Low-light image enhancement method based on combination of retinex and wavelet transform

Abstract

The application relates to the field of image processing and computer vision, and discloses a low-light image enhancement method based on the combination of Retinex and wavelet transform, aiming to improve the visual quality and detail definition of images in low-light environments. The method effectively handles noise, uneven illumination and detail blur in low-light images by combining Retinex theory, wavelet transform and deep learning technology. Through the illumination estimator and the degradation restorer, the method can generate enhanced illumination information and repair various forms of degradation, including noise, artifacts, underexposure / overexposure and color distortion. In addition, the method also uses a gating mechanism and the self-attention mechanism of the Transformer to selectively fuse features, so as to retain edge and detail information and promote global enhancement and restoration. Experimental results show that the method is superior to existing advanced methods on multiple datasets.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing and computer vision, and particularly relates to a low-light image enhancement method based on the combination of Retinex and wavelet transform. BACKGROUND

[0002] With the advancement of technology, the demand for extracting clear and accurate information from images is constantly expanding. In the field of image processing and computer vision, low-light image enhancement is an important research direction. However, environmental factors often hinder the capture of high-quality images. Night or low-light images often encounter noise, uneven lighting, and blurred details, all of which can reduce visual quality and lead to incomplete information. These limitations not only impair visual perception, but also have a chain effect on advanced computer vision tasks such as object recognition, autonomous driving, and image classification, ultimately reducing the accuracy of downstream processing tasks. To address these challenges, various low-light image enhancement algorithms have been proposed, including basic methods such as gamma correction and histogram equalization. However, these methods often result in over-enhancement and image distortion, as their success is largely dependent on the accuracy of manually set priors. In real-world scenarios, the complexity of lighting conditions makes the determination of low-light factors complex. Traditional cognitive methods, such as the Retinex theory, are inspired by the human visual system and decompose images into two components - illumination and reflectance - to simulate how human perception interprets color and brightness under different lighting conditions. According to the Retinex theory, the purpose of low-light enhancement is to mitigate the effects of low-light illumination, amplify the reflectance component, and restore details and colors. However, traditional Retinex methods require complex parameter adjustments and often introduce significant noise and artifacts, especially in low-light conditions.

[0003] With the advancement of deep learning, convolutional neural networks (CNNs) and Transformer-based models have set new benchmarks for low-light image enhancement. CNNs effectively capture local image features such as edges and textures, which are crucial for detail restoration and noise suppression in low-light environments. However, in low-light conditions, image details and textures are often lost, while accurate separation of illumination information remains crucial. Traditional convolutional operations struggle to balance detail preservation with accurate illumination estimation. In addition, traditional convolutional up-sampling and down-sampling mechanisms can degrade details, as texture and edge occlusion or blurring are common in low-light images. Direct convolutional down-sampling can mix noise with details, exacerbating this problem, especially in dim and degraded regions where noise is often amplified incorrectly.

[0004] In feature fusion during upsampling and downsampling, standard channel concatenation is usually used. This technique connects features along the channel dimension, lacks the ability to selectively emphasize relevant features, and often leads to the accumulation of redundant information, which blurs key details and hinders the model's ability to effectively capture key features. In addition, since CNN mainly captures local features, relying solely on these features may not be sufficient to address global lighting defects in low-light images. Although the Transformer model can achieve a global perspective and model long-range dependencies more effectively with its self-attention mechanism, it can enhance the details and structure recovery in low-light images. However, the application of the original Transformer architecture is computationally intensive and involves a complex training process, making it challenging to adopt in real-time low-light image enhancement tasks.

[0005] Therefore, it is necessary to provide a low-light image enhancement method based on the combination of Retinex and wavelet transform to solve the above technical problems. SUMMARY

[0006] Since images taken in low-light conditions are often disturbed by noise and uneven illumination distribution, resulting in poor image quality and blurred details. Existing methods often amplify noise and produce artifacts during the enhancement process, making it difficult to ensure the integrity and clarity of image structure. In order to make up for the shortcomings of existing methods, the present application provides a low-light image enhancement method based on the combination of Retinex and wavelet transform, aiming to improve the visual quality and detail clarity of images in low-light environments.

