Low-illumination image enhancement method and device based on improved normalized stream

Abstract

The embodiment of the invention discloses a low-illumination image enhancement method and device based on an improved normalized stream. The method comprises the following steps: acquiring a low-illumination image to be enhanced; performing normalization processing on the to-be-enhanced low-illumination image to obtain a normalized spectrum; performing edge extraction processing on the normalized spectrum to obtain an edge spectrum; performing feature fusion on the to-be-enhanced low-illumination image, the normalized spectrum and the edge spectrum to obtain a feature fusion image; performing feature extraction on the feature fusion image by adopting a feature extraction module to obtain a feature extraction image; performing feature mapping on the feature extraction image by adopting a feature mapping module to obtain a feature mapping image; and inputting the feature mapping image into the normalized stream reversible network, and outputting a normal illumination image. By enhancing the low-illumination image through the method, the enhancement effect can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a low-illumination image enhancement method and device based on improved normalized flow. Background Art

[0002] Due to the limitation of lighting conditions, low-light images usually face problems such as low contrast, high noise, and loss of details, which seriously affect the visual effect of the image and the accuracy of subsequent image analysis tasks. Therefore, it is necessary to enhance low-light images to improve their quality and increase the usability of images.

[0003] Traditional image enhancement methods, such as histogram equalization and gamma correction, can improve the visual effects of images to a certain extent, but they often have many limitations. For example, the histogram equalization method may introduce artifacts, and simple gamma correction cannot effectively restore the detailed information of the image. In addition, most of these methods are based on the global statistical characteristics of the image, and it is difficult to handle local details and regional features, especially when processing high dynamic range images, it is difficult to balance the highlights and dark details in the image. With the development of deep learning technology, methods based on deep neural networks have made significant progress in low-light image enhancement. However, existing deep learning-based methods, such as those based on convolutional neural networks (CNNs), can learn complex mapping relationships of images, but there are still some challenges. For example, CNN may have limitations in processing detailed features, resulting in loss of details or increased artifacts in the enhanced image. Summary of the invention

[0004] Based on this, it is necessary to propose a low-illumination image enhancement method and device based on improved normalized flow to address the above problems.

[0005] First aspect

[0006] A low-illumination image enhancement method based on improved normalized flow, the method comprising:

[0007] Acquire a low-light image to be enhanced;

[0008] Performing normalization processing on the low-illuminance image to be enhanced to obtain a normalized spectrum;

[0009] Performing edge extraction processing on the normalized spectrum to obtain an edge spectrum;

[0010] Performing feature fusion on the low-illuminance image to be enhanced, the normalized spectrum and the edge spectrum to obtain a feature fused image;

[0011] A feature extraction module is used to extract features from the feature fusion image to obtain a feature extraction image; wherein the feature extraction module includes a two-dimensional convolution layer, a maximum pooling layer and a residual dense block;

[0012] A feature mapping module is used to perform feature mapping on the feature extraction image to obtain a feature mapping image; wherein the feature mapping module includes a two-dimensional convolution layer and a group cross attention block, and the group cross attention block includes a grouping module, an average pooling layer, a feature fusion layer, a two-dimensional convolution layer, a weight redistribution layer and a group normalization layer;

[0013] The feature map image is input into a normalized flow reversible network, and a normal illumination image is output; wherein the normalized flow reversible network includes multiple flow steps and a two-dimensional compression layer.

[0014] Preferably, the normalization process is specifically:

[0015]

[0016] In the formula, x norm represents the normalized spectrum, x l represents the low-light image to be enhanced, ||x l ||1 represents the 1-norm of the low-light image to be enhanced.

[0017] Preferably, the edge extraction process is specifically:

[0018]

[0019] In the formula, x edge represents the edge spectrum, represents the gradient value of the normalized spectrum in the horizontal direction, Represents the gradient value of the normalized spectrum in the vertical direction.

