Method, device, equipment, medium and product for defogging non-uniform real foggy images

Abstract

The present application discloses a method, device, equipment, medium and product for defogging a non-uniform real foggy image, which relates to the field of image processing technology. A real foggy image is input into a trained non-uniform foggy image defogging network to obtain a defogged image; the non-uniform foggy image defogging network includes a two-dimensional discrete wavelet transform branch, a multi-level patch network branch and an image feature fusion module. The two-dimensional discrete wavelet transform branch and the multi-level patch network branch are used to perform image processing on the real foggy image respectively, and then the images output by each branch are fused to obtain a more accurate defogging image, which can improve the defogging accuracy and defogging effect.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a method, device, equipment, medium and product for defogging non-uniform real foggy images. Background Art

[0002] Environmental perception is the front end and foundation of the autonomous driving system. When the system perceives the surrounding environment, it is easily affected by bad weather (such as rain, fog, snow, etc.), which causes the collected image data to be blurred and distorted, which is not conducive to subsequent high-level machine vision tasks such as semantic segmentation and target recognition. Foggy weather is one of the most common adverse weather conditions, and fog image clarity helps improve the robustness of the autonomous driving perception system in foggy environments.

[0003] Existing image defogging algorithms can be divided into two categories: model-based and model-free. Among them, the model-based defogging algorithm makes assumptions about some conditions or parameters in the model through prior knowledge, converting the ill-posed problem into a well-posed problem. As one of the most widely used physical models, the atmospheric scattering model estimates the atmospheric light value and transmittance based on prior knowledge, and then introduces it into the model to deduce a fog-free image. However, the model-based algorithm only considers the ideal situation of uniform distribution of fog. When the prior knowledge or assumed parameters do not match the actual conditions, the problem of being unable to remove the fog in the image will occur. The model-free defogging algorithm is not restricted by the model, has a wider range of application scenarios, and is more flexible in image processing. Image enhancement technology is a commonly used processing method for early model-free image defogging algorithms, and typical representatives include Retinex, histogram equalization technology, bilateral filtering, homomorphic filtering, and wavelet transform. Image enhancement technology does not consider the reasons for the degradation of foggy images. It only improves the quality of degraded images from the perspective of improving visual effects. It uses methods such as improving image contrast to improve the quality of degraded images, but it may cause image detail loss and color distortion. The use scenarios of model-free defogging networks based on deep learning are more flexible, but they require a large amount of image data for training to learn the mapping between foggy images and corresponding clear images. At present, the foggy image datasets used in the training of defogging networks based on deep learning are mostly synthetic uniform foggy image datasets, and rarely consider the non-uniformity of fog in real foggy environments. Therefore, facing the non-uniform and discontinuous characteristics of the distribution of fog concentration in real foggy images, the defogging accuracy of existing image defogging algorithms is poor. Summary of the invention

[0004] The purpose of this application is to provide a method, device, equipment, medium and product for defogging non-uniform real foggy images, which takes into account the non-uniform and discontinuous characteristics of the distribution of fog concentration in real foggy images, and can improve the defogging accuracy and defogging effect of real foggy images.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides a method for defogging a non-uniform real foggy image, comprising:

[0007] Get real foggy images;

[0008] The real foggy image is input into a trained non-uniform foggy image defogging network to obtain a defogged image; the non-uniform foggy image defogging network comprises a two-dimensional discrete wavelet transform branch, a multi-level patch network branch and an image feature fusion module;

[0009] The two-dimensional discrete wavelet transform branch is used to process the low-frequency features of the real foggy image;

[0010] The multi-level patch network branch is used to process the high-frequency features of the real foggy image;

[0011] The image feature fusion module is used to fuse the image output by the two-dimensional discrete wavelet transform branch and the image output by the multi-level patch network branch to obtain the defogged image.

[0012] In a second aspect, the present application provides a non-uniform real foggy image defogging device, comprising:

[0013] An image acquisition sub-device, used for acquiring real foggy images;

[0014] An image defogging sub-device, used for inputting the real foggy image into a non-uniform foggy image defogging network to obtain a defogged image; the non-uniform foggy image defogging network comprises a two-dimensional discrete wavelet transform branch, a multi-level patch network branch and an image feature fusion module; the two-dimensional discrete wavelet transform branch is used for processing the low-frequency features of the real foggy image;

[0015] The multi-level patch network branch is used to process the high-frequency features of the real foggy image;

[0016] The image feature fusion module is used to fuse the image output by the two-dimensional discrete wavelet transform branch and the image output by the multi-level patch network branch to obtain the defogged image.

[0017] In a third aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned non-uniform real foggy image defogging method.

[0018] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned non-uniform real foggy image defogging method.

[0019] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the above-mentioned non-uniform real foggy image defogging method.

[0020] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0021] The present application provides a method, device, equipment, medium and product for defogging a non-uniform real foggy image. A real foggy image is input into a trained non-uniform foggy image defogging network to obtain a defogged image. The non-uniform foggy image defogging network includes a two-dimensional discrete wavelet transform branch, a multi-level patch network branch and an image feature fusion module. Among them, the two-dimensional discrete wavelet transform branch is used to focus on processing the low-frequency features of the image. This branch extracts feature information from different frequency domains in the image and performs targeted processing on the feature information in different frequency domains to retain more space and detail information; the multi-level patch network branch focuses on processing the high-frequency features of the image, and aggregates the image feature information from fine to coarse processing to retain the high-frequency detail features of the image. Finally, the images output by each branch are fused to obtain a more accurate defogging image, which can improve the defogging accuracy and defogging effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] Figure 1 This is an application environment diagram of a non-uniform real foggy image defogging method in an embodiment of the present application;