[0007] To achieve the purpose of the present application, the present application proposes a RetinexWT architecture, which is a novel method combining Retinex theory and wavelet transform for robust low-light image enhancement. RetinexWT mainly includes an illumination estimator and a degradation restorer. On the basis of the traditional Retinex model, the present application introduces a perturbation term to the reflectance and illumination components to accurately simulate the degradation that usually occurs in low-light conditions. The illumination estimator is combined with the wavelet transform to generate enhanced illumination information for low-light images, while the degradation restorer repairs and suppresses various forms of degradation, including noise, artifacts, under / overexposure, and color distortion. By integrating a wavelet feature decomposer into the downsampling module of the degradation restorer, the model can enhance brightness and suppress noise of different frequency components, respectively. In addition, a gating mechanism is used to selectively fuse the features of downsampling and upsampling, enabling the model to suppress noise while preserving edge and detail information. The self-attention mechanism of the Transformer is further utilized to capture long-range dependencies within the image, thereby facilitating accurate global enhancement and restoration.

[0008] The low-light image enhancement method based on the combination of Retinex and wavelet transform provided by the present application comprises the following steps:

[0009] S1, decomposing the input low-light image into a reflectance component and an illumination component using Retinex theory;

[0010] S2, introducing a perturbation term to simulate degradation under low-light conditions, generating a degraded reflectance component and an illumination component;

[0011] S3, performing frequency domain analysis on the illumination component through wavelet transform to obtain illumination features, and combining the illumination features and the wavelet transform results to perform enhancement processing on the low-light image;

[0012] S4, using a gating mechanism to selectively fuse down-sampled and up-sampled features to suppress noise and preserve edge and detail information;

[0013] S5, applying the self-attention mechanism of the Transformer to capture long-range dependencies within the image, achieving global enhancement and restoration.

[0014] Preferably, the wavelet transform is Haar wavelet transform, which is used to decompose the input features into high-frequency and low-frequency components.

[0015] Preferably, the gating mechanism includes a GatedFusion Module, which is used to adaptively determine which features are more important and control the contribution of different channel features.

[0016] Preferably, the self-attention mechanism of the Transformer is implemented through a hybrid attention network based on the Transformer to ensure effective modeling of long-range dependencies during enhancement.

[0017] Preferably, the up-sampling and down-sampling processes are divided into two levels, and each level of down-sampling consists of an Enhanced Illumination-Guided Attention Model (EIGAM) and a Wavelet Transform Feature Decomposer Downsampling (WTFDown); each level of up-sampling consists of a deconv 2x2 (stride = 2) and a conv 1x1 and an EIGAM.

[0018] Preferably, the EIGAM includes the following steps:

[0019] The input features are first subjected to illumination-guided attention (IGA), and the IGA also uses the illumination features generated in the IE as input to guide the calculation of attention;

[0020] Nonlinear activation freedom (NAF) is used to reduce computational complexity, helping to simplify the model architecture and reduce the consumption of computing resources;

[0021] After IGA and NAF, residual connections are performed to alleviate the gradient disappearance and retain the original features. Finally, the output features are obtained through a layer normalization (LN) and a feedforward network (FFN).

[0022] Preferably, the NAF comprises the following steps:

[0023] Use a point-by-point convolution (1x1 conv) to adjust the number of channels of the input feature and achieve linear transformation of the channel dimension;

[0024] Depth-wise Convolution performs independent convolution on each channel to better extract spatial features such as edges and details;

[0025] After simplified channel attention (SCA) and SimpleGate, the number of channels is restored by point-by-point convolution (1x1 conv) and connected with the input feature residual to obtain the first part of the output;

[0026] The first part of the output is convolved and gated to obtain the second part of the output, which is added to the first part of the output through a residual connection to form the final output.

[0027] Preferably, the WTFDown comprises the following steps:

[0028] The input features first undergo a 1*1conv operation to improve nonlinearity, and then the spatial features are converted into four frequency domain components through Haar wavelet transform, namely a low-frequency component A and three high-frequency components, where the high-frequency components can be divided into horizontal high-frequency components H, vertical high-frequency components V and diagonal high-frequency components D;

[0029] Low-frequency features are represented by convolution to learn features and extract global structural information. High-frequency features are concatenated by concatenating the horizontal, vertical, and diagonal high-frequency components, and then dimensionality reduction and nonlinear operations are performed through point-by-point convolution (1×1conv) and BatchNorm.

