A Super-Resolution Reconstruction Training Data Generation Method Based on Generalizable Degradation Representations

Abstract

The present invention discloses a method for generating super-resolution reconstruction training data based on generalizable degradation representations. The method for generating super-resolution reconstruction training data provided by the present invention learns degradation representations for unlabeled real low-resolution images with different degradation characteristics, and uses the degradation representations as conditional information to be input into a degradation network to generate low-resolution images with more consistent multiple degradation characteristics. Thus, low-resolution images-high-resolution images with consistent degradation characteristics can be used as data pairs to train a super-resolution network, obtaining a better image super-resolution reconstruction effect.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly to a method for generating super-resolution reconstruction training data based on generalizable degradation representations. Background Art

[0002] The training data of a super-resolution reconstruction model requires paired high-resolution images and low-resolution images. However, in reality, high-resolution images and low-resolution images are often not paired. In the prior art, a two-stage training method of a degradation network - super-resolution network is used to perform super-resolution reconstruction on real low-resolution images. Specifically, a degradation network is first trained to simulate the degradation characteristics of a real low-resolution image dataset. The trained degradation network is used to generate low-resolution images corresponding to high-resolution images, and then the low-resolution images generated by the degradation network and the corresponding high-resolution images are used for the training of the super-resolution network.

[0003] However, a significant defect of the prior art is that when the real low-resolution image dataset contains multiple degradation characteristics, due to the lack of input of degradation condition information, the degradation network cannot generate images with a degradation distribution consistent with the low-resolution image dataset, that is, there is a large domain gap between the generated low-resolution images and the real low-resolution images. Due to the existence of the domain gap, the super-resolution network trained based on the generated low-resolution images will not be able to generalize well to real low-resolution images, and the super-resolution reconstruction performance often drops sharply.

[0004] Therefore, the prior art still needs to be improved and enhanced. Summary of the Invention

[0005] In view of the above-mentioned defects of the prior art, the present invention provides a method for generating super-resolution reconstruction training data based on generalizable degradation representations, aiming to solve the problem that there is a domain gap between the generated super-resolution reconstruction training data and real data, resulting in poor super-resolution reconstruction effect of the super-resolution network obtained by training.

[0006] To solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0007] In a first aspect of the present invention, there is provided a method for generating super-resolution reconstruction training data based on generalizable degradation representations. The super-resolution reconstruction training data includes a sample high-resolution image and a sample low-resolution image corresponding to the sample high-resolution image. The method includes:

[0008] Perform different types of degradation processing on the first high-score image to obtain a synthesized low-score image, and input the synthesized low-score image and the first real low-score image into a degradation representation learning network. The degradation representation learning network includes a feature extraction network and a classifier. The feature extraction network is used to extract the degradation representation corresponding to the synthesized low-score image and the degradation representation corresponding to the first real low-score image. The classifier is used to output the classification result of the synthesized low-score image and the real low-score image based on the degradation representation corresponding to the input synthesized low-score image and the degradation representation corresponding to the first real low-score image;

[0009] Update the parameters of the feature extraction network and the classifier based on the true Gaussian distribution, the classification results corresponding to the synthesized low-score image and the first real low-score image, and the degradation representation until the parameters converge, so that the degradation representation output by the feature extraction network after the parameters converge follows a Gaussian distribution;

[0010] Sample in the Gaussian distribution followed by the degradation representation output by the feature extraction network after the parameters converge to obtain a target sampled degradation representation, and input the target sampled degradation representation and the sample high-score image into a trained degradation network to obtain the sample low-score image corresponding to the sample high-score image output by the degradation network.

[0011] The method for generating super-resolution reconstruction training data based on generalizable degradation representation, wherein sampling in the Gaussian distribution followed by the degradation representation output by the feature extraction network to obtain a target sampled degradation representation includes:

[0012] Sample in the Gaussian distribution followed by the degradation representation output by the feature extraction network to obtain an initial sampled degradation representation;

[0013] Input the initial sampled degradation representation into the classifier to obtain the classification result and confidence level output by the classifier;

[0014] When the confidence level output by the classifier is higher than a preset threshold, use the initial sampled degradation representation as the target sampled degradation representation.

[0015] The method for generating super-resolution reconstruction training data based on generalizable degradation representation, wherein the degradation network includes at least one degradation attention module and at least one feature extraction module, and each degradation attention module includes a convolution kernel reconstruction layer and an attention weight layer; obtaining the sample low-score image corresponding to the sample high-score image output by the degradation network includes:

[0016] In each degradation attention module, perform the following operations:

[0017] Use the target sampling degradation representation as the first input data of the degradation attention module, and use the output of the previous module in the degradation network as the second input data of the degradation attention module. The second input data of the first degradation attention module in the degradation network is the output data after the sample high-resolution image passes through a feature extraction module;

[0018] Input the first input data into the convolution kernel reconstruction layer to obtain a target convolution kernel, and perform convolution on the second input data according to the target convolution kernel to obtain first output data;

[0019] Input the first input data into the attention weight layer to obtain a target weight, and perform weighting on the second input data according to the target weight to obtain second output data;

[0020] Obtain the output data of the degradation attention module based on the first output data and the second output data.

[0021] The method for generating super-resolution reconstruction training data based on a generalizable degradation representation, wherein the training process of the degradation network training data is as follows:

[0022] Select a second real low-resolution image;

[0023] Input the second real low-resolution image into the feature extraction network after the parameters converge to obtain the degradation representation corresponding to the second real low-resolution image;

[0024] Input the second real low-resolution image and the degradation representation corresponding to the second real low-resolution image into the degradation network to train the degradation network and update the parameters of the degradation network;

[0025] Repeat the step of selecting the second real low-resolution image until the parameters of the degradation network converge.