[0024] Figure 2 A schematic diagram of a flow chart of a method for defogging a non-uniform real foggy image provided in one embodiment of the present application;

[0025] Figure 3 A schematic diagram of the structure of a non-uniform foggy image defogging network provided in an embodiment of the present application;

[0026] Figure 4 A schematic diagram of the structure of a multi-level patch network branch provided in an embodiment of the present application;

[0027] Figure 5 A schematic diagram of an encoder-decoder structure in a multi-level patch network branch provided in an embodiment of the present application;

[0028] Figure 6A schematic diagram of the processing process of a non-uniform foggy image defogging network provided in an embodiment of the present application;

[0029] Figure 7 A schematic diagram for comparing gamma correction provided in an embodiment of the present application;

[0030] Figure 8 A schematic diagram of functional modules of a foggy image defogging device provided in another embodiment of the present application;

[0031] Fig. 9 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0033] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0034] The non-uniform real foggy image defogging method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the terminal communicates with the server through the network. The data storage system can store the data that the server needs to process. The data storage system can be set up separately, integrated on the server, or placed on the cloud or other servers. The terminal can send a real foggy image to the server. After the server receives the real foggy image to be processed, for the real foggy image to be processed, the server inputs the real foggy image into the trained non-uniform foggy image defogging network to obtain the defogged image. The server can feed back the obtained defogged image to the terminal. In addition, in some embodiments, the non-uniform real foggy image defogging method can also be implemented by a server or a terminal alone, such as the terminal can directly apply the trained non-uniform foggy image defogging network to the real foggy image to be processed for defogging, or the server can obtain the real foggy image to be processed from the data storage system, and apply the trained non-uniform foggy image defogging network to perform defogging.

[0035] The terminals may be, but are not limited to, various desktop computers, laptops, smart phones, tablet computers, IoT devices and portable wearable devices. IoT devices may be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server may be implemented as an independent server or a server cluster consisting of multiple servers, or a cloud server.

[0036] In an exemplary embodiment, Figure 2 As shown, a non-uniform real foggy image defogging method is provided, specifically a lightweight non-uniform real foggy image defogging method for autonomous driving. The method is executed by a computer device, specifically a computer device such as a terminal or a server, and can also be executed by a terminal and a server together. In the embodiment of the present application, the method is applied to Figure 1 The server in is taken as an example to illustrate, including the following steps 101 to 102.

[0037] Step 101, obtaining a real foggy image.

[0038] Step 102, input the real foggy image into the trained non-uniform foggy image defogging network to obtain a defogged image; Figure 3 and Figure 4 As shown, the non-uniform foggy image defogging network includes a two-dimensional discrete wavelet transform branch, a multi-level patch network branch and an image feature fusion module. The two-dimensional discrete wavelet transform branch is used to process the low-frequency features of the real foggy image; the multi-level patch network branch is used to process the high-frequency features of the real foggy image; the image feature fusion module is used to fuse the image output by the two-dimensional discrete wavelet transform branch and the image output by the multi-level patch network branch to obtain the defogged image. Among them, the high-frequency features in the image correspond to the edges and details in the image, that is, the parts with drastic changes in brightness or grayscale; the low-frequency features are the areas with uniform brightness distribution and large area in the image, such as the non-uniformly distributed fog and sky background areas in the foggy image.

[0039] The two-dimensional discrete wavelet transform branch includes first convolution modules (corresponding to Figure 3 The first basic convolution module in ), wavelet transform downsampling module, the second convolution module (corresponding to Figure 3 The second basic convolution module in ), the first residual addition layer and the inverse wavelet transform upsampling module (corresponding to Figure 3The first convolution module has an input terminal which is the real foggy image; the first residual addition layer has an input terminal which is also connected to the output terminal of the wavelet transform down-sampling module; the inverse wavelet transform up-sampling module has an input terminal which is also connected to the output terminal of the wavelet transform down-sampling module; the output terminal of the inverse wavelet transform up-sampling module is connected to the input terminal of the image feature fusion module.

[0040] The multi-level patch network branch includes a first-level patch module, a second-level patch module and a third-level patch module.

[0041] The first-level patch module includes a second residual addition layer, a first encoder, a third residual addition layer and a first decoder which are connected in series in sequence.

[0042] The second-level patch module includes a first image segmentation layer, a fourth residual addition layer, a second encoder, a fifth residual addition layer, a first splicing layer and a second decoder which are sequentially connected in series.

[0043] The third-level patch module includes a second image segmentation layer, a third encoder, a second splicing layer and a third decoder which are sequentially connected in series.

[0044] The input end of the second residual addition layer is connected to the output end of the second decoder; the input of the second addition layer also includes the real foggy image; the input end of the third residual addition layer is also connected to the output end of the first splicing layer; the output end of the first decoder is connected to the input end of the image feature fusion module; the output of the image feature fusion module is the defogged image.

[0045] The input of the first image segmentation layer is the real foggy image; the input end of the fourth residual addition layer is also connected to the output end of the third decoder; the input end of the fifth residual addition layer is also connected to the output end of the second splicing layer.

[0046] The input of the second image segmentation layer is the real foggy image; the second image segmentation layer is used to divide the real foggy image into a number of patch blocks according to a number of horizontal dividing lines and a number of vertical dividing lines to obtain a first patch block divided image; the second splicing layer is used to splice adjacent patch blocks in the first splicing direction in the first patch block divided image to obtain a first spliced ​​image.

[0047] The first image segmentation layer is used to divide the real foggy image into patches according to the patch division method in the first spliced ​​image to obtain a second patch division image.

[0048] The first splicing layer is used to splice adjacent patch blocks in a second splicing direction in the image output by the fifth residual addition layer to obtain a second spliced ​​image; when the first splicing direction is a horizontal direction, the second splicing direction is a vertical direction; when the first splicing direction is a vertical direction, the second splicing direction is a horizontal direction.

