Low-light image enhancement method based on progressive feature-aware recurrent deep network

Through a progressive semantic progressive loop deep network based on Retinex theory, the lighting and color information of low-illumination images are gradually restored, and the confusion problem of sync learning of lighting and color information in the existing technology is solved, achieving better robustness and Adaptability ensures image enhancement effect.

CN115147300BActive Publication Date: 2024-11-15TAIYUAN UNIVERSITY OF TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210706850.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-06-21
Publication Date
2024-11-15
Estimated Expiration
2042-06-21

AI Technical Summary

Technical Problem

Existing low-illumination image enhancement methods based on deep learning are prone to confusion when learning lighting and color information synchronously, resulting in problems such as color distortion, overexposed or insufficient exposure, and insufficient robustness and generalization capabilities.

Method used

Using a progressive semantic progressive cyclic deep network based on Retinex theory, the lighting and color information are gradually restored through the light compensation subnet, the color compensation subnet and the refined restoration atomic network, and the coordinated attention mechanism and the LSTM cyclic convolution module are used to perform features. Semantic learning, set appropriate loss functions for iterative training.

Benefits of technology

Good adaptability on multiple public low-light image data sets is achieved, the robustness and universality of image enhancement is improved, color distortion is avoided, and accurate recovery of lighting and color information is ensured.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115147300B_ABST
    Figure CN115147300B_ABST
Patent Text Reader

Abstract

The present invention discloses a low-light image enhancement method based on a progressive feature-aware recurrent deep network. In the low-light image enhancement task, simultaneously enhancing the illumination and color information is likely to cause confusion between the two, resulting in color distortion, overexposure and underexposure in the enhancement results. To this end, we propose a progressive feature-aware semantic low-light enhancement recurrent network with a coordinated attention mechanism to gradually restore the illumination and color information of the image. Specifically, the illumination is adaptively enhanced through a pixel-level attention recurrent subnetwork. Under the guidance of the illumination information, the color enhancement subnetwork uses a channel attention mechanism to balance the enhancement of the HSV color space channels. Finally, the preliminary restored image is further enhanced through a dual attention recurrent refinement subnetwork. This progressive learning mode enables each subnetwork to focus on the learning of specific feature semantics, and the entire network achieves multi-semantic cumulative learning.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of low-light image enhancement, and in particular relates to a low-light image enhancement method based on a progressive feature-aware recurrent deep network. Background Art

[0002] Images are the basis of vision and the source of data for computer vision systems to complete autonomous perception and understanding of the environment. In outdoor unstructured visual perception systems, the lighting conditions are relatively complex, and the images obtained by the visual perception system are prone to image quality degradation due to insufficient lighting, which seriously affects the robustness and environmental adaptability of many high-level visual tasks of outdoor robots, such as target recognition, scene understanding, autonomous driving visual perception, image segmentation, etc., making it difficult for the reliability of the visual system to meet actual requirements. Therefore, the problem of autonomous perception of robots in complex lighting environments has become increasingly prominent. How to restore clear image data from a large amount of visual observation data under low illumination and improve the robot's autonomous perception and adaptability to the environment has become a research hotspot in the fields of robotics, pattern recognition, and computer vision.

[0003] Image low-light enhancement technology only obtains normal-light images through low-light images, which is an ill-posed problem in mathematical modeling and solution. At present, this technology is mainly divided into two categories: optimization enhancement methods based on prior information and enhancement methods based on deep learning. Image restoration technology based on prior optimization mainly obtains the best estimate of the clear image under the constraints of prior information. These methods have a common limitation: their effectiveness depends largely on the accuracy of prior information, and usually can only process some degraded images that meet specific priors. Image restoration technology based on deep learning uses deep networks to establish a mapping relationship from degraded low-light images to clear images. At present, these deep learning-based methods are usually better than prior optimization methods in terms of enhancement effect.

