Weak light image enhancement method and system based on dual-tree complex wavelet Retinex decomposition

Abstract

The invention discloses a weak light image enhancement method and system based on dual-tree complex wavelet Retinex decomposition. The method comprises the following steps: firstly, decomposing an input image into a low-frequency illumination component and a high-frequency reflection component by using dual-tree complex wavelet transformation; for low-frequency components, a global illumination smoothing module including expansion convolution and global average pooling is designed to accurately estimate and smooth illumination. For high-frequency components, a directional attention mechanism is introduced to enhance texture and edge details. And finally, performing diffusion adjustment and refinement processing on the illumination component and the reflection component to realize self-adaptive reconstruction of the image. According to the method, the brightness, the contrast ratio and the detail visibility of the low-light image can be effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and artificial intelligence technology, and relates to a low-light image enhancement method and system based on dual-tree complex wavelet Retinex decomposition. Background Technology

[0002] Low-light image enhancement technology aims to computationally restore the brightness, contrast, and detail of underexposed, low signal-to-noise ratio images while suppressing the side effects of amplifying imaging noise and artifacts during the computation process. It can be applied to fields such as autonomous driving, scene monitoring, and night photography in low-light environments. Traditional methods mainly employ classic techniques such as histogram equalization, gamma correction, and the Retinex model. Histogram equalization enhances the global contrast of an image by adjusting the grayscale distribution of image pixels. However, this method can easily lead to the loss of detail in overly dark or bright areas of the image while improving overall contrast, and its effect on enhancing local details is limited. Gamma correction is mainly used to non-linearly adjust the brightness of an image to improve visual effects. However, it also faces the problem of potentially over-amplifying noise while enhancing brightness, and it is difficult to achieve a good balance between global brightness adjustment and local detail enhancement. The Retinex model, based on the characteristics of human vision, simulates the human perception mechanism of object color and detail under different lighting conditions to solve the problems of color constancy and detail enhancement in images. However, these traditional methods all suffer from the trade-off between global brightness adjustment and local detail enhancement, resulting in poor visual quality of low-light image enhancement results and poor generalization when dealing with different low-light scenes.

[0003] With the rapid development of deep learning, diffusion models employing probabilistic generative theory have achieved high-quality noise modeling and image detail enhancement in low-light enhancement tasks. However, low-light images inherently have low signal-to-noise ratios, and strong noise characteristics still negatively impact the enhancement results. Even in frequency domain decomposition, the enhanced image is still prone to directional stripe artifacts, which may be distributed along a certain texture direction or illumination direction, interfering with visual effects. Furthermore, there is the problem of over-enhanced edge features, making image edges too abrupt and lacking a natural feel.

[0004] Existing low-light enhancement techniques based on Retinex decomposition have two limitations: 1) Low-light enhancement methods based on Retinex decomposition have difficulty accurately distinguishing between illumination and reflection components, resulting in directional stripe artifacts in the enhanced image, which interfere with visual effects; 2) Diffusion models are highly dependent on the distribution and quality of training data, and diffusion methods are sensitive to the diffusion conditions and image features of the encoded input. Furthermore, the generation illusions and feature distortions of diffusion models have a significant impact on subsequent visual tasks. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problems of directional stripe artifacts and image feature sensitivity in existing low-light image enhancement methods, and to provide a low-light image enhancement method and system based on dual-tree complex wavelet Retinex decomposition.

[0006] To achieve the above objectives, the present invention employs the following technical solution: The first aspect of this invention provides a low-light image enhancement method based on dual-tree complex wavelet Retinex decomposition, comprising the following steps: The image is decomposed using a dual-tree complex wavelet to obtain low-frequency and high-frequency features. Low-frequency features are input into a global illumination smoothing module for feature extraction to obtain illumination features. The global illumination smoothing module includes two cascaded 2D convolutional layers, a ReLU layer, a dilated convolutional layer, a global average pooling layer, a 2D convolutional layer, and a ReLU activation function. The illumination features include those of normal light. and the characteristics of low light illumination ; A directional attention mechanism is used to enhance the directional features of high-frequency features to obtain reflection features; the reflection features include the reflection features of normal light. and the reflective characteristics of weak light ; The illumination characteristics are adjusted by illumination diffusion to obtain the illumination component; the reflection characteristics are refined by reflection diffusion to obtain the reflection component. The image is reconstructed based on the illumination component and the reflection component.

