Partitioned exposure correction method based on deep learning

Abstract

The application discloses a kind of partition exposure correction methods based on deep learning, comprising: constructing partition exposure correction network based on deep learning: input original image into illumination attention estimation network, and obtain estimated illumination attention map;Original image is input into partition convolution enhancement network, and partition convolution enhancement network is according to the illumination attention map to carry out partition enhancement to different exposure regions of image;After inputting the image after partition enhancement into high-frequency feature fine network, output is obtained;After training data is input into the illumination attention estimation network and is frozen, it is input into partition convolution enhancement network and high-frequency feature fine network again and is trained, and loss function is constructed to carry out repeated iteration, and partition exposure correction network is trained to obtain.The application solves the problem of image exposure and detail loss caused by poor light.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to a partition exposure correction method based on deep learning. BACKGROUND

[0002] Uneven illumination often brings negative effects to images, easily causes overexposure or underexposure problems of images, brings negative visual feelings to people, and also causes loss of image details.

[0003] Traditional methods for estimating illumination attention maps of images often have problems of not adapting to multiple scenes, inaccurate illumination estimation, or non-structured illumination mapping. For example, some traditional methods for calculating illumination maps, such as gray level histogram and Retinex theory. The method of calculating the gray level histogram as the illumination map often cannot find a good balance standard to divide different degrees of illumination, and usually is a relatively unified division standard, which often cannot well handle the case that the training data set contains different scenes or the images have large differences in illumination conditions. In this case, the illumination mapping obtained by this method often cannot reflect the real illumination of the image, and cannot better adapt to multiple image scenes. The illumination mapping obtained by the Retinex theory method is usually not adaptive to the image structure, and the estimated illumination attention map often presents a square shape, so the image recovered according to the illumination map obtained by this method often has some bright spot problems in local regions, thereby causing uneven illumination of the image.

[0004] Moreover, overexposed regions and underexposed regions are often two opposite problems, one of which needs to suppress illumination, and the other of which needs to enhance brightness. Existing methods often cannot well handle both problems at the same time, such as solving the overexposure problem but still having the underexposure problem, or solving the underexposure problem but still having the overexposure problem, easily enhancing insufficiently or overenhancing, and being difficult to obtain a relatively balanced result that takes both problems into account. SUMMARY

[0005] (I) Technical problems to be solved

[0006] Based on the above problems, the present application provides a partition exposure correction method based on deep learning, which solves the problems of image exposure and detail loss caused by poor illumination, and can not only solve the single underexposure problem or overexposure problem of images, but also solve the overexposure problem and underexposure problem of images existing at the same time.

[0007] (II) Technical solutions

[0008] Based on the above technical problems, the present application provides a partition exposure correction method based on deep learning, comprising:

[0009] S 1, construct a deep learning-based partition exposure correction network comprising a light attention estimation network, a partition convolution enhancement network, and a high-frequency feature fine network;

[0010] S 11, input the original image into the light attention estimation network to obtain an estimated light attention map, the light attention map dividing different exposure regions;

[0011] S 12, input the original image into the partition convolution enhancement network, and the partition convolution enhancement network performs partition enhancement on different exposure regions of the image according to the light attention map;

[0012] S13, input the partition-enhanced image into the high-frequency feature fine network to output, the high-frequency feature fine network being used for restoring high-frequency information;

[0013] S2, after inputting the training data into the light attention estimation network for training, freezing the light attention estimation network, inputting the partition convolution enhancement network and the high-frequency feature fine network for training, and constructing a loss function for repeated iteration, the training is ended when the maximum iteration number is reached, and a trained partition exposure correction network is obtained.