[0004] Although existing deep learning-based low-light enhancement methods can obtain images with normal lighting, the robustness of data-driven deep networks is imperfect because low-light images do not have unique corresponding normal-light images. In addition, these deep learning low-light enhancement methods can easily confuse lighting and color information by learning them simultaneously, resulting in color distortion, overexposure, or underexposure. Summary of the invention

[0005] At present, low-light enhancement algorithms simultaneously enhance illumination and color information, which can easily cause confusion between the two. In view of the above technical deficiencies, the present invention proposes a progressive semantic progressive recurrent network with coordinated attention for low-light enhancement. It progressively restores illumination and color information and further enhances the preliminary results. This progressive learning mode enables each sub-network to focus on specific feature semantic learning, and the entire network can achieve multi-feature semantic accumulation learning. The coordinated attention mechanism can adaptively recalibrate the feature responses of the three sub-networks to appropriately restore illumination and color information. This method has good robustness, strong generalization ability, and fast processing speed, and can show good adaptability on multiple public low-light image datasets.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A low-light image enhancement method based on a semantic progressive recurrent deep network comprises the following steps:

[0008] Step 1: Based on the Retinex theory, a low-light image decomposition model is established to decompose the low-light image into the product of the normal illumination image and the residual illumination image, namely:

[0009] I l (x,y)=I n (x,y)*L r (x,y)

[0010] Among them, I l (x, y) represents the low-light image, I n (x, y) represents the normal lighting image, L r (x,y) represents the resulting residual illumination image;

[0011] Step 2: Input the low-light image into the illumination compensation subnetwork to extract the illumination information features of the image and obtain an image with normal illumination;

[0012] Step 3: Connect the low-light image in series with the image with normal illumination obtained in step 2 and input them into the color compensation subnetwork to obtain a preliminary restored image with normal illumination and color.

[0013] Step 4, inputting the preliminary restored image obtained in step 3 into the refined restoration atomic network to further enhance the preliminary restored image;

[0014] Step 5: Set the loss functions of the illumination compensation subnetwork, color compensation subnetwork, and refined complex atomic network. Under the constraints of the loss functions, iteratively train each subnetwork by learning the network model parameters on the real data set until the network converges.

[0015] Furthermore, the illumination compensation subnetwork in step 2 is: two front-end convolutional layers connected in series, a first-stage LSTM circular convolutional module, and two back-end convolutional layers. The first-stage LSTM circular convolutional module is composed of three groups of LSTM convolutional layers and a pixel attention layer connected in series.

[0016] Furthermore, in step 3, the low-light image is converted from the RGB space to the HSV space.

[0017] Furthermore, the color compensation subnetwork in step 3 is: a front-end convolutional layer connected in series, a second-stage LSTM circular convolutional module, and two back-end convolutional layers, and the second-stage LSTM circular convolutional module is composed of three groups of LSTM convolutional layers and channel attention layers connected in series.

[0018] Furthermore, the refined complex atomic network in step 4 is: a front-end convolutional layer connected in series, a third-stage LSTM recurrent convolutional module, and two back-end convolutional layers, and the third-stage LSTM recurrent convolutional module is composed of three groups of LSTM convolutional layers connected in series and a dual attention layer.

[0019] Furthermore, it is characterized in that the LSTM convolutional layer is composed of a forget gate f t , an input gate i t , an output gate o t and a cell state C t The interaction between the state gates along the time dimension is defined as follows:

[0020] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0021] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0022] C t =f t *C t-1 +i t *tanh(W c ·[h t-1 ,x t ]+b c )

[0023] o t =σ(W o ·[h t-1 ,x t ]+b o)

[0024] h t =o t *tanh(C t )

[0025] Among them, x t is the input convolutional layer L in The generated features, C t is the cell state fed to the next LSTM convolutional layer, h t is the output feature of the LSTM convolutional layer, * represents element-by-element multiplication, [h t-1 ,x t ] represents the cascade operation, σ represents the Sigmoid function, W f represents the forget gate f t The convolution operation coefficient, W i Represents the input gate i t The convolution operation coefficient, W c Indicates the unit state C t The convolution operation coefficient, W o Represents the output gate o t The convolution operation coefficient, b f represents the forget gate f t The bias of b i Represents the input gate i t The bias of b c Indicates the unit state C t The bias of b o Represents the output gate o t The bias.