[0007] Furthermore, the step of inputting low-frequency features into the global illumination smoothing module for feature extraction to obtain illumination features specifically involves: Low-frequency features are processed through two cascaded two-dimensional convolutional layers and a ReLU layer to generate fine features; Small features are input into the dilated convolutional layer to generate multi-scale convolutional features; the dilated convolutional layer includes multiple parallel dilated convolutions; the multi-scale convolutional features generated by each dilated convolution are batch normalized and then concatenated to generate low-frequency multi-scale hybrid features. The low-frequency multi-scale hybrid features are processed sequentially by a global average pooling layer, a two-dimensional convolution, and a ReLU activation function to generate channel weights, which are then synthesized by an inverse dual-tree complex wavelet and converted into spatial domain features. Spatial domain features are obtained by using two-dimensional convolution and the Sigmoid activation function to obtain illumination features.

[0008] Furthermore, the high-frequency features are enhanced using a directional attention mechanism to obtain reflection features; specifically: Projecting features from different directions in the high-frequency features yields mapped features; the mapped features include diagonal features and off-diagonal features. The diagonal features are used as values ​​and keys in the directional attention mechanism, and the off-diagonal features are used as query vectors in the directional attention memory. The diagonal-enhanced latent features are obtained by cross-product of the query vector with the key in the directional attention mechanism and then with the value in the directional attention mechanism. Hidden features enhanced by diagonal through nonlinear transformation After being projected onto the original space, linear layer processing is performed followed by residual connection to generate new diagonal features; Multi-scale convolution is used to extract high-frequency information in each direction from the new diagonal features, and the concatenation is performed along the channel dimension to obtain enhanced high-frequency features. The enhanced high-frequency features are subjected to inverse dual-tree complex wavelet transform to obtain the reflection features.

[0009] Furthermore, the diagonal-enhanced hidden features The expression is:

[0010] in, It is a diagonal feature; It is a non-diagonal feature; The Query transformation matrix is ​​for off-diagonal feature mappings; The key transformation matrix; is the dimension of the Key matrix for the diagonal feature mapping.

[0011] Furthermore, the expression for the new diagonal feature is:

[0012] in, To enhance the diagonal feature, All are learnable weights. FFN() represents the linear transformation matrix output by the linear layer after the dot product of the query vector and the key in the directional attention mechanism. .

[0013] Furthermore, the reflection characteristics are refined by refining the reflection diffusion, specifically as follows: The reflection characteristics of normal light Precoding is performed as input. As a conditional image, noise is gradually added to the pre-coded normal light reflection features according to the time step to generate a noisy image; Based on the noisy image and the conditional image, the noise in the noisy image is gradually reduced at each time step to obtain the reflection component.

[0014] Furthermore, the light diffusion of the aforementioned light characteristics is adjusted, specifically as follows: The illumination characteristics of normal light Precoding is performed as input. As a conditional image, noise is gradually added to the pre-coded illumination features according to the time step to generate a noisy image; Based on the noisy image and the conditional image, the noise in the noisy image is gradually reduced at each time step to obtain the illumination component.

[0015] Furthermore, the distribution of the noisy image is as follows:

[0016] in, These are preset noise intensity parameters. It is the identity matrix; the current noisy image is The noisy image from the previous time step is .

[0017] Furthermore, the image reconstruction based on the illumination component and the reflection component specifically includes:

[0018] Among them, the illumination component is The reflection component is ; Represents the reconstructed image.