[0030] The two parts are added together to obtain global and local composite features, which retains the overall information while enhancing edges and details and balancing the influence of different frequency components in the image.

[0031] Compared with related technologies, the low-light image enhancement method based on the combination of Retinex and wavelet transform provided by the present invention has the following beneficial effects:

[0032] 1. The present application proposes a hybrid attention network based on Transformer for low-light image enhancement, ensuring effective modeling of long-range dependencies in the enhancement process. The method of the present application utilizes the frequency domain information obtained by wavelet transform in combination with the Retinex model to obtain more accurate illumination information.

[0033] 2. The present application avoids traditional downsampling and instead introduces a Haar wavelet decomposer to preserve information. By decomposing the input features into high-frequency and low-frequency components through wavelet transform, the structural and detail information of the image is preserved and separated during downsampling, thereby improving the enhancement performance. In addition, by providing frequency domain features, the model can utilize these rich information during the reconstruction process, and perform more targeted noise suppression, edge refinement, noise removal, etc. according to the frequency domain features.

[0034] 3. The present application introduces a gating mechanism to selectively fuse features from upsampling and downsampling. Compared with direct channel concatenation, this mechanism can adaptively determine which features are more important. By controlling the contribution of different channel features, a better balance between noise suppression and detail preservation is achieved. During feature fusion, convolution and activation operations are applied to further refine and enhance the information flow, reducing the risk of information loss during upsampling and downsampling.

[0035] 4. Qualitative and quantitative experiments show that the RetinexWT of the present application is superior to all previous Retinex-based deep learning methods and achieves results superior to the state-of-the-art (SOTA) method on multiple datasets. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a schematic diagram of feature decomposition using Haar wavelet transform in the present application;

[0037] Figure 2 is an architecture diagram of a low-light image enhancement method based on Retinex combined with wavelet transform provided by the present application;

[0038] Figure 3 is a detail diagram of IGTW in the present application;

[0039] Figure 4 is (a) a gating fusion module (GFM) and (b) a nonlinear activation free (NAF);

[0040] Figure 5 is a schematic diagram of the WTFDown module of the present application introducing frequency domain feature mapping into the network for downsampling;

[0041] Figure 6 is a qualitative experimental result diagram of the SOTA method on LOL-v1 (52) and LOL-v2-synthetic (45) datasets;

[0042] Figure 7 Figures for qualitative experimental results of SOTA methods on LOL-v2-real (45) dataset;

[0043] Figure 8 Figures for visual results of LIME (11), ExDark (19), VV (34), MEF (25), DICM (19) of the present application;

[0044] Figure 9 Figures for visual comparison of the impact of the method of the present application on low-light object detection. DETAILED DESCRIPTION

[0045] The present application is further described below in conjunction with the accompanying drawings and embodiments.

[0046] Figure 2 The comprehensive architecture of the method of the present application is described. As shown in Figure 2 (a), the RetinexWT proposed by the present application mainly consists of two parts, Illumination Estimator and Corruption Restorer. Among them, the Illumination Estimator of the present application is influenced by the traditional Retinex model, and a disturbance term combined with frequency domain information is introduced on this basis. In the estimation of illumination features, wavelet transform is also combined. The design of Corruption Restorer is based on IGTW with Wavelet (IGTW), as shown in Figure 2 (b), the basic unit of IGTW is Enhanced Illumination-Guided Attention Model (EIGAM), which consists of Illumination-Guided Attention (IGA), Nonlinear Activation Free Block (NAF), normalization (LN) and feedforward network (FFN). Figure 3 (a) shows the details of IGTW.