[0014] Further, in step S12, the partition convolution enhancement network comprises: an encoder part comprising three sub-regional convolution enhancement modules connected in sequence, and a decoder comprising three convolution enhancement modules connected in sequence, and a skip connection is added to each layer of the encoder and the decoder to connect the features of the same dimension of the encoder and the decoder; the sub-regional convolution enhancement module comprises two branches, after one convolution is performed on the two branches respectively, one branch connects the light attention map and the feature map in the channel, and the other branch connects the light attention map after being inverted and the feature map in the channel, after one convolution, one group normalization and one activation function are performed on the two branches respectively, the two branches are connected and then connected in residual after one convolution, then, the feature map is connected after maximum pooling and average pooling are performed on the feature map respectively, the corresponding weight is obtained through one convolution, the corresponding feature is weighted in the channel, and then output through the Sigmoid activation function; the difference between the convolution enhancement module and the sub-regional convolution enhancement module is only that the light attention map is not included.

[0015] Furthermore, in step S11, the illumination attention estimation network includes: the input image is processed by a 3×3 convolution, the feature map is sequentially fed into the encoder and decoder, and then output as an illumination attention map by a 3×3 convolution; the encoder includes a residual module, a downsampling module, a residual module, a downsampling module and a residual module connected in sequence, and the decoder includes a residual module, an upsampling module, a residual module, an upsampling module and a residual module connected in sequence, and each layer of the encoder and decoder connects features of the same dimension of the encoder and decoder through skip connections.

[0016] Furthermore, the residual module sequentially includes two 3×3 convolutions, a ReLU activation function, and a channel attention mechanism (SE). The result of the first 3×3 convolution and the result of the channel attention mechanism (SE) are residually concatenated.

[0017] Furthermore, in step S13, the high-frequency feature fine network includes: first, performing a two-dimensional discrete wavelet transform (DWT) on the feature image, and obtaining high-frequency components including three high-frequency features: vertical edge features, horizontal edge features, and diagonal features. The three high-frequency features are then passed through a 3×3 convolution, and then the feature information in the vertical edge features and horizontal edge features is supplemented to the diagonal features through two self-attention modules (SAM). The vertical edge features, horizontal edge features, and the supplemented diagonal features are then enhanced by a convolutional block structure (CB), and finally output after inverse discrete wavelet transform (IDWT).

[0018] Furthermore, the self-attention module SAM includes taking the vertical edge features or the horizontal edge features through a convolution to obtain Q, taking the diagonal features through a convolution to obtain K and V respectively, multiplying Q and K, passing them through a Softmax activation function and a Dropout layer, and then multiplying them with V before outputting.

[0019] Furthermore, the convolutional block structure CB includes 5 connected convolutional structures and a ReLU activation function, with an inflation rate d = {1, 2, 3, 2, 1}.

[0020] Furthermore, in step S2, the loss function is:

[0021] L = L rec +λ1L VGG +λ2L HD

[0022]

[0023] Where L is the total loss, L rec For reconstruction loss, L VGG To perceive loss, L HDfor high frequency loss, and λ1 and λ2 are weight parameters, respectively represent the vertical edge feature V, the horizontal edge feature H and the diagonal feature D enhanced by the high frequency feature fine network, and correspond to represent the high frequency feature of the normal illumination image, and k represents the feature scale.

[0024] The application also discloses a partition exposure correction system based on deep learning, which comprises:

[0025] at least one processor; and at least one memory connected to the processor in communication, wherein:

[0026] The memory stores program instructions executable by the processor, and the processor calling the program instructions can execute the method.

[0027] The application also discloses a non-transitory computer readable storage medium storing computer instructions, which make the computer execute the method.

[0028] (Three) beneficial effects

[0029] The above technical solutions of the application have the following advantages:

[0030] (1) The illumination attention estimation network is adopted in the application to estimate the illumination mapping of the image, which can not only adaptively estimate the appropriate illumination attention mapping according to the images of different scenes, but also can obtain the illumination mapping that can fit the structural boundary of the image, so that the partition can be realized according to the different illumination; and the partition convolution enhancement network is adopted to guide the estimated illumination attention mapping, and the overexposed area and the underexposed area are respectively recovered and processed, and more image feature information is obtained through the maximum pooling operation and the average pooling operation, so that the features of the partition are enhanced, so that the overexposure and underexposure problems of the image can be processed at the same time, and the detailed features of the image are enhanced.