[0019] A second aspect of the present invention provides a low-light image enhancement system based on dual-tree complex wavelet Retinex decomposition, comprising: The dual-tree complex wavelet decomposition module performs Retinex decomposition on the image using dual-tree complex wavelets to obtain low-frequency and high-frequency features. A global illumination smoothing module is used to extract low-frequency features from the input to obtain illumination features. The global illumination smoothing module includes two cascaded 2D convolutional layers, a ReLU layer, a dilated convolutional layer, a global average pooling layer, a 2D convolutional layer, and a ReLU activation function. The illumination features include those of normal light. and the characteristics of low light illumination ; The directional feature enhancement module uses a directional attention mechanism to enhance the directional features of high-frequency features, thereby obtaining reflection features; the reflection features include the reflection features of normal light. and the reflective characteristics of weak light ; The diffusion adjustment and refinement module adjusts the illumination diffusion of the illumination characteristics to obtain the illumination component; and refines the reflection diffusion of the reflection characteristics to obtain the reflection component. The image reconstruction module reconstructs the image based on the illumination component and the reflection component.

[0020] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses a low-light image enhancement method based on dual-tree complex wavelet Retinex decomposition. By introducing dual-tree complex wavelet transform for image decomposition, its translation invariance and multi-directional selectivity enable more accurate separation of high-frequency details and low-frequency illumination in the image. Combined with a subsequent directional attention mechanism to enhance high-frequency features, it can specifically strengthen texture and edge information in different directions, effectively avoiding edge blurring or detail loss problems common in traditional methods, resulting in a more detailed and clearer image. A global illumination smoothing module, by combining dilated convolution to expand the receptive field and global average pooling to capture global contextual information of the image, achieves more accurate and smooth global illumination estimation. This effectively avoids local overexposure or underexposure, and solves the halo artifacts and uneven illumination problems easily generated in traditional Retinex methods, resulting in a more uniform brightness distribution and a more natural visual experience in the adjusted image. Within the dual-tree complex wavelet domain, the distribution characteristics of image signals and noise differ at different scales and directions. This method, when processing the high-frequency reflection component, selectively enhances useful structural information (such as edges and textures) through a directional attention mechanism, while suppressing irrelevant high-frequency noise to some extent, thus achieving better noise control while enhancing the image. By adjusting the illumination diffusion and refining the reflection diffusion of the illumination and reflection components respectively, independent and refined optimization of the two components is achieved. This divide-and-conquer strategy results in a significantly improved brightness, contrast, color fidelity, and overall visual quality in the final reconstructed image. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the overall structure of a low-light image enhancement method based on dual-tree complex wavelet Retinex decomposition according to the present invention. Figure 2 This is a schematic diagram of the global illumination smoothing module; Figure 3 This is a schematic diagram of the directional feature enhancement module; Figure 4This is a schematic diagram of the directional attention mechanism of the directional feature enhancement module; Figure 5 This is a schematic diagram of the structure of the light diffusion adjustment branch and the reflection diffusion refinement branch. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0026] The present invention will now be described in further detail with reference to the accompanying drawings: See Figure 1 This invention discloses a low-light image enhancement method based on dual-tree complex wavelet Retinex decomposition, which is implemented according to the following steps: Step 1: Perform Retinex decomposition on the input image through dual-tree complex wavelet transform to convert it from the image domain to the frequency domain. Input the low-frequency features into the global illumination smoothing module, use progressively dilated convolution groups to extract illumination features, batch normalize the low-frequency features at different scales and then fuse them to obtain multi-scale illumination features. Step 1.1: Perform dual-tree complex wavelet decomposition on the input image, and obtain the low-frequency features after decomposition. and high frequency characteristics As shown in formula (1):

[0027] (1); in, and It is a low-pass filter. and It is a high-pass filter; Indicates the number of decomposition layers.

[0028] Step 1.2: Extract low-frequency features After two 2D convolutions and the ReLU function, dilated convolution is used to construct multi-granularity convolutional features. For the first... One dilated convolution, kernel size and expansion rate The only constraint is shown in formula (2): (2); in, It is the receptive field size, ensuring that the receptive field of each convolutional branch does not exceed the predetermined range.

[0029] Step 1.3: After batch normalization of the multi-scale features extracted by different convolution branches, concat the features and perform global average pooling on the low-frequency multi-scale mixed features to obtain channel-level statistical information. Then, generate channel weights using the ReLU activation function. Step 1.4: Perform inverse dual-tree complex wavelet transform on the refined low-frequency features, and then normalize them using two-dimensional convolution and the Sigmoid function to obtain the illumination features. and . The characteristics of normal light illumination. Characteristics of weak light illumination.

