An Image Dehazing Method and System Based on Meta-Assisted Learning

Abstract

The present invention relates to an image defogging method and system based on meta-assisted learning. The method includes: constructing a data set, which includes an artificially synthesized fog image data set and a natural fog image data set, and the natural fog image data set includes multiple real foggy day images; constructing a convolutional network model, which includes an encoder, a decoder, a meta-learning unit, and an auxiliary learning unit; using the convolutional network model to extract features from the synthesized fog images and the real foggy day images to obtain image feature data; using the image feature data and the clear images to optimize the convolutional network model to obtain an optimized convolutional network model; and using the optimized convolutional network model to perform defogging processing on the foggy images to be measured. By introducing an auxiliary learning unit, the present invention makes the data distributions of the synthesized fog images and the real foggy day images consistent, and improves the defogging performance of the convolutional network model.

Description

Technical Field

[0001] The present invention relates to the field of image dehazing, and particularly to an image dehazing method and system based on meta-aided learning. Background Art

[0002] Image dehazing is one of the core research issues in the field of image enhancement. The purpose is to restore a blurred hazy image into a clear image and improve the performance of other computer vision tasks. Image dehazing is a basic task in computer vision, and its application scenarios are very extensive. It can assist other computer vision tasks, such as object detection, person re-identification, image segmentation, image classification, etc. Therefore, the research on image dehazing technology is extremely urgent.

[0003] Existing image dehazing technologies are mainly divided into two categories. One is the machine learning method based on statistical priors, and the other is the deep learning method based on end-to-end. The methods based on statistical priors mainly include methods such as dark channel prior, color prior, and contrast prior. These methods analyze a large amount of data and combine the atmospheric scattering model for dehazing. The deep learning method based on end-to-end directly inputs the image to be dehazed into the network model, and the network model directly outputs a clear image. However, the deep learning method based on end-to-end is too dependent on the artificially synthesized paired hazy image dataset. The data distribution of the artificially synthesized hazy images is inconsistent with that of real foggy images. The network model trained only with the artificially synthesized dataset performs poorly on real data. Some methods introduce natural datasets in a self-supervised manner, but there are many noises, resulting in suboptimal dehazing performance. Summary of the Invention

[0004] The purpose of the present invention is to provide an image dehazing method and system based on meta-aided learning to solve the problem of slightly poor dehazing performance of the existing image dehazing methods.

[0005] To achieve the above purpose, the present invention provides the following solutions:

[0006] An image dehazing method based on meta-aided learning includes:

[0007] Constructing a dataset; the dataset includes an artificially synthesized hazy image dataset and a natural hazy image dataset; the artificially synthesized hazy image dataset includes a clear image and multiple synthetic hazy images artificially synthesized from the clear image; the natural hazy image dataset includes multiple real foggy images;

[0008] Constructing a convolutional network model; the convolutional network model includes an encoder, a decoder, a meta-learning unit, and an auxiliary learning unit;

[0009] Use a convolutional network model to extract features from the synthesized fog image and the real foggy-day image to obtain image feature data; the image feature data includes first image feature data and second image feature data;

[0010] Use the image feature data and the clear image to optimize the convolutional network model to obtain an optimized convolutional network model;

[0011] Use the optimized convolutional network model to perform defogging processing on the foggy image to be measured.

[0012] Optionally, before using the convolutional network model to extract features from the synthesized fog image and the real foggy-day image to obtain image feature data, it further includes:

[0013] Perform normalization processing on the synthesized fog image and the real foggy-day image to obtain a normalized synthesized fog image and a normalized real foggy-day image.

[0014] Optionally, the using the image feature data and the clear image to optimize the convolutional network model to obtain an optimized convolutional network model specifically includes:

[0015] Use the first image feature data and the clear image to optimize the meta-learning unit of the convolutional network model to obtain an optimized meta-learning unit;

[0016] Use the second image feature data to optimize the auxiliary learning unit of the convolutional network model to obtain an optimized auxiliary learning unit;

[0017] Determine the optimized convolutional network according to the optimized meta-learning unit and the optimized auxiliary learning unit.