[0026] The method for generating super-resolution reconstruction training data based on a generalizable degradation representation, wherein updating the parameters of the feature extraction network and the classifier until the parameters converge based on the true Gaussian distribution, the classification results corresponding to the synthetic low-resolution image and the first real low-resolution image, and the degradation representation, so that the degradation representation output by the feature extraction network follows a Gaussian distribution, includes:

[0027] Select a target training batch, where the target training batch includes multiple first real low-resolution images and multiple synthetic low-resolution images, and input the target training batch into the feature extraction network to extract the corresponding degradation representation and the corresponding classification result;

[0028] Generate a first adversarial loss according to the classification result corresponding to the first real low-resolution image;

[0029] Generate a first classification loss according to the classification result corresponding to the synthesized low-resolution image and the classification label corresponding to the synthesized low-resolution image;

[0030] Generate a second adversarial loss according to the degradation representations corresponding to the synthesized low-resolution image and the first real low-resolution image respectively and the true Gaussian distribution;

[0031] Update the parameters of the feature extraction network and the classifier according to the first adversarial loss, the first classification loss and the second adversarial loss;

[0032] Re-execute the step of selecting the target training batch until the parameters of the feature extraction network and the classifier converge.

[0033] The method for generating super-resolution reconstruction training data based on generalizable degradation representations, wherein generating a first adversarial loss according to the classification result corresponding to the first real low-resolution image includes:

[0034] Obtain the first adversarial loss according to the difference between the classification result corresponding to the first real low-resolution image and the preset classification space of the real low-resolution image.

[0035] The method for generating super-resolution reconstruction training data based on generalizable degradation representations, wherein generating a second adversarial loss according to the degradation representations corresponding to the synthesized low-resolution image and the first real low-resolution image respectively and the true Gaussian distribution includes:

[0036] Sample in the degradation representations corresponding to the synthesized low-resolution image and the first real low-resolution image respectively to obtain a first sampling result;

[0037] Sample in the true Gaussian distribution to obtain a second sampling result;

[0038] Obtain the second adversarial loss according to the first sampling result and the second sampling result.

[0039] In a second aspect of the present invention, there is provided a device for generating super-resolution reconstruction training data based on generalizable degradation representations, wherein the super-resolution reconstruction training data includes a sample high-resolution image and a sample low-resolution image corresponding to the sample high-resolution image, and the device includes:

[0040] A degradation representation learning module, which includes a model calculation unit and a training unit;

[0041] Among them, the model calculation unit is used to perform different types of degradation processing on the first high-score image to obtain a synthesized low-score image, and input the synthesized low-score image and the first real low-score image into the degradation characterization learning machine. The degradation characterization learning machine includes a feature extraction network and a classifier. The feature extraction network is used to extract the degradation characterization corresponding to the synthesized low-score image and the degradation characterization corresponding to the first real low-score image. The classifier is used to output the classification results of the synthesized low-score image and the real low-score image based on the input degradation characterization corresponding to the synthesized low-score image and the degradation characterization corresponding to the first real low-score image;

[0042] The training unit is used to update the parameters of the feature extraction network and the classifier based on the true Gaussian distribution, the classification results corresponding to the synthesized low-score image and the first real low-score image, and the degradation characterization until the parameters converge, so that the degradation characterization output by the feature extraction network after the parameters converge follows the Gaussian distribution;

[0043] A data generation module, which is used to sample in the Gaussian distribution followed by the degradation characterization output by the feature extraction network after the parameters converge to obtain a target sampled degradation characterization, and input the target sampled degradation characterization and the sample high-score image into the trained degradation network to obtain the sample low-score image corresponding to the sample high-score image output by the degradation network.

[0044] In the third aspect of the present invention, a terminal is provided. The terminal includes a processor and a computer-readable storage medium communicatively connected to the processor. The computer-readable storage medium is adapted to store a plurality of instructions, and the processor is adapted to call the instructions in the computer-readable storage medium to execute the steps of implementing the super-resolution reconstruction training data generation method based on generalizable degradation characterization described in any one of the above.

[0045] In the fourth aspect of the present invention, a computer-readable storage medium is provided. The computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the super-resolution reconstruction training data generation method based on generalizable degradation characterization described in any one of the above.

[0046] Compared with the prior art, the present invention provides a method for generating super-resolution reconstruction training data based on generalizable degradation representations. In the method for generating super-resolution reconstruction training data based on generalizable degradation representations provided by the present invention, different synthetic low-resolution images are generated and input together with real low-resolution images into a feature extraction network for extracting degradation representations, and classification is performed based on the degradation representations. The feature extraction network and the classifier are trained through the classification results, such that the feature space of the degradation representations output by the feature extraction network follows a Gaussian distribution. In this way, the degradation representations of unlabeled real low-resolution images with different degradation characteristics can be learned. Inputting the degradation representations into a degradation network can generate low-resolution images with various degradation characteristics consistent with actual low-resolution images as training data to train a super-resolution network, thereby obtaining a better image super-resolution reconstruction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flowchart of an embodiment of the method for generating super-resolution reconstruction training data based on generalizable degradation representations provided by the present invention;

[0048] Figure 2 is a schematic diagram of the network framework of an existing super-resolution reconstruction method;

[0049] Figure 3 is a schematic diagram of the network framework of the method for generating super-resolution reconstruction training data based on generalizable degradation representations provided by the present invention;

[0050] Figure 4 is a schematic diagram of the model of the degradation representation learner in the method for generating super-resolution reconstruction training data based on generalizable degradation representations provided by the present invention;

[0051] FIG. 5(a) is a schematic diagram of the model of the degradation network in the method for generating super-resolution reconstruction training data based on generalizable degradation representations provided by the present invention;

[0052] FIG. 5(b) is a schematic diagram of the residual group in the degradation network in the method for generating super-resolution reconstruction training data based on generalizable degradation representations provided by the present invention;

[0053] FIG. 5(c) is a schematic diagram of the degradation attention module in the model of the degradation network in the method for generating super-resolution reconstruction training data based on generalizable degradation representations provided by the present invention;

[0054] Figure 6 is the effect verification of the method for generating super-resolution reconstruction training data based on generalizable degradation representations provided by the present invention Figure 1 ;

[0055] Figure 7 is the effect verification of the method for generating super-resolution reconstruction training data based on generalizable degradation representations provided by the present invention Figure 2 ;

[0056] Figure 8 Effect verification of the super-resolution reconstruction training data generation method based on generalizable degradation representation provided by the present invention Figure 3 ;

