Image de-degradation methods, apparatuses, devices, media, and products

Abstract

The application relates to an image de-degradation method, device, equipment, medium and product, the method comprising: acquiring a bright field degradation image and a dark field degradation image; for each iteration training process, performing de-degradation estimation on random noise based on a self-supervised de-degradation model to obtain a bright field estimation image and de-degradation estimation data; performing de-degradation processing on the dark field degradation image according to the de-degradation estimation data to obtain a dark field estimation image; determining a comprehensive degradation loss; performing self-supervised training on the self-supervised de-degradation model according to the comprehensive degradation loss; and taking the minimization of the comprehensive degradation loss as an objective, determining the bright field estimation image output in the last iteration as a bright field de-degradation image, and determining the dark field estimation image output in the last iteration as a dark field de-degradation image. The application can perform de-degradation processing on the degradation image collected by a low-cost gene chip scanner, thereby improving the accuracy and reliability of subsequent gene site identification.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an image de-degradation method, apparatus, device, medium, and product. Background Technology

[0002] Based on in-situ synthesis of high-density chip hybridization technology, it is possible to quickly detect all known respiratory system diseases and tumor mutation sites. The core equipment used is a gene chip scanner, which is used to detect a large number of mutation sites on gene chips with high sensitivity through microscopic imaging technology.

[0003] The microscopic imaging technology of gene chip scanners uses simulation software and precision motion devices to accurately adjust multiple spherical lenses and mirror assemblies to ensure image quality and result analysis. This involves dozens of independent aspherical lenses and mirrors, resulting in gene chip scanners costing over a million dollars, making them difficult to popularize in communities and small research institutions. Therefore, low-cost gene chip scanners, employing simple lens imaging, computational imaging, and deep learning technologies, meet the market application requirements for low-cost, automated, high-throughput, and high-precision gene mutation site detection.

[0004] To simplify design, low-cost gene chip scanners often employ simpler single-lens or multi-lens arrays, supplemented by computational imaging techniques to compensate for optical limitations. However, this simplified optical approach inevitably introduces various image degradations, such as motion blur caused by minute vibrations during scanning or defocusing due to inaccurate focusing, as well as superimposed or multiplicative artifacts caused by contamination such as dust or oil on the lens surface. These image degradations severely impact the accuracy and reliability of subsequent gene locus identification, becoming a key bottleneck restricting the performance of miniaturized scanners. Summary of the Invention

[0005] Therefore, it is necessary to provide an image de-degradation method, apparatus, device, medium, and product to address the aforementioned technical problems, which can perform de-degradation processing on degraded images acquired by low-cost gene chip scanners, thereby improving the accuracy and reliability of subsequent gene locus identification.

[0006] In a first aspect, this application provides an image de-degradation method, comprising:

[0007] Acquire bright-field and dark-field degraded images; the bright-field and dark-field degraded images are gene chip images acquired in the same area by the same gene chip scanner;

[0008] For each iteration of the training process, based on the self-supervised de-degradation model, the random noise is de-degraded and estimated to obtain the bright field estimated image and the de-degradation estimated data.

[0009] Based on the de-degradation estimation data, the dark-field degraded image is degraded to obtain the dark-field estimated image;

[0010] Determine the comprehensive degradation loss based on at least one of the bright-field degraded image, the bright-field estimated image, the dark-field degraded image, and the dark-field estimated image;

[0011] Based on the comprehensive degradation loss, the self-supervised de-degradation model is trained under self-supervised conditions.

[0012] With the goal of minimizing the overall degradation loss, the bright field estimated image output by the last iteration is determined as the bright field de-degradation image, and the dark field estimated image output by the last iteration is determined as the dark field de-degradation image.

[0013] In one embodiment, based on a self-supervised de-degradation model, random noise is de-degraded to obtain a bright-field estimated image and de-degradation estimated data. This includes: using a shared coding network in the self-supervised de-degradation model to encode features of random noise, obtaining noise coding features of random noise for different task networks; wherein, the task networks include a deblurring network corresponding to the deblurring task and a decontamination network corresponding to the decontamination task; the de-degradation estimated data includes deblurred estimated data and decontamination estimated data; based on the deblurring network, feature processing is performed on the corresponding noise coding features to obtain a bright-field estimated image and deblurred estimated data; based on the decontamination network, feature processing is performed on the corresponding noise coding features to obtain decontamination estimated data.

[0014] In one embodiment, performing de-degradation processing on a dark-field degraded image based on de-degradation estimation data to obtain a dark-field estimated image includes: using decontamination estimation data to perform decontamination processing on the dark-field degraded image to obtain a decontamination estimated image; and using deblurring estimation data to perform deblurring processing on the decontamination estimated image to obtain a dark-field estimated image.

[0015] In one embodiment, determining a comprehensive degradation loss based on at least one of a bright-field degraded image, a bright-field estimated image, a dark-field degraded image, and a dark-field estimated image includes: determining a bright-field reconstruction loss based on the bright-field degraded image and the bright-field estimated image; and determining a dark-field reconstruction loss based on the dark-field degraded image and the dark-field estimated image; wherein the comprehensive degradation loss includes the bright-field reconstruction loss and / or the dark-field reconstruction loss.

[0016] In one embodiment, determining the brightfield reconstruction loss based on the brightfield degraded image and the brightfield estimated image includes: performing degradation processing on the brightfield estimated image to obtain the brightfield degraded estimated image; and determining the brightfield reconstruction loss based on the brightfield degraded estimated image and the brightfield degraded image.

[0017] In one embodiment, determining the dark field reconstruction loss based on the dark field degraded image and the dark field estimated image includes: performing degradation processing on the dark field estimated image to obtain the dark field degraded estimated image; and determining the dark field reconstruction loss based on the dark field degraded estimated image and the dark field degraded image.

[0018] In one embodiment, determining a comprehensive degradation loss based on at least one of the bright-field degraded image, the bright-field estimated image, the dark-field degraded image, and the dark-field estimated image further includes: determining a bright-field similarity loss based on the similarity between image patches in the bright-field estimated image; and determining a dark-field similarity loss based on the similarity between image patches in the dark-field estimated image; wherein the comprehensive degradation loss further includes the bright-field similarity loss and / or the dark-field similarity loss.