[0018] Optionally, the using the first image feature data to optimize the meta-learning unit of the convolutional network model to obtain an optimized meta-learning unit specifically includes:

[0019] Input the first image feature data into the meta-learning unit of the convolutional network model to generate a first pseudo-fog-free image;

[0020] Calculate the loss function of the meta-learning unit according to the clear image and the first pseudo-fog-free image;

[0021] Optimize the meta-learning unit using the backpropagation algorithm according to the loss function of the meta-learning unit to obtain an optimized meta-learning unit.

[0022] Optionally, the using the second image feature data to optimize the auxiliary learning unit of the convolutional network model to obtain an optimized auxiliary learning unit specifically includes:

[0023] Input the second image feature data into the auxiliary learning unit of the convolutional network model to obtain third image feature data;

[0024] According to the third image feature data and the first haze-free image, obtain fourth image feature data;

[0025] Determine a second haze-free image according to the fourth image feature data;

[0026] Input the second haze-free image into the discriminator network of the auxiliary learning unit to calculate the loss function of the auxiliary learning unit;

[0027] Optimize the auxiliary learning unit according to the loss function of the auxiliary learning unit using the backpropagation algorithm to obtain an optimized auxiliary learning unit.

[0028] An image dehazing system based on meta-aided learning, comprising:

[0029] A dataset construction module for constructing a dataset; the dataset includes an artificially synthesized foggy image dataset and a natural foggy image dataset; the artificially synthesized foggy image dataset includes a clear image and multiple synthetic foggy images artificially synthesized from the clear image; the natural foggy image dataset includes multiple real foggy day images;

[0030] A convolutional network model construction module for constructing a convolutional network model; the convolutional network model includes an encoder, a decoder, a meta-learning unit, and an auxiliary learning unit;

[0031] A feature extraction module for extracting features from the synthetic foggy images and the real foggy day images using the convolutional network model to obtain image feature data; the image feature data includes first image feature data and second image feature data;

[0032] A model optimization module for optimizing the convolutional network model using the image feature data and the clear image to obtain an optimized convolutional network model;

[0033] A dehazing module for dehazing a foggy image to be measured using the optimized convolutional network model.

[0034] Optionally, it further includes:

[0035] A normalization processing module for performing normalization processing on the synthetic foggy images and the real foggy day images to obtain normalized synthetic foggy images and normalized real foggy day images.

[0036] Optionally, the model optimization module specifically includes:

[0037] A meta-learning unit optimization sub-module, configured to optimize the meta-learning unit of the convolutional network model by using the first image feature data and the clear image, so as to obtain an optimized meta-learning unit;

[0038] An auxiliary learning unit optimization sub-module, configured to optimize the auxiliary learning unit of the convolutional network model by using the second image feature data, so as to obtain an optimized auxiliary learning unit;

[0039] A model determination sub-module, configured to determine an optimized convolutional network model according to the optimized meta-learning unit and the optimized auxiliary learning unit.

[0040] Optionally, the meta-learning unit optimization sub-module specifically includes:

[0041] A first haze-free image generation unit, configured to input the first image feature data into the meta-learning unit of the convolutional network model to generate a first pseudo-haze-free image;

[0042] A first loss function calculation unit, configured to calculate the loss function of the meta-learning unit according to the clear image and the first pseudo-haze-free image;

[0043] A first optimization unit, configured to optimize the meta-learning unit according to the loss function of the meta-learning unit by using the backpropagation algorithm to obtain an optimized meta-learning unit.