[0049] By implementing the above steps 101 to 102, the non-uniform foggy image defogging network suitable for real scenes proposed by the present invention is mainly divided into two parts: a discrete wavelet transform branch and a multi-level patch network branch. The two branches focus on processing the low-frequency and high-frequency components in the foggy image respectively. First, for the low-frequency part in the image, the two-dimensional discrete wavelet transform branch focuses on processing the low-frequency features of the image. The image can be reconstructed by wavelet inverse transform (i.e., wavelet inverse transform upsampling module). The feature information of different frequency domains of the image is extracted by frequency domain learning. The feature information of different frequency domains is processed in a targeted manner to retain more space and detail information; secondly, for the high-frequency part of the image, the multi-level patch network focuses on processing the high-frequency part of the image. The multi-level patch network branch constructed based on the encoding-decoding structure processes the input non-uniform foggy image from fine to coarse (i.e., the processing process of the third-level patch module-the second-level patch module-the first-level patch module), aggregates the image feature information, and retains the high-frequency detail features of the image, so that the real scene of the foggy image is restored after the two branches are fused.

[0050] In another exemplary embodiment of the present application, in step 102, the real foggy image is input into a trained non-uniform foggy image defogging network to obtain a defogged image, specifically including:

[0051] (1) The real foggy image is input into the two-dimensional discrete wavelet transform branch in the non-uniform foggy image defogging network to obtain a defogged image after wavelet transform.

[0052] (2) Inputting the real foggy image into the multi-level patch network branch in the non-uniform foggy image defogging network to obtain a defogged image after multi-level patch processing.

[0053] (3) Inputting the dehazed image after wavelet transformation and the dehazed image after multi-level patch processing into the image feature fusion module to obtain the dehazed image.

[0054] In another exemplary embodiment of the present application, in step (1), the wavelet transform in the two-dimensional discrete wavelet transform branch is often used to analyze unstable signals, and the image is suitable for wavelet transform as a two-dimensional plane signal. The processing process of the two-dimensional discrete wavelet transform branch: (1) First, a one-dimensional wavelet transform is performed on each row of the image to obtain the low-frequency component and high-frequency component of the original image in the horizontal direction, and then a one-dimensional wavelet transform is performed on each column of the transformed data to obtain the low-frequency component of the original image in the horizontal and vertical directions, the low-frequency component in the horizontal direction and the high-frequency component in the vertical direction, the high-frequency component in the horizontal direction and the low-frequency component in the vertical direction, and the high-frequency component in the horizontal and vertical directions. Figure 5 As shown in , the wavelet transform is applied to decompose the image into low-frequency and high-frequency components and process them separately, retaining the frequency domain information in the feature map, which is conducive to the restoration of image texture details. The two-dimensional discrete wavelet transform can be used as downsampling to extract the feature information of the image. (2) As Figure 3 As shown in the figure, the high-frequency component information extracted by the two-dimensional discrete wavelet transform (corresponding to the wavelet transform downsampling module) is directly transmitted to the wavelet inverse transform upsampling module through skipping, and the low-frequency component information is combined with the convolution and then input into the wavelet inverse transform upsampling module through residual summation, so as to reconstruct a clear image. The two-dimensional discrete wavelet inverse transform replaces upsampling to reconstruct the image, fuse the features and denoise the image. Among them, residual learning and skipping are used to reduce the network structure parameters while retaining a large amount of detail information.

[0055] Therefore, in step (1), the real foggy image is input into the two-dimensional discrete wavelet transform branch in the non-uniform foggy image defogging network to obtain the defogging image after wavelet transform, which specifically includes:

[0056] (1-1) Inputting the real foggy image into the first convolution module to obtain a first convolution image.

[0057] (1-2) Inputting the first convolution image into the wavelet transform downsampling module to obtain a downsampled image.

[0058] (1-3) Input the downsampled image into the second convolution module to obtain a second convolution image.

[0059] (1-4) Input the downsampled image and the second convolution image into the first residual addition layer to obtain an added image.

[0060] (1-5) Inputting the residual addition image and the down-sampled image into the wavelet inverse transform up-sampling module to obtain an up-sampled image; the up-sampled image is the defogging image after the wavelet transform.

[0061] In another exemplary embodiment of the present application, since it is difficult to restore non-uniform foggy images using only wavelet transform, a multi-level patch network branch is introduced to process foggy images with uneven fog concentration and distribution, and the non-uniform foggy images are processed by fine-to-coarse block processing and from local to overall. The multi-level patch network branch is divided into a three-layer network structure. The third-level patch module and the second-level patch module divide the input image into several patches of the same size for processing, and the extracted image information is transmitted upward from the third layer to the first layer.

[0062] In the multi-level patch network branch, the input image is segmented into different numbers of patches from the coarsest to the finest. The first-level patch module processes the image without patch segmentation, and the second-level patch module processes the image after vertical or horizontal segmentation. For example, Figure 4 An image P2 containing two vertical patches is shown; the third-level patch module processes the patch block division image after horizontal or vertical cutting based on the second layer, for example Figure 4 A type of Each patch module in the multi-layer patch network branch includes an encoder and a decoder, and the information transmission is from bottom to top (the third layer ), from the finest to the coarsest, Figure 4 It is a structural diagram of the multi-level patch network branch. Figure 4 In the example, the input of the first-level patch module is the original foggy image without cutting; in the second-level patch module, the original foggy image is vertically cut into and In the third-level patch module, the image vertically segmented in the second-level patch module is horizontally segmented again and divided into 4 patch blocks, namely , , and Starting from the third-level patch module, each level of patch module encodes the input patch block and the feature information in the previous layer to obtain new image features, adds the feature images in adjacent spaces, and then adds the decoded features to the previous layer through residual addition to achieve multi-level feature information fusion.