[0030] Step 2: The high-frequency feature input direction feature enhancement module uses the direction cross attention mechanism to use information from the off-diagonal direction to enhance the features in the diagonal direction, thereby obtaining detailed reflection features.

[0031] Step 2.1, High-frequency characteristics It includes high-frequency subbands in six directions: 15°, -15°, 45°, -45°, 75°, and -75°, using depthwise separable convolution pairs for the high-frequency coefficients in different directions. and Feature extraction is performed to obtain feature maps. and ,in , .

[0032] Step 2.2, Diagonal Features The value matrix is ​​obtained through dot product operation. and key matrix The non-diagonal features are used to obtain the query matrix through dot product operations. The softmax function is used to map each score to the [0,1] region. The key matrix and query matrix are multiplied by a dot product, and then the result is cross-multiplied by the value matrix to obtain the diagonally enhanced latent features. The calculation process is shown in formula (3): (3) in, is the vector dimension of the key matrix.

[0033] Step 2.3: Transform the latent features using nonlinear transformation. Projecting onto the original space, a fully connected feedforward network is used to further refine the global information, and residual connections are used to add the input sequence, introducing learnable coefficients into the branches of each residual connection. , , and The adaptive learning of data from different branches is performed as shown in formula (4): (4) in, This represents the output weight matrix preceding the linear layer (FFN). .

[0034] Step 2.4: The diagonal features enhanced by 15° and 75° directions are concatenated to form new diagonal features. Multi-scale convolution is used to extract high-frequency information in each direction, and the features are concatenated in the channel dimension to capture detailed information at different scales.

[0035] Step 2.5: Perform inverse dual-tree complex wavelet transform on the enhanced high-frequency features for synthesis, use a Reformer denoising network for noise suppression, and normalize using a Sigmoid function to obtain the detailed enhanced reflection features. and . This reflects the characteristics of normal light. It is a characteristic of weak light reflection.

[0036] Step 3: After frequency domain enhancement, the illumination features are adjusted through the illumination diffusion branch to obtain the enhanced illumination component, and the reflection features are refined through the reflection diffusion branch to obtain the enhanced reflection component. Step 3.1: Reflection characteristics Precoding is performed as input. As a conditional input, from the time step Beginning, at the time step At this point, the reflection features after precoding Add Gaussian noise. Throughout the forward diffusion process, gradually increase the noise at certain time steps to simulate the degradation process of data from ordered to disordered. At any time step... Noisy image The data distribution can be represented as ,in These are preset noise intensity parameters used to control the amount of noise added. It is an identity matrix. The purpose of this step is to build a probabilistic model of the diffusion process from the original data to the noisy data.

[0037] Step 3.2: During the reflection and diffusion process, from the time step... Initially, noise is gradually reduced to restore the data. At each time step... Based on the current noisy image and conditional images Using a denoising model To generate The probability distribution. The denoising model is a deep neural network with parameters of... Based on the current noise data Condition information and time step To predict the distribution of the denoised data, i.e. ,in It is the mean of the model's predictions. This is the preset variance. The purpose of this step is to use the model to learn how to recover the original data from noisy data, and through gradual denoising, ultimately... When the reflection component is obtained .

[0038] Step 3.3: Illumination characteristics Precoding is performed as input. As a conditional input, the illumination component is obtained through the same forward and reverse diffusion processes as in steps 3.1 and 3.2. ; Step 3.4: Separate the light components after the light diffusion adjustment branching process. and the reflection component after reflection diffusion refinement branching Perform a reshaping operation and Retinex reconstruction, as shown in formula (5): (5) Among them, the illumination component is The reflection component is ; Represents the reconstructed image.