[0020] Preferably, the step of using a feature extraction module to extract features from the feature fusion image to obtain a feature extraction image specifically includes:

[0021] Input the feature fusion image into the first two-dimensional convolution layer and the second two-dimensional convolution layer in sequence, and output the first convolution image;

[0022] Input the first convolution image into the maximum pooling layer and the residual dense block in sequence, and output a first image;

[0023] The first convolution image and the first image are summed to obtain a feature extraction image.

[0024] Preferably, the adopting a feature mapping module to perform feature mapping on the feature extraction image to obtain a feature mapping image specifically includes:

[0025] Upsampling the feature extraction image to obtain an upsampled image;

[0026] Inputting the upsampled image into a third two-dimensional convolutional layer, and outputting a first feature map image;

[0027] Input the feature extraction image into a first grouped cross attention block, and output a second feature mapping image;

[0028] Downsampling the feature extraction image to obtain a first downsampled image;

[0029] Input the first down-sampled image into a fourth two-dimensional convolutional layer, and output a second image;

[0030] Input the second image into a second grouped cross attention block, and output a third feature map image;

[0031] Downsampling the second image to obtain a second downsampled image;

[0032] The second down-sampled image is sequentially input into the fifth two-dimensional convolutional layer and the third grouped cross attention block, and a fourth feature map image is output.

[0033] Preferably, the step of inputting the feature extraction image into a first grouped cross attention block and outputting a second feature mapping image specifically includes:

[0034] Performing group processing on the feature extraction images to obtain a first group image, a second group image and a third group image;

[0035] Performing average pooling in a horizontal direction on the second group of images to obtain a second pooled image;

[0036] Performing average pooling in a vertical direction on the third grouped images to obtain a third pooled image;

[0037] splicing the second pooled image and the third pooled image to obtain a spliced ​​image;

[0038] Inputting the stitched image into a sixth two-dimensional convolutional layer to obtain a third image;

[0039] Redistributing weights of the third image and the first grouped image to obtain a redistributed image;

[0040] performing group normalization processing on the reallocated image to obtain a normalized image;

[0041] Input the normalized image into a first average pooling layer, and output a fourth image;

[0042] Inputting the first grouped image into a seventh two-dimensional convolutional layer to obtain a fifth image;

[0043] multiplying the fourth image and the fifth image to obtain a sixth image;

[0044] Input the fifth image into a second average pooling layer, and output a seventh image;

[0045] multiplying the seventh image and the normalized image to obtain an eighth image;

[0046] The first grouped image, the sixth image and the eighth image are weighted redistributed to output a second feature map image.

[0047] Preferably, the flow steps include activation normalization, 1×1 reversible convolutional layers and affine coupled transformation layers.

[0048] Preferably, the normalized flow reversible network further includes a two-dimensional decompression layer.

[0049] Preferably, the step of inputting the feature map image into a normalized flow reversible network and outputting a normal illumination image specifically includes:

[0050] The feature map image is sequentially input into the flow step and the 2D compression layer, and the normal illumination image is output.

[0051] Second aspect

[0052] A low-illumination image enhancement device based on improved normalized flow, the device comprising:

[0053] A low-light image acquisition module, used to acquire a low-light image to be enhanced;

[0054] A normalized spectrum acquisition module, used for performing normalization processing on the low-illuminance image to be enhanced to obtain a normalized spectrum;

[0055] An edge spectrum acquisition module, used for performing edge extraction processing on the normalized spectrum to obtain an edge spectrum;

[0056] A feature fusion image acquisition module is used to perform feature fusion on the low-illuminance image to be enhanced, the normalized spectrum and the edge spectrum to obtain a feature fusion image;

[0057] A feature extraction image acquisition module, used to extract features from the feature fusion image using a feature extraction module to obtain a feature extraction image; wherein the feature extraction module includes a two-dimensional convolution layer, a maximum pooling layer and a residual dense block;

[0058] A feature mapping image acquisition module, used to perform feature mapping on the feature extraction image using a feature mapping module to obtain a feature mapping image; wherein the feature mapping module includes a two-dimensional convolution layer and a group cross attention block, and the group cross attention block includes a grouping module, an average pooling layer, a feature fusion layer, a two-dimensional convolution layer, a weight redistribution layer and a group normalization layer;

[0059] The normal illumination image acquisition module is used to input the feature map image into a normalized flow reversible network and output a normal illumination image; wherein the normalized flow reversible network includes multiple flow steps and a two-dimensional compression layer.