[0063] by Figure 4 The patch block division method shown in FIG. 1 is used as an example to illustrate the processing process of the multi-level patch network branch: First, in the third-level patch module, the original foggy image is Split into 4 patches of equal size ( , , , , ), and input to the encoder In this paper, a new feature representation is generated. . It is expressed by formula (1).

[0064] .

[0065] Secondly, connect the two adjacent feature maps in the vertical direction to generate a new feature representation , , the size is the same as the feature representation of the second-level patch module and is expressed by formula (2).

[0066] .

[0067] Finally, the new feature representation Input to decoder , generate the image , where the image Includes two patches, .

[0068] Similarly, in the second-level patch module, the encoder The input is an image and images The image after residual summation, where the image Contains two vertically split patches ;Decoder The input is the newly generated feature representation in the second-level patch module , see formula (3).

[0069] .

[0070] The encoder The generated feature representations are connected to obtain the new feature representation generated in the second-level patch module , see formula (4), through the decoder After decoding, the image is obtained .

[0071] .

[0072] In the first-level patch module, the encoder The input is an image (Original foggy image) and image , For encoder The output feature representation is The expression is shown in formula (5), after the decoder The clear image restored after decoding is . The expression is shown in formula (6).

[0073] .

[0074] .

[0075] Based on the above description, in step (2), the real foggy image is input into the multi-level patch network branch in the non-uniform foggy image defogging network to obtain the defogged image after multi-level patch processing, which specifically includes:

[0076] (2-1) Input the real foggy image into the second image segmentation layer to obtain the first patch block segmentation image (corresponding to ).

[0077] (2-2) Inputting the first patch block divided image into the third encoder In the first coded image (including feature representation ).

[0078] (2-3) Input the first encoded image into the second splicing layer to obtain a first splicing image (including feature representation , ).

[0079] (2-4) Inputting the first spliced ​​image into the third decoder The first decoded image (corresponding to ).

[0080] (2-5) Input the real foggy image into the first image segmentation layer to obtain the second patch block segmentation image (corresponding to ).

[0081] (2-6) Inputting the second patch block divided image and the first decoded image into the fourth residual addition layer to obtain a first residual addition image (corresponding to ).

[0082] (2-7) Inputting the first residual addition image to the second encoder The second coded image (corresponding to ).

[0083] (2-8) The second coded image and the first spliced ​​image (including the feature representation ) is input to the fifth residual addition layer to obtain the second residual addition image (including feature representation ).

[0084] (2-9) Inputting the second residual addition image into the first splicing layer to obtain a second splicing image (corresponding to ).

[0085] (2-10) Inputting the second spliced ​​image into the second decoder The second decoded image (corresponding to ).

[0086] (2-11) The second decoded image and the real foggy image Input to the second residual addition layer to obtain the third residual addition image (corresponding to ).

[0087] (2-12) Inputting the third residual addition image into the first encoder The third coded image (corresponding to ).

[0088] (2-13) The third coded image and the second stitched image (corresponding to ) is input into the third residual addition layer to obtain the fourth residual addition image (corresponding to ).

[0089] (2-14) Inputting the fourth residual addition image into the first decoder The third decoded image (corresponding to ); The third decoded image is the defogged image after the multi-level patch processing.

[0090] The multi-level patch network branch is different from the dehazing process performed on the original size image. The multi-patch blocks can better process the local non-uniform fog image in space and limit the receptive field of the specific local area. The pixel output of each patch block is only affected by the receptive field of the corresponding patch block, rather than the global image.

[0091] The multi-layer patch network includes an encoding-decoding structure and a residual learning structure. The encoding-decoding structure of each layer in the network consists of 12 residual blocks and 5 convolutions. The structures of the encoder and decoder are similar, except that the convolution used in the decoder is a transposed convolution. Figure 5 As shown in the figure, each residual block consists of two convolutions, a ReLU and a residual link. By using residual learning, more image information is retained while making the network structure lighter.

[0092] In another exemplary embodiment of the present application, before executing step 102 "inputting the real foggy image into the trained non-uniform foggy image defogging network to obtain the defogged image", the foggy image defogging method further includes: training the non-uniform foggy image defogging network, such as Figure 6 As shown, the specific training process is:

[0093] (a) Obtain a dataset of different foggy image samples.

[0094] (b) The gamma correction method based on RGB value is used to preprocess each foggy image sample data set to obtain each preprocessed image sample data set.

[0095] The sample dataset of real foggy scene images is small in scale, and overfitting is prone to occur during network training. However, there are differences between the RGB values ​​of images in different sample datasets, and it is not possible to simply splice and synthesize the sample dataset. Data preprocessing technology is required. In the present invention, the NH-HAZE dataset is used as the original dataset, and the DENSE-HAZE dataset is the dataset to be corrected. The RGB values ​​of the dataset images are made closer through gamma correction, reducing the difference between the target dataset image and the dataset image to be corrected, thereby achieving the purpose of expanding the dataset scale. Based on this idea, the present invention uses gamma correction based on RGB values ​​to preprocess the image. Different from grayscale gamma correction, the present invention corrects the R, G, and B channels of images in different datasets separately, calculates the average pixel value of the RGB channels of images in different datasets, and compares the average RGB values ​​of the images in each dataset, determines the gamma coefficient to adjust the dataset, so that the RGB value of the corrected dataset is closer to the dataset to be fused, thereby achieving the effect of enriching the dataset.