[0044] Optionally, the auxiliary learning unit optimization sub-module specifically includes:

[0045] A first feature extraction unit, configured to input the second image feature data into the auxiliary learning unit of the convolutional network model to obtain third image feature data;

[0046] A second feature extraction unit, configured to obtain fourth image feature data according to the third image feature data and the first pseudo-haze-free image;

[0047] A second haze-free image generation unit, configured to determine a second pseudo-haze-free image according to the fourth image feature data;

[0048] A second loss function calculation unit, configured to input the second pseudo-haze-free image into the discriminator network of the auxiliary learning unit to calculate the loss function of the auxiliary learning unit;

[0049] A second optimization unit, configured to optimize the auxiliary learning unit according to the loss function of the auxiliary learning unit by using the backpropagation algorithm to obtain an optimized auxiliary learning unit.

[0050] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0051] An image dehazing method and system based on meta-assisted learning according to the present invention, the method comprising: constructing a data set; normalizing the synthetic foggy images and real foggy images in the data set to obtain normalized synthetic foggy images and normalized real foggy images; using a convolutional network to extract features from the normalized synthetic foggy images and the normalized real foggy images to obtain image feature data; the convolutional network comprising an encoder, a decoder, a meta-learning unit and an auxiliary learning unit; using the image feature data to optimize the convolutional network, i.e., optimizing the meta-learning unit and the auxiliary learning unit, to obtain an optimized convolutional network; using the optimized convolutional network to perform dehazing processing on the foggy image to be measured to obtain a clear image. By introducing an auxiliary learning unit, the meta-learning unit and the auxiliary learning unit share a part of the parameters of the convolutional network, and the auxiliary learning unit is trained on the natural foggy image data set in a self-supervised manner to fine-tune the network, so that the data distributions of the synthetic foggy images and the real foggy images are consistent, and the dehazing performance of the convolutional network model is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention, and for those of ordinary skill in the art, other drawings can be obtained according to these drawings without creative efforts.

[0053] Figure 1 It is a flowchart of an image dehazing method based on meta-assisted learning provided by the present invention;

[0054] Figure 2 It is a structural diagram of a convolutional network model provided by the present invention;

[0055] Figure 3 It is a structural diagram of a discriminator network in the auxiliary learning unit provided by the present invention;

[0056] Figure 4 It is a structural diagram of an image dehazing system based on meta-assisted learning provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0058] The object of the present invention is to provide an image dehazing method and system based on meta-assisted learning to solve the problem of slightly poor dehazing performance of the existing image dehazing methods.

[0059] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] In order to optimize the dehazing performance, the present invention proposes an image dehazing method and system based on meta-assisted learning. In order to enable the convolutional network model to learn more information, an auxiliary learning unit is added to the convolutional network model. The auxiliary learning unit shares a part of the parameters of the meta-learning unit. The auxiliary learning unit is trained on the natural fog image dataset in a self-supervised manner to fine-tune the network. By introducing the auxiliary learning unit, the convolutional network model can better adapt to the real dehazing scenario.

[0061] Figure 1 The flowchart of an image dehazing method based on meta-assisted learning provided by the present invention is as Figure 1 shown. An image dehazing method based on meta-assisted learning includes:

[0062] Step 101: Construct a dataset. The dataset includes an artificially synthesized fog image dataset and a natural fog image dataset; the artificially synthesized fog image dataset includes a clear image and multiple synthetic fog images artificially synthesized from the clear image; the natural fog image dataset includes multiple real foggy day images.

[0063] The natural fog image dataset is real foggy day images taken in real scenarios. The difference from the artificially synthesized fog image dataset is that the natural fog image dataset only contains real foggy day images. The clear image is a fog-free image manually selected.

[0064] In practical applications, data collection and preprocessing are first performed. The collected data includes an artificially synthesized fog image dataset and a natural fog image dataset. The collected fog image dataset is randomly divided into three subsets, including a training set, a test set, and a validation set, with a ratio of 8:1:1, and the fog images are standardized.

[0065] Step 102: Construct a convolutional network model. The convolutional network model includes an encoder, a decoder, a meta-learning unit, and an auxiliary learning unit.