[0026] Furthermore, the pixel attention layer consists of two compressed 1*1 convolutional layers, expressed as:

[0027] M p (F) = σ(f conv1*1 (δ(f conv1*1 (F))))

[0028] Where δ and σ represent the ReLU function and the Sigmoid function, f conv1*1 represents a 1*1 convolution operation, and F represents the input feature map.

[0029] Furthermore, the channel attention layer adopts an improved SE attention module to obtain channel attention weights through compression operation and excitation operation. The compression operation uses global average pooling and maximum pooling to convert global spatial information into channel statistics. The cth statistic z c The calculation formula is:

[0030]

[0031] where u c represents the cth feature map output by the LSTM convolutional layer, W and H represent the width and height of the feature map respectively;

[0032] The excitation operation adopts a multi-layer perceptron threshold mechanism, which first compresses the output channel and then upgrades the dimension to the number of channels before compression. The above operation is described as:

[0033] s=f ex (z,w)=σ(W2(W1z))

[0034] Where σ represents the Sigmoid function, W1 and W2 are the parameters of the multi-layer perceptron, and the output of the channel attention layer is:

[0035] u o =f scale (u c ,s c )=u c ·s c

[0036] Among them, f scale (·) represents the LSTM convolutional layer output u c With scalar s c The channel-level multiplication between .

[0037] Furthermore, the illumination loss function in step 5 is:

[0038]

[0039] Where N represents the total number of training data, I v and Represent the real illumination value and predicted illumination value of the normal illumination image respectively. and Represent the weights of the illumination channel balancing absolute loss and SSIM loss respectively;

[0040] The color loss function is:

[0041]

[0042] Among them, I hsv and Respectively represent the true value and predicted value of the image in the HSV space, and Represent the weights of HSV space balanced absolute loss and SSIM loss respectively;

[0043] The loss function of the refined complex atomic network is:

[0044]

[0045] Among them, I and They represent the real normal illumination image value and the predicted normal illumination image value in RGB space, respectively, al and γ ssim Represent the weights of RGB space balanced absolute loss and SSIM loss respectively.

[0046] Furthermore, the iterative training method in step 5 is:

[0047] First, the illumination compensation sub-network is trained separately using the low-illumination image. Then, the estimated normal illumination map is connected in series with the HSV color channel of the low-illumination image to train the color compensation network separately. Finally, the entire network is optimized to obtain the final optimization result.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] 1. The method of the present invention establishes a low-illumination image decomposition method based on the Retinex theory, decomposes the low-illumination image into a normal illumination image and a residual illumination image, describes the relationship between the various imaging factors of the low-illumination image, and effectively uses the Retinex theory to guide the image restoration process.

[0050] 2. The method of the present invention establishes a cumulative learning model that coordinates attention mechanism and progressive feature semantics.

[0051] 3. The method of the present invention establishes an adaptive illumination compensation subnetwork, which realizes adaptive illumination restoration of areas with illumination differences in the image by using pixel-level attention.

[0052] 4. The method of the present invention establishes a color compensation subnetwork of the HSV color space. Under the guidance of normal illumination, the network uses a channel attention mechanism to enable the color subnetwork to adaptively calibrate the compensation ratio of each channel in the HSV color space, so that the restored image does not have color distortion.

[0053] 5. The method of the present invention establishes a refined enhancement sub-network with dual attention, which allows further enhancement of the initial restored image and ultimately achieves multi-feature semantic progressive accumulation learning.

[0054] 6. The illumination loss function and HSV space color loss function proposed by the method of the present invention can better guide the training of the network and obtain the final normal illumination image.

[0055] 7. The present invention adopts a cyclic structure to gradually restore the illumination, color and other characteristic information of the low-light image.

[0056] 8. The final image enhancement result of the method of the present invention is better. Compared with other methods, the method of the present invention has better universality and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is the overall deep network framework of the present invention;

[0058] Figure 2 It is the LSTM module used in the present invention;

[0059] Figure 3 is the pixel and channel attention module used in the present invention;

[0060] Figure 4 It is the result of local low-light image enhancement;

[0061] Figure 5 It is the global low-illumination image enhancement result. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. The specific examples mentioned herein are only used to explain the present invention and do not limit the present invention.