[0057] Figure 9 Effect verification of the super-resolution reconstruction training data generation method based on generalizable degradation representation provided by the present invention Figure 4 ;

[0058] Figure 10 Effect verification diagram five of the super-resolution reconstruction training data generation method based on generalizable degradation representation provided by the present invention;

[0059] Figure 11 Effect verification of the super-resolution reconstruction training data generation method based on generalizable degradation representation provided by the present invention Figure 6 ;

[0060] Figure 12 Effect verification of the super-resolution reconstruction training data generation method based on generalizable degradation representation provided by the present invention Figure 7 ;

[0061] Figure 13 Effect verification of the super-resolution reconstruction training data generation method based on generalizable degradation representation provided by the present invention Figure 8 ;

[0062] Figure 14 Effect verification of the super-resolution reconstruction training data generation method based on generalizable degradation representation provided by the present invention Figure 9 ;

[0063] Figure 15 Effect verification of the super-resolution reconstruction training data generation method based on generalizable degradation representation provided by the present invention Figure 10 ;

[0064] Figure 16 Structural schematic diagram of an embodiment of the super-resolution reconstruction training data generation device based on generalizable degradation representation provided by the present invention;

[0065] Figure 17 Schematic diagram of the principle of an embodiment of the terminal provided by the present invention. Detailed implementation manners

[0066] To make the objectives, technical solutions and effects of the present invention clearer and more definite, the following further describes the present invention in detail with reference to the accompanying drawings and by way of examples. It should be understood that the specific examples described herein are only used to explain the present invention and are not used to limit the present invention.

[0067] The method for generating super-resolution reconstruction training data based on generalizable degradation representations provided by the present invention can be applied to terminals with computing capabilities. The terminal can execute the method for generating super-resolution reconstruction training data based on generalizable degradation representations provided by the present invention to generate training data for the super-resolution network. The terminal can be, but is not limited to, various computers, mobile terminals, smart home appliances, wearable devices, etc.

[0068] Embodiment 1

[0069] As Figure 1 shown, in an embodiment of the method for generating super-resolution reconstruction training data based on generalizable degradation representations, the method includes the steps:

[0070] S100. Perform different types of degradation processing on the first high-resolution image to obtain a synthesized low-resolution image, and input the synthesized low-resolution image and the first real low-resolution image into the degradation representation learner.

[0071] In the real world, the high-resolution images and low-resolution images in the dataset that can be used for super-resolution model training are often unpaired. In the prior art, as Figure 2 shown, super-resolution reconstruction is performed on real low-resolution images through a two-stage training method of a degradation network - super-resolution network. Among them, the degradation network aims to simulate the degradation characteristics of the low-resolution image dataset. However, when the low-resolution images in the dataset contain multiple degradation characteristics, due to the lack of input of degradation condition information in the prior art, the degradation network cannot generate images with degradation characteristics consistent with the real low-resolution images in the dataset, resulting in a large domain gap between the generated low-resolution images and the real low-resolution images, affecting the super-resolution reconstruction performance. To solve this problem, the method provided in this embodiment, as Figure 3 shown, designs a generalizable degradation representation learner to learn degradation representations for unlabeled real low-resolution images with different degradation characteristics, and uses these degradation representations as conditional information to be input into the degradation network to generate low-resolution images with more consistent multiple degradation distributions. Thus, it can effectively use the low-resolution image - high-resolution image pairs with consistent degradation distributions to train the super-resolution network and obtain better image super-resolution reconstruction effects.

[0072] The first high-resolution image is a real high-resolution image in an existing dataset, and the first real low-resolution image is a real low-resolution image in an existing dataset. Perform multiple degradation processes on the first high-resolution image to obtain a synthesized low-resolution image, and input the synthesized low-resolution image and the first real low-resolution image into the degradation representation learner to train the degradation representation learner, so that the degradation representation learner can learn the degradation representations corresponding to multiple types of degradation processes, and these degradation representations are consistent with the degradation representations of the real low-resolution images.

[0073] Specifically, the degradation characterization learner includes a feature extraction network and a classifier. The feature extraction network is used to extract the degradation characterization corresponding to the synthesized low-resolution image and the degradation characterization corresponding to the first real low-resolution image. The classifier is used to output the classification results of the synthesized low-resolution image and the real low-resolution image based on the input degradation characterization corresponding to the synthesized low-resolution image and the degradation characterization corresponding to the first real low-resolution image.

[0074] In order to be able to learn the degradation characterizations corresponding to multiple different degradation characteristics, in this embodiment, it is required that the feature space of the degradation characterizations extracted by the degradation characterization learner follows a Gaussian distribution. In this way, multiple corresponding degradation characterizations can be sampled from this distribution to generate diversified degraded images. At the same time, the degradation characterizations extracted by the degradation characterization learner also need to be able to accurately reflect different degradation types, and the extracted degradation characterizations can also reflect the degradation characteristics of the real low-resolution images. To achieve this goal, the method provided in this embodiment updates the parameters of the feature extraction network and the classifier based on the real Gaussian distribution, the classification results corresponding to the synthesized low-resolution image and the first real low-resolution image, and the degradation characterizations. Specifically, it includes the steps:

[0075] S200. Update the parameters of the feature extraction network and the classifier based on the real Gaussian distribution, the classification results corresponding to the synthesized low-resolution image and the first real low-resolution image, and the degradation characterizations until the parameters converge, so that the degradation characterizations output by the feature extraction network after the parameters converge follow a Gaussian distribution.

[0076] The updating of the parameters of the feature extraction network and the classifier based on the real Gaussian distribution, the classification results corresponding to the synthesized low-resolution image and the first real low-resolution image, and the degradation characterizations until the parameters converge, so that the degradation characterizations output by the feature extraction network follow a Gaussian distribution, includes:

[0077] Select a target training batch, where the target training batch includes multiple first real low-resolution images and multiple synthesized low-resolution images, and input the target training batch into the feature extraction network to extract the corresponding degradation characterizations and corresponding classification results;

[0078] Generate a first adversarial loss according to the classification result corresponding to the first real low-resolution image;

[0079] Generate a first classification loss according to the classification result corresponding to the synthesized low-resolution image and the classification label corresponding to the synthesized low-resolution image;

[0080] Generate a second adversarial loss according to the degradation characterizations corresponding to the synthesized low-resolution image and the first real low-resolution image respectively and the real Gaussian distribution;

[0081] Update the parameters of the feature extraction network and the classifier according to the first adversarial loss, the first sub-loss, and the second adversarial loss;

[0082] Re-execute the step of selecting the target training batch, except that the parameters of the feature extraction network and the classifier converge.