[0039] Step 4: Design the loss function, which includes: reconstruction loss. Light smoothing loss Consistency loss and back diffusion loss ; Step 4.1: In the dual-tree complex wavelet Retinex decomposition stage (Step 1.1), design a reconstruction loss to measure the difference between the image reconstructed from the illumination component and the reflection component after decomposition and the original image, thereby ensuring that the decomposition result can better restore the original image. The calculation formula is shown in (6): (6); in It's a hyperparameter. The cross-multiplication loss term represents the illumination and reflectance diagrams for weak and normal light; Step 4.2: Design a consistency loss for the reflection map (directional feature enhancement module) in the dual-tree complex wavelet Retinex decomposition stage to maintain the stability and consistency of the reflection map obtained by decomposition in different processing or transformation processes. Its calculation formula is shown in (7): (7); Step 4.3: Construct the illumination smoothness loss (global illumination smoothing module) in the dual-tree complex wavelet Retinex decomposition stage to constrain the smoothness of the illumination map and avoid overly abrupt illumination changes in the illumination map. Its calculation formula is (8): (8); in and It's a weighting factor, set and , It is a constraint factor. Represents a gradient filter; Step 4.4: Design diffusion loss in the illumination diffusion adjustment branch and the reflection diffusion refinement branch to optimize the learning process of each step of the denoising diffusion model so that it can accurately restore the original image. The calculation formula is shown in (9): (9); in, For normal light image components, To predict noise, The actual noise added. and Used to control the process of adding and removing noise; Step 4.5: Design the consistency loss in the light diffusion adjustment branch and the reflection diffusion refinement branch, as shown in formula (10): (10) in, This represents the noise predicted by the denoising network.

[0040] One embodiment of the present invention provides a low-light image enhancement method based on dual-tree complex wavelet Retinex decomposition, which is implemented according to the following steps: Step 1 is attached. Figure 1 and attached Figure 2 As shown, the specific process is as follows: Step 1.1: Input a two-dimensional array of images, define low-pass filters H0 and G0, high-pass filters H1 and G2 for dual-tree complex wavelets, decompose the number of layers J=3, and use the dwt function of the pywt library to perform wavelet decomposition. It is necessary to decompose the two trees separately to obtain low-frequency and high-frequency features. Step 1.2: Low-frequency features are processed by two two-dimensional convolution kernels with ReLU function. Dilated convolutional layers are defined. The first convolutional layer uses a kernel size of 7×7 with padding of 3. The second convolutional layer uses a kernel size of 5×5 with padding of 3 and a dilation rate of 4. The third convolutional layer uses a kernel size of 3×3 with padding of 1 and a dilation rate of 1.

[0041] Step 1.3: Perform batch normalization on the multi-scale features extracted by different convolution branches, and concatenate the features using torch.cat([x[:,i,:,:].unsqueeze(1) for i in range(x.shape[1])],dim=1), perform global average pooling using adaptive_avg_pool2d(x_concat, (1, 1)), and use the ReLU activation function to generate channel weights; Step 1.4: Use the inverse dual-tree complex wavelet transform synthesis function idwt to iteratively perform inverse transform (high_freq_layer[0], high_freq_layer[1], high_freq_layer[2]) at each decomposition layer, and perform normalization operations through two-dimensional convolution and the Sigmoid function to obtain illumination features. and .

[0042] Step 2 is attached. Figure 3 and attached Figure 4 As shown, the specific process is as follows: Step 2.1: Define depthwise separable convolution to extract features in each direction: high_freqFeatures[:, 0:1, :, :], high_freqFeatures[:, 1:2, :, :], high_freqFeatures[:, 2:3, :, :], high_freqFeatures[:, 3:4, :, :], high_freqFeatures[:, 4:5, :, :], high_freqFeatures[:, 5:6, :, :]; Step 2.2: Construct the value matrix (value = self.to_value(diagonal_theta)) and key matrix (key = self.to_key(diagonal_theta)) for diagonal features, and construct the query matrix (query = self.to_query(non_diagonal_theta)) for non-diagonal features. Calculate the attention score using the built-in functions torch.matmul and dots.softmax of the PyTorch framework to obtain the diagonally enhanced latent features. ; Step 2.3, Latent Features The input sequence is added using a linear layer `self.ffn(z_i)` and a residual connection, with learnable coefficients introduced in each branch of the residual connection. , , and , self.alpha * z_proj + self.beta *input_sequence, self.eta * residual + self.varphi * ffn_output, project back into the original space to obtain diagonal enhanced features [diagonal_enhanced15], [diagonal_enhanced75].