[0019] In one embodiment, determining the bright-field similarity loss based on the similarity between image patches in the bright-field estimated image includes: dividing the bright-field estimated image into a plurality of first image patches; for a first image patch, determining the bright-field similarity sub-loss of the first pixel patch based on the similarity between the first image patch and its neighboring first pixel patches; and determining the bright-field similarity loss based on the bright-field similarity sub-loss of each first pixel patch.

[0020] In one embodiment, determining the bright-field similarity loss based on the similarity between image patches in the bright-field estimated image includes: dividing the dark-field estimated image into a plurality of second image patches; for each second image patch, determining the dark-field similarity sub-loss of the second pixel patch based on the similarity between the second image patch and its neighboring second pixel patches; and determining the dark-field similarity loss based on the dark-field similarity sub-loss of each second pixel patch.

[0021] In one embodiment, determining the comprehensive degradation loss based on at least one of the bright-field degraded image, the bright-field estimated image, the dark-field degraded image, and the dark-field estimated image further includes: determining the bright-field gradient loss of the bright-field estimated image based on the pixel gradient corresponding to each pixel in the bright-field estimated image; and determining the dark-field gradient loss of the dark-field estimated image based on the pixel gradient corresponding to each pixel in the dark-field estimated image; wherein the comprehensive degradation loss further includes the bright-field gradient loss and / or the dark-field gradient loss.

[0022] In one embodiment, determining a comprehensive degradation loss based on at least one of a bright-field degraded image, a bright-field estimated image, a dark-field degraded image, and a dark-field estimated image further includes: determining a regularization loss for the degraded estimation data; wherein the comprehensive degradation loss further includes a regularization loss.

[0023] In one embodiment, the de-degradation estimation data includes decontamination estimation data; determining the regularization loss of the de-degradation estimation data includes: determining a first regularization loss based on the pixel value change gradient between each pixel and its neighboring pixels in the decontamination estimation data; and determining a second regularization loss of the decontamination estimation data based on a first preset norm function; wherein the regularization loss includes the first regularization loss and / or the second regularization loss.

[0024] In one embodiment, the dedegenerate estimation data includes defuzzified estimation data; correspondingly, determining the regularization loss of the dedegenerate estimation data includes: determining a third regularization loss of the defuzzified estimation data based on a second preset norm function; applying a center constraint to the defuzzified estimation data according to a first preset Gaussian weight matrix, and determining a fourth regularization loss of the center constraint result based on the third preset norm function; wherein the regularization loss includes the third regularization loss and / or the fourth regularization loss.

[0025] Secondly, this application also provides an image de-degradation apparatus, comprising:

[0026] The image acquisition module is used to acquire bright-field degraded images and dark-field degraded images; the bright-field degraded images and dark-field degraded images are gene chip images acquired in the same area by the same gene chip scanner;

[0027] The first de-degradation module is used to perform de-degradation estimation on random noise based on a self-supervised de-degradation model for each iteration of training, so as to obtain the bright field estimated image and de-degradation estimated data.

[0028] The second de-degradation module is used to perform de-degradation processing on the dark field degraded image based on the de-degradation estimation data to obtain the dark field estimated image;

[0029] The first determining module is used to determine the comprehensive degradation loss based on at least one of the bright field degraded image, the bright field estimated image, the dark field degraded image, and the dark field estimated image;

[0030] The model training module is used to perform self-supervised training on the self-supervised de-degradation model based on the comprehensive degradation loss.

[0031] The second determining module is used to determine the bright field estimated image output by the last iteration as the bright field de-degradation image and the dark field estimated image output by the last iteration as the dark field de-degradation image, with the goal of minimizing the overall degradation loss.

[0032] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method provided in the first aspect.

[0033] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method provided in the first aspect.

[0034] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the method provided in the first aspect.

[0035] The aforementioned image de-degradation methods, apparatuses, devices, media, and products, after acquiring bright-field degraded images and dark-field degraded images, achieve self-supervised training of a self-supervised de-degradation model through multiple training iterations. In each training iteration, the self-supervised de-degradation model is used to estimate random noise for de-degradation, obtaining a bright-field estimated image and de-degradation estimation data. Based on the de-degradation estimation data, the dark-field degraded image is then degraded. Since dark-field degraded images typically have a low signal-to-noise ratio, their de-degradation difficulty is much greater than that of bright-field degraded images. This embodiment utilizes the complementarity between bright-field degraded images with high contrast characteristics and dark-field degraded images with high signal-to-noise ratio characteristics. The de-degradation estimation data from the training iteration process is used as prior knowledge to perform de-degradation processing on dark-field degraded images with low signal-to-noise ratios, reducing the de-degradation difficulty of dark-field degraded images and improving the de-degradation effect. Moreover, after minimizing the overall degradation loss, the training iteration ends. The bright-field estimated image output by the last iteration is the expected bright-field de-degraded image, and the dark-field estimated image output by the last iteration is the expected dark-field de-degraded image. This achieves the de-degradation processing of the bright-field and dark-field degraded images, thereby improving the accuracy and reliability of subsequent gene locus identification and enhancing the performance of the gene chip scanner. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart illustrating an image de-degradation method in one embodiment;

[0038] Figure 2 This is a flowchart illustrating the de-degradation estimation steps for random noise in one embodiment;

[0039] Figure 3 This is a flowchart illustrating the de-degradation process for a dark-field degraded image in one embodiment.

[0040] Figure 4 This is a flowchart illustrating the steps for determining the overall degradation loss in one embodiment;

[0041] Figure 5 This is a flowchart illustrating the steps for determining similarity loss in one embodiment;

[0042] Figure 6 This is a flowchart illustrating the gradient loss determination steps in one embodiment;

[0043] Figure 7 This is a flowchart illustrating the regularization loss determination steps in one embodiment;

[0044] Figure 8 This is a flowchart illustrating the regularization loss determination steps in one embodiment;

[0045] Figure 9 This is a flowchart illustrating the regularization loss determination steps in one embodiment;

[0046] Figure 10 This is a structural block diagram of an image de-degradation device in one embodiment;

[0047] Figure 11 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0049] In one exemplary embodiment, an image de-degradation method is provided, see [link to relevant documentation]. Figure 1 The methods include:

[0050] S110, acquire bright-field degraded image and dark-field degraded image.

[0051] Among them, the bright-field degradation image and the dark-field degradation image are gene chip images acquired in the same area by the same gene chip scanner.

[0052] Understandably, a degraded image is an image that exhibits degradation issues such as blurring and / or contamination.