[0083] The synthesized low-resolution image is generated by using different types of degradation processing. That is to say, the synthesized low-resolution image has a label of the degradation type, but the specific degradation type of the first real low-resolution image is unknown and there is no degradation type label. In order to enable the degradation feature extractor extracted after the training to accurately reflect the characteristics of various degradation types (including theoretical degradation types and real degradation types), and the feature space of the degradation representation can follow a Gaussian distribution. In this embodiment, when training the degradation feature extractor, as Figure 4 shown, three types of losses are calculated. Each type of loss is described in detail below.

[0084] The generation of the first adversarial loss according to the classification result corresponding to the first real low-resolution image includes:

[0085] Obtain the first adversarial loss according to the difference between the classification result corresponding to the first real low-resolution image and the preset classification space of the real low-resolution image.

[0086] Specifically, in this embodiment, the classifier classifies the real low-resolution image and the synthesized low-resolution image into different classification spaces. That is to say, when the degradation representation of the synthesized low-resolution image is input to the classifier, the classification result output by the classifier is input to the first classification space, for example, classified into the one-hot space of the first 6 (0-5) categories, and when the degradation representation of the first real low-resolution image is input to the classifier, the classification result output by the classifier belongs to the second classification space, for example, classified into the one-hot space of 2 (6-7) categories different from the first classification space. Subsequently, an adversarial method is used to obtain the loss. Specifically, samples are taken in the one-hot space different from the first classification space as positive samples, and the classification result of the real low-resolution image output by the classifier is used as a negative sample to obtain the first adversarial loss, so that the degradation representations of the real low-resolution image and the synthesized low-resolution image are closer. That is to say, even for the synthesized low-resolution image, the feature extraction network can extract a degradation representation consistent with the degradation characteristics of the real low-resolution image, so that the feature extraction network can learn the degradation characteristics of the real low-resolution image. Specifically, when the first real low-resolution image is input to the degradation feature extractor and is not classified into the preset feature space of the real low-resolution image, an adversarial loss will be generated.

[0087] Generating a first sub-loss according to the classification result corresponding to the synthesized low-resolution image and the classification label corresponding to the synthesized low-resolution image includes:

[0088] Obtaining the first sub-loss according to the difference between the classification result corresponding to the synthesized low-resolution image and the classification label corresponding to the synthesized low-resolution image.

[0089] Updating the parameters of the degradation representation learning network according to the first sub-loss can enable the degradation representation extracted by the degradation representation learning network to accurately reflect different degradation types.

[0090] Generating a second adversarial loss according to the degradation representations corresponding to the synthesized low-resolution image and the first real low-resolution image and the real Gaussian distribution includes:

[0091] Sampling in the degradation representations corresponding to the synthesized low-resolution image and the first real low-resolution image respectively to obtain a first sampling result;

[0092] Sampling in the real Gaussian distribution to obtain a second sampling result;

[0093] Obtaining the second adversarial loss according to the first sampling result and the second sampling result.

[0094] Each time, a batch including multiple images is input into the degradation representation learning network for training. Sampling is performed in the degradation representations corresponding to all the images in a batch to obtain a first sampling result. The second adversarial loss is generated according to the first sampling result and the second sampling result obtained by sampling in the real Gaussian distribution, and then the parameters of the degradation representation learning network are updated according to the second adversarial loss. In this way, the degradation representation extracted by the degradation representation learning network will ultimately follow a Gaussian distribution.

[0095] After the degradation representation learning network is trained, the degradation representation is obtained according to the degradation representation learning network and used as part of the input of the degradation network, so that a lower-resolution image closer to the real lower-resolution image can be generated when generating a lower-resolution image.

[0096] As Figure 1 shown, the method provided in this embodiment further includes the step of:

[0097] S300. Sampling in the Gaussian distribution followed by the degradation representation output by the feature extraction network after the parameters converge to obtain a target sampled degradation representation. Input the target sampled degradation representation and the sample high-resolution image into the trained degradation network, and obtain the sample low-resolution image corresponding to the sample high-resolution image output by the degradation network.

[0098] Specifically, the degradation network includes at least one degradation attention module and at least one feature extraction module. Each degradation attention module includes a convolution kernel reconstruction layer and an attention weight layer; obtaining the sample low-resolution image corresponding to the sample high-resolution image output by the degradation network includes:

[0099] In each degradation attention module, perform the following operations:

[0100] Use the target sampling degradation representation as the first input data of the degradation attention module, and use the output of the previous module in the degradation network as the second input data of the degradation attention module. The second input data of the first degradation attention module in the degradation network is the output data after the sample high-resolution image passes through one feature extraction module;

[0101] Input the first input data into the convolution kernel reconstruction layer to obtain a target convolution kernel, and perform convolution on the second input data according to the target convolution kernel to obtain first output data;

[0102] Input the first input data into the attention weight layer to obtain a target weight, and perform weighting on the second input data according to the target weight to obtain second output data;

[0103] Based on the first output data and the second output data, obtain the output data of the degradation attention module.

[0104] As shown in FIGS. 5(a), 5(b), and 5(c), the degradation network includes at least one convolutional layer (Conv) and at least one residual group. Each residual group includes at least one degradation attention module (DAblock). The role of the degradation attention module is to integrate degradation condition information (degradation representation) into the degradation network through dynamic convolution. Specifically, on the one hand, the degradation representation passes through the convolution kernel reconstruction layer to reconstruct a convolution kernel, that is, the target convolution kernel, and performs depthwise separable convolution operation on the image features according to the target convolution kernel. On the other hand, the degradation representation passes through the attention weight layer to calculate the channel attention weight of the image features and weight the image features. Finally, the features output by the two aspects are added to obtain the output data of the degradation attention module. The convolution kernel reconstruction layer and the attention weight layer can both be a fully connected layer.

