 6 It is the experimental collection scene graph;

[0030] Figure 7 It is a diagram of the data set;

[0031] Figure 8 This is the experimental result of different algorithms comparison 1;

[0032] Figure 9 This is the experimental result of comparing different algorithms2. DETAILED DESCRIPTION

[0033] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0034] 1 Overview

[0035] To address the issues of low brightness, uneven illumination, and high noise in gated images, this paper uses a low-light enhancement algorithm to improve the quality of gated images. Since it is impossible to obtain paired images with normal illumination for supervised training, this paper improves the zero-shot model RRDnet by combining the characteristics of gated images and proposes a gated image low-light enhancement network RG_RRDnet.

[0036] 2 RG_RRDnet network

[0037] 2.1 Overall Network Design

[0038] This paper proposes a range-gated image low-light enhancement network (RG_RRDnet) based on an improved RRDnet network. The network structure is mainly divided into two parts: a decomposition network and a light and detail enhancement network. The specific structure is as follows: Figure 1 shown.

[0039] The network first inputs a low-light image , decomposed by the three-branch residual network, the reflection component is obtained , illumination component and noise ;

[0040] illuminance component The illumination is adaptively adjusted through the adaptive illumination correction module to obtain the illumination component ;

[0041] Reflection component and Perform weighted fusion to obtain the reflection component ,here It is a low-light image, the illumination component and noise Calculated;

[0042] Will and After multiplication, the detail enhancement module is used to enhance the edge details of the image to obtain the output image .

[0043] like Figure 1 As shown, the red dotted line in the figure is an improvement solution proposed by the present invention to solve the problems of low brightness, uneven illumination and high noise in the strobed image.

[0044] 2.2 Residual Decomposition Network

[0045] The RRDnet network only uses five ordinary convolutions when performing image decomposition, and the gated image is usually a grayscale image and lacks rich color information. This network structure is prone to information loss and insufficient feature extraction when decomposing the reflection component, illumination component and noise component, resulting in the extracted components affecting the effectiveness of subsequent processing. In order to better extract deep features in the gated image, the present invention proposes a solution to deepen the network structure by introducing more convolutional layers for deep feature learning. However, deep networks are prone to gradient vanishing or exploding problems. Therefore, a residual block is added to the network. The residual structure can effectively avoid the gradient vanishing problem by introducing jump connections, enhance the nonlinear modeling ability of the network, thereby improving the accuracy of reflection, illumination and noise decomposition, and effectively improving the effect of gated image decomposition.

[0046] The specific structure of the decomposed network of RG_RRDnet is as follows Figure 2 As shown. The network input is a gated image According to Retinex theory, the image can be decomposed into three parts: reflection component, illumination component and noise , the mathematical expression is as follows:

[0047]

[0048] The top branch is used to estimate the reflection component R1(x). It consists of two convolutional layers, 6 residual blocks and sigmoid blocks connected in series. The initial convolutional layer is used to extract low-level features from the input image. This layer can usually capture the basic information of the image, such as edges, textures, etc., which is especially important for low-light strobed images because the input image has low brightness and the information is relatively blurred. The role of the residual block is to gradually extract high-level features in the image, effectively retain useful features, adapt to the complex local changes of the reflection component in the image, and avoid information loss. The structure of the residual block is as follows: Figure 3 As shown in the figure, a convolutional layer is used after all residual blocks to further process the learned high-level features and obtain the final output of the reflection component. The sigmoid activation function is used in the last layer to limit the output value of the reflection component to [0,1], ensuring that the output is within a reasonable range and avoiding the output exceeding the actual brightness range. The middle and lower branches are used to estimate the illumination component. and noise , the illumination component estimation branch structure is the same as the reflection component, and the last layer of the noise branch uses the tanh activation function to reduce the noise value to [-1,1].