[0053] Understandably, bright-field imaging works by imaging through transmitted or reflected light, primarily reflecting the physical morphology and surface topology of the gene chip. Bright-field images can clearly show the microstructure and contaminants (such as dust and fingerprints) on the gene chip surface. Bright-field images have high contrast, but their signal-to-noise ratio may be limited by the uniformity of the light source.

[0054] Understandably, dark-field imaging works by using fluorescence excitation, primarily reflecting the fluorescence signal emitted after gene probe hybridization. Bright-field images, on the other hand, can display gene loci with high sensitivity, and the background is typically dark. Dark-field images have sparse signals and low signal-to-noise ratios, making direct blind deblurring and decontamination difficult.

[0055] It is evident that bright-field degraded images are images with degradation issues obtained based on the principle of transmitted or reflected light imaging, while dark-field degraded images are images with degradation issues obtained based on the principle of fluorescence excitation imaging.

[0056] After obtaining the bright-field and dark-field degraded images, multiple iterative training processes are performed through subsequent steps to achieve self-supervised training of the self-supervised degraded model. After training is completed, the image output by the last iteration of the training process is the degraded image.

[0057] S120: For each iteration of the training process, based on the self-supervised de-degradation model, the random noise is de-degraded to obtain the bright field estimated image and the de-degradation estimated data.

[0058] That is, random noise is input into the self-supervised de-degradation model, which performs a series of upsampling and convolution operations on the random noise to gradually generate a high-resolution bright-field estimation image, and at the same time obtains de-degradation estimation data.

[0059] The de-degradation estimation data refers to the de-degradation parameters used in the de-degradation estimation process. This data may include defuzzification estimation data and decontamination estimation data. Defuzzification estimation data is equivalent to defuzzification estimation parameters, and decontamination estimation data is equivalent to decontamination estimation parameters. The defuzzification estimation parameters and / or decontamination estimation parameters can be matrices or other forms, which are not limited here.

[0060] Understandably, the de-degradation estimation data obtained in the last training iteration is used as de-degradation data, i.e., de-degradation parameters. The de-degradation data includes deblurred data, i.e., deblurred parameters, and decontaminated data, i.e., decontaminated parameters.

[0061] Among them, the bright field estimated image is the estimated image of the final expected bright field de-degradation image.

[0062] S130, Based on the de-degradation estimation data, perform de-degradation processing on the dark field degraded image to obtain the dark field estimated image.

[0063] Physically, gene chip images acquired by the same gene chip scanner in the same area have the same deblurred and decontamination data. Therefore, in S130, the dark field degraded image is degraded based on the de-degradation estimation data obtained in S120, thereby obtaining the dark field estimated image.

[0064] Among them, the dark field estimated image is the estimated image of the dark field de-degradation image that is finally expected to be obtained.

[0065] S140, determine the comprehensive degradation loss based on at least one of the bright-field degraded image, the bright-field estimated image, the dark-field degraded image, and the dark-field estimated image.

[0066] Among them, the comprehensive degradation loss can reflect the comprehensive loss in the degradation process.

[0067] In practical scenarios, the comprehensive loss can include reconstruction loss, and in addition to reconstruction loss, it can also include at least one of similarity loss, gradient loss, and regularization loss, and of course, other types of loss, which are not limited here. The total loss value is obtained by summing or weighted summing the various losses. In section S150, the self-supervised de-degradation model is trained using the comprehensive degradation loss, i.e., the self-supervised de-degradation model is trained using the total loss value. Section S160 aims to minimize the comprehensive degradation loss, i.e., minimize the total loss value.

[0068] Understandably, determining the comprehensive degradation loss through multiple loss methods allows these types of losses to work synergistically, making the iterative process more stable, effectively avoiding trivial solutions, and ensuring the high accuracy and reliability of the final de-degraded image. Furthermore, by employing a weighted summation method for various losses, the weights of each loss can be adjusted according to the needs of the actual application scenario, enabling the iterative process to focus on optimization in different aspects.

[0069] S150, based on the comprehensive degradation loss, performs self-supervised training on the self-supervised de-degradation model.

[0070] In other words, self-supervised training is actually a process of adjusting the parameters of the self-supervised degenerate model. By adjusting the parameters in the model, the corresponding comprehensive degenerate loss becomes smaller and smaller. When the comprehensive degenerate loss is less than the preset loss threshold, the training converges and the iteration process ends.

[0071] S160, with the goal of minimizing the overall degradation loss, determines the bright field estimated image output by the last iteration as the bright field de-degradation image, and the dark field estimated image output by the last iteration as the dark field de-degradation image.

[0072] That is, after the iteration process ends, the bright field estimated image output by the last iteration is used as the bright field de-degradation image, the dark field estimated image output by the last iteration is used as the dark field de-degradation image, and the de-degradation estimated data output by the last iteration is used as the de-degradation data.

[0073] Understandably, by using random noise as input, the self-supervised de-degradation model can generate images from scratch, rather than transforming the input image. Thus, the image generation process is entirely controlled by the network parameters, which are learned through the optimization process. Because the network structure of the self-supervised de-degradation model has prior knowledge, the generated images naturally tend towards natural images. Therefore, during optimization, the network structure first fits the natural parts of the image, and then fits the noise and degradation (because the noise and degradation do not conform to the prior knowledge of natural images). During optimization, the parameters can be adjusted based on the comprehensive degradation loss, where the regularization loss can suppress noise.

[0074] This embodiment does not require pre-training the network with a large amount of data. Instead, it randomly initializes the network and then optimizes the network parameters to fit a degraded image. Because the network structure has prior knowledge, it tends to generate natural images, thus enabling the recovery of the degraded image during the fitting process.