[0047] Retinex-based Framework

[0048] In the field of low-light image enhancement, Retinex theory is used to simulate the perception of human visual system to brightness and color. The traditional Retinex algorithm decomposes the input image I into the reflection component R and the illumination component L, which is expressed as

[0049]

[0050] denotes the multiplication of elements. This method can effectively deal with the illumination changes and color distortion in the image, but in low light conditions, since the method does not take into account the noise and artifacts generated by insufficient light and other factors, it may be amplified in the enhancement process. To overcome these deficiencies, the present application adopts the disturbance modeling proposed in [2], unlike the original formula which regards I as undamaged, this study introduces a disturbance term through the illumination component L and the reflection component R to simulate the loss generated in low light conditions

[0051]

[0052] and denotes the loss, where R is regarded as an image with good exposure. Next, the illumination map F is obtained by feature extraction through convolution Using element-wise multiplication to light up the low light image I, where The formula is expressed as

[0053]

[0054] which can be simplified as

[0055]

[0056] where I lu is the lightened image, C represents the overall loss term, that is, the noise and artifacts amplified in the enhancement process and the exposure problems and color distortion caused in the enhancement process The sum. Therefore, the RetinexWT of the present application can be expressed as

[0057] (I lu ,F lu )=IE(I,L p ),I en =CR(I lu ,F lu ) (5)

[0058] where IE denotes Illumination Estimator, and CR denotes Corruption Restorer. L p =mean c (I), mean c denotes the operation of the average value of each pixel along the channel dimension, used to evaluate the overall illumination level of the image. IE takes I and L p as input, and the output result is the lightened image I lu and the illumination feature F lu . CR takes Ilu and F lu As input, to process the noise and distortion introduced in the image, and finally generate the repaired image I en In addition, in order to improve the performance of the model, the model uses a convolutional neural network (CNN) to extract illumination features and process illumination mapping, combined with image information, to effectively enhance the image under complex illumination conditions.

[0059] IlluminationEstimator

[0060] As Figure 2 (a) shows that the original low-light image I and the illumination prior L p is obtained by calculating the average value of the pixel dimension of I channel are merged and the channel dimension is increased as input. First, 1×1 convolution is used to fuse I and L p , that is, the illumination prior is applied to the low-light image. Then a depth separable 5×5 convolution is used to upsample the input and further extract features to generate the preliminary illumination feature. It is known that the low-frequency information of wavelet transform mainly contains the overall illumination and structural information of the image, and L p ignores the details of color components and focuses on the overall brightness and illumination information of the image, so the low-frequency information of the illumination prior L p focuses on the essence of illumination and focuses on illumination information. Therefore, in order to generate more accurate illumination features and make them more robust under low-light and complex illumination conditions, the present application performs wavelet transform on the illumination prior L p , combines the obtained low-frequency component with the preliminary obtained illumination feature to generate the final illumination feature F lu , where the feature dimension n feat is set to 40. Finally, another 1×1 convolution layer is used for downsampling to restore the 3-channel illumination map , and then multiplied element by element with the original low-light image I to obtain the point light image I lu .

[0061] Illumination GuidedTransformerwithWavelet

[0062] In the RetinexWT framework, the CorruptionRestorer (IGTW) is composed of an encoder and a decoder based on IlluminationGuidedTransformerwithWavelet, which introduces a gating fusion mechanism. The encoder represents the downsampling process, while the decoder represents the upsampling process. As Figure 2(b) is shown. Both up-sampling and down-sampling processes are divided into two levels. First, the point light image I lu is down-sampled by conv 3x3 (stride=2) to match the dimension of the illumination feature F lu for subsequent operations. Next, two-level down-sampling is performed to extract deep features, and each level of down-sampling consists of an Enhanced Illumination-Guided Attention Model and a Wavelet Transform Feature Decomposer Down (WTFDown). Down-sampling with WTFDown can effectively preserve the high-frequency information of the image, improve the detail recovery ability of IGTW, and effectively suppress noise amplification in the enhancement process. After each WTFDown, the width and height of the image are halved, and the feature dimension is doubled, i.e., the input feature dimension is C, the feature dimension after one level of down-sampling is 2C, and the deepest feature dimension after two levels of down-sampling is 4C. After extracting features through down-sampling, IGTW needs to continue to up-sample to restore the image. Like down-sampling, up-sampling also has two levels, each of which consists of a deconv 2x2 (stride=2), a conv 1x1, and an EIGAM. After each deconv, the size of the image shape doubles, while the feature dimension is halved. Then, the output of the deconv is fused with the output of the EIGAM in the corresponding level of down-sampling through a Gated Fusion Module (GFM), which aims to reduce the loss of image information during down-sampling, flexibly and efficiently connect the up-sampling and down-sampling features, and achieve the effects of noise suppression, detail enhancement, and information fusion. Finally, the image is convolved by a conv 3x3 (stride=2) to reduce the feature dimension, and restored to a three-channel RGB format. The restored image and the point light image I lu are connected to obtain the final enhanced image I en .