[0031] (2) The high frequency feature fine network is used to process the high frequency details of the image by taking advantage of the discrete wavelet transform to process the details of the image, so as to further enhance the detailed features of the image and fully recover the feature information of the image; and the high frequency loss is considered to improve the performance of the training model. BRIEF DESCRIPTION OF DRAWINGS

[0032] The features and advantages of the application will be more clearly understood through reference to the accompanying drawings, which are schematic and should not be understood as limiting the application, and in which:

[0033] Figure 1A flowchart of the partition exposure correction method based on deep learning of the embodiment of the present application;

[0034] Figure 2 A whole network structure diagram of the partition exposure correction network of the embodiment of the present application;

[0035] Figure 3 A comparison diagram of illumination estimation mapping of the embodiment of the present application and a traditional method;

[0036] Figure 4 A structure diagram of the high-frequency feature fine network of the embodiment of the present application;

[0037] Figure 5 A comparison of test results of the LCDP data set of the embodiment of the present application and an existing algorithm. DETAILED DESCRIPTION

[0038] The specific embodiments of the present application will be further described in detail below in combination with the drawings and embodiments. The following embodiments are used to illustrate the present application, but not to limit the scope of the present application.

[0039] The embodiment of the present application is a partition exposure correction method based on deep learning, as shown in Figure 1 The method comprises the following steps:

[0040] S 1, constructing a partition exposure correction network based on deep learning comprising an illumination attention estimation network, a partition convolution enhancement network and a high-frequency feature fine network;

[0041] S 11, inputting an original image into the illumination attention estimation network to obtain an estimated illumination attention mapping diagram, wherein the illumination attention mapping diagram divides different exposure regions;

[0042] Considering the limitations of the traditional method for illumination estimation, the present application adopts an illumination attention estimation network for estimating the illumination mapping of the image, which can not only adaptively estimate a suitable illumination attention mapping diagram according to the images of different scenes, but also obtain an illumination mapping diagram that can fit the structural boundaries of the image. Specifically, the structure of the illumination attention estimation network of the present application is improved on the basis of the structure of U-Net, and its special U-shaped structure and skip connection can well extract the edge structural features of the image. In the illumination attention mapping estimated by the present application, the relationship between the overall structural boundary of the image and the illumination is also considered. When the light source illuminates the object or a specific region, we can observe that the illumination intensity is not uniform, especially in the presence of obstructions or uneven surfaces. These factors will cause differences in illumination intensity between different regions, in other words, the illumination intensity will be different according to the situation of the region, therefore, we utilize the characteristics of the U-shaped structure network that is good at capturing boundaries and structures to design the illumination attention estimation network of the present application.Figure 2 The structure of the illumination attention estimation network is shown in detail, and specifically includes:

[0043] First, the input image is first passed through a 3x3 convolution, and then the feature map is sent to the encoder part and the decoder part, and an illumination attention map A is output through a 3x3 convolution; the encoder includes a residual module, a down-sampling module, a residual module, a down-sampling module, and a residual module connected in turn, and the decoder includes a residual module, an up-sampling module, a residual module, an up-sampling module, and a residual module connected in turn; the encoder is composed of a residual module RB and a down-sampling module Down, and three feature maps of different scales are obtained through two down-sampling operations, and the feature information of each scale is obtained through the residual block. The specific design of the residual module RB is shown in the lower right corner of Figure 2 : it includes two 3x3 convolutions Conv, a ReLU activation function, and a channel attention mechanism SE in turn, and the results of the first 3x3 convolution and the channel attention mechanism SE are connected in residual. The residual block RB can extract and learn the spatial features of the image, and can better obtain the shape, texture and other feature information of the image. The channel attention mechanism SE dynamically adjusts the weight of the feature map channel through deep learning, so as to better focus on important feature channels and improve the feature acquisition ability and model performance. Secondly, in the decoder part, we also extract the feature information of the image through the residual module RB, and add an up-sampling module Up operation to adjust the size and channel number of the feature. It is worth noting that we add a skip connection to connect the features of the same dimension of the encoder and the decoder, which helps to fuse the feature information of each scale. Finally, through a 3x3 convolution layer, the number of feature map channels is converted to 3, and the final illumination attention map A is obtained.