[0060] The embodiments of the present invention have the following beneficial effects:

[0061] The present invention uses normalization and edge extraction techniques to pre-process and enhance the input image. This step not only enhances the detail representation of the image, but also highlights the edge information of the image, providing a solid foundation for the subsequent image processing stage. In this way, the present invention can significantly improve the visual quality and structural integrity of the image. Furthermore, the present invention improves the traditional flow-based generation model, and by introducing an encoder with a multi-scale attention mechanism, it can effectively extract deep features from low-light images. This feature extraction method can capture the essential details of the image at different scales, thereby enhancing the image quality during the image generation process and achieving clearer and more detailed image output. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0063] in:

[0064] Figure 1 is a flow chart of a low illumination image enhancement method based on an improved normalized flow in one embodiment;

[0065] Figure 2 The improved normalized flow low illumination enhancement network structure diagram in one embodiment;

[0066] Figure 3 A visualization diagram of some intermediate results in an embodiment;

[0067] Figure 4 The structure block diagram of a low-illumination image enhancement device based on improved normalized flow in one embodiment. DETAILED DESCRIPTION

[0068] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] Figure 1 FIG. 1 is a flow chart of a low-light image enhancement method based on an improved normalized flow in one embodiment. Figure 1 As shown, in one embodiment, a low-light image enhancement method based on improved normalized flow is provided. The method specifically includes S1100 to S1700:

[0070] S1100: Acquire a low-light image to be enhanced.

[0071] Low-illumination image enhancement technology is an important research direction in the field of image processing, which aims to improve the quality of images captured in low-light environments.

[0072] Image data are inherently discrete because they consist of a fixed number of pixels, each of which has a finite value, and can therefore be described by a discrete distribution. However, despite the discrete nature of images, it is often more convenient and flexible to use continuous distributions to model image data. p(y) = ∫ x p θ (y|x)p(x)dx is a continuous probability distribution that can theoretically fit any distribution, where p(x) is usually a standard Gaussian distribution and p θ (y|x) can be any conditional distribution.

[0073] In low-light image enhancement, our goal is to use a set of low-light x i and the normal illumination version y i The paired image set {x i ,y i} to estimate the conditional distribution p θ The parameters θ of (y|x) are usually estimated by maximum likelihood estimation, which involves solving the equation

[0074] argmaxΕ x～y [logp θ (y|x)]

[0075] However, due to the continuous conditional distribution p θ (y|x) is usually in integral form, and it is unknown whether the above equation can be solved.

[0076] The normalized flow introduces a reversible neural network fθ To parameterize p θ , f θ Map the image to an intermediate variable z=f θ (y; x), for each input x, there is a unique z corresponding to it in the latent space, and vice versa, which ensures that the network is strictly reversible. The normal lighting image y can be obtained by (i.e., inverse network) generation. By defining a simple distribution p in the latent space z z (e.g. standard Gaussian distribution), and use a reversible network f θ Mapping samples from this distribution into data space, we can calculate p by simple variable substitution: θ (y|x):

[0077]

[0078] Build p θ After the calculation formula of (y|x), we introduce negative log-likelihood as the loss function of model parameter θ:

[0079]

[0080] At the same time, in order to further improve the calculation efficiency and stability, f θ Decompose into N reversible layers The output of the nth layer is denoted as h n , considering the input and output, we can get h 0 =y,h N =y, the transformation is written as:

[0081]

[0082] After introducing the chain rule, the negative log-likelihood can be written as:

[0083]

[0084] Figure 2 FIG. 1 is a diagram showing the structure of a normalized low-light enhancement network in an embodiment. The low-light enhancement network designed in this embodiment mainly includes two main components: a conditional attention encoder, which first performs a l Processing is performed to produce an enhanced intermediate result x, and a reversible flow network that projects the intermediate result image x into a latent space z and reconstructs a normal lighting image y from this space under the guidance of the Ground Truth image, i.e., the output.

[0085] The steps of this embodiment are described in detail below with reference to the accompanying drawings:

[0086] S1200: performing normalization processing on the low-illuminance image to be enhanced to obtain a normalized spectrum.

[0087] The input image is processed using a conditional attention encoder: Figure 2 As shown in Figure 1, the conditional attention encoder is mainly used to extract image features and generate results at multiple scales to guide the image generation of the flow network. The first step of the encoder is to transform the low-light image x l Normalization is performed to ensure that the pixel values ​​are within a continuous range. This step can be written as:

[0088]

[0089] In the formula, x norm represents the normalized spectrum, x l represents the low-light image to be enhanced, ||x l ||1 represents the 1-norm of the low-light image to be enhanced.

[0090] S1300, performing edge extraction processing on the normalized spectrum to obtain an edge spectrum.

[0091] After the image is normalized, the next step is to calculate the edges within the image to highlight areas where the intensity changes rapidly. This is done by calculating the gradient values ​​in the horizontal and vertical directions at each pixel, that is, calculating the difference between adjacent pixels. The magnitude of the gradient indicates the strength of the edge, while the direction of the gradient indicates the direction of the edge, expressed as:

[0092]

[0093] In the formula, x edge represents the edge spectrum, represents the gradient value of the normalized spectrum in the horizontal direction, Represents the gradient value of the normalized spectrum in the vertical direction.

[0094] Figure 3 Figure 2 is a visualization diagram of some intermediate results in one embodiment. The visualization results of the normalized spectrum and the edge spectrum are shown in Figure 2. Figure 3 shown.

[0095] S1400, performing feature fusion on the low-illuminance image to be enhanced, the normalized spectrum, and the edge spectrum to obtain a feature fused image.

[0096] After obtaining the normalized spectrum and edge spectrum, these features are concatenated with the original low-light image in the channel dimension to obtain a feature fusion image. In this way, the feature dimension of the image is increased, providing richer information for subsequent feature extraction and image enhancement.

[0097] S1500, using a feature extraction module to extract features from the feature fusion image to obtain a feature extracted image.

[0098] Next, the feature extraction process is further refined using residual dense blocks (RRDBs), which enhance the model’s ability to capture complex image structures through dense connections and residual learning.

[0099] In some examples, the feature extraction module includes a two-dimensional convolutional layer, a maximum pooling layer, and a residual dense block.

[0100] In some examples, a feature extraction module is used in S1500 to extract features from the feature fusion image to obtain a feature extraction image, which specifically includes S1501 to S1503:

[0101] S1501, input the feature fusion image into the first two-dimensional convolution layer (1) and the second two-dimensional convolution layer (2) in sequence, and output a first convolution image (3).

[0102] S1502, input the first convolution image (3) into the maximum pooling layer (4) and the residual dense block (5) in sequence, and output the first image (6).

[0103] S1503, summing the first convolution image (3) and the first image (6) to obtain a feature extraction image (7).

[0104] S1600, using a feature mapping module to perform feature mapping on the feature extraction image to obtain a feature mapping image.

[0105] Grouped criss-cross attention (PCxA) refines feature maps by emphasizing salient channels and suppressing less informative channels, thereby improving the encoder's sensitivity to relevant image attributes.