[0063] EIGAM. The structure of the Enhanced Illumination-Guided Attention Model (EIGAM) is shown in Figure 3 (a). In EIGAM, the input feature F in is first subjected to Illumination-Guided Attention (IGA), and at the same time, IGA also uses the illumination feature F luAs the input to guide attention, the calculation is followed by Nonlinear Activation Free (NAF) to reduce computational complexity, helping to simplify the model architecture and reduce the consumption of computing resources. After IGA and NAF, residual connection is performed to alleviate gradient vanishing and preserve original detail information. Finally, an output feature F is obtained through a Layer Normalization (LN) and a Feed-Forward Network (FFN) out . Wherein, IGA is used to process the illumination feature F lu and guide the calculation of multi-head self-attention, as Figure 3 (b). And in order to solve the problem of huge computational cost of global multi-head self-attention in Transformer, IGA adjusts the input feature to k head tokens X (HW*c), and further divides it into k heads X i (HW*dk), for each head, through three fully connected layers to calculate and

[0064]

[0065] Wherein W Q,i , W K,i and W V,i , i is the learnable parameter of the fc layer, and T represents the matrix transpose. Then, the illumination feature F lu encodes the illumination information, provides the global illumination condition, and adjusts it to k head tokens Y (HW*c) and also divides it into k heads Y i (HW*dk) to guide the calculation of self-attention of each head:

[0066]

[0067] Wherein α i is a scaling parameter in the scientific system, used to adjust the calculation result. Finally, the k head features are obtained through a fully connected layer and position encoding to obtain the output feature.

[0068] GFM. GFM (Gated Fusion Module) is used to fuse features in up-sampling and down-sampling in IGTW (Illumination Guided Transformer with Wavelet), and its structure is as Figure 4 (a). GFM (firstly, up-sampling feature F u (H*W*C) and down-sampling F d(H*W*C) are concatenated along the channel dimension for information integration, followed by a Point-wise Convolution for channel information mixing, integrating feature information and using Depth-wise Convolution to process local spatial information, preserving spatial structure, especially edges and details. After convolution processing, GFM divides the mixed features along the channel dimension into two parts, F gate (H*W*C) and F content (H*W*C). Where F gate After processing by the GELU activation function, it is used as a gating feature to control the information flow. Then F content and F gate Hadamard Product is performed to achieve adaptive filtering of feature information. Finally, the gated feature is combined with the original input feature ResidualConnection to obtain the output feature. Compared with the direct channel connection for detail preservation, GFM will further extract and strengthen the information in the flow process when fusing the features in the down-sampling and up-sampling process, reducing the loss of information in the down-sampling and up-sampling process, and maintaining the consistency of the features, improving the representativeness of the features, which can help the model better learn and express the brightness and color distribution in the low-light image, thereby improving the enhancement quality.

[0069] NAF. The present application uses a module NAF (Nonlinear Activation Free) which removes the traditional nonlinear activation function, aiming to reduce the computational complexity and improve the performance. As shown in Figure 4(b)As shown, the NAF first uses a pointwise convolution (1x1 conv) to adjust the channel number of the input feature, realizing linear transformation of the channel dimension, and then performs independent convolution on each channel through Depth-wise Convolution, which can better extract spatial features such as edges and details. Then, after passing through the simplified channel attention (SCA) and SimpleGate, the channel number is restored by the pointwise convolution (1x1 conv), and the first part of the output is obtained by connecting the residual connection with the input feature. Finally, the first part of the output is convolved and gated to obtain the second part of the output, which is added to the first part of the output through residual connection to form the final output. The core feature of the NAF module is that it does not contain traditional nonlinear activation functions (such as ReLU, Sigmoid, etc.), but replaces the channel attention / GELU with the simplified channel attention (SCA) and SimpleGate. The NAF module removes the nonlinear activation and uses simple element-wise operations, which not only reduces the computational complexity, but also preserves the feature information, making it an efficient network structure design. In addition, the depthwise convolution and SCA of the NAFBlock can enhance the regions with significant differences in brightness and contrast in the feature map, i.e., the features at the edges and contours are given additional attention and enhancement. Through a series of efficient feature processing operations, the model can better restore the edge and boundary information of the image in low-light environments.