[0066] Figure 2 The structural diagram of the convolutional network model provided by the present invention is as Figure 2As shown, in practical applications, then design the network structure shared by the meta-learning unit and the auxiliary learning unit. The shared network structure includes an image encoder and a feature decoder. The encoder consists of 7 modules, and the decoder consists of 6 modules. The encoder is used to extract multi-dimensional feature data of the foggy image. During the process of extracting feature data, the feature fusion module is used to fuse and convolve features with inconsistent dimensions to obtain more image information. After the encoder extracts the feature data, the decoder performs an up-convolution operation on the extracted feature data to restore the clear image. This shared network structure can extract image features to the maximum extent.

[0067] Then design the meta-learning module. After passing through the encoder and decoder, perform a second convolution operation on the feature data to obtain a clear restored image, and calculate the loss with the real fog-free image as the target, and backpropagate to optimize the network model.

[0068] Finally, design the auxiliary learning module. In order to enable the convolutional network model to learn more information, an auxiliary learning unit is introduced. The auxiliary learning unit mainly fine-tunes the network in a self-supervised manner. In addition to adding a new branch to the original convolutional network model, the auxiliary learning unit also newly introduces a discriminator network to supervise the auxiliary learning unit. Figure 3 This is the structure diagram of the discriminator network in the auxiliary learning unit provided by the present invention. The discriminator network is as Figure 3 shown.

[0069] Step 103: Use the convolutional network model to extract features from the synthesized foggy image and the real foggy day image to obtain image feature data. The image feature data includes first image feature data and second image feature data.

[0070] In a specific embodiment, before the step 103, it further includes:

[0071] Perform normalization processing on the synthesized foggy image and the real foggy day image to obtain a normalized synthesized foggy image and a normalized real foggy day image. In practical applications, perform data preprocessing on the synthesized foggy image and the real foggy day image, that is, before the synthesized foggy image and the real foggy day image are input into the convolutional network model, normalize the pixel values of the synthesized foggy image and the real foggy day image to between [0, 1].

[0072] In practical applications, use the encoder and decoder to extract features from the input image of the convolutional network model. The input images include the synthesized foggy images in the artificially synthesized foggy image dataset and the real foggy day images in the natural foggy image dataset.

[0073] First, input the normalized synthesized foggy image and the normalized real foggy day image into the convolutional network model in sequence. Taking the normalized synthesized foggy image as an example, input the normalized synthesized foggy image into the encoder. The steps are as follows:

[0074] The normalized synthesized fog image is input into the first convolutional module (Conv block1) to obtain the first synthesized fog image feature data.

[0075] The first synthesized fog image feature data is input into the second residual module (Resnet block2) to obtain the second synthesized fog image feature data, that is, the output data of the first convolutional module is convolved again to obtain the second synthesized fog image feature data, J = φ relu (I + ρ(I)), where J represents the second synthesized fog image feature data, I represents the first synthesized fog image feature data, and φ relu represents the ReLU activation function, and ρ represents the convolution operation.

[0076] The second synthesized fog image feature data is input into the third convolutional module (Conv block3) for downsampling to obtain the third synthesized fog image feature data. The stride of the third convolutional module is 2.

[0077] The first synthesized fog image feature data and the third synthesized fog image feature data are input into the fourth fusion module (Feature block4) for feature fusion to obtain the fourth synthesized fog image feature data. In order to fuse feature data of different scales, all the feature data to be fused are resampled to the same scale before fusion.

[0078] The fourth synthesized fog image feature data is input into the fifth residual module (Resnet block5) for feature extraction to obtain the fifth synthesized fog image feature data.

[0079] The fifth synthesized fog image feature data is fed into the sixth convolutional module (Conv block6) for downsampling to obtain the sixth synthesized fog image feature data.

[0080] The first synthesized fog image feature data, the fourth synthesized fog image feature data, and the sixth synthesized fog image feature data are input into the seventh fusion module (Feature block7) for feature fusion to obtain the seventh synthesized fog image feature data.

[0081] The seventh synthesized fog image feature data is input into the eighth residual module (Resnet block8) to obtain the eighth synthesized fog image feature data, and the eighth synthesized fog image feature data is input into the ninth residual module (Resnet block9) to obtain the ninth synthesized fog image feature data.