[0049] By using residual blocks for decomposition, the improved network can more effectively extract detailed information from the gated image. Furthermore, through a deeper network structure, it improves the ability to learn gated image features. This improvement not only improves the decomposition of low-light gated images, but also accelerates the training process and improves the network's ability to process gated images.

[0050] 2.3 Adaptive Lighting Correction Module

[0051] The gated image is a grayscale image with uneven illumination, resulting in some areas being too bright or too dark. The global gamma correction method used in the RRDnet network can adjust the overall brightness of the image when processing the illumination component of the image. However, because it does not consider local illumination variations, this global correction may cause some areas to be too bright or too dark, thus affecting image quality.

[0052] To address the above issues, the present invention adopts an adaptive Gamma correction method that dynamically adjusts the local illumination of each pixel, thereby avoiding the brightness imbalance problem caused by traditional global correction. This method calculates the local mean of the image and adaptively adjusts the Gamma parameter based on this value to ensure that the local brightness of the image is properly corrected, making the illumination distribution more uniform while preserving image details. This enables the network to improve the overall brightness of the image while avoiding the appearance of overly bright or dark areas, ensuring a more natural visual effect. The specific formula is as follows:

[0053]

[0054]

[0055] Where mean is the local mean of the input image, and 128 is the normalized baseline value. Adaptive adjustment of γ ensures adaptive balance of image brightness. When the local brightness of the image is greater than 128, γ is appropriately reduced to reduce areas of excessive brightness. When the local brightness of the image is less than 128, γ is increased to enhance dark details.

[0056] Ultimately, the brightness value at each pixel in the corrected image is calculated based on the adjusted gamma, achieving a more balanced global and local illumination. This method not only improves the overbrightness or underbrightness of traditional global gamma correction in low-light images, but also enhances detail in the gated image by dynamically adjusting the brightness range, providing more accurate illumination components for subsequent gated image enhancement.

[0057] 2.4 Improved calculation method of reflection component

[0058] In the RRDnet network, the reflectance component is obtained by subtracting noise from the low-light image. However, this method relies too heavily on noise estimation, especially in gated images, making the distinction between the subject and background unclear in the original image. Due to inaccurate noise estimation, using this method can lead to loss of image details and even blurring of image edges, directly affecting image quality.

[0059] In order to solve the above problems, the present invention proposes an improved method for calculating the reflection component. Specifically, the reflection component The reflection component is obtained by subtracting the noise from the low-light image and dividing it by the illumination component. The reflection component obtained by the decomposition network Perform weighted fusion to calculate. The calculation formula of weighted fusion is as follows:

[0060]

[0061] Here, a is an adaptive weighting coefficient ranging from [0 to 1], which is adaptively adjusted based on the quality of the reflection component. By balancing the contributions of the two reflection components, the final reflection component retains the global features learned by the decomposition network while also incorporating detailed information from the noise-subtracted low-light image.

[0062] The improved reflectance component calculation method effectively reduces reliance on noise estimation by adaptively adjusting the weighting coefficient a, thereby avoiding damage to image detail caused by inaccurate noise estimation. Furthermore, the fused reflectance components better preserve image detail and significantly improve edge clarity in the gated image, avoiding detail loss and blurring. This provides more accurate reflectance components and higher-quality input for subsequent image processing.

[0063] 2.5 Detail Enhancement Module

[0064] Due to its own imaging characteristics, the gated image usually has problems such as low brightness and uneven illumination, high noise and blurred edges. The RRDnet network directly multiplies the reflection component and the adjusted illumination component. Although it can obtain a preliminary enhancement result, due to the loss of details in the edges and local textures of the gated image, the result after direct multiplication may not be able to fully restore the details, especially the edge of the image is still blurred. The present invention proposes to further introduce a detail enhancement module on the intermediate result obtained by multiplying the reflection component and the adjusted illumination component, and through the detail enhancement convolution operation, effectively restore the edge details and texture information of the image, thereby significantly improving the clarity and detail performance of the image. The specific structure of the detail enhancement module is as follows Figure 4 shown.