[0106] The conditional attention encoder finally outputs a processed intermediate result x and its multiple downsampled features x ×2 , x ×4 , x ×8 . These features are designed to serve different flow steps within the network, corresponding to their downsampling scales. This design ensures that each step of the network can access features at the corresponding level of detail to guide the generation of enhanced images. The four images correspond to different feature levels and resolutions, which will be fed into different levels of the normalized flow reversible network, mainly for feature fusion to improve details and quality.

[0107] In some examples, the feature mapping module includes a two-dimensional convolution layer and a grouped cross-attention block, and the grouped cross-attention block includes a grouping module, an average pooling layer, a feature fusion layer, a two-dimensional convolution layer, a weight redistribution layer, and a group normalization layer.

[0108] The input feature map is divided into multiple groups in the channel dimension. The purpose of grouping is to apply different processing strategies on different channel groups to highlight important features and suppress unimportant information.

[0109] The feature maps of each group are processed through different paths. X average pooling and Y average pooling correspond to average pooling in the horizontal and vertical directions respectively, and then concatenated.

[0110] After grouping, the weights of the feature maps are redistributed to adjust the contributions of different groups, involving the learned weight matrix, which is adjusted according to the importance of the features. This is implemented by torch.matmul and reshape.

[0111] In some examples, S1600 uses a feature mapping module to perform feature mapping on the feature extraction image to obtain a feature mapping image, which specifically includes S1601 to S1608:

[0112] S1601, up-sample the feature extraction image (7) to obtain an up-sampled image (8).

[0113] S1602, input the upsampled image (8) into the third two-dimensional convolutional layer (9), and output a first feature map image (10).

[0114] S1603, input the feature extraction image (7) into the first grouped cross attention block (11), and output the second feature mapping image (12).

[0115] In S1603, the feature extraction image is input into the first group cross attention block, and the second feature mapping image is output, which specifically includes S101 to S113:

[0116] S101, performing grouping processing on the feature extraction image (7) to obtain a first group image (701), a second group image (702) and a third group image (703).

[0117] S102, performing average pooling in the horizontal direction on the second grouped image (702) to obtain a second pooled image (704).

[0118] S103, performing average pooling in the vertical direction on the third grouped image (703) to obtain a third pooled image (705).

[0119] S104, stitching the second pooled image (704) and the third pooled image (705) to obtain a stitched image (706).

[0120] S105, input the stitched image (706) into the sixth two-dimensional convolutional layer (707) to obtain a third image (708).

[0121] S106, redistribute the weights of the third image (708) and the first grouped image (701) to obtain a redistributed image (709).

[0122] S107, performing group normalization processing on the reallocated image (709) to obtain a normalized image (710).

[0123] S108, input the normalized image (710) into the first average pooling layer (711), and output a fourth image (712).

[0124] S109, input the first grouped image (701) into the seventh two-dimensional convolutional layer (713) to obtain a fifth image (714).

[0125] S110, multiply the fourth image (712) and the fifth image (714) to obtain a sixth image (715).

[0126] S111, the fifth image (714) is input into the second average pooling layer (716), and the seventh image (717) is output.

[0127] S112, multiplying the seventh image (717) and the normalized image (710) to obtain an eighth image (718).

[0128] S113, redistribute the weights of the first grouped image (701), the sixth image (715) and the eighth image (718), and output a second feature map image (12).

[0129] S1604, down-sample the feature extraction image (7) to obtain a first down-sampled image (13).

[0130] S1605, input the first down-sampled image (13) into a fourth two-dimensional convolutional layer (14), and output a second image (15).

[0131] S1606, input the second image (15) into the second grouped cross attention block (16), and output a third feature map image (17).

[0132] S1607, down-sample the second image (15) to obtain a second down-sampled image (18).

[0133] S1608, input the second down-sampled image (18) into the fifth two-dimensional convolutional layer (19) and the third grouped cross attention block (20) in sequence, and output a fourth feature map image (21).