[0043] Step 2.4: Concatenate the enhanced diagonal features torch.cat([diagonal_enhanced15,diagonal_enhanced75], dim=1), use multi-scale convolution to extract high-frequency information in each direction, and concatenate them along the channel dimension torch.cat([conv1_out, conv2_out], dim=1); Step 2.5: Perform inverse dual-tree complex wavelet idwt transform on the enhanced high-frequency features for synthesis, use the Reformer denoising network for noise suppression (self.reformer_denoise), and normalize using Sigmoid to obtain the detailed enhanced reflection features. and .

[0044] Step 3 is attached. Figure 5 As shown, the specific process is as follows: Step 3.1: Reflection characteristics Precoding is performed as input. As a conditional input, in the total number of diffusion steps Within the noise intensity sequence, `self.beta = torch.linspace(0.0001, 0.02, 1000)` is used for pre-calculation. and For any time , Perform forward noise addition Obtain a noisy image Distribution; Step 3.2: During the reflection and diffusion process, time t and conditions. Input the denoising network self.eps_model, output the noise estimate Predict noise and inversely deduce the distribution of the original image based on the reparameterization formula. ,set up Parameters (0.0001, 0.02, 1000); Step 3.3, the light diffusion adjustment process is the same as steps 3.1 and 3.2. The forward diffusion process is defined as forward_diffusion.forward_process(Z_norm, t), and the reverse diffusion process is defined as reverse_diffusion.reverse_process(xt_Z, t, Z_0); Step 3.4: Separate the light components after the light diffusion adjustment branching process. and the reflection component after reflection diffusion refinement branching Perform a reshaping operation and Retinex reconstruction, where I_reconstructed = I_illumination * I_reflection; The total loss function of the low-light image enhancement method based on dual-tree complex wavelet Retinex decomposition in this invention is: + , specifically: By calculating the reconstruction loss under normal light Calculate the reconstruction loss under low light conditions Calculate the cross-component loss Set hyperparameters The final total reconstruction loss was The reconstruction loss is used for dual-tree complex wavelet decomposition.

[0045] Build reflection in the direction feature enhancement module Figure 1 The loss of efficacy is: .

[0046] Through calculation and The gradients are: horizontal gradient = gradient_x = torch.roll(x, shifts=-1, dims=2) - x, vertical gradient = gradient_y = torch.roll(x, shifts=-1, dims=3) - x, constraint factor is set to 0.5, and the total loss of the global illumination smoothing module is... ;; To optimize each step of the denoising diffusion model's learning process, a diffusion loss was designed. .

[0047] Design the consistency loss of diffusion content for the overall light diffusion process and the reflection diffusion process. .

[0048] All experiments in this embodiment were conducted on an NVIDIA GeForce RTX 4090 GPU using the PyTorch framework. During encoder training, the number of DTCWT decomposition layers was set. , , , Learning rate The batch size is 32, and the total number of training iterations is 300. During decoder training, the diffusion step size is set to... The batch size is 24, and 500,000 iterations are performed. This paper trains on the LOL-v1 and VE-LOL-L datasets. The LOL-v1 dataset contains 485 pairs of training images and 15 pairs of test images. The VE-LOL-L dataset contains more complex real-world scenes and more diverse image noise, and is divided into 1500 data pairs in a 7:1:2 ratio.

[0049] The peak signal-to-noise ratio (PSNR), structural similarity (SSIM), learned perceptual patch similarity (LPIPS), and natural image quality evaluator (NIQE) were used to evaluate the present invention and existing low-light enhancement methods. Test results were obtained on the LOL-v1 and VE-LOL-L datasets, as shown in Tables 1 and 2, respectively. Specifically, on the LOL-v1 dataset, the present invention achieved a PSNR of 24.36 dB, 0.53 dB higher than the existing method URetinex-Net++; SSIM reached 0.898 dB, 0.001 dB higher than the SNRNet method; LPIPS reached 0.032 dB, 0.054 dB lower than the HVI method; and NIQE reached 4.301 dB, 0.153 dB lower than the URetinex-Net++ method. On the VE-LOL-L dataset, the present invention also demonstrated significant performance improvements, with each metric showing a significant improvement over the best existing method.