[0070] Wavelet Transform Feature Decomposer Downsampling

[0071] Wavelet Transform is a signal processing technique that can decompose a signal into sub-signals of different frequencies for analysis in time and frequency domains. In image processing, Wavelet Transform is used to decompose an image into multiple sub-bands, including low-frequency components and high-frequency components. The low-frequency components preserve the overall brightness and structure of the image, facilitating the enhancement of global brightness; the high-frequency components contain detailed information such as edges and textures, as well as noise. Therefore, in the task of low-light image enhancement, the low-frequency and high-frequency parts can be processed separately to enhance the brightness while preserving the delicate texture and edge information of the image. However, previous studies often directly apply Wavelet Transform in the down-sampling layer of the network, replacing the spatial domain features with frequency domain features, which may cause the loss of spatial information, making the enhanced image blurred or distorted. To solve this problem, the present invention introduces a method that combines frequency domain and spatial domain information, namely Wavelet Transform Feature Decomposer Downsampling (WTFDown), to introduce frequency domain feature mapping for down-sampling in the low-light enhancement network. Its structure is as followsFigure 5 In WTFDown, the input feature is first passed through a 1*1 conv operation to improve the nonlinearity, and then the spatial feature is converted into four frequency components through Haar wavelet transform, which are a low-frequency component A, and three high-frequency components. Among them, the high-frequency components can be divided into horizontal high-frequency components H, vertical high-frequency components V and diagonal high-frequency components D. The low-frequency feature is learned through convolution for feature representation, and the global structure information is extracted. The high-frequency feature is spliced with the horizontal, vertical and diagonal high-frequency components, and the dimension reduction processing and non-linear operation are performed through point-by-point convolution (1*1conv) and BatchNorm. This processing method enhances the edge and local detail information in the high-frequency component, which helps to retain the details in the low-light image. Finally, the two parts are added to obtain the composite feature of the global and local, which retains the overall information while enhancing the edge and details, balancing the influence of different frequency components in the image. WTFDown uses Haar wavelet transform to simply and efficiently decompose the image signal, introduces frequency information into the network, and realizes more comprehensive feature representation. Figure 1 The process of Haar wavelet transform is shown. Among them, A(X) represents a low-pass filter, and the original data is low-pass filtered. D(X) represents a high-pass filter, and the original data is high-pass filtered. cA represents a low-frequency component, and cH, cV, and cD represent a horizontal high-frequency component, a vertical high-frequency component, and a diagonal high-frequency component, respectively. After each filter, the feature size is reduced by half.

[0072] Specifically, the input feature X (H*W*C) is first passed through a 1*1 conv operation to improve the nonlinearity, and then the spatial feature is converted into four frequency components through Haar wavelet transform, which are a low-frequency component A, and three high-frequency components. Among them, the high-frequency components can be divided into horizontal high-frequency components H, vertical high-frequency components V and diagonal high-frequency components D. For each channel X c (H*W), the wavelet transform is calculated as follows:

[0073]

[0074] Among them, A c (i,j) and D c (i,j) represent the low-frequency approximation coefficient and the high-frequency approximation coefficient of channel c, respectively, i is the row index, ranging from 1 to H. j is the column index, ranging from 1 to Next, the Haar wavelet transform is applied to the approximation coefficients and detail coefficients of each column, denoted as:

[0075]

[0076] Among them, A c is the low-frequency approximation coefficient of a single channel, Hc , V c , D c are the high-frequency detail coefficients in horizontal, vertical and diagonal directions, respectively. Then, the three high-frequency components are concatenated and reduced dimension using point-wise convolution to remove unimportant information (noise, etc.) and preserve key features (edge and detail information), resulting in the final high-frequency feature F l and the low-frequency feature F h , whose formula is:

[0077]

[0078] where, denotes the batch normalization operation. Finally, the obtained high-frequency and low-frequency features are added to provide a down-sampled feature map containing frequency domain features.