[0075] The aforementioned image de-degradation method, after acquiring the bright-field degraded image and the dark-field degraded image, implements self-supervised training of the self-supervised de-degradation model through multiple training iterations. In each training iteration, the self-supervised de-degradation model is used to estimate random noise for de-degradation, obtaining the estimated bright-field image and de-degradation estimation data. Based on the de-degradation estimation data, the dark-field degraded image is then degraded. Since dark-field degraded images typically have a lower signal-to-noise ratio, their de-degradation difficulty is much greater than that of bright-field degraded images. This embodiment utilizes the complementarity between bright-field degraded images with high contrast characteristics and dark-field degraded images with high signal-to-noise ratio characteristics. The de-degradation estimation data from the training iteration process is used as prior knowledge to perform de-degradation processing on dark-field degraded images with low signal-to-noise ratios, reducing the de-degradation difficulty of dark-field degraded images and improving the de-degradation effect. Furthermore, after minimizing the overall degradation loss, the training iteration ends. The bright-field estimated image output in the last iteration is the desired bright-field de-degraded image, and the dark-field estimated image output in the last iteration is the desired dark-field de-degraded image. This achieves de-degradation processing for both bright-field and dark-field degraded images, thereby improving the accuracy and reliability of subsequent gene locus identification and enhancing the performance of the gene chip scanner. Moreover, the training process in the above image de-degradation method is a self-supervised training process that does not require pre-training, avoiding errors introduced by manual labels and reducing dependence on the training dataset, effectively addressing the technical challenge of acquiring gene chip images. Furthermore, this embodiment uses random noise as input to fit the image, rather than directly using the degraded image as input, thus avoiding overfitting. Therefore, this embodiment provides an image de-degradation method for low-cost gene chip scanners to overcome image quality problems caused by simplified optical systems and meet the application requirements of high-precision gene locus detection.

[0076] Based on the technical solutions provided in the above embodiments, an optional embodiment is provided, in which the de-degradation estimation step of random noise in S120 is refined.

[0077] See Figure 2 The de-degradation estimation steps for random noise include:

[0078] S210, based on the shared coding network in the self-supervised de-degradation model, performs feature encoding on random noise to obtain the noise coding features of random noise for different task networks.

[0079] The task network includes the deblurring network corresponding to the deblurring task and the decontamination network corresponding to the decontamination task; the dedegradation estimation data includes the deblurring estimation data and the decontamination estimation data.

[0080] The shared coding network can include a shared expert layer and a task-specific expert layer. The shared expert layer is used to extract features from random noise to obtain shared noise coding features, which are general features. The task-specific expert layer is used to extract the noise coding-specific features required by each task network from the shared noise coding features, and then outputs the noise coding-specific features to the task network.

[0081] S220, based on a deblurring network, performs feature processing on the corresponding noise coding features to obtain the bright field estimated image and the deblurring estimated data.

[0082] Among them, the corresponding noise coding features are the noise coding-specific features required by the deblurring network.

[0083] S230, based on the decontamination network, performs feature processing on the corresponding noise coding features to obtain decontamination estimation data.

[0084] Among them, the corresponding noise coding features are the noise coding-specific features required for the decontamination network.

[0085] In this embodiment, random noise is feature-encoded using a shared coding network to obtain noise coding features corresponding to the deblurring network and the decontamination network. Then, the deblurring network is used to process the corresponding noise coding features to obtain a bright-field estimated image and deblurred estimation data. Finally, the decontamination network is used to process the corresponding noise coding features to obtain decontamination estimation data. Thus, the bright-field estimated image and de-degradation estimation data are obtained. It can be seen that this embodiment implements the corresponding de-degradation task based on a multi-task network, thereby obtaining the corresponding processing results, avoiding interference, conflicts, or negative transfer between de-degradation tasks, and improving the reliability of each processing result.

[0086] Based on the technical solutions provided in the above embodiments, an optional embodiment is provided, in which the de-degradation processing steps of the dark field degraded image in S130 are refined.

[0087] See Figure 3 The de-degradation processing steps for dark-field degraded images include:

[0088] S310. Using the decontamination estimation data, the dark field degraded image is decontaminated to obtain the decontamination estimation image.

[0089] The decontamination treatment can be achieved using the following calculation formula:

[0090]

[0091] In the formula, To estimate the image for decontamination, To estimate the pollution levels, This is a degraded image in dark conditions.

[0092] Of course, other methods can also be used to decontaminate dark-field degraded images, which are not limited here.

[0093] S320: Using the deblurred estimation data, the decontamination estimation image is deblurred to obtain the dark field estimation image.

[0094] Specifically, the decontamination estimation image can be input into a network module configured with deblurring estimation data. The network module then uses the deblurring estimation data to deblur the decontamination estimation image, thereby obtaining the dark field estimation image.

[0095] In this embodiment, during the de-degradation processing of the dark-field degraded image, the dark-field degraded image is first decontaminated, and then the deblurred estimation data is used to deblur the decontaminated result, thus achieving a cascaded de-degradation process. Furthermore, performing decontamination before deblurring improves the de-degradation effect of the dark-field estimated image compared to deblurring before decontamination. Moreover, after the iteration, both blurring and surface contamination issues are resolved simultaneously, avoiding error accumulation that may result from multi-step cascaded processing, thereby improving processing efficiency and final image quality.

[0096] Based on the technical solutions provided in the above embodiments, an optional embodiment is provided. In this optional embodiment, the overall degradation loss is refined to include bright field reconstruction loss and / or dark field reconstruction loss, and the overall degradation loss determination step in S140 is further refined.

[0097] See Figure 4 The steps for determining the overall degradation loss include:

[0098] S410, determine the bright-field reconstruction loss based on the bright-field degraded image and the bright-field estimated image.

[0099] The reconstruction loss can be understood as the difference between the original image and the reconstructed image.

[0100] In one optional implementation, the bright-field reconstruction loss determination step in S410 includes:

[0101] S1A performs degradation processing on the bright-field estimated image to obtain a degraded bright-field estimated image.

[0102] The following degradation mathematical model can be used to degrade the bright-field estimated image:

[0103]

[0104] In the formula, The result of the degradation processing is shown in this step as the bright-field degradation estimation image, where K represents the deblurred estimation data and S represents the decontamination estimation data. The image to be degraded is, in this step, the bright-field estimation image. It is additive white Gaussian noise. This is the symbol for the convolution operation. This is the element-wise dot product symbol.

[0105] The aforementioned degraded data model achieves fuzzing through convolutional degradation and contamination through nonlinear degradation.

[0106] S2A determines the bright-field reconstruction loss based on the bright-field degradation estimation image and the bright-field degradation image.

[0107] That is, the difference between the estimated bright-field degradation image and the degraded bright-field image is calculated, and this difference is used as the bright-field reconstruction loss.

[0108] In the above implementation, the bright-field estimated image is degraded to obtain a reconstructed bright-field degraded estimated image. Then, the difference between the bright-field degraded estimated image and the bright-field degraded image is calculated to obtain the accurate bright-field reconstruction loss.

[0109] S420, determine the dark field reconstruction loss based on the dark field degradation image and the dark field estimation image.

[0110] In one optional implementation, the dark field reconstruction loss determination step in S420 includes:

[0111] S1B performs degradation processing on the dark field estimated image to obtain a dark field degradation estimated image.