[0134] S1700, input the feature map image into a normalized flow reversible network, and output a normal illumination image.

[0135] like Figure 2 As shown in the network on the right, the normalized flow reversible network consists of several flow steps and compression (decompression layers). The flow steps include activation normalization, 1×1 reversible convolution, and affine coupling transformation.

[0136] Activation normalization is used to ensure that the activation values ​​of each layer of the neural network maintain a standardized distribution. The specific goal is to make the mean zero and the variance one. By stabilizing the distribution of activation values, the activation normalization layer reduces the possibility of gradient vanishing or gradient exploding.

[0137] The 1×1 reversible convolution class uses 1×1 convolution to implement permutations. In reversible neural networks, especially in normalized flows, permutations refer to a reordering of data that is both reversible and differentiable. The weight matrix W is initialized using QR decomposition to ensure that it is orthogonal. An orthogonal matrix has the property that its inverse is equal to its transpose, which simplifies the inverse transformation process. In the forward propagation, the input is convolved with the weight matrix W; in the reverse propagation, the inverse of W is W. -1 The formula Calculate, where det(W) is the determinant of W.

[0138] The affine coupling layer is a key component in the normalized flow architecture, which facilitates the modeling of complex distributions through reversible transformations. This layer works by splitting the input data x into two parts x a and x b , then for x b Apply affine transformation to achieve this, that is, x b The transformation of x a This transformation can be expressed mathematically as:

[0139] x' b =α(x a ,c)·x b +β(x a ,c)

[0140] Among them, x' b represents the transformation result, α is the scaling factor, β is the translation factor, and c is the conditional information. The transformation is obviously reversible and can be expressed as:

[0141]

[0142] The design of the affine coupling layer ensures the reversible differentiability of the transformation, allowing end-to-end training using gradient descent-based optimization. The inclusion of conditional information c provides the flexibility to adapt to different contexts and enhances the model's ability to capture diverse data patterns.

[0143] The compression (decompression) layer is used as an effective dimensionality reduction and feature enhancement mechanism within the normalized flow framework. It is used to reduce the spatial dimension while expanding the channel dimension of the input data. During forward propagation, this layer compresses the input through a squeeze2d function, and during backward propagation, it uses an unsqueeze2d function to restore the original spatial dimension, which helps the model capture and express features more effectively when processing image data while reducing the computational burden.

[0144] The GROUND TRUTH image is used as a reference standard to guide the network to generate images that are closer to real lighting conditions. In the network, the GROUND TRUTH image is used to evaluate the difference between the generated image and the real image. The specific process includes comparing the GROUND TRUTH image with the image generated by the network, calculating the loss, and then updating the network parameters through back propagation to gradually reduce the difference between the generated image and the real image.

[0145] In S1700, the feature map image is input into the normalized flow reversible network, and a normal illumination image is output, which specifically includes:

[0146] The feature map image is sequentially input into the flow step and the 2D compression layer, and the normal illumination image is output.

[0147] This embodiment uses normalization and edge extraction techniques to pre-process and enhance the input image. This step not only enhances the detail representation of the image, but also highlights the edge information of the image, providing a solid foundation for the subsequent image processing stage. In this way, this embodiment can significantly improve the visual quality and structural integrity of the image. Furthermore, this embodiment improves the traditional flow-based generation model, and by introducing an encoder with a multi-scale attention mechanism, it can effectively extract deep features from low-light images. This feature extraction method can capture the essential details of the image at different scales, thereby enhancing the image quality during the image generation process and achieving clearer and more detailed image output.

[0148] Figure 4 FIG. 1 is a structural block diagram of a low-light image enhancement device based on an improved normalized flow in one embodiment. Figure 4 As shown, the device comprises:

[0149] The low-illumination image acquisition module 4100 is used to acquire the low-illumination image to be enhanced.

[0150] The normalized spectrum acquisition module 4200 is used to perform normalization processing on the low-illuminance image to be enhanced to obtain a normalized spectrum.