[0065] First, the image is subjected to feature extraction through an ordinary convolution layer to effectively extract features from the image, which serves as the input layer in the image detail enhancement module to prepare for subsequent enhancement steps. The detail enhancement convolution is the core part of this module, which includes an ordinary convolution and center differential convolution CDC, angle differential convolution ADC, horizontal differential convolution HDC and vertical differential convolution VDC in parallel. It is specifically used to enhance the details of the image by calculating the difference between the input feature map and its feature map after processing with different convolution kernels to capture finer details. The ReLU activation function introduces nonlinearity to help the network learn more complex features, enhance significant details in the image, improve image clarity, remove negative values, and retain positive information. After DEConv and ReLU, the image passes through a convolution layer again to further process and refine the extracted features. The main purpose of this convolution layer is to further extract higher-level features and ensure that the details of the image are enhanced in further convolution operations. The structure of the core module detail enhancement convolution is as follows: Figure 5 shown.

[0066] CDC convolution can effectively highlight the central detail features in the image by calculating the difference between the pixel point in the image and its surrounding pixels, which helps to strengthen the details of the central area of ​​the image, especially the fine structures and textures. ADC convolution enhances the detailed features of the image in the angular direction by calculating the angular difference between the pixel point in the image and its adjacent pixels, which helps to enhance the edge of the image in a specific angular direction. HDC convolution focuses on the details in the horizontal direction of the image by calculating the difference between the pixel point in the image and its horizontal adjacent pixels, effectively enhancing the horizontal edges and horizontal textures in the image. VDC convolution is similar to HDC, but focuses on the details in the vertical direction of the image. By calculating the difference between the pixel point and its vertical adjacent pixels, it emphasizes the restoration of vertical edges and textures.

[0067] The detail enhancement convolution module combines four differential convolutions—CDC, ADC, HDC, and VDC—with regular convolution to enhance details in different directions and structures within the image. CDC emphasizes details in the central area of ​​the image, ADC processes angular differences, and HDC and VDC enhance details in the horizontal and vertical directions, respectively. In this way, detail enhancement convolution effectively restores edge and texture details in low-light images, enhancing visual clarity and providing richer structural information, especially in situations where image detail loss is severe.

[0068] 3 Experimental Design and Dataset

[0069] 3.1 Data Collection

[0070] The dataset for this experiment is a gated image dataset. Since there is currently no publicly available gated image dataset, the training and test data used in this experiment were collected by ourselves. To enhance the diversity of the dataset, the data was collected from multiple scenes, as shown in Figure 6.

[0071] The image size in the dataset is 1600×1088. 800 images are obtained from 800 sets of gated sequence images according to the range gated image reconstruction algorithm. The laser pulse width and gate width are 150ns. Some datasets are as follows Figure 7 shown.

[0072] 3.2 Experimental Design and Training Details

[0073] This experiment uses the Ubuntu 18.04 operating system, the CPU is Intel(R) Xeon(R) CPU E5-1620 v2 @3.70GHz, the GPU is NVIDIA Quadro RTX6000 graphics card, and the running memory is 24G.

[0074] The network of the present invention can directly use one picture for iterative training, and the training cycle is set to 1000 rounds.

[0075] 4 Experimental results and analysis

[0076] 4.1 Comparative Experiment

[0077] 1) Quantitative analysis

[0078] In order to ensure the fairness of the experimental results, the algorithms involved in the comparative experiments were trained using the gated image dataset, and HE [1] , MSR [2] , LIME [3] , EnlightenGAN [4] , SCI [5] ,Zero DCE [6] , Zero DCE++ [7] and RRDnet [8] For comparison, HE, MSR, and LIME are traditional methods, EnlightenGAN, SCI, Zero DCE, and Zero DCE++ use unsupervised training, and RRDnet and the proposed algorithm use zero-shot training. Because gated images lack corresponding normal-light images for comparison, supervised training is not possible and, therefore, is not compared with supervised methods.