[0082] The extracted ninth synthesized fog image feature data is fed into the decoder, and the steps are as follows:

[0083] Input the ninth synthetic fog image feature data into the tenth upsampling convolutional module (Dconv block10) for upsampling to obtain the tenth synthetic fog image feature data.

[0084] Input the ninth synthetic fog image feature data and the tenth synthetic fog image feature data into the eleventh fusion module (Feature block11) for feature fusion operation, and merge the data after feature fusion with the fifth synthetic fog image feature data to obtain the eleventh synthetic fog image feature data, K = D(concat(L, D(ρ(M)+N))), where K represents the eleventh synthetic fog image feature data, L represents the fifth synthetic fog image feature data, M represents the ninth synthetic fog image feature data, N represents the tenth synthetic fog image feature data, D represents the upsampling convolutional operation, and concat represents the feature merging operation.

[0085] Input the eleventh synthetic fog image feature data into the twelfth residual module (Resnet block12) to obtain the twelfth synthetic fog image feature data.

[0086] Input the twelfth synthetic fog image feature data into the thirteenth upsampling convolutional module (Dconv block13) for upsampling to obtain the thirteenth synthetic fog image feature data.

[0087] Input the ninth synthetic fog image feature data, the eleventh synthetic fog image feature data, and the thirteenth synthetic fog image feature data into the fourteenth fusion module (Feature block14) for feature fusion operation, and merge the data after feature fusion with the second synthetic fog image feature data to obtain the fourteenth synthetic fog image feature data, O = D(concat(U, D(ρ q (V)+ρ s (W)+X))), where O represents the fourteenth synthetic fog image feature data, U represents the second synthetic fog image feature data, V represents the ninth synthetic fog image feature data, W represents the eleventh synthetic fog image feature data, and X represents the thirteenth synthetic fog image feature data. The fourteenth synthetic fog image feature data is the first image feature data.

[0088] Input the normalized real foggy image into the encoder and the decoder, perform the above operations, and obtain the fourteenth real fog-free feature data. The fourteenth real fog-free feature data is the second image feature data.

[0089] Step 104: Optimize the convolutional network model using the image feature data and the clear image to obtain an optimized convolutional network model.

[0090] In a specific embodiment, step 104 specifically includes:

[0091] Optimize the meta - learning unit of the convolutional network model by using the first image feature data and the clear image to obtain an optimized meta - learning unit.

[0092] Optimize the auxiliary learning unit of the convolutional network model by using the second image feature data to obtain an optimized auxiliary learning unit.

[0093] Determine an optimized convolutional network according to the optimized meta - learning unit and the optimized auxiliary learning unit.

[0094] In a specific embodiment, the step of optimizing the meta - learning unit of the convolutional network model by using the first image feature data to obtain an optimized meta - learning unit specifically includes:

[0095] Input the first image feature data into the meta - learning unit of the convolutional network model to generate a first pseudo - haze - free image.

[0096] Calculate the loss function of the meta - learning unit according to the clear image and the first pseudo - haze - free image.

[0097] Optimize the meta - learning unit according to the loss function of the meta - learning unit by using the backpropagation algorithm to obtain an optimized meta - learning unit.

[0098] In practical applications, input the fourteenth synthesized haze - image feature data output by the decoder into the fifteenth residual module (resnet block15) for feature extraction to obtain the fifteenth synthesized haze - image feature data.

[0099] Send the fifteenth synthesized haze - image feature data into the sixteenth convolutional module (conv block16) to obtain the finally generated first pseudo - haze - free image.