[0112] In this process, the dark field estimated image can be used as the image to be degraded and input into the aforementioned degradation mathematical model to obtain the dark field degradation estimated image.

[0113] S2B determines the dark field reconstruction loss based on the dark field degradation estimation image and the dark field degradation image.

[0114] That is, the difference between the dark field degradation estimated image and the dark field degradation image is calculated, and this difference is used as the dark field reconstruction loss.

[0115] In the above implementation, the dark field estimated image is degraded to obtain a reconstructed dark field degraded estimated image. Then, the difference between the dark field degraded estimated image and the dark field degraded image is calculated to obtain an accurate dark field reconstruction loss.

[0116] In this embodiment, bright field reconstruction loss and / or dark field reconstruction loss are used as reconstruction loss. Reconstruction loss is the core term in the comprehensive degradation loss, which aims to represent the difference between the original image and the reconstructed image. Training is then performed based on this difference to ensure that the difference between the original image and the reconstructed image becomes smaller and smaller, thus ensuring the effectiveness of training.

[0117] Based on the technical solutions provided in the above embodiments, an optional embodiment is provided. In this optional embodiment, the comprehensive degradation loss is further refined to include bright field similarity loss and / or dark field similarity loss, and the comprehensive degradation loss determination step is further refined to include a similarity loss determination step.

[0118] See Figure 5 The steps for determining similarity loss include:

[0119] S510, determine the bright field similarity loss based on the similarity between image patches in the bright field estimated image.

[0120] In real-world scenarios, inspired by image self-similarity, it is assumed that local patches in the bright-field estimated image or dark-field estimated image are repetitive in their respective images. Therefore, based on the similarity between each image patch in the bright-field estimated image, the bright-field similarity loss is determined, and based on the similarity between each image patch in the dark-field estimated image, the dark-field similarity loss is determined. The bright-field similarity loss and / or dark-field similarity loss are used as the similarity prior loss.

[0121] In one optional implementation, the bright-field similarity loss determination step in S510 includes:

[0122] S1C divides the bright-field estimated image into multiple first image blocks.

[0123] For example, the bright-field estimated image is uniformly divided into 16×16 image blocks, which are referred to here as the first image block in order to distinguish them from the image blocks of the dark-field estimated image.

[0124] S2C, for the first image block, determines the bright-field similarity sub-loss of the first pixel block based on the similarity between the first image block and the neighboring first pixel blocks.

[0125] Specifically, for each first image block, its neighboring image blocks can be found first. Then, from the neighboring image blocks, image blocks with a similarity higher than a preset value to the first image block can be found as neighboring similar image blocks. Then, the square of the L2 norm of the difference between the first image block and the neighboring similar image blocks can be calculated as the bright field similarity loss of the first pixel block.

[0126] S3C determines the bright-field similarity loss based on the bright-field similarity sub-loss of each first pixel block.

[0127] The bright-field similarity loss can be calculated using the following formula:

[0128]

[0129] In the formula, For similarity loss, in this step it is bright field similarity loss; The image to be estimated is the bright-field estimated image in this step. This refers to the i-th image patch in the image to be estimated, which in this step is the i-th first image patch. In this step, the k-th neighboring similar image block of the i-th first image block is... The weight of the k-th neighboring similar image block (which is also the first image block in this step).

[0130] The weights of neighboring similar image patches can be calculated, but are not limited to, using a Gaussian kernel function.

[0131] Of course, other methods can be used to calculate the bright-field similarity loss, which are not limited here.

[0132] S520, determine the dark field similarity loss based on the similarity between image patches in the dark field estimated image.

[0133] In one optional implementation, the dark field similarity loss determination step in S520 may include:

[0134] S1D divides the dark field estimated image into multiple second image blocks.

[0135] For example, the dark field estimated image is uniformly divided into 16×16 image blocks, which are referred to here as the second image block in order to distinguish them from the image blocks of the bright field estimated image.

[0136] S2D, for each second image block, determines the dark field similarity loss of the second pixel block based on the similarity between the second image block and its neighboring second pixel blocks.

[0137] Specifically, for each second image block, its neighboring image blocks can be found first. Then, from the neighboring image blocks, image blocks with a similarity higher than a preset value to the second image block can be found as neighboring similar image blocks. Then, the square of the L2 norm of the difference between the second image block and the neighboring similar image blocks can be calculated as the bright field similarity sub-loss of the second pixel block.

[0138] S3D determines the dark field similarity loss based on the dark field similarity sub-loss of each second pixel block.

[0139] In this step, the dark field similarity loss can be calculated using the aforementioned similarity loss formula. For dark field similarity loss, This step involves estimating the dark field image. For the i-th second image patch, In this step, the k-th neighboring similar image block is the i-th second image block. The weight is the weight of the k-th neighboring similar image block (which is also the second image block in this step).

[0140] The weights of neighboring similar image patches can be calculated, but are not limited to, using a Gaussian kernel function.

[0141] Of course, other methods can be used to calculate the dark field similarity loss, which are not limited here.

[0142] In real-world scenarios, inspired by image self-similarity, it is assumed that image patches in the bright-field estimated image or dark-field estimated image are repetitive within their respective images. Therefore, based on the similarity between image patches in the bright-field estimated image, a bright-field similarity loss is determined, and based on the similarity between image patches in the dark-field estimated image, a dark-field similarity loss is determined. The bright-field similarity loss and / or dark-field similarity loss are used as prior similarity losses.

[0143] Moreover, the aforementioned similarity prior loss is a content-aware prior loss, which effectively constrains the training iterations and can thus address the pathological nature of the de-degradation problem, namely, the lack of a unique solution.

[0144] In this embodiment, the similarity loss is calculated based on the repetition of image patches in the bright-field estimated image or the dark-field estimated image. The corresponding similarity loss serves as a supplement to the reconstruction loss, so that when calculating the comprehensive degradation loss, the bright-field similarity loss and / or dark-field similarity loss are taken into account on the basis of the reconstruction loss, thereby making the comprehensive degradation loss more comprehensive and accurate.

[0145] In real-world scenarios, prior loss can include not only similarity loss but also self-guided multi-scale prior loss. One method for calculating self-guided multi-scale prior loss is as follows: using a randomly initialized network module, inputting random noise into the randomly initialized network module to generate an ideal de-degradation prior image, subtracting the desired de-degradation image from the de-degradation prior image, calculating the L2 norm, and squaring the result to obtain the multi-scale prior loss.