[0105] The network parameters in the degradation network are optimized and determined through training. The following is an explanation of the training process of the degradation network:

[0106] The training process of the training data of the degradation network is as follows:

[0107] Select the second real low-resolution image;

[0108] Input the second real low-resolution image into the feature extraction network after parameter convergence to obtain the degradation representation corresponding to the second real low-resolution image;

[0109] Input the second real low-resolution image and the degradation representation corresponding to the second real low-resolution image into the degradation network to train the degradation network and update the parameters of the degradation network;

[0110] Repeat the step of selecting the second real low-resolution image until the parameters of the degradation network converge.

[0111] In a possible implementation manner, during the training process of the degradation network, a high-resolution image is input and a low-resolution image is output. However, in this embodiment, in order to enable the degradation network to retain the features of the low-resolution image, a method similar to image reconstruction is used to train the degradation network, and the parameters of the degradation network are optimized by inputting a real low-resolution image and outputting a low-resolution image as well. Specifically, during the training process, each time a second real low-resolution image is selected. The second real low-resolution image may be the same as or different from the first real low-resolution image. Input the second real low-resolution image into the feature extraction network in the trained degradation representation learning device to obtain the degradation representation of the second real low-resolution image output by the feature extraction network. Input the degradation representation of the second real low-resolution image and the second real low-resolution image into the degradation network. After being processed by the degradation network, obtain the image output by the degradation network. Calculate the training loss based on the image output by the degradation network to update the parameters of the degradation network, complete one training, repeat the training multiple times until the parameters of the degradation network converge, and the training of the degradation network is completed.

[0112] After the degradation network is trained, diverse low-resolution images are generated based on the degradation network as super-resolution reconstruction training data. Specifically, sampling is performed on the Gaussian distribution followed by the degradation representation output by the feature extraction network after training to obtain a target sampled degradation feature. The target sampled degradation feature and the sample high-resolution image are input into the trained degradation network, and the image output by the degradation network is used as the sample low-resolution image corresponding to the sample high-resolution image, thus obtaining a training pair in the super-resolution reconstruction training data. By randomly sampling on the Gaussian distribution followed by the degradation representation output by the feature extraction network after training, low-resolution images with various degradation characteristics can be generated, producing richer super-resolution reconstruction training data. To further reduce the domain gap between the generated sample low-resolution images and real low-resolution images, in this embodiment, the sampled degradation representation is verified after sampling on the Gaussian distribution. Specifically, sampling in the Gaussian distribution followed by the degradation representation output by the feature extraction network to obtain a target sampled degradation representation includes:

[0113] Sampling in the Gaussian distribution followed by the degradation representation output by the feature extraction network to obtain an initial sampled degradation representation;

[0114] Inputting the initial sampled degradation representation into the classifier to obtain the classification result and confidence level output by the classifier;

[0115] When the confidence level output by the classifier is higher than the preset threshold, the initial sampled degradation representation is used as the target sampled degradation representation.

[0116] When the classifier outputs a classification result, it also outputs the confidence level of the classification result. Only when the confidence level output by the classifier is higher than the preset threshold, the sampled degradation representation is used to generate the sample low-resolution image.

[0117] After obtaining the sample high-resolution image and the corresponding sample low-resolution image, they can be used to train the super-resolution reconstruction network. The super-resolution reconstruction network can select ESRGAN, which is mainly composed of convolutional layers and secret residual blocks, and is trained based on relative GAN, capable of generating images with realistic details and reducing network overfitting to specific degradations.

[0118] The method provided in this embodiment significantly improves the quality of the generated low-score sample images in more complex scenarios. Therefore, the super-resolution reconstruction model trained based on the low-score sample images can also achieve good generalization ability on real low-score images. Experimental verification shows that, whether in qualitative or quantitative analysis and comparison, the method provided in this embodiment achieves better super-resolution reconstruction results than the prior art. In the experiment, six low-resolution images with blur and noise degradation with different parameters were synthesized as the base class to learn their degradation characteristics. To better verify the generalization ability of the degradation characteristics, two experiments were set up in the new category, namely the synthetic low-resolution image dataset and the real low-resolution image dataset. Both sets of image data contain two degradation distributions. Among them, the synthetic low-resolution image dataset contains two degradation distributions, namely the camera sensor noise simulated by the algorithm and JPEG compression. The real low-resolution image dataset uses the image datasets of the 2019 AIM real image super-resolution reconstruction competition and the 2020 NTIRE real image super-resolution reconstruction track competition. Among them, the low-resolution image dataset of 2019 AIM contains compression and blur degradation, and the low-resolution image dataset of 2020 NTIRE contains high-frequency and inconsistent noise.

[0119] Quantitative analysis: Figure 6 and Figure 7 show the performance comparison results between the method provided in this embodiment (the "ours" column in the figure) and other technologies of the same type on two standard test data. Figure 6 are the performance comparison results of the generated low-score images. Figure 7 are the performance comparison results of the high-score images obtained by super-resolution reconstruction. In the experiment, quantitative comparisons were made on multiple standard evaluation metrics (Degradation GAN, FSSR, DASR, Impressionism). The method provided in this embodiment leads significantly in three of these metrics.

[0120] Qualitative analysis: Figures 8 - 11 shows that the method provided in this embodiment achieves better performance in the generated low-score images compared with other technologies of the same type. Figures 12 - 15 shows that the method provided in this embodiment achieves better performance in the super-resolution reconstruction images compared with other technologies of the same type. Specifically, Figures 8 - 9 shows the qualitative performance comparison results between the method provided in this embodiment and the synthetic low-score images generated by the image degradation network in other technologies of the same type. Figures 10 - 11 shows the qualitative performance comparison results between the method provided in this embodiment and the real low-score images generated by the image degradation network in other technologies of the same type. Figures 12 - 13 shows the qualitative performance comparison results between the method provided in this embodiment and the super-resolution reconstruction results of the synthetic low-score images in other technologies of the same type.Figures 14 - 15 It shows the qualitative performance comparison results of the method provided in this embodiment and other technologies of the same type for the super-resolution reconstruction of real low-resolution images.