[0100] Calculate the loss of the meta - learning unit with the first pseudo - haze - free image and the clear image as the targets, loss1 = αL1(I a ,C)+βL2(I a ,C)+γL perceptual (∑ρ i (I a ,C)), where loss1 is the loss function of the meta - learning unit, L1 is the mean absolute error between the first pseudo - haze - free image and the clear image, L2 is the mean square error between the first pseudo - haze - free image and the clear image, L perceptual is the perceptual loss, C is the clear image, I a is the first pseudo - haze - free image, ρ i is the error function for calculating the error of the features of C and I a at the i - th layer, and α, β, γ are the weight parameters of the meta - learning unit.

[0101] Optimize the meta - learning unit using the backpropagation algorithm according to the loss function of the meta - learning unit.

[0102] In a specific embodiment, optimizing the auxiliary learning unit of the convolutional network model using the second image feature data to obtain an optimized auxiliary learning unit specifically includes:

[0103] Input the second image feature data into the auxiliary learning unit of the convolutional network model to obtain third image feature data.

[0104] According to the third image feature data and the first pseudo - haze - free image, obtain fourth image feature data.

[0105] Determine the second pseudo - haze - free image according to the fourth image feature data.

[0106] Input the second pseudo - haze - free image into the discriminator network of the auxiliary learning unit to calculate the loss function of the auxiliary learning unit.

[0107] Optimize the auxiliary learning unit using the backpropagation algorithm according to the loss function of the auxiliary learning unit to obtain an optimized auxiliary learning unit.

[0108] In practical applications, an auxiliary learning unit is introduced into the convolutional network model. The auxiliary learning unit and the meta - learning unit share the parameters of the encoder and decoder. The training of the auxiliary learning unit uses a natural fog image dataset. Through a self - supervised form, the convolutional network model learns more information from the natural fog image dataset. The specific implementation steps are as follows:

[0109] Input the fourteenth real foggy - day image feature data finally output by the decoder into the seventeenth residual module (resnetblock17) for feature extraction to obtain the seventeenth real foggy - day image feature data.

[0110] Input the seventeenth real foggy - day image feature data into the eighteenth convolutional module (conv block18) to obtain the eighteenth real foggy - day image feature data.

[0111] Input the first pseudo - haze - free image I obtained by the meta - learning unit a and the eighteenth real foggy - day image feature data jointly into the nineteenth residual module (resnet block19) to obtain the nineteenth real foggy - day image feature data.

[0112] Input the nineteenth real foggy - day image feature data into the twentieth convolutional module (conv block20) to obtain the second pseudo - haze - free image I b and the obtained second pseudo - haze - free image I bInput it into the discriminator network, calculate the loss function of the auxiliary learning unit, and use the backpropagation algorithm to optimize the parameters of the auxiliary learning unit. loss2 is the loss function for the auxiliary learning task, L t is the total variation loss, L d is the dark channel loss, L GAN is the generative adversarial loss, and μ, σ, τ are the weight parameters of the auxiliary learning unit; N u represents the number of image pairs in the training dataset; represents the differential operation matrix in the horizontal direction of the second pseudo-fog-free image; represents the differential operation matrix in the vertical direction of the second pseudo-fog-free image; I i represents the i-th foggy image in the input; I b represents the second pseudo-fog-free image; represents the second pseudo-fog-free image I b 's dark channel matrix.

[0113] Step 105: Use the optimized convolutional network model to perform defogging on the foggy image to be measured.

[0114] Figure 4 is the structural diagram of an image defogging system based on meta-aided learning provided by the present invention. As Figure 4 shown, an image defogging system based on meta-aided learning includes:

[0115] A dataset construction module 401 for constructing a dataset. The dataset includes an artificially synthesized foggy image dataset and a natural foggy image dataset; the artificially synthesized foggy image dataset includes a clear image and multiple synthetic foggy images artificially synthesized from the clear image; the natural foggy image dataset includes multiple real foggy day images.

[0116] A convolutional network model construction module 402 for constructing a convolutional network model. The convolutional network model includes an encoder, a decoder, a meta-learning unit, and an auxiliary learning unit.

[0117] A feature extraction module 403 for using the convolutional network model to extract features from the synthetic foggy images and the real foggy day images to obtain image feature data. The image feature data includes first image feature data and second image feature data.