[0063] Example 1

[0064] like Figure 1 As shown, a low-light image enhancement method based on a semantic progressive recurrent deep network in this embodiment includes the following steps:

[0065] Step 1: Based on the Retinex theory, a low-light image decomposition model is established to decompose the low-light image into the product of the normal illumination image and the residual illumination image, namely:

[0066] I l (x,y)=I n (x,y)*L r (x,y)

[0067] Among them, I l (x, y) represents the low-light image, I n (x, y) represents the normal lighting image, L r (x,y) represents the resulting residual illumination image;

[0068] Retinex theory is an image enhancement method based on scientific experiments and human visual perception. The image imaging theory based on Retinex believes that any image can be modeled as the product of the reflection map R(x, y) and the illumination map L(x, y), which can be expressed as:

[0069] I(x,y)=R(x,y)*L(x,y)

[0070] Where * indicates pixel-by-pixel multiplication. The reflectance map R(x,y) describes the inherent properties of the reflected background object, which is considered to be consistent under any brightness conditions. The illumination map L(x,y) represents the brightness of each pixel in the image. In low-light images, darkness and uneven weak light distribution usually appear. Similarly, the mapping function in the Retinex-based low-light image restoration method is modeled as:

[0071] I l (x,y)=R(x,y)*L(x,y)*L r (x,y)

[0072] Among them I l (x, y) represents the low-light image, L r (x, y) represents the residual illumination image. Because the solution of the above model is an ill-posed problem, there are many ways to decompose the image in the actual restoration process. Sometimes part of the residual illumination is mistakenly decomposed into a reflection map, which will cause the color of the restoration result to be distorted. In this case, the mapping function of the image restoration method is:

[0073] I lh (x,y)=R(x,y)*R r1 (x,y)*L(x,y)*L r1 (x,y)

[0074] Where L r (x,y)=R r1 (x,y)*L r1 (x, y) represents a pseudo image. In order to avoid the generation of pseudo images during the decomposition process, the present invention progressively restores the low-illuminance image in the illumination layer and the reflection layer, and then performs fine restoration on the preliminary restored image.

[0075] Step 2: Input the low-light image into the illumination compensation subnetwork, extract the illumination information features of the image, and obtain an image with normal illumination. The illumination compensation subnetwork is composed of two front-end convolutional layers connected in series, a first-stage LSTM circular convolutional module, and two back-end convolutional layers. The first-stage LSTM circular convolutional module consists of three groups of LSTM convolutional layers connected in series and a pixel attention layer.

[0076] The front-end convolution block is used to increase the number of channels and feature fusion (category features and low-level features), and its output is sent to the first-stage LSTM circular convolution module. The LSTM circular convolution module retains the features that are conducive to illumination restoration and forgets the features that are irrelevant to illumination. The back-end convolution module reduces the dimension and fuses the features output by the LSTM circular convolution module. Finally, an image with normal illumination is output.

[0077] like Figure 2 As shown, the LSTM convolutional layer consists of a forget gate f t , an input gate i t , an output gate o t and a cell state C t The interaction between the state gates along the time dimension is defined as follows:

[0078] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0079] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0080] C t =f t *C t-1 +i t *tanh(W c ·[h t-1 ,x t ]+b c )

[0081] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0082] h t =o t *tanh(C t )

[0083] Among them, x t is the input convolutional layer L in The generated features, C t is the cell state fed to the next LSTM convolutional layer, h t is the output feature of the LSTM convolutional layer, * represents element-by-element multiplication, [h t-1 ,x t ] represents the cascade operation, σ represents the Sigmoid function, W f represents the forget gate f t The convolution operation coefficient, W i Represents the input gate i t The convolution operation coefficient, W c Indicates the unit state C tThe convolution operation coefficient, W o Represents the output gate o t The convolution operation coefficient, b f represents the forget gate f t The bias of b i Represents the input gate i t The bias of b c Indicates the unit state C t The bias of b o Represents the output gate o t The bias.