[0079] The proposed algorithm uses the no-reference evaluation metrics NIQE and LOE. Lower values ​​indicate better image quality. Because the test images lack paired reference images, PSNR and SSIM cannot be used for evaluation. The experimental results are shown in Table 1.

[0080] Table 1 Comparative experimental results of different algorithms

[0081]

[0082] The data in the table above show that deep learning methods significantly outperform traditional methods in terms of NIQE scores. Zero DCE++, EnlightenGAN, and RRDnet have lower scores, indicating that these methods have significant advantages in generating natural and high-quality images. The proposed method's performance is 0.38 lower than that of the RRDnet algorithm, indicating that it can effectively restore details in low-light images while avoiding over-enhancement or excessive noise. In contrast, traditional enhancement methods HE, MSR, and LIME perform poorly, with higher NIQE scores. This indicates that traditional methods have certain problems when enhancing images, often resulting in unnatural brightness transitions or loss of detail.

[0083] In terms of LOE, deep learning algorithms also achieved good values. SCI and RRDnet showed advantages in brightness restoration and detail preservation, with their indicators significantly lower than those of other methods. The method proposed in this paper reduced the indicator by 59.4 points compared to the RRDnet algorithm, indicating that the method proposed in this paper not only effectively enhanced the brightness of the image, but also well preserved the details and textures in the gated image, adapting to the complex conditions of gated images in low-light conditions. In contrast, the traditional methods HE, MSR, and LIME achieved lower LOE scores, indicating that these methods performed only moderately well in brightness enhancement.

[0084] 2) Qualitative analysis

[0085] In order to demonstrate the visual effect of the algorithm of the present invention, two distinctive strobed images were selected as test images. The first image has rich content but blurred edge details, low overall image brightness, and uneven illumination, which can test the improvement effect of the algorithm of the present invention on uneven illumination and blurred edges. The second image has low overall brightness and a single object, which can effectively test the algorithm's brightness improvement effect. In order to ensure the fairness of the experimental results, all models were trained using the strobed data set. The effects of the two test images are as follows: Figure 8 and 9 shown.

[0086] Comparing the results of different algorithms in the first test image shows that traditional methods suffer from over-enhancement issues in enhanced images. HE suffers from overexposure, MSR suffers from some distortion, and LIME improves overall brightness but with noticeable noise. Among the deep learning methods, Zero DCE and Zero DCE++ significantly improve brightness but alter image color. EnlightenGAN and RRDnet perform better, but RRDnet lacks adaptive adjustment based on light intensity, resulting in partial overexposure in the center of the image. Our proposed algorithm, RG_RRDnet, builds on this by adaptively adjusting illumination, resulting in a more natural overall illumination. It also enhances detail and makes image edges more distinct, demonstrating that our proposed algorithm outperforms other algorithms in visual quality.

[0087] In the second picture, the enhanced images using the HE and MSR algorithms show obvious artifacts and over-enhancement. The noise of the LIME algorithm is particularly obvious. Zero DCE, Zero DCE++, and RRDnet improve brightness but have obvious noise. EnlightenGAN has less noise but the brightness improvement is not obvious. The algorithm RG_RRDnet proposed in this paper effectively suppresses noise while improving brightness, and its visual effect is better than other algorithms.

[0088] 4.2 Ablation Experiment

[0089] In order to verify the effectiveness of each module added to the RG_RRDnet network, the present invention conducts the following ablation experiments, as shown in Table 2.

[0090] Use the RRDnet network as the baseline;

[0091] Improvement 1: Change the decomposition network in the baseline to a residual decomposition network;

[0092] Improvement 2: Based on Improvement 1, the global gamma transform is replaced with the adaptive illumination correction module;

[0093] Improvement 3: Based on Improvement 2, the calculation method of the reflection component is improved;

[0094] Improvement 4: Add a detail enhancement module based on Improvement 3.