[0146] Based on the technical solutions provided in the above embodiments, an optional embodiment is provided. In this optional embodiment, the comprehensive degradation loss is further refined to include bright field gradient loss and / or dark field gradient loss, and the comprehensive degradation loss determination step is further refined to include a gradient loss determination step.

[0147] See Figure 6 The gradient loss determination steps include:

[0148] S610, determine the bright field gradient loss of the bright field estimated image based on the pixel gradient corresponding to each pixel in the bright field estimated image.

[0149] S620: Determine the dark field gradient loss of the dark field estimated image based on the pixel gradient corresponding to each pixel in the dark field estimated image.

[0150] Among them, the bright field gradient loss or dark field gradient loss can be understood as the sum of the differences between the pixel values ​​of each pixel and its neighboring pixels in the corresponding estimated image.

[0151] The following formula can be used to calculate the bright-field gradient loss or the dark-field gradient loss:

[0152]

[0153] In the formula, for the bright-field estimation image, It is a bright-field gradient loss. To estimate the pixel value of the pixel in the (i+1)th row and jth column of the bright field image, To estimate the pixel value of the pixel in the i-th row and j-th column of the bright field image, This represents the pixel value of the pixel in the i-th row and j+1-th column of the bright-field estimated image. For the dark-field estimated image, It is dark field gradient loss. Estimate the pixel value of the pixel in the (i+1)th row and jth column of the image in the dark field. To estimate the pixel value of the pixel in the i-th row and j-th column of an image in a dark field. Estimate the pixel value of the pixel in the i-th row and j+1-th column of the image for dark scenes.

[0154] The above calculation is equivalent to the Total Variation (TV) regularization term, which encourages the recovered de-degradation image to have piecewise smoothness and effectively remove noise.

[0155] Of course, other methods can be used to calculate the gradient loss, which are not limited here.

[0156] In this embodiment, the gradient loss of the corresponding estimated image is calculated based on the pixel gradient of each pixel in the bright field estimated image or the dark field estimated image. The gradient loss serves as a supplement to the reconstruction loss, so that when calculating the comprehensive degradation loss, the bright field gradient loss and / or dark field gradient loss are taken into account on the basis of the reconstruction loss, thereby making the comprehensive degradation loss more comprehensive and accurate.

[0157] Based on the technical solutions provided in the above embodiments, an optional embodiment is provided. In this optional embodiment, the comprehensive degradation loss is further refined to include regularization loss, and the comprehensive degradation loss determination step is further refined to include a regularization loss determination step.

[0158] See Figure 7 The steps for determining the regularization loss include:

[0159] S710, determine the regularization loss for de-degradation estimation data.

[0160] The regularization loss can be understood as the loss obtained by regularizing the de-degenerate estimation data.

[0161] In one alternative implementation, the dedegraded estimation data includes the decontaminated estimation data, and the corresponding regularization loss includes a first regularization loss and / or a second regularization loss. See also Figure 8 The steps for determining the regularization loss in S710 include:

[0162] S810 determines the first regularization loss based on the pixel value change gradient between each pixel and its neighboring pixels in the decontamination estimation data.

[0163] S820, based on the first preset norm function, determines the second regularization loss of the decontamination estimation data.

[0164] The first preset norm function can be either absolute value or norm.

[0165] For example, the regularization loss can be calculated using the following formula:

[0166]

[0167] In the formula, The regularization loss is calculated based on the first and second regularization losses. and For weights, To estimate the pollution levels, To estimate the data in the (i+1)th row and jth column of the decontamination data, To estimate the data in the i-th row and j+1-th column of the decontamination estimation data, This refers to the data in the i-th row and j-th column of the decontamination estimation data.

[0168] As can be seen, the first term on the right-hand side of the above formula is the first regularization loss, and the second term is the second regularization loss. The first regularization loss is used to encourage smoothness, and the second regularization loss is used to encourage sparsity.

[0169] The above implementation method uses a first regularization loss based on the gradient of pixel value changes to help improve the smoothness of the de-degraded image, and a second regularization loss based on a first preset norm function to help improve the sparsity of the subsequent de-degraded image. Thus, the comprehensive degradation loss including the first regularization loss and / or the second regularization loss can improve the quality of the subsequently obtained de-degraded image.

[0170] In one optional implementation, the dedegradation estimation data includes deblurred estimation data, and the corresponding regularization loss includes a third regularization loss and / or a fourth regularization loss. See also Figure 9 The steps for determining the regularization loss in S710 include:

[0171] S910, based on the second preset norm function, determines the third regularization loss of the defuzzified estimated data.

[0172] The second and third predefined norm functions can be the same, for example, both being L2 norm functions. Of course, they can also be different, which is not limited here. L2 norm is the Euclidean norm.

[0173] S920, based on the first preset Gaussian weight matrix, center constraints are applied to the defuzzified estimation data, and based on the third preset norm function, the fourth regularization loss of the center constraint result is determined.

[0174] For example, the formula for calculating regularization loss is as follows:

[0175]

[0176] In the formula, The regularization loss is calculated based on the third and fourth regularization losses. To defuzzify the estimated data, This is the first preset Gaussian weight matrix.

[0177] As can be seen, the first term on the right-hand side of the above formula is the third regularization loss, and the second term is the fourth regularization loss. The third regularization loss is calculated using the L2 norm function, which encourages pixel values ​​in the deblurred estimation data to be close to zero, giving it a certain degree of sparsity. The fourth regularization loss achieves a central constraint by introducing a first preset Gaussian weight matrix and minimizing its dot product with the deblurred estimation data, thereby ensuring the reasonableness of the deblurred estimation data gradually increases with the number of training iterations.

[0178] The above implementation, based on the third regularization loss calculated using the second preset norm function, ensures that the sparsity of the deblurred estimation data gradually increases with the number of training iterations. The fourth regularization loss, calculated using the third preset norm function and the first preset Gaussian weight matrix, ensures that the reasonableness of the deblurred estimation data gradually increases with the number of training iterations. Therefore, by using regularization losses including the third and / or fourth regularization losses, the quality of the deblurred estimation data obtained in subsequent training iterations can be improved.

[0179] In this embodiment, a regularization loss is determined for the de-degradation estimation data. The regularization loss can ensure the physical rationality of the de-degradation estimation parameters, thereby improving the rationality of the degradation estimation parameters obtained in subsequent training iterations.