[0050] The proposed invention was evaluated against existing low-light enhancement methods using Learning Perceptual Image Patch Similarity (LPIPS) and Natural Image Quality Evaluator (NIQE). Test results were obtained on both the LOL-v1 and VE-LOL-L datasets, as shown in Tables 1 and 2, respectively. Table 1 compares the PSNR, SSIM, LPIPS, and NIQE metrics of the proposed invention and existing methods on the LOL-v1 dataset; Table 2 compares the PSNR, SSIM, LPIPS, and NIQE metrics of the proposed invention and existing methods on the VE-LOL-L dataset. It can be seen that on the LOL-v1 dataset, the proposed invention achieved a PSNR of 24.36 dB, which is 0.53 dB higher than the existing method URetinex-Net++; an SSIM of 0.898 dB, which is 0.001 dB higher than the SNRNet method; an LPIPS of 0.032 dB, which is 0.054 dB lower than the HVI method; and an NIQE of 4.301 dB, which is 0.153 dB lower than the URetinex-Net++ method. On the VE-LOL-L dataset, this invention also demonstrates significant performance improvements, with each metric showing a significant improvement over the existing best methods.

[0051] Table 1

[0052] Table 2

[0053] One embodiment of the present invention provides a low-light image enhancement system based on dual-tree complex wavelet Retinex decomposition, comprising: The dual-tree complex wavelet decomposition module performs Retinex decomposition on the image using dual-tree complex wavelets to obtain low-frequency and high-frequency features. The global illumination smoothing module extracts low-frequency features to obtain illumination features. The global illumination smoothing module includes two cascaded two-dimensional convolutional layers, a ReLU layer, a dilated convolutional layer, a global average pooling layer, a two-dimensional convolutional layer, and a ReLU activation function. The directional feature enhancement module uses a directional attention mechanism to enhance the directional features of high-frequency features and obtain reflection features. The diffusion adjustment and refinement module adjusts the illumination diffusion of the illumination characteristics to obtain the illumination component; and refines the reflection diffusion of the reflection characteristics to obtain the reflection component. The image reconstruction module reconstructs the image based on the illumination component and the reflection component.

[0054] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A low-light image enhancement method based on dual-tree complex wavelet Retinex decomposition, characterized in that, The method comprises the following steps: perform Retinex decomposition on the image through a dual-tree complex wavelet to obtain a low-frequency feature and a high-frequency feature; The low-frequency feature is input into a global light smoothing module for feature extraction to obtain a light feature; the global light smoothing module comprises two series two-dimensional convolution layers, a ReLU layer, an expansion convolution layer, a global average pooling layer, a two-dimensional convolution, and a ReLU activation function connected in sequence; the light feature comprises a normal light feature and a weak light feature . ​ The high-frequency features are enhanced by a directional attention mechanism to obtain reflection features; the reflection features include reflection features of normal light and reflection features of weak light ​​ perform illumination diffusion adjustment on the illumination feature to obtain an illumination component; and perform reflection diffusion refinement on the reflection feature to obtain a reflection component; reconstruct the image according to the illumination component and the reflection component.

2. The dual-tree complex wavelet Retinex decomposition based low-light image enhancement method according to claim 1, characterized in that, The low-frequency feature is input into a global illumination smoothing module to extract a feature to obtain the illumination feature, specifically as follows: the low-frequency feature is subjected to two two-dimensional convolution layers and ReLU layers in series to generate a fine feature; the fine feature is input into a dilated convolution layer to generate a multi-scale convolution feature; the dilated convolution layer comprises a plurality of parallel dilated convolutions; the multi-scale convolution features generated by each dilated convolution are subjected to batch normalization and then spliced to generate a low-frequency multi-scale mixed feature; the low-frequency multi-scale mixed feature is sequentially subjected to a global average pooling layer, a two-dimensional convolution and an ReLU activation function to generate a channel weight, and is converted into a spatial domain feature through inverse dual-tree complex wavelet synthesis; the spatial domain feature is subjected to a two-dimensional convolution and a Sigmoid activation function to obtain the illumination feature.