[0084] After each set of LSTM convolutional layers and pixel attention layers in series, the output of the current LSTM convolutional layer and pixel attention layer in series is sent to the next set, and the intermediate state of each set is input into the feature semantic extraction subnetwork of the next stage of the progressive recurrent network, so that the feature semantic learning of the next stage can be carried out on the basis of the learning of the previous stage.

[0085] The pixel attention layer uses Figure 3 The structure shown in a is composed of two layers of compressed 1*1 convolutional layers. The first 1*1 convolutional layer extracts the global features of the input image by running through the channel dimension. The second 1*1 convolutional layer reduces the dimension and fuses the information together, which can be expressed as:

[0086] M p (F) = σ(f conv1*1 (δ(f conv1*1 (F))))

[0087] Where δ and σ represent the ReLU function and the Sigmoid function, f conv1*1 represents a 1*1 convolution operation, and F represents the input feature map.

[0088] Step 3: Convert the low-light image from RGB space to HSV space, connect it in series with the image with normal illumination obtained in step 2, and input it into the color compensation subnetwork to obtain a preliminary restored image with normal illumination and color. The color compensation subnetwork is composed of a front-end convolution layer, a second-stage LSTM circular convolution module, and two back-end convolution layers in series. The second-stage LSTM circular convolution module consists of three groups of LSTM convolution layers and a channel attention layer in series.

[0089] Compared with normal illumination images, there are some areas with drastic illumination changes in low illumination images, and such drastic changes will cause a certain loss of color information in the image. In the image RGB space, the effects of illumination changes on the three channels are basically the same. If the image is restored directly from the RGB space, the network is likely to only focus on the common changes in the image and ignore the differences (color). Even with channel attention, the imbalance of the three-channel attention can easily lead to color distortion. On the contrary, in the HSV color space, the effects of illumination changes on the three channels are quite different, and the HSV space can better reflect the differences in illumination and color loss in low illumination images. In the HSV color space, hue H is measured in angles, ranging from 0° to 360°. Saturation S indicates the degree to which the color is close to the spectral color. The greater the proportion of the spectral color, the higher the saturation of the color. Usually, S ranges from 0% to 100%. Brightness V indicates the brightness of the color, usually ranging from 0% (black) to 100% (white).

[0090] Based on the above analysis, the color of the image is restored in the HSV space on the basis of preliminary restoration of the illumination information. First, the result of the illumination sub-network restoration is connected to the chromaticity channel (H) and saturation channel (S) of the low-light image as the brightness information (V) of the image and input into the color compensation sub-network. It first extracts features of the low-light image in the HSV space, and then sends the low-level features to three groups of LSTM convolutional layers and channel attention layers connected in series. Because the effect of illumination on the three HSV channels is inconsistent, the channel attention mechanism is used to give different attention to each channel after each group of LSTM convolutional layers and channel attention layers connected in series.

[0091] The channel attention layer structure is as follows Figure 3 As shown in (b), the improved SE attention module is used to obtain the channel attention weight through compression operation and excitation operation. The compression operation uses global average pooling and maximum pooling to convert the global spatial information into channel statistics. The cth statistic z c The calculation formula is:

[0092]

[0093] where u c represents the cth feature map output by the LSTM convolutional layer, W and H represent the width and height of the feature map respectively;

[0094] The excitation operation adopts a multi-layer perceptron threshold mechanism, which first compresses the output channel and then upgrades the dimension to the number of channels before compression. The above operation is described as:

[0095] s=f ex (z,w)=σ(W2(W1z))

[0096] Where σ represents the Sigmoid function, and W1 and W2 are the parameters of the multilayer perceptron.

[0097] The output of the channel attention layer is:

[0098] u o =f scale (u c ,s c )=u c ·s c

[0099] Among them, f scale (·) represents the LSTM convolutional layer output u c With scalar s c The channel-level multiplication between .