[0044] In Figure 3 , the illumination estimation method of the present embodiment is compared with the traditional illumination estimation method, Figure 3 , Figure (a) is the original image, Figure 3 , Figures (b) and (c) are the attention maps estimated by the traditional method, Figure 3 , Figure (d) is the attention map estimated by the method of the present embodiment, it can be seen that, Figure 3 , Figure (d) in accurately distinguishes the overexposed area and the underexposed area in the image. Among them, the overexposed area of the image is represented by a smaller value, while the underexposed area is represented by a larger value. And Figure 3 , (b) and (c) in are the estimated attention maps obtained by the traditional method, which often cannot well estimate the different degrees of illumination of the image. Careful observation can see that the light spot on the face and the overexposed area in the background are not well divided.

[0045] S 12. Input the original image into the partitioned convolutional enhancement network. The partitioned convolutional enhancement network performs partitioned enhancement on different exposure areas of the image according to the illumination attention map.

[0046] Considering that poorly exposed images often suffer from severe information loss in both overexposed and underexposed areas, different operations should be applied to these two types of regions to recover their corresponding features. This is because these two types of regions face opposite problems: one requires suppressing illumination, while the other requires enhancing brightness. Therefore, we will utilize the illumination attention map A to perform regional enhancement on different exposure areas of the image. The illumination attention map A can effectively divide the image into different illumination regions, and to better capture different features at various scales, we employ a three-layer U-shaped network structure combined with the illumination attention map A to enhance different regions.

[0047] like Figure 2As shown, the partitioned convolutional enhancement network structure of the present application takes the original image as the input image, and is still a U-shaped structure. The encoder part includes three partitioned convolutional enhancement modules connected in turn, and the decoder includes three convolutional enhancement modules CEM connected in turn. The encoder part is composed of three partitioned convolutional enhancement modules, and the three partitioned convolutional enhancement modules PCEM can extract features of different scales in the image respectively. The partitioned convolutional enhancement module includes two branches, which respectively process the overexposed area and the underexposed area: the feature map is first subjected to a convolution operation; then in one branch, the light attention map A obtained in step S2 is connected with the feature map in the channel, so that it can play a key attention role for the recovery of the overexposed area; and in the other branch, the light attention map A is connected with the feature map in the channel after being inverted, so that it pays more attention to the recovery of the underexposed area; then a convolution, a group normalization and an activation function are connected, which are used to extract feature information of each area; after the two branches are connected and subjected to a convolution, a residual connection is added; then, the feature map is subjected to a maximum pooling operation and an average pooling operation, both of which can often pay attention to different types of features, wherein the maximum pooling operation is more sensitive to the feature texture of the image, and the background information obtained by the average pooling operation will be more rich, in general, both branches are to obtain more feature information in the image; then a convolution is performed to obtain corresponding weights for channel weighting of the corresponding features, highlight the features, and then output through a Sigmoid activation function. The decoder part does not include the guidance of the light attention map, and the rest of the structure is the same as that of the encoder part, that is, the difference between the convolutional enhancement module and the partitioned convolutional enhancement module is only that the guidance of the light attention map is not included, including: the feature map is first subjected to a convolution, and then a convolution, a group normalization and an activation function are connected, which are used to extract feature information of each area; after a convolution, a residual connection is added; then, the feature map is subjected to a maximum pooling operation and an average pooling operation, and more feature information in the image is obtained; then a convolution is performed to obtain corresponding weights for channel weighting of the corresponding features, highlight the features, and then output through a Sigmoid activation function. Similarly, we add a jump connection to each layer of the encoder and the decoder to connect the features of the same dimension of the encoder and the decoder, so as to better recover and fuse the image feature information.