[0079] Compared with the related art, the low-light image enhancement method based on the combination of Retinex and wavelet transform provided by the present application has the following technical effects:

[0080] The present application compares the proposed method with various deep learning-based SOTA methods listed in Table 1, including SID, 3DLUT, DeepUPE, SCI, RetinecNet, etc. The data sets used for comparison include synthetic data of LOL-v1 and real data and synthetic data of LOL-v2. In order to compare fairly, the present application uses the official pre-trained model and its public code of each method to obtain quantitative results.

[0081] Quantitative analysis: The evaluation indicators for comparison are PSNR and SSIM, where PSNR reflects the overall enhancement quality, and the higher the value, the better the performance. SSIM measures the preservation of high-frequency details and structural information, and the higher the value, the better the preservation effect of image content. The RetinexWT method proposed by the present application shows a significant performance advantage over the above SOTA methods on the LOL data set.

[0082] In particular, compared to other Retinex-based deep learning SOTA methods, including SID, DeepUPE, SCI, LIME, RetinexNet, RUAS, FIDE, KinD, and Retinexformer, the method of the present invention achieves significant improvements in PSNR and SSIM on LOL-vl and LOL-v2 datasets. In terms of PSNR, RetinexWT achieves an enhancement of 0.27 dB, 0.07 dB, and 0.35 dB on LOL-vl, LOLv2-real, and LOL-v2-synthetic datasets, respectively. Similarly, SSIM improvements of 0.007 dB, 0.02 dB, and 0.001 dB are observed on the same datasets, which highlights the superior ability of the method of the present invention in balancing enhanced quality and detail preservation. As shown in Table 4.1, the effectiveness of the method of the present invention is further highlighted.

[0083] Quantitative analysis: To provide a more comprehensive and intuitive comparison, the present invention analyzed RetinexWT

[0084] Methods were visually evaluated against other state-of-the-art (SOTA) methods. Figure 6 and Figure 7 Taken from the LOL dataset (LOL-vl and LOL-v2), where the input consists of severely degraded low-light images. The results reveal several limitations of existing methods: noise amplification (e.g., RetinexNet in Figure 7 ), under- or over-exposure (e.g., LEDNet and RUAS in Figure 6 ), color distortion (e.g., Restormer in Figure 7 ), and the introduction of black spots and artifacts (e.g., EnGAN and Kind in Figure 7 ). By comparing the results with Figure 6 , Figure 7 , the clear advantages of the present method are shown, which, upon closer inspection, achieves superior visual results.

[0085]

[0086] Table 1: Performance comparison of various methods on LOL(v1(52)and v2(45))datasets in terms of PSNR and SSIM. Retinex indicates whether the method is based on Retinex theory. The bold-red font indicates the best result, and the second-best result is formatted with bold-blue font.

[0087] In contrast, the method of the present invention exhibits significant improvements in both global and local enhancement. It can effectively manage exposure levels, significantly improve visibility and contrast, and minimize noise, thereby providing outstanding visual quality. Furthermore, the present invention demonstrates unpaired benchmark results, where the method of the present invention exhibits precise exposure control and vivid color restoration. This further highlights its outstanding generalization ability in various low-light image scenarios. Figure 8

[0088] Low-light Object Detection:

[0089] In this section, the present invention will investigate the impact of various preprocessing methods on the efficiency of object detection under low-light conditions. Specifically, the present invention conducts experiments on the ExDark dataset, which contains 7,363 real-world nighttime images covering 12 object classes. To evaluate the impact of preprocessing, the present invention first applies the RetinexWT method proposed by the present invention as well as several comparative methods, and then uses YOLO-v3 as the object detection model to evaluate the detection performance.

[0090] Quantitative Analysis: In Table 2, the present invention demonstrates the Average Precision (AP) scores obtained by different methods used as a preprocessing step for object detection. RetinexWT stands out, not only achieving the highest overall average AP score, but also obtaining the highest AP in five specific classes: bicycle, cat, dog, person, and table. Furthermore, it achieves the second-highest AP in the ship, bus, and car classes, demonstrating strong performance. These results highlight the effectiveness and robustness of RetinexWT in enhancing the performance of object detection for different classes.