[0180] In real-world scenarios, deblurring estimation data can be in matrix form, or it can be pre-modeled as a function with a few parameters, rather than a massive matrix. For example, for isotropic defocus blur, the deblurring estimation data can be approximated as a disk function or a two-dimensional Gaussian function, with its only independent variable being the standard deviation. For linear motion blur, the deblurring estimation data can be determined by two parameters: motion length and motion angle. Therefore, the deblurring estimation data is constructed as a function with motion length and motion angle as independent parameters. In this case, the deblurring estimation data output by the deblurring network is no longer a complete kernel matrix, but rather these independent parameters. In this situation, complex regularization terms are no longer needed to constrain the physical plausibility of the deblurring estimation data, because the deblurring estimation data itself is obtained by parameterizing the physical model.

[0181] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0182] Based on the same inventive concept, this application also provides an image de-degradation apparatus for implementing the image de-degradation method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more image de-degradation apparatus embodiments provided below can be found in the limitations of the image de-degradation method described above, and will not be repeated here.

[0183] In one exemplary embodiment, an image de-degradation apparatus is provided, see [link to relevant documentation]. Figure 10 The device includes an image acquisition module 1010, a first de-degradation module 1020, a second de-degradation module 1030, a first determination module 1040, a model training module 1050, and a second determination module 1060, wherein:

[0184] The image acquisition module 1010 is used to acquire bright-field degraded images and dark-field degraded images; the bright-field degraded images and dark-field degraded images are gene chip images acquired in the same area by the same gene chip scanner;

[0185] The first de-degradation module 1020 is used to perform de-degradation estimation on random noise based on a self-supervised de-degradation model for each iteration of training, so as to obtain the bright field estimated image and de-degradation estimated data.

[0186] The second de-degradation module 1030 is used to perform de-degradation processing on the dark field degraded image based on the de-degradation estimation data to obtain the dark field estimated image;

[0187] The first determining module 1040 is used to determine the comprehensive degradation loss based on at least one of the bright field degraded image, the bright field estimated image, the dark field degraded image, and the dark field estimated image;

[0188] Model training module 1050 is used to perform self-supervised training on the self-supervised de-degradation model based on the comprehensive degradation loss;

[0189] The second determining module 1060 is used to determine the bright field estimated image output by the last iteration as the bright field de-degradation image and the dark field estimated image output by the last iteration as the dark field de-degradation image, with the goal of minimizing the overall degradation loss.

[0190] In one embodiment, the first de-degradation module includes: a noise coding unit, configured to perform feature encoding on random noise based on a shared coding network in a self-supervised de-degradation model, to obtain noise coding features of random noise for different task networks; wherein, the task networks include a deblurring network corresponding to a deblurring task and a decontamination network corresponding to a decontamination task; the de-degradation estimation data includes deblurring estimation data and decontamination estimation data; a first processing unit, configured to perform feature processing on the corresponding noise coding features based on the deblurring network, to obtain a bright-field estimated image and deblurring estimation data; and a second processing unit, configured to perform feature processing on the corresponding noise coding features based on the decontamination network, to obtain decontamination estimation data.

[0191] In one embodiment, the second de-degradation module includes: a third processing unit, configured to perform de-degradation processing on the dark field degraded image using decontamination estimation data to obtain a decontamination estimation image; and a fourth processing unit, configured to perform deblurring processing on the decontamination estimation image using deblurring estimation data to obtain a dark field estimation image.

[0192] In one embodiment, the first determining module includes: a first determining unit, configured to determine a bright-field reconstruction loss based on the bright-field degraded image and the bright-field estimated image; and a second determining unit, configured to determine a dark-field reconstruction loss based on the dark-field degraded image and the dark-field estimated image; wherein the combined degradation loss includes the bright-field reconstruction loss and / or the dark-field reconstruction loss.

[0193] In one embodiment, the first determining unit is configured to: perform degradation processing on the bright-field estimated image to obtain a bright-field degradation estimated image; and determine the bright-field reconstruction loss based on the bright-field degradation estimated image and the bright-field degradation image.

[0194] In one embodiment, the second determining unit is configured to: perform degradation processing on the dark field estimated image to obtain a dark field degradation estimated image; and determine the dark field reconstruction loss based on the dark field degradation estimated image and the dark field degradation image.

[0195] In one embodiment, the first determining module further includes: a third determining unit, configured to determine a bright-field similarity loss based on the similarity between image patches in the bright-field estimated image; and a fourth determining unit, configured to determine a dark-field similarity loss based on the similarity between image patches in the dark-field estimated image; wherein the comprehensive degradation loss further includes bright-field similarity loss and / or dark-field similarity loss.

[0196] In one embodiment, the third determining unit is specifically used to: divide the bright field estimated image into multiple first image blocks; for each first image block, determine the bright field similarity sub-loss of the first pixel block based on the similarity between the first image block and its neighboring first pixel blocks; and determine the bright field similarity loss based on the bright field similarity sub-loss of each first pixel block.

[0197] In one embodiment, the fourth determining unit is specifically used to: divide the dark field estimated image into a plurality of second image blocks; for each second image block, determine the dark field similarity sub-loss of the second pixel block based on the similarity between the second image block and its neighboring second pixel blocks; and determine the dark field similarity loss based on the dark field similarity sub-loss of each second pixel block.

[0198] In one embodiment, the first determining module further includes: a fifth determining unit, configured to determine the bright field gradient loss of the bright field estimated image based on the pixel gradient corresponding to each pixel in the bright field estimated image; and a sixth determining unit, configured to determine the dark field gradient loss of the dark field estimated image based on the pixel gradient corresponding to each pixel in the dark field estimated image; wherein the comprehensive degradation loss further includes the bright field gradient loss and / or the dark field gradient loss.

[0199] In one embodiment, the first determining module further includes: a seventh determining unit, used to determine the regularization loss of the de-degradation estimation data; wherein the comprehensive degradation loss also includes the regularization loss.

[0200] In one embodiment, the de-degradation estimation data includes decontamination estimation data; the seventh determining unit is specifically used to: determine a first regularization loss based on the pixel value change gradient between each pixel and its neighboring pixels in the decontamination estimation data; and determine a second regularization loss of the decontamination estimation data based on a first preset norm function; wherein the regularization loss includes the first regularization loss and / or the second regularization loss.