[0100] Step 4, input the preliminary restored image obtained in step 3 into the refined complex sub-network to further enhance the preliminary restored image. The refined complex sub-network is: a front-end convolutional layer in series, a third-stage LSTM circular convolutional module, and two back-end convolutional layers. The third-stage LSTM circular convolutional module is composed of three groups of series-connected LSTM convolutional layers and a dual attention layer. The convolutional module realizes dimensionality reduction and information fusion, and finally outputs an image with normal lighting and color.

[0101] The structure of the enhanced refinement subnetwork is similar to that of the color compensation subnetwork, but the feature extraction and restoration of low-light images in the enhanced subnetwork are both performed in the RGB space.

[0102] Step 5: Set the loss functions of the illumination compensation subnetwork, color compensation subnetwork, and refined complex atomic network. Under the constraints of the loss functions, iteratively train each subnetwork by learning the network model parameters on the real data set until the network converges.

[0103] In order to avoid the square penalty in the square loss amplifying the difference between large errors and small errors, which will cause the restored image to be overly smooth. The network uses the mean absolute error L al As a loss function, in order to make the restored normal lighting image more in line with human visual effects, SSIM loss is introduced to constrain the learning of network parameters.

[0104] The illumination loss function is:

[0105]

[0106] Where N represents the total number of training data, I v and Represent the real illumination value and predicted illumination value of the normal illumination image respectively. and Represent the weights of the illumination channel balancing absolute loss and SSIM loss respectively;

[0107] The color loss function is:

[0108]

[0109] Among them, I hsv and Respectively represent the true value and predicted value of the image in the HSV space, and Represent the weights of HSV space balanced absolute loss and SSIM loss respectively;

[0110] The loss function of the refined complex atomic network is:

[0111]

[0112] Among them, I and They represent the real normal illumination image value and the predicted normal illumination image value in RGB space, respectively, al and γ ssim Represent the weights of RGB space balanced absolute loss and SSIM loss respectively.

[0113] The iterative training method is:

[0114] Firstly, the illumination compensation sub-network is trained separately using the low-illuminance image. Then, the estimated normal illumination map is connected in series with the HSV color channel of the low-illuminance image to train the color compensation network separately. Finally, the entire network is optimized to obtain the final optimization result.

[0115] Figure 4 and Figure 5 is the low-light enhancement result of the method of the present invention on different data sets. The method of the present invention can process local low-light images ( Figure 4 ), and can also process global low-light images ( Figure 5 ), achieving a balance between over-enhancement and under-enhancement of lighting, while ensuring that colors are not distorted.

Claims

1. A low-light image enhancement method based on a progressive feature-aware recurrent deep network. It is characterized in that The following steps are involved: Step 1: Based on the Retinex theory, a low-light image decomposition model is established to decompose the low-light image into the product of the normal illumination image and the residual illumination image, namely: I l (x,y)=I n (x,y)*L r (x,y) Among them, Il(x,y) represents the low-light image, In(x,y) represents the normal-light image, and L r (x,y) represents the resulting residual illumination image; Step 2: Input the low-light image into the illumination compensation subnetwork to extract the illumination information features of the image and obtain an image with normal illumination; Step 3: Connect the low-light image in series with the image with normal illumination obtained in step 2 and input them into the color compensation subnetwork to obtain a preliminary restored image with normal illumination and color. Step 4, inputting the preliminary restored image obtained in step 3 into the refined restoration atomic network to further enhance the preliminary restored image; Step 5: Set the loss functions of the illumination compensation sub-network, the color compensation sub-network, and the refined complex atomic network. Under the constraints of the loss functions, iteratively train each sub-network by learning the network model parameters on the real data set until the network converges. The illumination compensation subnetwork in step 2 is: two front-end convolutional layers connected in series, a first-stage LSTM circular convolutional module, and two back-end convolutional layers, wherein the first-stage LSTM circular convolutional module is composed of three groups of LSTM convolutional layers connected in series and a pixel attention layer; The color compensation subnetwork in step 3 is: a front-end convolutional layer, a second-stage LSTM circular convolutional module, and two back-end convolutional layers connected in series, wherein the second-stage LSTM circular convolutional module is composed of three groups of LSTM convolutional layers and channel attention layers connected in series; The refined complex atomic network in step 4 is: a front-end convolutional layer, a third-stage LSTM recurrent convolutional module, and two back-end convolutional layers connected in series, wherein the third-stage LSTM recurrent convolutional module is composed of three groups of LSTM convolutional layers connected in series and a dual attention layer; The illumination loss function in step 5 is: Where N represents the total number of training data, I v and Represent the real illumination value and predicted illumination value of the normal illumination image respectively. and Represent the weights of the illumination channel balancing absolute loss and SSIM loss respectively; The color loss function is: Among them, I hsv and Respectively represent the true value and predicted value of the image in the HSV space, and Represent the weights of HSV space balanced absolute loss and SSIM loss respectively; The loss function of the refined complex atomic network is: Among them, I and They represent the real normal illumination image value and the predicted normal illumination image value in RGB space, respectively, al and γ ssim Represent the weights of RGB space balanced absolute loss and SSIM loss respectively.