[0151] The edge spectrum acquisition module 4300 is used to perform edge extraction processing on the normalized spectrum to obtain an edge spectrum.

[0152] The feature fusion image acquisition module 4400 is used to perform feature fusion on the low-illuminance image to be enhanced, the normalized spectrum and the edge spectrum to obtain a feature fusion image.

[0153] The feature extraction image acquisition module 4500 is used to extract features from the feature fusion image using a feature extraction module to obtain a feature extraction image; wherein the feature extraction module includes a two-dimensional convolution layer, a maximum pooling layer and a residual dense block.

[0154] The feature mapping image acquisition module 4600 is used to perform feature mapping on the feature extraction image using the feature mapping module to obtain a feature mapping image; wherein the feature mapping module includes a two-dimensional convolution layer and a group cross attention block, and the group cross attention block includes a grouping module, an average pooling layer, a feature fusion layer, a two-dimensional convolution layer, a weight redistribution layer and a group normalization layer.

[0155] Normal illumination image acquisition 4700 is used to input the feature map image into a normalized flow reversible network and output a normal illumination image; wherein the normalized flow reversible network includes multiple flow steps and a two-dimensional compression layer.

[0156] It should be noted here that the above description of the embodiment of the low-light image enhancement device based on the improved normalized flow is similar to the above method description and has the same beneficial effects as the method embodiment. For technical details not disclosed in the embodiment of the low-light image enhancement device based on the improved normalized flow of the present invention, those skilled in the art should refer to the description of the method embodiment of the present invention for understanding.

[0157] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0158] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0159] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A low-light image enhancement method based on improved normalized flow, characterized in that: The method comprises: Acquire a low-light image to be enhanced; Performing normalization processing on the low-illuminance image to be enhanced to obtain a normalized spectrum; Performing edge extraction processing on the normalized spectrum to obtain an edge spectrum; Performing feature fusion on the low-illuminance image to be enhanced, the normalized spectrum and the edge spectrum to obtain a feature fused image; A feature extraction module is used to extract features from the feature fusion image to obtain a feature extraction image; wherein the feature extraction module includes a two-dimensional convolution layer, a maximum pooling layer and a residual dense block; A feature mapping module is used to perform feature mapping on the feature extraction image to obtain a feature mapping image; wherein the feature mapping module includes a two-dimensional convolution layer and a group cross attention block, and the group cross attention block includes a grouping module, an average pooling layer, a feature fusion layer, a two-dimensional convolution layer, a weight redistribution layer and a group normalization layer; The feature map image is input into a normalized flow reversible network, and a normal illumination image is output; wherein the normalized flow reversible network includes multiple flow steps and a two-dimensional compression layer.

2. The low-illumination image enhancement method based on improved normalized flow according to claim 1, characterized in that: The normalization process is specifically as follows: In the formula, x norm represents the normalized spectrum, x l represents the low-light image to be enhanced, ||x l ||1 represents the 1-norm of the low-light image to be enhanced.

3. The low-illumination image enhancement method based on improved normalized flow according to claim 1, characterized in that: The edge extraction process is specifically as follows: In the formula, x edge represents the edge spectrum, represents the gradient value of the normalized spectrum in the horizontal direction, Represents the gradient value of the normalized spectrum in the vertical direction.

4. The low-illumination image enhancement method based on improved normalized flow according to claim 1, characterized in that: The adopting a feature extraction module to extract features from the feature fusion image to obtain a feature extraction image specifically includes: Input the feature fusion image into the first two-dimensional convolution layer and the second two-dimensional convolution layer in sequence, and output the first convolution image; Input the first convolution image into the maximum pooling layer and the residual dense block in sequence, and output a first image; The first convolution image and the first image are summed to obtain a feature extraction image.