[0121] In summary, this embodiment provides a method for generating super-resolution reconstruction training data based on generalizable degradation representation. By generating different synthetic low-resolution images and inputting them together with real low-resolution images into a feature extraction network for extracting degradation representation, and classifying based on the degradation representation, the feature extraction network and the classifier are trained through the classification results, so that the feature space of the degradation representation output by the feature extraction network follows a Gaussian distribution. In this way, the degradation representations of unlabeled real low-resolution images with different degradation characteristics can be learned. Inputting this degradation representation into a degradation network can generate low-resolution images with degradation characteristics consistent with actual low-resolution images as training data to train a super-resolution network, obtaining a better image super-resolution reconstruction effect.

[0122] It should be understood that although the steps in the flowcharts given in the accompanying drawings of the present invention are shown in sequence according to the indication of the arrows, these steps do not necessarily have to be executed in the order indicated by the arrows. Unless there is a clear indication in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages do not necessarily have to be executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages does not necessarily have to be sequential, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0123] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. 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 methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided by the present invention can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0124] Embodiment 2

[0125] Based on the above embodiments, the present invention also correspondingly provides a super-resolution reconstruction training data generation device based on a generalizable degradation representation, as Figure 16 shown. The super-resolution reconstruction training data generation device based on a generalizable degradation representation includes:

[0126] A degradation representation learning module, which includes a model calculation unit and a training unit;

[0127] Among them, the model calculation unit is used to perform different types of degradation processing on a first high-resolution image to obtain a synthesized low-resolution image, and input the synthesized low-resolution image and the first real low-resolution image into a degradation representation learner. The degradation representation learner includes a feature extraction network and a classifier. The feature extraction network is used to extract the degradation representation corresponding to the synthesized low-resolution image and the degradation representation corresponding to the first real low-resolution image. The classifier is used to output the classification result of the synthesized low-resolution image and the real low-resolution image based on the input degradation representation corresponding to the synthesized low-resolution image and the degradation representation corresponding to the first real low-resolution image, specifically as described in Embodiment 1;

[0128] The training unit is used to update the parameters of the feature extraction network and the classifier based on the true Gaussian distribution, the classification results corresponding to the synthesized low-resolution image and the first true low-resolution image, and the degradation representation until the parameters converge, so that the degradation representation output by the feature extraction network after the parameters converge follows a Gaussian distribution, as specifically described in Embodiment 1;

[0129] A data generation module, which is used to sample in the Gaussian distribution followed by the degradation representation output by the feature extraction network after the parameters converge to obtain a target sampled degradation representation, and input the target sampled degradation representation and the sample high-resolution image into the trained degradation network to obtain the sample low-resolution image corresponding to the sample high-resolution image output by the degradation network, as specifically described in Embodiment 1.

[0130] Embodiment 3

[0131] Based on the above embodiments, the present invention also correspondingly provides a terminal, as Figure 17 shown, the terminal includes a processor 10 and a memory 20. Figure 17 Only some components of the terminal are shown, but it should be understood that it is not required to implement all the shown components, and more or fewer components can be alternatively implemented.

[0132] The memory 20 may be an internal storage unit of the terminal in some embodiments, such as the hard disk or memory of the terminal. The memory 20 may also be an external storage device of the terminal in other embodiments, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. equipped on the terminal. Further, the memory 20 may also include both the internal storage unit and the external storage device of the terminal. The memory 20 is used to store application software installed on the terminal and various types of data. The memory 20 may also be used to temporarily store data that has been output or will be output. In one embodiment, a super-resolution reconstruction training data generation program 30 based on a generalizable degradation representation is stored on the memory 20, and the super-resolution reconstruction training data generation program 30 based on a generalizable degradation representation can be executed by the processor 10, thereby implementing the super-resolution reconstruction training data generation method based on a generalizable degradation representation in the present application.

[0133] The processor 10 may be a central processing unit (CPU), a microprocessor or other chips in some embodiments, and is used to run the program code stored in the memory 20 or process data, such as executing the above-mentioned method for target-based multi-modal named entity recognition, etc.

[0134] In one embodiment, when the processor 10 executes the super-resolution reconstruction training data generation program 30 based on the generalizable degradation representation in the memory 20, the following steps are implemented:

[0135] Perform different types of degradation processing on the first high-resolution image to obtain a synthetic low-resolution image, and input the synthetic low-resolution image and the first real low-resolution image into a degradation representation learning machine. The degradation representation learning machine includes a feature extraction network and a classifier. The feature extraction network is used to extract the degradation representation corresponding to the synthetic low-resolution image and the degradation representation corresponding to the first real low-resolution image. The classifier is used to output the classification results of the synthetic low-resolution image and the real low-resolution image based on the input degradation representation corresponding to the synthetic low-resolution image and the degradation representation corresponding to the first real low-resolution image;

[0136] Update the parameters of the feature extraction network and the classifier based on the true Gaussian distribution, the classification results corresponding to the synthetic low-resolution image and the first real low-resolution image, and the degradation representation until the parameters converge, so that the degradation representation output by the feature extraction network after the parameters converge follows the Gaussian distribution;

[0137] Sample in the Gaussian distribution followed by the degradation representation output by the feature extraction network after the parameters converge to obtain a target sampled degradation representation, and input the target sampled degradation representation and the sample high-resolution image into a trained degradation network to obtain the sample low-resolution image corresponding to the sample high-resolution image output by the degradation network.

[0138] Among them, the sampling in the Gaussian distribution followed by the degradation representation output by the feature extraction network to obtain a target sampled degradation representation includes:

[0139] Sample in the Gaussian distribution followed by the degradation representation output by the feature extraction network to obtain an initial sampled degradation representation;

[0140] Input the initial sampled degradation representation into the classifier to obtain the classification result and confidence level output by the classifier;

[0141] When the confidence level output by the classifier is higher than a preset threshold, use the initial sampled degradation representation as the target sampled degradation representation.