[0095] The ablation experiment results are shown in Table 3:

[0096] Table 2 Ablation experiment

[0097]

[0098] Table 3 Ablation experiment results

[0099]

[0100] Comparing the metrics after each improvement in the table shows that, as a control group, the RRDnet network provides baseline performance in both NIQE and LOE, achieving moderate scores. This demonstrates that the network can effectively restore image brightness and detail in low-light image enhancement tasks. In Improvement 1, the traditional decomposition network is replaced with a residual decomposition network. Results show that this improvement effectively improves image quality, particularly with a significant reduction in LOE. This demonstrates that the residual decomposition network is better able to process image detail. The introduction of the adaptive illumination correction module results in a slight reduction in both NIQE and LOE scores, demonstrating that adaptive illumination correction effectively adjusts the illumination distribution in the image and avoids over-enhancement. In Improvement 3, the calculation method of the reflectance component is modified. This improvement results in reductions in both NIQE and LOE. This optimization significantly reduces image noise, resulting in more natural and high-quality images. The addition of the detail enhancement module further improves image quality, particularly with reductions in NIQE and LOE, demonstrating that the detail enhancement module addresses the edge blurring issue in the gated image to some extent.

[0101] 5. Summary

[0102] This paper proposes a novel low-light enhancement network, RG_RRDnet, targeting the characteristics of gated images. First, a residual decomposition network is proposed to separate the image into reflectance, noise, and illumination components. A two-dimensional adaptive gamma correction method is then used to locally adjust the illumination component, improving the uneven illumination problem in gated images. An adaptive weighted fusion method is also proposed to fuse the reflectance component generated by the decomposition network with the calculated reflectance component, effectively reducing the impact of noise estimation errors. Finally, a detail enhancement module is introduced to enhance the image's edge and texture information using the DEConv convolution operation. Experimental results demonstrate that the proposed method outperforms current mainstream low-light enhancement algorithms in both objective and subjective visual performance on gated image datasets.

[0103] References

[0104] [1]Pizer S M. Contrast-limited adaptive histogram equalization: Speedand effectiveness stephen m. pizer, r. eugene johnston, james p. ericksen,bonnie c. yankaskas, keith e. muller medical image display research group[C] / / Proceedings of the first conference on visualization in biomedicalcomputing, Atlanta,Georgia. 1990, 337: 1.

[0105] [2]Rahman Z, Jobson D J, Woodell G A. Multi-scale retinex for colorimage enhancement[C] / / Proceedings of 3rd IEEE international conference onimageprocessing. IEEE, 1996, 3: 1003-1006.

[0106] [3]Guo, Xiaojie, Yu Li, and Haibin Ling. "LIME: Low-light imageenhancement via illumination map estimation." IEEE Transactions on imageprocessing26.2 (2016): 982-993.

[0107] [4]Jiang Y, Gong X, Liu D, et al. Enlightengan: Deep lightenhancement without paired supervision[J]. IEEE transactions on imageprocessing, 2021, 30: 2340-2349.

[0108] [5]Long M, Teng M, et al. "Toward fast, flexible, and robust low-light image enhancement." Proceedings of the IEEE / CVF conference on computervision and pattern recognition. 2022.

[0109] [6]Guo C, Li C, Guo J, et al. Zero-reference deep curve estimationfor low-light image enhancement[C] / / Proceedings of the IEEE / CVF conference oncomputer vision and pattern recognition. 2020: 1780-1789.

[0110] [7]Li C, Guo C, Loy C C. Learning to enhance low-light image viazero-reference deep curve estimation[J]. IEEE transactions on patternanalysis and machine intelligence, 2021, 44(8): 4225-4238.