2. According to claim 1, a low-light image enhancement method based on a progressive feature-aware recurrent deep network, It is characterized in that In step 3, the low-light image is converted from the RGB space to the HSV space.

3. According to claim 1, a low-light image enhancement method based on a progressive feature-aware recurrent deep network, It is characterized in that The LSTM convolutional layer consists of a forget gate f t , an input gate i t , an output gate o t and a cell state C t The interaction between the state gates along the time dimension is defined as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) C t =f t *C t-1 +i t *tanh(W c ·[h t-1 ,x t ]+b c ) the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t *tanh(C t ) Among them, x t is the input convolutional layer L in The generated features, C t is the cell state fed to the next LSTM convolutional layer, h t is the output feature of the LSTM convolutional layer, * represents element-by-element multiplication, [h t-1 ,x t ] represents the cascade operation, σ represents the Sigmoid function, W f represents the forget gate f t The convolution operation coefficient, W i Represents the input gate i t The convolution operation coefficient, W c Indicates the unit state C t The convolution operation coefficient, W o Represents the output gate o t The convolution operation coefficient, b f represents the forget gate f t The bias of b i Represents the input gate i t The bias of b c Indicates the unit state C t The bias of b o Represents the output gate o t The bias.

4. According to claim 1, a low-light image enhancement method based on a progressive feature-aware recurrent deep network, It is characterized in that The pixel attention layer consists of two compressed 1*1 convolutional layers, expressed as: M p (F)=σ(f conv1*1 (δ(f conv1*1 (F)))) Where δ and σ represent the ReLU function and the Sigmoid function, f conv1*1 represents a 1*1 convolution operation, and F represents the input feature map.

5. The low-light image enhancement method based on a progressive feature-aware recurrent deep network according to claim 1, It is characterized in that The channel attention layer adopts an improved SE attention module to obtain channel attention weights through compression operation and excitation operation. The compression operation uses global average pooling and maximum pooling to convert global spatial information into channel statistics. The cth statistic z c The calculation formula is: where u c represents the cth feature map output by the LSTM convolutional layer, W and H represent the width and height of the feature map respectively; The excitation operation adopts a multi-layer perceptron threshold mechanism, which first compresses the output channel and then upgrades the dimension to the number of channels before compression. The above operation is described as: s=f ex (with,in)=σ(In 2 (IN 1 With)) Where σ represents the Sigmoid function, W 1 and W 2 is the parameter of the multi-layer perceptron, and the output of the channel attention layer is: in o =f scale (in c ,with c )=in c ·with c Among them, f scale (·) represents the LSTM convolutional layer output u c With scalar s c The channel-level multiplication between .

6. The low-light image enhancement method based on a progressive feature-aware recurrent deep network according to claim 1, It is characterized in that The method of iterative training in step 5 is: Firstly, the illumination compensation sub-network is trained separately using the low-illuminance image. Then, the estimated normal illumination map is connected in series with the HSV color channel of the low-illuminance image to train the color compensation network separately. Finally, the entire network is optimized to obtain the final optimization result.