[0096] Determine the gamma coefficient: Usually, when performing gamma correction, the image can be corrected by setting the gamma value and substituting it into the following formula (7). Generally, the image is converted into a grayscale image for processing. The present invention corrects the R, G, and B channels of the image respectively. The reference formula for gamma correction is shown in the following formula (7). Before manual correction, the known RGB value of the target image is compared with the RGB value of the image to be corrected, and the γ coefficient is inferred in the formula (7). At this time, an approximate value is obtained; the final γ coefficient is determined by comparison with the influence of the transformation of the γ value range on the image brightness.

[0097] Dataset preprocessing process: 1. Calculate the average pixel value of the image in the target dataset, then calculate the average pixel value of the image in the dataset to be corrected, and compare the average pixel values ​​of different datasets. 2. Usually when the γ value is less than 1, the contrast of the low grayscale area of ​​the image is enhanced, making the dark image brighter; when the γ value is greater than 1, the contrast of the high grayscale area of ​​the image is enhanced, making the bright image darker. Figure 7 (1) The images in the NH-HAZE dataset are generally darker. Figure 7 (2) The images in the DENSE-HAZE dataset are brighter overall. In order to make the image color distribution in the DENSE-HAZE dataset closer to the NH-HAZE dataset, the specific value of γ is determined through manual correction and adaptive gamma correction, thereby correcting the dataset images. Figure 7 (3) is the corrected image. After correction, it can be seen by naked eye Figure 7 The overall brightness of (3) is Figure 7 (1) is similar.

[0098] The gamma correction formula is shown in formula (7).

[0099] .

[0100] Where O and I are the output and input pixel intensities, respectively (∈[0, 255]); γ is the gamma coefficient; the subscripts R, G, and B indicate that the values ​​of different channels are unique.

[0101] In the present invention, in order to supplement the training data set, the gamma correction based on RGB value is used to pre-process the existing real foggy data set to expand the training data set and avoid overfitting of the network.

[0102] (c) The preprocessed image sample data sets are combined into a total sample set; the total sample set includes a training sample set.

[0103] (d) Using the pre-defogging image samples in the training sample set as input and the corresponding real-life images as labels, an initial non-uniform foggy image defogging network is trained.

[0104] (e) calculating a total loss error based on the predicted image samples after defogging output by the non-uniform foggy image defogging network during training, the corresponding image samples before defogging, and the corresponding real-life image samples; the total loss error includes a smoothing L1 loss error term, a perceptual loss error term, and an MS-SSIM loss error term.

[0105] Regarding the calculation of loss error, in terms of image reconstruction, smooth L1 loss error is more robust than MSE loss. The calculation formula of smooth L1 loss error is shown in formula (8) and formula (9):

[0106] .

[0107] .

[0108] in, and They are respectively represented as the pixel value of the i-th pixel of the predicted dehazed image sample e and the real image sample g without haze; N is the total number of pixels.

[0109] In addition to pixel supervision, perceptual loss is also used to optimize the dehazing effect. Perceptual loss extracts the features of the input and target images through a pre-trained network, calculates their distance in the feature space (i.e., the Euclidean distance between the two), and minimizes this distance to obtain the optimal result. The VGG16 network pre-trained on ImageNet is used as the backbone network of the loss network, and the network training is adjusted through the loss function to make the restored image closer to the real-life image sample. The perceptual loss function is defined in formula (10):

[0110] .

[0111] In the formula, They are represented as channels, height and width in the feature map respectively; M represents the number of feature layers of the pre-trained VGG16 network; Represents the feature representation of the predicted image sample after dehazing or the real image sample after dehazing in the jth layer of the pre-trained VGG16 network; Represents the square of the L2 norm, that is, the square of the Euclidean distance.

[0112] MS-SSIM loss error It introduces multiple scales based on the SSIM (structural similarity) loss error to better capture the detailed information of the image. From the perspective of image structure, it calculates the difference in brightness, contrast and structure between different regions of the image to obtain the similarity between the images. The MS-SSIM loss error is a weighted average of the SSIM errors of each scale (see formula (11)) to obtain the final MS-SSIM loss error value.

[0113] .

[0114] In formula (11), d and b represent two windows centered on the i-th pixel in the dehazed predicted image sample and the actual dehazed image sample, respectively; They represent the mean values ​​obtained by Gaussian filtering of window d and window b respectively; , They represent the standard deviations of Gaussian filtering of window d and window b respectively; represents the covariance obtained by Gaussian filtering of window d and window b; and Represents two variables and has the function of stabilizing the division. ;in <<1 is a constant, usually 0.01; The general value is 0.03. L is the dynamic range of grayscale, which is determined by the data type of the image. If the data is uint8 type, then L=255.

[0115] The weighted sum of the above three parts of smooth L1 loss error, perceptual loss error, and MS-SSIM loss error is the total loss error, as shown in formula (12), which supervises the training of the non-uniform foggy image dehazing network:

[0116] .

[0117] in, α and β represents the weight; as an example, α =0.001, β =0.2.

[0118] (f) adjusting the network parameters of the non-uniform foggy image dehazing network according to the total loss error back propagation.

[0119] (g) When the total loss error converges or the number of training iterations reaches a maximum number of iterations, a trained non-uniform foggy image defogging network is obtained.

[0120] The technical effects of this embodiment are as follows: (1) Fog is one of the common adverse weather conditions, which leads to reduced visibility and reduced image quality captured by visual sensors. The image defogging algorithm of the present invention improves image quality, provides high-quality image data for subsequent advanced visual tasks, and ensures that the autonomous driving vehicle can perceive the surrounding environment accurately, in real time, and robustly.

[0121] (2) The present invention analyzes the characteristics of foggy images and designs a defogging algorithm suitable for autonomous driving scenarios in real foggy environments based on a model-free deep learning method, based on the fact that fog in real foggy environments is non-uniformly distributed. This makes the defogging algorithm not limited to training data and performs well in real foggy scenes, thereby improving the generalization ability of the algorithm.