[0048] S13, input the partitioned enhanced image into a high-frequency feature fine network to output, the high-frequency feature fine network is used for recovering high-frequency information;

[0049] The poor exposure problem image has been properly adjusted by the steps S11-S12 and the corresponding region of the image has been enhanced in features. However, in order to obtain more image detail information, a high-frequency feature fine network is added after the partition enhancement for restoring more high-frequency information of the image.

[0050] In the high-frequency feature fine network HFRM, the image is subjected to two-dimensional discrete wavelet transform DWT, the first discrete wavelet transform decomposes the image into high-frequency h1 and low-frequency 1 1, and the second discrete wavelet transform is to obtain h2 and 12 by performing discrete wavelet transform on the low-frequency feature 1 1 after the first transform. In the high-frequency feature fine network, only the high-frequency part is subjected to corresponding enhancement processing, and the low-frequency part is not processed. The enhancement operation of the obtained two high-frequency parts h1 and h2 is specifically implemented as shown in Figure 4 .

[0051] The feature image is subjected to two-dimensional discrete wavelet transform DWT first, and the obtained high-frequency components include three parts: vertical edge feature Vertical, horizontal edge feature Horizontal, and diagonal feature Diagonal. The three obtained high-frequency features first pass through a 3x3 convolution respectively, and then pass through two self-attention modules (SAM) to supplement the details in the diagonal feature by using the feature information in the vertical edge feature and the horizontal edge feature respectively, so as to achieve feature enhancement. Then, the vertical edge feature, the horizontal edge feature and the supplemented diagonal feature are subjected to local feature enhancement in the convolution block structure CB by using dilated convolution. The dilated convolution can adjust the receptive field by adjusting the dilated rate d, so as to better capture the features. In the present embodiment, the convolution block structure CB includes 5 groups of connected convolution structures and ReLU activation functions, and the dilated rate d = {1, 2, 3, 2, 1}. Finally, a residual connection is added and the discrete wavelet inverse transform IDWT is performed to obtain the final output I'.

[0052] The self-attention module SAM includes that the vertical edge feature or the horizontal edge feature passes through a convolution to obtain Q, the diagonal feature passes through a convolution to obtain K and V respectively, Q and K are multiplied, then after the Softmax activation function and the Dropout layer, V is multiplied to output.

[0053] S2, after the training data is input into the illumination attention estimation network for training, the illumination attention estimation network is frozen, the partition convolution enhancement network and the high-frequency feature fine network are input for training, and a loss function is constructed for repeated iteration until the maximum iteration number is reached to end the training, and a trained partition exposure correction network is obtained;

[0054] To better supervise our network, we not only use the basic reconstruction loss (L1 loss) and perceptual loss (VGG loss) in image processing, but also add a high-frequency detail loss. In the high-frequency detail loss, the L2 loss of the high-frequency feature maps in two scales is calculated, and the specific formula is as follows:

[0055]

[0056] wherein, respectively represent the vertical edge feature (Vertical), the horizontal edge feature (Horizontal) and the diagonal feature (Diagonal) enhanced by the high-frequency feature fine network, and the corresponding represents the high-frequency feature of the normal illumination image, and k represents the feature scale.

[0057] The total loss is shown in formula 4.2:

[0058] L = L rec + λ1L VGG + λ2L HD (2)

[0059] wherein, L rec is the reconstruction loss, L VGG is the perceptual loss, L HD is the high-frequency loss, and λ1 and λ2 are weight parameters.