5. The low-illumination image enhancement method based on improved normalized flow according to claim 1, characterized in that: The adopting a feature mapping module to perform feature mapping on the feature extraction image to obtain a feature mapping image specifically includes: Upsampling the feature extraction image to obtain an upsampled image; Inputting the upsampled image into a third two-dimensional convolutional layer, and outputting a first feature map image; Input the feature extraction image into a first grouped cross attention block, and output a second feature mapping image; Downsampling the feature extraction image to obtain a first downsampled image; Input the first down-sampled image into a fourth two-dimensional convolutional layer, and output a second image; Input the second image into a second grouped cross attention block, and output a third feature map image; Downsampling the second image to obtain a second downsampled image; The second down-sampled image is sequentially input into the fifth two-dimensional convolutional layer and the third grouped cross attention block, and a fourth feature map image is output.

6. The low-illumination image enhancement method based on improved normalized flow according to claim 5, characterized in that: The step of inputting the feature extraction image into a first grouped cross attention block and outputting a second feature mapping image specifically includes: Performing group processing on the feature extraction images to obtain a first group image, a second group image and a third group image; Performing average pooling in a horizontal direction on the second group of images to obtain a second pooled image; Performing average pooling in a vertical direction on the third grouped images to obtain a third pooled image; splicing the second pooled image and the third pooled image to obtain a spliced ​​image; Inputting the stitched image into a sixth two-dimensional convolutional layer to obtain a third image; Redistributing weights of the third image and the first grouped image to obtain a redistributed image; Performing group normalization processing on the reallocated image to obtain a normalized image; Input the normalized image into a first average pooling layer, and output a fourth image; Inputting the first grouped image into a seventh two-dimensional convolutional layer to obtain a fifth image; multiplying the fourth image and the fifth image to obtain a sixth image; Input the fifth image into a second average pooling layer, and output a seventh image; multiplying the seventh image and the normalized image to obtain an eighth image; The first grouped image, the sixth image and the eighth image are weighted redistributed to output a second feature map image.

7. The low-light image enhancement method based on improved normalized flow according to claim 1, characterized in that: The flow steps include activation normalization, 1×1 reversible convolutional layers, and affine coupled transformation layers.

8. The low-illumination image enhancement method based on improved normalized flow according to claim 1, characterized in that: The normalized flow reversible network also includes a two-dimensional decompression layer.

9. The low-illumination image enhancement method based on improved normalized flow according to claim 1, characterized in that: The step of inputting the feature map image into a normalized flow reversible network and outputting a normal illumination image specifically includes: The feature map image is sequentially input into the flow step and the 2D compression layer, and the normal illumination image is output.

10. A low-light image enhancement device based on improved normalized flow, characterized in that: The device comprises: A low-light image acquisition module, used to acquire a low-light image to be enhanced; A normalized spectrum acquisition module, used for performing normalization processing on the low-illuminance image to be enhanced to obtain a normalized spectrum; An edge spectrum acquisition module, used for performing edge extraction processing on the normalized spectrum to obtain an edge spectrum; A feature fusion image acquisition module is used to perform feature fusion on the low-illuminance image to be enhanced, the normalized spectrum and the edge spectrum to obtain a feature fusion image; A feature extraction image acquisition module, used to extract features from the feature fusion image using a feature extraction module to obtain a feature extraction image; wherein the feature extraction module includes a two-dimensional convolution layer, a maximum pooling layer and a residual dense block; A feature mapping image acquisition module, used to perform feature mapping on the feature extraction image using a feature mapping module to obtain a feature mapping image; wherein the feature mapping module includes a two-dimensional convolution layer and a group cross attention block, and the group cross attention block includes a grouping module, an average pooling layer, a feature fusion layer, a two-dimensional convolution layer, a weight redistribution layer and a group normalization layer; The normal illumination image acquisition module is used to input the feature map image into a normalized flow reversible network and output a normal illumination image; wherein the normalized flow reversible network includes multiple flow steps and a two-dimensional compression layer.

Images