[0122] (3) The multi-level patch network branch adopted in the present invention limits the local receptive field, so that the pixel output of each patch block is only affected by the receptive field of the corresponding patch block, which is more suitable for processing non-uniform fog images; the encoding-decoding structure processes image information from fine to coarse and from local to global, and the residual learning method can retain more high-frequency detail information of the image; considering the demand of autonomous driving vehicles for lightweight networks, a lightweight defogging network is constructed based on the encoding-decoding structure, which reduces network parameters while ensuring the defogging effect, facilitating subsequent embedding into mobile devices.

[0123] (4) In order to avoid overfitting in network training, this application uses gamma correction based on RGB values ​​to process the data set, thereby achieving the purpose of enriching the data set. By calculating the average pixel value of the RGB channels of the images in the different data sets used, and comparing the average RGB values ​​of the images in each data set, the gamma coefficient is determined, and the data set is adjusted so that the RGB value of the corrected data set is closer to the target data set to be fused, reducing the differences between different data sets, and achieving the purpose of enriching the real foggy data set;

[0124] (5) The present invention selects the weighted sum of the smooth L1 loss error, the perceptual loss error, and the MS-SSIM loss error as the total loss error function, which can better optimize the prediction of the dehazed image, so that the dehazed image is close to the corresponding real-life image sample, and is more robust in image reconstruction.

[0125] (6) Inspired by spatial pyramid matching, the image is divided into small grids of the same size, features are extracted from each grid, and then connected into a large feature representation. Because the amount of information reflected in each local area of ​​the image is different, dividing the image into different grids instead of directly inputting the whole image for processing is conducive to the network paying more attention to the local information of different areas of the image and capturing more hierarchical information; at the same time, dividing the image into grids for processing can better process non-uniform fog images, limit the receptive field of specific areas, and the pixel output of each grid is only affected by the receptive field of the corresponding grid, without affecting the global information of the image, and can improve the defogging effect of the image while maintaining fast operation. Therefore, starting from the analysis of the characteristics of foggy images, the present invention performs targeted processing on fog in different areas and different concentrations, thereby reasonably allocating the computing resources of the defogging network, improving the network's operating efficiency, and improving the image defogging quality.

[0126] In order to verify the effectiveness of the defogging method proposed in the present invention in specific applications, the defogging network proposed in the present invention is tested with several existing defogging methods such as DCP, AOD-Net, GCA-Net, FFA-Net, and DMPHN in different real foggy image data sets for quantitative comparison. The superiority of the algorithm proposed in the present invention is demonstrated through comparative experiments. The quantitative comparison uses two objective evaluation indicators, structural similarity SSIM and peak signal-to-noise ratio PSNR, to compare the algorithm proposed in the present invention with the existing algorithms. The quantitative comparison data is shown in Table 1.

[0127] Table 1 Quantitative comparison data on NHHAZE and DENSE-HAZE datasets

[0128]

[0129] From the comparison data in Table 1, it can be seen that the defogging algorithm proposed in the present invention is superior to other defogging methods in the quantitative comparison results based on the evaluation indicators PSNR and SSIM on the non-uniform fog image dataset NH-HAZE and the dense fog image dataset DENSE-HAZE, and has achieved the best non-uniform fog removal effect. On the non-uniform fog image dataset NH-HAZE, the evaluation index results PSNR and SSIM are 0.68dB and 0.036 higher than the second-ranked FFA-Net defogging method, and 2.44dB and 0.065 higher than the third-ranked defogging method based on a single deep multi-patch structure DMPHN; on the dense fog image dataset DENSE-HAZE, the evaluation index results PSNR and SSIM are 1.26dB and 0.009 higher than the second-ranked FFA-Net defogging algorithm, and 1.63dB and 0.049 higher than the third-ranked defogging algorithm based on a single deep multi-patch structure DMPHN. The superiority of the defogging method proposed in the present invention can be proved by quantitative comparison of experimental data. Compared with the existing defogging methods, the defogging method of the present invention has the best defogging performance for real non-uniform foggy images and has higher practical application value.

[0130] The present application also provides an application scenario, which applies the above-mentioned foggy image defogging method. Specifically: The foggy image defogging method provided in this embodiment can be applied to the defogging of embedded devices, such as autonomous driving vehicles, video surveillance, mobile devices, etc. When the above-mentioned embedded devices encounter foggy weather during use (such as autonomous driving vehicles running in foggy environments, video surveillance in foggy environments, etc.), and the information collected by the front-end visual sensor cannot be recognized and applied, the method of the present invention can be used to restore the collected foggy image to a clear image, provide more favorable scene information, help subsequent advanced machine vision tasks, and enrich the environmental applicability of the above-mentioned embedded devices.

[0131] Based on the same inventive concept, the embodiment of the present application also provides a non-uniform real foggy image defogging device for implementing the non-uniform real foggy image defogging method involved above. The implementation solution provided by the device to solve the problem is similar to the implementation solution recorded in the above method, so the specific limitations in one or more non-uniform real foggy image defogging device embodiments provided below can refer to the limitations of the non-uniform real foggy image defogging method above, and will not be repeated here.

[0132] In an exemplary embodiment, Figure 8 As shown, a non-uniform real foggy image defogging device is provided, comprising:

[0133] The image acquisition sub-device M1 is used to acquire real foggy images.

[0134] The image defogging sub-device M2 is used to input the real foggy image into the non-uniform foggy image defogging network to obtain the defogged image; the non-uniform foggy image defogging network includes a two-dimensional discrete wavelet transform branch, a multi-level patch network branch and an image feature fusion module;

[0135] The two-dimensional discrete wavelet transform branch is used to process the low-frequency features of the real foggy image.