[0142] Among them, the degradation network includes at least one degradation attention module and at least one feature extraction module. Each degradation attention module includes a convolution kernel reconstruction layer and an attention weight layer; the obtaining of the sample low-resolution image corresponding to the sample high-resolution image output by the degradation network includes:

[0143] In each degradation attention module, perform the following operations:

[0144] Use the target sampling degradation representation as the first input data of the degradation attention module, and use the output of the previous module in the degradation network as the second input data of the degradation attention module. The second input data of the first degradation attention module in the degradation network is the output data after the sample high-resolution image passes through a feature extraction module;

[0145] Input the first input data into the convolution kernel reconstruction layer to obtain a target convolution kernel, and perform convolution on the second input data according to the target convolution kernel to obtain first output data;

[0146] Input the first input data into the attention weight layer to obtain a target weight, and perform weighting on the second input data according to the target weight to obtain second output data;

[0147] Obtain the output data of the degradation attention module based on the first output data and the second output data.

[0148] Among them, the training process of the training data of the degradation network is as follows:

[0149] Select a second real low-resolution image;

[0150] Input the second real low-resolution image into the feature extraction network after the parameters converge, and obtain the degradation representation corresponding to the second real low-resolution image;

[0151] Input the second real low-resolution image and the degradation representation corresponding to the second real low-resolution image into the degradation network to train the degradation network and update the parameters of the degradation network;

[0152] Repeat the step of selecting the second real low-resolution image until the parameters of the degradation network converge.

[0153] Among them, updating the parameters of the feature extraction network and the classifier based on the classification results, degradation representations corresponding to the true Gaussian distribution, the synthetic low-resolution image, and the first real low-resolution image until the parameters converge, so that the degradation representation output by the feature extraction network follows a Gaussian distribution, includes:

[0154] Select a target training batch, where the target training batch includes multiple first real low-resolution images and multiple synthetic low-resolution images, and input the target training batch into the feature extraction network to extract corresponding degradation representations and corresponding classification results;

[0155] Generate a first adversarial loss according to the classification result corresponding to the first real low-resolution image;

[0156] Generate a first sub-loss based on the classification result corresponding to the synthesized low-resolution image and the classification label corresponding to the synthesized low-resolution image;

[0157] Generate a second adversarial loss based on the degradation representations corresponding to the synthesized low-resolution image and the first real low-resolution image respectively and the real Gaussian distribution;

[0158] Update the parameters of the feature extraction network and the classifier according to the first adversarial loss, the first sub-loss, and the second adversarial loss;

[0159] Re-execute the step of selecting the target training batch until the parameters of the feature extraction network and the classifier converge.

[0160] Among them, generating the first adversarial loss according to the classification result corresponding to the first real low-resolution image includes:

[0161] Obtain the first adversarial loss according to the difference between the classification result corresponding to the first real low-resolution image and the preset classification space of the real low-resolution image.

[0162] Among them, generating the second adversarial loss according to the degradation representations corresponding to the synthesized low-resolution image and the first real low-resolution image respectively and the real Gaussian distribution includes:

[0163] Sample in the degradation representations corresponding to the synthesized low-resolution image and the first real low-resolution image respectively to obtain a first sampling result;

[0164] Sample in the real Gaussian distribution to obtain a second sampling result;

[0165] Obtain the second adversarial loss according to the first sampling result and the second sampling result.

[0166] Example 4

[0167] The present invention also provides a computer-readable storage medium, in which one or more programs are stored, and the one or more programs can be executed by one or more processors to implement the steps of the method for generating super-resolution reconstruction training data based on generalizable degradation representations as described above.

[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating super-resolution reconstruction training data based on generalizable degradation representations, where the super-resolution reconstruction training data includes sample high-resolution images and corresponding sample low-resolution images of the sample high-resolution images, characterized in that, the method includes: Performing different types of degradation processing on a first high-resolution image to obtain a synthetic low-resolution image, and inputting the synthetic low-resolution image and a first real low-resolution image into a degradation representation learning device. The degradation representation learning device includes a feature extraction network and a classifier. The feature extraction network is used to extract the degradation representation corresponding to the synthetic low-resolution image and the degradation representation corresponding to the first real low-resolution image, and the classifier is used to output a classification result of the synthetic low-resolution image and the real low-resolution image based on the input degradation representation corresponding to the synthetic low-resolution image and the degradation representation corresponding to the first real low-resolution image; Updating the parameters of the feature extraction network and the classifier based on a true Gaussian distribution, the classification results corresponding to the synthetic low-resolution image and the first real low-resolution image, and the degradation representations until the parameters converge, so that the degradation representations output by the feature extraction network after the parameters converge follow a Gaussian distribution; The updating the parameters of the feature extraction network and the classifier based on a true Gaussian distribution, the classification results corresponding to the synthetic low-resolution image and the first real low-resolution image, and the degradation representations until the parameters converge, so that the degradation representations output by the feature extraction network follow a Gaussian distribution includes: Selecting a target training batch, where the target training batch includes multiple first real low-resolution images and multiple synthetic low-resolution images, and inputting the target training batch into the feature extraction network to extract corresponding degradation representations and corresponding classification results; Generating a first adversarial loss according to the classification result corresponding to the first real low-resolution image; Generating a first classification loss according to the classification result corresponding to the synthetic low-resolution image and the classification label corresponding to the synthetic low-resolution image; Generating a second adversarial loss according to the degradation representations corresponding to the synthetic low-resolution image and the first real low-resolution image and the true Gaussian distribution; Updating the parameters of the feature extraction network and the classifier according to the first adversarial loss, the first classification loss, and the second adversarial loss; Re-executing the step of selecting the target training batch until the parameters of the feature extraction network and the classifier converge; Sampling in the Gaussian distribution followed by the degradation representations output by the feature extraction network after the parameters converge to obtain a target sampled degradation representation, and inputting the target sampled degradation representation and the sample high-resolution image into a trained degradation network to obtain the sample low-resolution image corresponding to the sample high-resolution image output by the degradation network.