[0118] A model optimization module 404 for using the image feature data and the clear image to optimize the convolutional network model to obtain an optimized convolutional network model.

[0119] A defogging module 405 for using the optimized convolutional network model to perform defogging on the foggy image to be measured.

[0120] In a specific embodiment, the image dehazing system based on meta-assisted learning further includes:

[0121] A normalization processing module, configured to perform normalization processing on the synthesized foggy image and the real foggy-day image to obtain a normalized synthesized foggy image and a normalized real foggy-day image.

[0122] In a specific embodiment, the model optimization module 404 specifically includes:

[0123] A meta-learning unit optimization sub-module, configured to optimize the meta-learning unit of the convolutional network model by using the first image feature data and the clear image to obtain an optimized meta-learning unit.

[0124] An auxiliary learning unit optimization sub-module, configured to optimize the auxiliary learning unit of the convolutional network model by using the second image feature data to obtain an optimized auxiliary learning unit.

[0125] A model determination sub-module, configured to determine an optimized convolutional network model according to the optimized meta-learning unit and the optimized auxiliary learning unit.

[0126] In a specific embodiment, the meta-learning unit optimization sub-module specifically includes:

[0127] A first fog-free image generation unit, configured to input the first image feature data into the meta-learning unit of the convolutional network model to generate a first pseudo fog-free image.

[0128] A first loss function calculation unit, configured to calculate the loss function of the meta-learning unit according to the clear image and the first pseudo fog-free image.

[0129] A first optimization unit, configured to optimize the meta-learning unit according to the loss function of the meta-learning unit by using the backpropagation algorithm to obtain an optimized meta-learning unit.

[0130] In a specific embodiment, the auxiliary learning unit optimization sub-module specifically includes:

[0131] A first feature extraction unit, configured to input the second image feature data into the auxiliary learning unit of the convolutional network model to obtain third image feature data.

[0132] A second feature extraction unit, configured to obtain fourth image feature data according to the third image feature data and the first pseudo fog-free image.

[0133] A second fog-free image generation unit, configured to determine a second pseudo fog-free image according to the fourth image feature data.

[0134] A second loss function calculation unit, configured to input the second fog-free image into a discriminator network of an auxiliary learning unit and calculate a loss function of the auxiliary learning unit.

[0135] A second optimization unit, configured to optimize the auxiliary learning unit according to the loss function of the auxiliary learning unit by using a backpropagation algorithm to obtain an optimized auxiliary learning unit.

[0136] The embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method part.

[0137] In this article, specific examples are used to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. An image dehazing method based on meta-assisted learning, characterized in that, Including: Construct a dataset; The dataset includes a synthetic fog image dataset and a natural fog image dataset; the synthetic fog image dataset includes a clear image and multiple synthetic fog images synthesized from the clear image; the natural fog image dataset includes multiple real foggy day images; Construct a convolutional network model; the convolutional network model includes an encoder, a decoder, a meta-learning unit, and an auxiliary learning unit; Use the convolutional network model to extract features from the synthetic fog images and the real foggy day images to obtain image feature data; the image feature data includes first image feature data and second image feature data; Use the image feature data and the clear image to optimize the convolutional network model to obtain an optimized convolutional network model; The step of using the image feature data and the clear image to optimize the convolutional network model to obtain an optimized convolutional network model specifically includes: Use the first image feature data and the clear image to optimize the meta-learning unit of the convolutional network model to obtain an optimized meta-learning unit; Use the second image feature data to optimize the auxiliary learning unit of the convolutional network model to obtain an optimized auxiliary learning unit; Determine the optimized convolutional network according to the optimized meta-learning unit and the optimized auxiliary learning unit; The step of using the first image feature data and the clear image to optimize the meta-learning unit of the convolutional network model to obtain an optimized meta-learning unit specifically includes: Input the first image feature data into the meta-learning unit of the convolutional network model to generate a first pseudo-fog-free image; Calculate the loss function of the meta-learning unit according to the clear image and the first pseudo-fog-free image; Optimize the meta-learning unit using the backpropagation algorithm according to the loss function of the meta-learning unit to obtain an optimized meta-learning unit; The step of using the second image feature data to optimize the auxiliary learning unit of the convolutional network model to obtain an optimized auxiliary learning unit specifically includes: Input the second image feature data into the auxiliary learning unit of the convolutional network model to obtain third image feature data; Obtain fourth image feature data according to the third image feature data and the first pseudo-fog-free image; Determine a second pseudo-fog-free image according to the fourth image feature data; Input the second pseudo-fog-free image into the discriminator network of the auxiliary learning unit to calculate the loss function of the auxiliary learning unit; Optimize the auxiliary learning unit using the backpropagation algorithm according to the loss function of the auxiliary learning unit to obtain an optimized auxiliary learning unit; Use the optimized convolutional network model to perform defogging processing on the foggy image to be measured.