[0060] The entire network is built in PyTorch and trained on an NVIDIA RTX 2080Ti. The network training stage is divided into two stages, the first stage is the training of the illumination estimation attention network, and the second stage is the training of the partition convolution enhancement network and the high-frequency feature fine network. In the first stage, the training of the illumination estimation attention network, the image size participating in the training is 512x512, the batch size is 8, and 100 iterations are performed. The ADAM optimizer is used, and the initial learning rate is set to 0.0001. After the training is completed, the illumination estimation attention network is frozen, and the partition convolution enhancement network and the high-frequency feature fine network are trained. The image size participating in the second stage training is also adjusted to 512x512, the batch size is set to 4, and 300 iterations of training are performed. The ADAM optimizer is used, and the initial learning rate is set to 0.0001. The learning rate is reduced by 0.05 at the 200th and 250th iteration of training. The loss function weight parameter is set to: λ1 = 0.5, λ2 = 0.1.

[0061] In order to better compare the visual effects of the method of the embodiment and the existing algorithm, the embodiment carries out experiments on the exposure correction dataset LCDP, Figure 5are three images with different scenes and different exposure levels selected from the LCDP dataset for effect comparison. In the comparison chart, the first image is the input image, the last image is the reference standard image, and the rest are the effect images of the existing algorithms LIME, IAT, URetinex-Net, Bread, MSEC, FourLLIE, LCDP and the effect image of the method of the embodiment. The LCDP dataset has the characteristics that each image has overexposed and underexposed regions, so it can well test the effect of our method. We selected an image with more overexposed regions (the first one), an image with more underexposed regions (the second one), and an image with dark and bright regions with little difference (the third one). By observation, we can find that our method can better restore the detail information and color of the underexposed and overexposed regions of the image, and also can adjust the overall exposure of the image to a suitable state. Therefore, a large number of experimental results and comparison experiments with existing algorithms have verified the effectiveness of the method of the embodiment.

[0062] Finally, it should be noted that the above method can be converted into software program instructions, which can be implemented by a system including a processor and a memory, or by computer instructions stored in a non-transitory computer readable storage medium. The integrated unit implemented in the form of software functional units described above can be stored in a computer readable storage medium. The software functional unit described above is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute part of the steps of the method described in each embodiment of the application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a magnetic disk or an optical disk, and various program code storage media.

[0063] In summary, through the above-mentioned partition exposure correction method based on deep learning, the following beneficial effects are achieved:

[0064] (1) The application adopts an illumination attention estimation network for estimating the illumination mapping of an image, which can not only adaptively estimate a suitable illumination attention mapping according to images of different scenes, but also can make the estimated illumination mapping fit the structural boundary of the image, so as to realize partition according to different illuminations; and then, according to different processing methods for overexposed areas and underexposed areas, a partition convolution enhancement network is adopted to guide the estimated illumination attention mapping, and the overexposed areas and underexposed areas are respectively processed, more image feature information is obtained through max-pooling operation and average-pooling operation, and the features of the partition are enhanced, so that the overexposure and underexposure problems of the image can be processed at the same time, and the detailed features of the image are enhanced.

[0065] (2) The application uses the discrete wavelet transform to process the high-frequency details of the image through a high-frequency feature fine network, further enhances the detailed features of the image, and fully restores the feature information of the image; and the high-frequency loss is considered, and the performance of the training model is improved.