[0136] The multi-level patch network branch is used to process the high-frequency features of the real foggy image.

[0137] The image feature fusion module is used to fuse the image output by the two-dimensional discrete wavelet transform branch and the image output by the multi-level patch network branch to obtain the defogged image.

[0138] The two-dimensional discrete wavelet transform branch includes a first convolution module, a wavelet transform downsampling module, a second convolution module, a first residual addition layer and a wavelet inverse transform upsampling module which are connected in series in sequence; the input of the first convolution module is the real foggy image; the input end of the first residual addition layer is also connected to the output end of the wavelet transform downsampling module; the input end of the wavelet inverse transform upsampling module is also connected to the output end of the wavelet transform downsampling module; the output end of the wavelet inverse transform upsampling module is connected to the input end of the image feature fusion module; the multi-level patch network branch includes a first-level patch module, a second-level patch module, a first residual addition layer and a wavelet inverse transform upsampling module; the first residual addition layer is ... first residual addition layer is connected to the output end of the wavelet transform downsampling module; the first residual addition layer is connected to the output end of the wavelet transform upsampling module; the first residual addition layer is connected to the output end of the wavelet transform downsampling module; the first residual addition layer is connected to the output end of the wavelet transform upsampling module; the first residual addition layer is connected to the output end of the wavelet transform downsampling module; the first residual addition layer is connected to the output end of the wavelet transform upsampling module; the first residual addition layer is connected to the output end of the wavelet transform downsampling module; the first residual addition layer is connected to the output end of the wavelet transform upsampling module; the first residual addition layer is connected to the output end of the wavelet transform downsampling module The patch module and the third-level patch module; the first-level patch module includes a second residual addition layer, a first encoder, a third residual addition layer and a first decoder connected in series in sequence; the second-level patch module includes a first image segmentation layer, a fourth residual addition layer, a second encoder, a fifth residual addition layer, a first splicing layer and a second decoder connected in series in sequence; the third-level patch module includes a second image segmentation layer, a third encoder, a second splicing layer and a third decoder connected in series in sequence; the input end of the second residual addition layer is connected to the output end of the second decoder; the input of the second residual addition layer also includes the real foggy sky image; the input end of the third residual addition layer is also connected to the output end of the first splicing layer; the output of the first decoder is the defogged image; the input of the first image segmentation layer is the real foggy image; the input end of the fourth residual addition layer is also connected to the output end of the third decoder; the input end of the fifth residual addition layer is also connected to the output end of the second splicing layer; the input of the second image segmentation layer is the real foggy image; the second image segmentation layer is used to divide the real foggy image into a number of patches according to a number of horizontal dividing lines and a number of vertical dividing lines to obtain a first patch divided image; the The second stitching layer is used to stitch adjacent patch blocks in the first stitching direction in the first patch block divided image to obtain a first stitching image; the first image segmentation layer is used to perform patch block division on the real foggy image according to the patch block division method in the first stitching image to obtain a second patch block divided image; the first stitching layer is used to stitch adjacent patch blocks in the second stitching direction in the image output by the fifth residual addition layer to obtain a second stitching image; when the first stitching direction is a horizontal direction, the second stitching direction is a vertical direction; when the first stitching direction is a vertical direction, the second stitching direction is a horizontal direction.

[0139] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Fig. 9As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store defogging image data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for defogging a non-uniform real foggy image is implemented.

[0140] Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0141] In an exemplary embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0142] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0143] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0144] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0145] 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 computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0146] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0147] 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.

[0148] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for defogging non-uniform real foggy images, characterized in that: include: Get real foggy images; Inputting the real foggy image into the trained non-uniform foggy image defogging network to obtain a defogged image; The non-uniform foggy image defogging network includes a two-dimensional discrete wavelet transform branch, a multi-level patch network branch and an image feature fusion module; The two-dimensional discrete wavelet transform branch is used to process the low-frequency features of the real foggy image; The multi-level patch network branch is used to process the high-frequency features of the real foggy image; The two-dimensional discrete wavelet transform branch includes a first convolution module, a wavelet transform downsampling module, a second convolution module, a first residual addition layer and a wavelet inverse transform upsampling module which are connected in series in sequence; the input of the first convolution module is the real foggy image; the input end of the first residual addition layer is also connected to the output end of the wavelet transform downsampling module; the input end of the wavelet inverse transform upsampling module is also connected to the output end of the wavelet transform downsampling module; the output end of the wavelet inverse transform upsampling module is connected to the input end of the image feature fusion module; the multi-level patch network branch includes a first-level patch module, a second-level patch module and a third-level patch module; The first-level patch module includes a second residual addition layer, a first encoder, a third residual addition layer and a first decoder connected in series in sequence; the second-level patch module includes a first image segmentation layer, a fourth residual addition layer, a second encoder, a fifth residual addition layer, a first splicing layer and a second decoder connected in series in sequence; the third-level patch module includes a second image segmentation layer, a third encoder, a second splicing layer and a third decoder connected in series in sequence; The input end of the second residual addition layer is connected to the output end of the second decoder; the input of the second residual addition layer also includes the real foggy image; the input end of the third residual addition layer is also connected to the output end of the first splicing layer; the output end of the first decoder is connected to the input end of the image feature fusion module; the output of the image feature fusion module is the defogged image; The input of the first image segmentation layer is the real foggy image; the input end of the fourth residual addition layer is also connected to the output end of the third decoder; the input end of the fifth residual addition layer is also connected to the output end of the second splicing layer; The input of the second image segmentation layer is the real foggy image; the second image segmentation layer is used to divide a plurality of patch blocks in the real foggy image according to a plurality of horizontal segmentation lines and a plurality of vertical segmentation lines to obtain a first patch block segmentation image; the second splicing layer is used to splice adjacent patch blocks in a first splicing direction in the first patch block segmentation image to obtain a first spliced ​​image; The first image segmentation layer is used to divide the real foggy image into patches according to the patch division method in the first spliced ​​image to obtain a second patch division image; The first splicing layer is used to splice adjacent patch blocks in a second splicing direction in the image output by the fifth residual addition layer to obtain a second splicing image; when the first splicing direction is a horizontal direction, the second splicing direction is a vertical direction; when the first splicing direction is a vertical direction, the second splicing direction is a horizontal direction; The image feature fusion module is used to fuse the image output by the two-dimensional discrete wavelet transform branch and the image output by the multi-level patch network branch to obtain the defogged image.