[0111] [8]Zhu A, Zhang L, Shen Y, et al. Zero-shot restoration ofunderexposed images via robust retinex decomposition[C] / / 2020 IEEEInternational Conference on Multimedia and Expo (ICME). IEEE, 2020: 1-6。

Claims

1. A low-light enhancement method for strobed images based on an improved RRDnet network, characterized by The low-light enhancement network RG_RRDnet model is used to process the collected low-light strobed image to obtain the enhanced image; The RG_RRDnet is a modification of the RRDnet. Its structure includes a decomposition network and an illumination and detail enhancement network. The decomposition network includes three branches, each of which is composed of multiple residual blocks connected in series. The illumination and detail enhancement network includes an adaptive illumination correction module and a detail enhancement module. In the RG_RRDnet network model, the steps for input strobe image processing include: 1) Strobe image After being processed by the three branches of the decomposition network, the reflection component is obtained , illumination component and noise , ; 2) Yes Remove the noise to get the corrected reflection component ; 3) The corrected illumination component is obtained after processing by the adaptive illumination correction module ; 4) and Multiply to get the corrected image, use the detail enhancement module to enhance the edge details of the image, and get the output image ; In step 2): 2.1) First, the reflection component obtained in step 1) Further processing yields :By minus Divide by The resulting reflection component ;2.2) Then, and Perform weighted fusion to obtain ; In step 3), an adaptive Gamma correction method is used to first calculate the local mean of the image, and then the Gamma parameter is adaptively adjusted based on the mean so that the local brightness of the image is appropriately corrected. In the detail enhancement module of step 4), 4.1) First, the features of the input image are extracted by the ordinary convolutional layer; 4.2) The features extracted in step 4.1) are processed by detail enhancement convolution DEConv; DEConv includes parallel center differential convolution CDC, angular differential convolution ADC, horizontal differential convolution HDC and vertical differential convolution VDC, as well as a normal convolution; among them, CDC convolution calculates the difference between a pixel in an image and its surrounding pixels; ADC convolution calculates the angular difference between a pixel in an image and its adjacent pixels; HDC convolution calculates the difference between a pixel in an image and its horizontally adjacent pixels; VDC convolution calculates the difference between a pixel and its vertically adjacent pixels; In DEConv, details are captured by calculating the difference between the input feature map and the feature map after the input features are processed by different convolution kernels in parallel; 4.3) The output of step 4.2) is activated and passed through a convolutional layer to further process and refine the extracted features.

2. The low-light enhancement method for gated images based on the improved RRDnet network according to claim 1, wherein In step 1), the three branches of the decomposition network are the same and are used to estimate the reflection component and illumination component The sigmoid activation function is used after the two branches to estimate the noise The tanh activation function is used after one branch of .

3. The low-light enhancement method for gated images based on the improved RRDnet network according to claim 1 or 2, characterized in that In step 1), any branch of the decomposition network is composed of the first and last two convolutional layers and the six residual blocks connected in series in the middle; The initial convolutional layer starts from Extract low-level features; each residual block gradually extracts Advanced features in All residual blocks are followed by a convolutional layer to further process the learned high-level features and obtain the final output of the reflection component.

4. The low-light enhancement method for gated images based on the improved RRDnet network according to claim 1 is characterized in that The formula for step 2.1) is: ; The calculation formula of weighted fusion in step 2.1) is: , Wherein, a is an adaptive weighting coefficient with a value range of [0,1], and a is adaptively adjusted according to the quality of the reflection component.

5. The low-light enhancement method for gated images based on the improved RRDnet network according to claim 1 is characterized in that The formula of the adaptive gamma correction method in step 3) is as follows: Among them, mean is the local mean of the input image, and 128 is the standardized reference value; By adaptively adjusting γ, the image brightness is adaptively balanced. When the local brightness of the image is greater than 128, γ is reduced to reduce the area with excessive brightness. When the local brightness of the image is less than 128, γ is increased to enhance the dark details. Finally, the brightness value of the rectified image at each pixel position is calculated based on the adjusted γ.

Images