[0066] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the embodiments of the present application are described in combination with the drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A deep learning-based method for zonal exposure correction, characterized in that, include: S1. Construct a deep learning-based partitioned exposure correction network, which includes an illumination attention estimation network, a partitioned convolutional enhancement network, and a high-frequency feature refinement network. S11. Input the original image into the illumination attention estimation network to obtain the estimated illumination attention map, which divides different exposure areas. S12. Input the original image into the partitioned convolutional enhancement network. The partitioned convolutional enhancement network performs partitioned enhancement on different exposure areas of the image according to the illumination attention map. In step S12, the partitioned convolutional enhancement network includes: an encoder part comprising three sequentially connected partitioned convolutional enhancement modules, and a decoder comprising three sequentially connected convolutional enhancement modules. Each layer of the encoder and decoder adds skip connections to connect features of the same dimension between the encoder and decoder. The partitioned convolutional enhancement module comprises two branches. After each branch performs a convolution, one branch performs a channel connection between the illumination attention map and the feature map, and the other branch performs a channel connection between the inverted illumination attention map and the feature map. The two branches are then connected by a convolution, a grouping normalization, and an activation function, respectively. The two branches are then connected and subjected to a convolution followed by a residual connection. Then, the feature maps are subjected to max pooling and average pooling respectively and connected. After a convolution, the corresponding weights are obtained, and the corresponding features are channel-weighted and then output through a Sigmoid activation function. The only difference between the convolutional enhancement module and the partitioned convolutional enhancement module is that the latter does not include the guidance of the illumination attention map. S13. The enhanced image after partitioning is input into a high-frequency feature fine network and then output. The high-frequency feature fine network is used to recover high-frequency information. In step S13, the high-frequency feature fine network includes: first, performing a two-dimensional discrete wavelet transform (DWT) on the feature image, and obtaining high-frequency components including three high-frequency features: vertical edge features, horizontal edge features, and diagonal features. The three high-frequency features are then passed through a 3×3 convolution, and then the feature information in the vertical edge features and horizontal edge features is supplemented to the diagonal features through two self-attention modules (SAM). The vertical edge features, horizontal edge features, and the supplemented diagonal features are then enhanced by a convolutional block structure (CB), and finally output after inverse discrete wavelet transform (IDWT). S2. After training the illumination attention estimation network by inputting the training data, freeze the illumination attention estimation network, and then input the partition convolutional enhancement network and the high-frequency feature refinement network for training. Construct a loss function and iterate repeatedly until the maximum number of iterations is reached, and then end the training to obtain the trained partition exposure correction network.

2. The deep learning-based partitioned exposure correction method according to claim 1, characterized in that, In step S11, the illumination attention estimation network includes: the input image is processed by a 3×3 convolution, the feature map is sequentially fed into the encoder and decoder, and then output as an illumination attention map by a 3×3 convolution; the encoder includes a residual module, a downsampling module, a residual module, a downsampling module and a residual module connected in sequence, and the decoder includes a residual module, an upsampling module, a residual module, an upsampling module and a residual module connected in sequence, and each layer of the encoder and decoder connects features of the same dimension of the encoder and decoder through skip connections.

3. The deep learning-based partitioned exposure correction method according to claim 2, characterized in that, The residual module consists of two 3×3 convolutions, a ReLU activation function, and a channel attention mechanism (SE). The result of the first 3×3 convolution and the result of the channel attention mechanism (SE) are residually concatenated.

4. The deep learning-based partitioned exposure correction method according to claim 1, characterized in that, The self-attention module SAM includes convolving vertical or horizontal edge features to obtain Q, convolving diagonal features to obtain K and V respectively, multiplying Q and K, passing them through a Softmax activation function and a Dropout layer, and then multiplying them with V before outputting the result.

5. The deep learning-based partitioned exposure correction method according to claim 1, characterized in that, The convolutional block structure CB includes 5 connected convolutional structures and a ReLU activation function, with an inflation rate d= .

6. The deep learning-based partitioned exposure correction method according to claim 1, characterized in that, In step S2, the loss function is Where L is the total loss, To reconstruct the loss, In order to perceive loss, For high-frequency loss, and For weight parameters, These represent the vertical edge feature V, the horizontal edge feature H, and the diagonal feature D, respectively, enhanced by a high-frequency feature refinement network. The high-frequency features of a normally illuminated image are represented by k, which represents the feature scale.

7. A deep learning-based zone exposure correction system, characterized in that, include: At least one processor; and at least one memory communicatively connected to the processor, wherein: The memory stores program instructions that can be executed by the processor, and the processor can execute the method as described in any one of claims 1 to 6 by calling the program instructions.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to perform the method as described in any one of claims 1 to 6.

Images