[0091]

[0092] ​Table 2: Low-light object detection results on ExDark(2) dataset using different preprocessing methods. The bold-red font indicates the best result, and the second-best result is formatted with bold-blue font.

[0093] Quantitative analysis: The present application visually compares the target detection results of the original low-light image and the image enhanced by the technology of the present application, as shown in Figure 9 In the detection results of the original low-light image, some categories are missed, and the overall detection result accuracy is greatly limited. In contrast, the enhanced image processed by the technology of the present application shows comprehensive category detection, and the detection accuracy is significantly improved.

[0094] The above only describes the embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.

Claims

1. A low-light image enhancement method based on the combination of Retinex and wavelet transform, characterized in that: The following steps are involved: S1. Decompose the input low-light image into reflection component and illumination component using Retinex theory; S2, introducing a perturbation term to simulate the degradation under low light conditions and generate degraded reflection and illumination components; S3, performing frequency domain analysis on the illumination component through wavelet transform to obtain illumination features, and combining the illumination features with the results of wavelet transform to enhance the low-light image; S4, using a gating mechanism to selectively fuse downsampled and upsampled features to suppress noise and preserve edge and detail information; S5. Apply the Transformer’s self-attention mechanism to capture long-range dependencies within the image and achieve global enhancement and restoration. The upsampling and downsampling processes are divided into two levels. Each downsampling level consists of Enhanced Illumination-Guided Attention Model (EIGAM) and Wavelet Transform Feature Decomposer Downsampling (WTFDown). Each upsampling level consists of a deconv 2×2, stride=2, a conv 1×1, and an EIGAM. The EIGAM includes the following steps: the input features are first passed through the lighting guided attention, and the lighting guided attention also uses the lighting features generated in the lighting estimator as input to guide the calculation of attention; the computational complexity is reduced through the nonlinear activation function module, which helps to simplify the model architecture and reduce the consumption of computing resources; after the lighting guided attention and nonlinear activation function modules, residual connections are performed to alleviate the gradient disappearance and retain the original gradient; finally, the output features are obtained through a layer normalization and feedforward network.

2. The low-light image enhancement method based on combining Retinex and wavelet transform according to claim 1, characterized in that: The wavelet transform is a Haar wavelet transform, which is used to decompose input features into high-frequency components and low-frequency components.

3. The low-light image enhancement method based on combining Retinex and wavelet transform as claimed in claim 1, characterized in that: The gating mechanism includes a gated fusion module to adaptively determine which features are more important and control the contribution of features from different channels.

4. The low-light image enhancement method based on combining Retinex and wavelet transform according to claim 1, wherein: The Transformer’s self-attention mechanism is implemented via a Transformer-based hybrid attention network to ensure effective modeling of long-range dependencies during the reinforcement process.

5. The low-light image enhancement method based on combining Retinex and wavelet transform as claimed in claim 1, characterized in that: The nonlinear activation function module comprises the following steps: Use a point-by-point convolution 1x1 conv to adjust the number of channels of the input feature and achieve linear transformation of the channel dimension; Depth-wise convolution is used to perform independent convolution on each channel to better extract spatial features; After simplifying the channel attention and SimpleGate, the number of channels is restored by point-by-point convolution 1x1 conv, and the first part of the output is obtained by connecting it with the input feature residual; The first part of the output is convolved and gated to obtain the second part of the output, which is added to the first part of the output through a residual connection to form the final output.

6. The low-light image enhancement method based on combining Retinex and wavelet transform as claimed in claim 1, characterized in that: The WTFDown process includes the following steps: The input features first undergo a 1*1conv operation to improve nonlinearity, and then the spatial features are converted into four frequency domain components through Haar wavelet transform, namely a low-frequency component A and three high-frequency components, where the high-frequency components are divided into horizontal high-frequency components H, vertical high-frequency components V and diagonal high-frequency components D; Low-frequency features are represented by convolution to learn features and extract global structural information; high-frequency features are concatenated by concatenating the horizontal, vertical, and diagonal high-frequency components, and dimensionality reduction and nonlinear operations are performed through point-by-point convolution 1×1conv and normalization. The two parts are added together to obtain global and local composite features, which retains the overall information while enhancing edges and details and balancing the influence of different frequency components in the image.

Images