2. The method for generating super-resolution reconstruction training data based on generalizable degradation representations according to claim 1, characterized in that, the sampling in the Gaussian distribution followed by the degradation representations output by the feature extraction network to obtain a target sampled degradation representation includes: Sampling in the Gaussian distribution followed by the degradation representations output by the feature extraction network to obtain an initial sampled degradation representation; Input the initial sampling degradation characterization into the classifier to obtain the classification result and confidence level output by the classifier; When the confidence level output by the classifier is higher than the preset threshold, use the initial sampling degradation characterization as the target sampling degradation characterization.

3. The method for generating super-resolution reconstruction training data based on generalizable degradation characterization according to claim 1, characterized in that, the degradation network includes at least one degradation attention module and at least one feature extraction module, and each degradation attention module includes a convolution kernel reconstruction layer and an attention weight layer; obtaining the sample low-resolution image corresponding to the sample high-resolution image output by the degradation network includes: In each degradation attention module, perform the following operations: Use the target sampling degradation characterization as the first input data of the degradation attention module, and use the output of the previous module in the degradation network as the second input data of the degradation attention module. The second input data of the first degradation attention module in the degradation network is the output data after the sample high-resolution image passes through a feature extraction module; Input the first input data into the convolution kernel reconstruction layer to obtain a target convolution kernel, and perform convolution on the second input data according to the target convolution kernel to obtain first output data; Input the first input data into the attention weight layer to obtain a target weight, and perform weighting on the second input data according to the target weight to obtain second output data; Based on the first output data and the second output data, obtain the output data of the degradation attention module.

4. The method for generating super-resolution reconstruction training data based on generalizable degradation characterization according to claim 3, characterized in that, the training process of the degradation network training data is: Select a second real low-resolution image; Input the second real low-resolution image into the feature extraction network after parameter convergence to obtain the degradation characterization corresponding to the second real low-resolution image; Input the second real low-resolution image and the degradation characterization corresponding to the second real low-resolution image into the degradation network to train the degradation network and update the parameters of the degradation network; Repeat the step of selecting the second real low-resolution image until the parameters of the degradation network converge.

5. The method for generating super-resolution reconstruction training data based on generalizable degradation characterization according to claim 1, characterized in that, generating the first adversarial loss according to the classification result corresponding to the first real low-resolution image includes: Obtain the first adversarial loss according to the difference between the classification result corresponding to the first real low-resolution image and the preset real low-resolution image classification space.

6. The method for generating super-resolution reconstruction training data based on generalizable degradation characterization according to claim 1, characterized in that, generating the second adversarial loss according to the degradation characterizations corresponding to the synthesized low-resolution image and the first real low-resolution image and the real Gaussian distribution includes: Perform sampling on the degradation characterizations corresponding to the synthesized low-resolution image and the first real low-resolution image respectively to obtain a first sampling result; Sampling is performed in a true Gaussian distribution to obtain a second sampling result; The second adversarial loss is obtained according to the first sampling result and the second sampling result.

7. A super-resolution reconstruction training data generation device based on a generalizable degradation representation, where the super-resolution reconstruction training data includes a sample high-resolution image and a sample low-resolution image corresponding to the sample high-resolution image, characterized in that, the device includes: a degradation representation learning module, where the degradation representation learning module includes a model calculation unit and a training unit; Among them, the model calculation unit is used to perform different types of degradation processing on a first high-resolution image to obtain a synthetic low-resolution image, and input the synthetic low-resolution image and a first true low-resolution image into a degradation representation learning machine. The degradation representation learning machine includes a feature extraction network and a classifier. The feature extraction network is used to extract the degradation representation corresponding to the synthetic low-resolution image and the degradation representation corresponding to the first true low-resolution image. The classifier is used to output a classification result of the synthetic low-resolution image and the true low-resolution image based on the input degradation representation corresponding to the synthetic low-resolution image and the degradation representation corresponding to the first true low-resolution image; The training unit is used to update the parameters of the feature extraction network and the classifier based on the true Gaussian distribution, the classification results corresponding to the synthetic low-resolution image and the first true low-resolution image, and the degradation representation until the parameters converge, so that the degradation representation output by the feature extraction network after the parameters converge follows a Gaussian distribution; The step of updating the parameters of the feature extraction network and the classifier based on the true Gaussian distribution, the classification results corresponding to the synthetic low-resolution image and the first true low-resolution image, and the degradation representation until the parameters converge, so that the degradation representation output by the feature extraction network follows a Gaussian distribution includes: Select a target training batch, where the target training batch includes multiple first true low-resolution images and multiple synthetic low-resolution images, and input the target training batch into the feature extraction network to extract the corresponding degradation representation and the corresponding classification result; Generate a first adversarial loss according to the classification result corresponding to the first true low-resolution image; Generate a first classification loss according to the classification result corresponding to the synthetic low-resolution image and the classification label corresponding to the synthetic low-resolution image; Generate a second adversarial loss according to the degradation representations corresponding to the synthetic low-resolution image and the first true low-resolution image and the true Gaussian distribution; Update the parameters of the feature extraction network and the classifier according to the first adversarial loss, the first classification loss, and the second adversarial loss; Re-execute the step of selecting the target training batch until the parameters of the feature extraction network and the classifier converge; a data generation module, where the data generation module is used to sample in the Gaussian distribution followed by the degradation representation output by the feature extraction network after the parameters converge to obtain a target sampled degradation representation, and input the target sampled degradation representation and the sample high-resolution image into a trained degradation network to obtain the sample low-resolution image corresponding to the sample high-resolution image output by the degradation network.

8. A terminal, It is characterized in that the terminal includes: a processor and a computer-readable storage medium communicatively connected to the processor. The computer-readable storage medium is adapted to store multiple instructions, and the processor is adapted to call the instructions in the computer-readable storage medium to execute the steps of implementing the method for generating super-resolution reconstruction training data based on a generalizable degradation representation according to any one of claims 1-6 above.

9. A computer-readable storage medium It is characterized in that the computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the method for generating super-resolution reconstruction training data based on a generalizable degradation representation according to any one of claims 1-6.

Images