2. The method for defogging non-uniform real foggy images according to claim 1, characterized in that: Inputting the real foggy image into the trained non-uniform foggy image defogging network to obtain a defogged image specifically includes: Inputting the real foggy image into the two-dimensional discrete wavelet transform branch in the non-uniform foggy image defogging network to obtain a defogged image after wavelet transformation; Inputting the real foggy image into the multi-level patch network branch in the non-uniform foggy image defogging network to obtain a defogged image after multi-level patch processing; The defogged image after wavelet transformation and the defogged image after multi-level patch processing are input into the image feature fusion module to obtain the defogged image.

3. The method for defogging non-uniform real foggy images according to claim 2, characterized in that: Inputting the real foggy image into the two-dimensional discrete wavelet transform branch in the non-uniform foggy image defogging network to obtain a defogged image after wavelet transformation, specifically comprising: Inputting the real foggy image into the first convolution module to obtain a first convolution image; Inputting the first convolution image into the wavelet transform downsampling module to obtain a downsampled image; Inputting the downsampled image into the second convolution module to obtain a second convolution image; Inputting the downsampled image and the second convolution image into the first residual addition layer to obtain a residual addition image; The residual addition image and the down-sampled image are input into the wavelet inverse transform up-sampling module to obtain an up-sampled image; the up-sampled image is the defogging image after the wavelet transform.

4. The method for defogging non-uniform real foggy images according to claim 2, characterized in that: Inputting the real foggy image into the multi-level patch network branch in the non-uniform foggy image defogging network to obtain a defogged image after multi-level patch processing, specifically comprising: Inputting the real foggy image into the second image segmentation layer to obtain the first patch block segmentation image; Inputting the first patch block divided image into the third encoder to obtain a first encoded image; Inputting the first encoded image into the second splicing layer to obtain a first splicing image; Inputting the first spliced ​​image into the third decoder to obtain a first decoded image; Inputting the real foggy image into the first image segmentation layer to obtain the second patch block segmentation image; Inputting the second patch block divided image and the first decoded image into the fourth residual addition layer to obtain a first residual addition image; Inputting the first residual addition image into the second encoder to obtain a second encoded image; Inputting the second coded image and the first spliced ​​image into the fifth residual addition layer to obtain a second residual addition image; Inputting the second residual addition image into the first splicing layer to obtain a second splicing image; Inputting the second spliced ​​image into the second decoder to obtain a second decoded image; Inputting the second decoded image and the real foggy image into the second residual addition layer to obtain a third residual addition image; Inputting the third residual addition image into the first encoder to obtain a third encoded image; Inputting the third coded image and the second spliced ​​image into the third residual addition layer to obtain a fourth residual addition image; The fourth residual addition image is input into the first decoder to obtain a third decoded image; the third decoded image is the defogging image after the multi-level patch processing.

5. The method for defogging non-uniform real foggy images according to claim 1, characterized in that: Before executing the step of "inputting the real foggy image into the trained non-uniform foggy image defogging network to obtain the defogged image", the non-uniform real foggy image defogging method further includes: training the non-uniform foggy image defogging network, specifically: Obtain different foggy image sample datasets; The gamma correction method based on RGB value is used to preprocess each foggy image sample data set, and each preprocessed image sample data set is obtained; The preprocessed image sample data sets are combined into a total sample set; the total sample set includes a training sample set; Taking the pre-defogging image samples in the training sample set as input and the corresponding real-life image samples as labels to train an initial non-uniform foggy image defogging network; Calculate the total loss error according to the predicted image samples after defogging output by the non-uniform foggy image defogging network during training, the corresponding image samples before defogging and the corresponding real-life image samples; the total loss error includes a smoothing L1 loss error term, a perceptual loss error term and an MS-SSIM loss error term; Adjusting the network parameters of the non-uniform foggy image defogging network according to the total loss error back propagation; When the total loss error converges or the number of training iterations reaches a maximum number of iterations, a trained non-uniform foggy image defogging network is obtained.

6. A non-uniform real foggy image defogging device based on the non-uniform real foggy image defogging method according to any one of claims 1 to 5, characterized in that: The non-uniform real foggy image defogging device comprises: An image acquisition sub-device, used for acquiring real foggy images; An image defogging sub-device, used for inputting the real foggy image into a non-uniform foggy image defogging network to obtain a defogged image; the non-uniform foggy image defogging network comprises a two-dimensional discrete wavelet transform branch, a multi-level patch network branch and an image feature fusion module; The two-dimensional discrete wavelet transform branch is used to process the low-frequency features of the real foggy image; The multi-level patch network branch is used to process the high-frequency features of the real foggy image; The image feature fusion module is used to fuse the image output by the two-dimensional discrete wavelet transform branch and the image output by the multi-level patch network branch to obtain the defogged image.

7. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the non-uniform real foggy image defogging method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the non-uniform real foggy image defogging method described in any one of claims 1 to 5 is implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the non-uniform real foggy image defogging method described in any one of claims 1 to 5 is implemented.

Images