3. The dual-tree complex wavelet Retinex decomposition based low-light image enhancement method according to claim 1, characterized in that, The high-frequency feature is subjected to a direction attention mechanism to enhance a direction feature to obtain the reflection feature, specifically as follows: different direction features in the high-frequency feature are projected to obtain a mapping feature; the mapping feature comprises a diagonal line feature and a non-diagonal line feature; the diagonal line feature is used as a value and a key in the direction attention mechanism, and the non-diagonal line feature is used as a query vector in the direction attention mechanism; the query vector is subjected to dot product with the key in the direction attention mechanism, and then is crossed multiplied with the value in the direction attention mechanism to obtain a diagonal line enhanced hidden feature; a multi-scale convolution is used on the new diagonal line feature to extract high-frequency information of each direction, and the high-frequency information is spliced in a channel dimension to obtain an enhanced high-frequency feature; Diagonal enhanced hidden features by nonlinear transformation After linear layer processing in the original space, residual connection is performed to generate new diagonal features the enhanced high-frequency feature is subjected to inverse dual-tree complex wavelet transformation to obtain the reflection feature. An expression of the new diagonal line feature is as follows:

4. The dual-tree complex wavelet Retinex decomposition based low-light image enhancement method according to claim 3, characterized in that, The diagonal enhanced hidden feature The expression is: wherein, is a diagonal feature; is a non-diagonal feature; is a Query transformation matrix for non-diagonal feature mapping; is a Key transformation matrix; is a Key matrix dimension for diagonal feature mapping.

5. The dual-tree complex wavelet Retinex decomposition based low-light image enhancement method according to claim 4, characterized in that, the reflection feature is subjected to reflection diffusion refinement, specifically as follows: wherein, are enhanced diagonal features, are learnable weights, denotes a linear transformation matrix of the dot product of the query vector and the keys in the directional attention mechanism output by the linear layer, and FFN() denotes a linear layer; .

6. The low-light image enhancement method based on dual-tree complex wavelet Retinex decomposition according to claim 1, characterized in that, a noise image is gradually reduced in noise at each time step according to the noise image and a conditional image to obtain the reflection component. reflective characteristics of normal light performing pre-encoding as input, as a conditional image, stepwise increasing noise to the pre-encoded reflective characteristics of normal light by time steps, generating a noise image; the illumination feature is subjected to illumination diffusion adjustment, specifically as follows:

7. The dual-tree complex wavelet Retinex decomposition based low-light image enhancement method according to claim 1, wherein, a noise image is gradually reduced in noise at each time step according to the noise image and a conditional image to obtain the illumination component. a light feature of normal light performing pre-encoding as input, as a conditional image, gradually increasing noise on the pre-encoded light feature by time steps to generate a noise image; A distribution of the noise image is as follows:

8. The dual-tree complex wavelet Retinex decomposition based low-light image enhancement method according to claim 6 or 7, characterized in that, the image is reconstructed according to the illumination component and the reflection component, specifically as follows: wherein, is a preset noise intensity parameter, is an identity matrix; the current noise image is ; the noise image of the previous time step is .

9. The dual-tree complex wavelet Retinex decomposition based low-light image enhancement method according to claim 1, wherein, comprises: wherein the illumination component is ; the reflection component is ; denotes the reconstructed image.

10. A low-light image enhancement system based on dual-tree complex wavelet Retinex decomposition, based on the low-light image enhancement method based on dual-tree complex wavelet Retinex decomposition according to claim 1, characterized in that, a dual-tree complex wavelet decomposition module that performs Retinex decomposition on the image through a dual-tree complex wavelet to obtain a low-frequency feature and a high-frequency feature; a diffusion adjustment and refinement module that performs illumination diffusion adjustment on the illumination feature to obtain an illumination component; and performs reflection diffusion refinement on the reflection feature to obtain a reflection component; A global illumination smoothing module is used to extract low-frequency features from the input to obtain illumination features. The global illumination smoothing module includes two cascaded 2D convolutional layers, a ReLU layer, a dilated convolutional layer, a global average pooling layer, a 2D convolutional layer, and a ReLU activation function. The illumination features include those of normal light. and the characteristics of low light illumination ; The directional feature enhancement module uses a directional attention mechanism to enhance the directional features of high-frequency features, thereby obtaining reflection features; the reflection features include the reflection features of normal light. and the reflective characteristics of weak light ; an image reconstruction module that reconstructs the image according to the illumination component and the reflection component. ​

Images