2. The image dehazing method based on meta-assisted learning according to claim 1, characterized in that, Before the step of using the convolutional network model to extract features from the synthetic fog images and the real foggy day images to obtain image feature data, it further includes: Perform normalization processing on the synthetic fog images and the real foggy day images to obtain normalized synthetic fog images and normalized real foggy day images.

3. An image dehazing system based on meta-assisted learning, characterized in that, Including: A dataset construction module for constructing a dataset; the dataset includes a synthetic fog image dataset and a natural fog image dataset; the synthetic fog image dataset includes a clear image and multiple synthetic fog images artificially synthesized from the clear image; the natural fog image dataset includes multiple real foggy day images; A convolutional network model construction module for constructing a convolutional network model; the convolutional network model includes an encoder, a decoder, a meta-learning unit, and an auxiliary learning unit; A feature extraction module for extracting features from the synthetic fog images and the real foggy day images using the convolutional network model to obtain image feature data; the image feature data includes first image feature data and second image feature data; A model optimization module for optimizing the convolutional network model using the image feature data and the clear image to obtain an optimized convolutional network model; The model optimization module specifically includes: A meta-learning unit optimization sub-module for optimizing the meta-learning unit of the convolutional network model using the first image feature data and the clear image to obtain an optimized meta-learning unit; An auxiliary learning unit optimization sub-module for optimizing the auxiliary learning unit of the convolutional network model using the second image feature data to obtain an optimized auxiliary learning unit; A model determination sub-module for determining the optimized convolutional network model based on the optimized meta-learning unit and the optimized auxiliary learning unit; The meta-learning unit optimization sub-module specifically includes: A first fog-free image generation unit for inputting the first image feature data into the meta-learning unit of the convolutional network model to generate a first pseudo fog-free image; A first loss function calculation unit for calculating the loss function of the meta-learning unit based on the clear image and the first pseudo fog-free image; A first optimization unit for optimizing the meta-learning unit using the backpropagation algorithm based on the loss function of the meta-learning unit to obtain an optimized meta-learning unit; The auxiliary learning unit optimization sub-module specifically includes: A first feature extraction unit for inputting the second image feature data into the auxiliary learning unit of the convolutional network model to obtain third image feature data; A second feature extraction unit for obtaining fourth image feature data based on the third image feature data and the first pseudo fog-free image; A second fog-free image generation unit for determining a second pseudo fog-free image based on the fourth image feature data; A second loss function calculation unit for inputting the second pseudo fog-free image into the discriminator network of the auxiliary learning unit to calculate the loss function of the auxiliary learning unit; A second optimization unit for optimizing the auxiliary learning unit using the backpropagation algorithm based on the loss function of the auxiliary learning unit to obtain an optimized auxiliary learning unit; A defogging module for defogging a foggy image to be measured using the optimized convolutional network model.

4. The image dehazing system based on meta-assisted learning according to claim 3, wherein, It further includes: A normalization processing module for normalizing the synthetic fog images and the real foggy day images to obtain normalized synthetic fog images and normalized real foggy day images.

Images