[0201] In one embodiment, the dedegradation estimation data includes defuzzification estimation data; the seventh determining unit is specifically used to: determine the third regularization loss of the defuzzification estimation data based on the second preset norm function; perform center constraints on the defuzzification estimation data according to the first preset Gaussian weight matrix, and determine the fourth regularization loss of the center constraint result based on the third preset norm function; wherein, the regularization loss includes the third regularization loss and / or the fourth regularization loss.

[0202] Each module in the aforementioned image de-degradation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0203] In one exemplary embodiment, a computer device is provided, the internal structure of which can be as shown in the figure. Figure 11 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements an image de-degradation method.

[0204] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0205] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the image de-degradation methods provided in the above embodiments.

[0206] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the image de-degradation methods provided in the above embodiments.

[0207] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the image de-degradation methods provided in the above embodiments.

[0208] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0209] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0210] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0211] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An image de-degradation method, characterized in that, include: Acquire bright-field degraded images and dark-field degraded images; The bright-field degraded image and the dark-field degraded image are gene chip images acquired in the same area by the same gene chip scanner; For each iteration of training, based on the shared coding network in the self-supervised de-degradation model, the random noise is feature-encoded to obtain the noise coding features of the random noise for different task networks; wherein, the task networks include the deblurring network corresponding to the deblurring task and the decontamination network corresponding to the decontamination task. Based on the deblurring network, feature processing is performed on the corresponding noise coding features to obtain the bright field estimated image and the deblurring estimated data. Based on the decontamination network, feature processing is performed on the corresponding noise coding features to obtain decontamination estimation data; Using the decontamination estimation data, the dark field degraded image is decontamination processed to obtain a decontamination estimation image; Using the deblurred estimation data, the decontamination estimation image is deblurred to obtain a dark field estimation image; The overall degradation loss is determined based on at least one of the bright-field degraded image, the bright-field estimated image, the dark-field degraded image, and the dark-field estimated image; Based on the comprehensive degradation loss, the self-supervised de-degradation model is trained under self-supervised conditions. With the goal of minimizing the overall degradation loss, the bright-field estimated image output in the last iteration is determined as the bright-field de-degradation image, and the dark-field estimated image output in the last iteration is determined as the dark-field de-degradation image.

2. The method according to claim 1, characterized in that, The step of determining the comprehensive degradation loss based on at least one of the bright-field degraded image, the bright-field estimated image, the dark-field degraded image, and the dark-field estimated image includes: The brightfield reconstruction loss is determined based on the brightfield degraded image and the brightfield estimated image; The dark field reconstruction loss is determined based on the dark field degraded image and the dark field estimated image; The overall degradation loss includes the bright-field reconstruction loss and / or the dark-field reconstruction loss.

3. The method according to claim 2, characterized in that, The step of determining the brightfield reconstruction loss based on the degraded brightfield image and the estimated brightfield image includes: The bright-field estimated image is subjected to degradation processing to obtain a bright-field degraded estimated image; The bright-field reconstruction loss is determined based on the bright-field degradation estimation image and the bright-field degradation image.

4. The method according to claim 2, characterized in that, The step of determining the dark field reconstruction loss based on the dark field degraded image and the dark field estimated image includes: The dark field estimated image is subjected to degradation processing to obtain a dark field degradation estimated image; The dark field reconstruction loss is determined based on the dark field degradation estimation image and the dark field degradation image.

5. The method according to claim 2, characterized in that, The step of determining the comprehensive degradation loss based on at least one of the bright-field degraded image, the bright-field estimated image, the dark-field degraded image, and the dark-field estimated image further includes: The bright-field similarity loss is determined based on the similarity between image patches in the estimated bright-field image. The dark field similarity loss is determined based on the similarity between image patches in the dark field estimated image; The overall degradation loss further includes the bright-field similarity loss and / or the dark-field similarity loss.

6. The method according to claim 5, characterized in that, The step of determining the bright-field similarity loss based on the similarity between image patches in the estimated bright-field image includes: The bright-field estimated image is divided into multiple first image blocks; For the first image patch, the bright-field similarity sub-loss of the first image patch is determined based on the similarity between the first image patch and its neighboring first image patches; The bright field similarity loss is determined based on the bright field similarity sub-loss of each first image block.

7. The method according to claim 5, characterized in that, The step of determining the dark field similarity loss based on the similarity between image patches in the dark field estimated image includes: The dark field estimated image is divided into multiple second image blocks; For each second image patch, the dark field similarity sub-loss of the second image patch is determined based on the similarity between the second image patch and its neighboring second image patches; The dark field similarity loss is determined based on the dark field similarity sub-loss of each second image block.

8. The method according to claim 2, characterized in that, The step of determining the comprehensive degradation loss based on at least one of the bright-field degraded image, the bright-field estimated image, the dark-field degraded image, and the dark-field estimated image further includes: The bright field gradient loss of the bright field estimated image is determined based on the pixel gradient corresponding to each pixel in the bright field estimated image. The dark field gradient loss of the dark field estimated image is determined based on the pixel gradient corresponding to each pixel in the dark field estimated image. The overall degradation loss further includes the bright field gradient loss and / or the dark field gradient loss.

9. The method according to claim 2, characterized in that, The step of determining the comprehensive degradation loss based on at least one of the bright-field degraded image, the bright-field estimated image, the dark-field degraded image, and the dark-field estimated image further includes: Determine the regularization loss for the de-degradation estimate data; The overall degradation loss also includes regularization loss.

10. The method according to claim 9, characterized in that, The de-degradation estimation data includes decontamination estimation data; The determination of the regularization loss for the de-degradation estimation data includes: The first regularization loss is determined based on the pixel value change gradient between each pixel and its neighboring pixels in the decontamination estimation data. Based on the first preset norm function, the second regularization loss of the decontamination estimation data is determined; The regularization loss includes the first regularization loss and / or the second regularization loss.

11. The method according to claim 9, characterized in that, The de-degradation estimation data includes deblurred estimation data; correspondingly, the regularization loss for determining the de-degradation estimation data includes: Based on the second preset norm function, the third regularization loss of the defuzzified estimated data is determined; Based on the first preset Gaussian weight matrix, the defuzzified estimation data is subjected to center constraints, and based on the third preset norm function, the fourth regularization loss of the center constraint result is determined. The regularization loss includes the third regularization loss and / or the fourth regularization loss.

12. An image de-degradation apparatus, characterized in that, include: The image acquisition module is used to acquire bright-field degraded images and dark-field degraded images; The bright-field degraded image and the dark-field degraded image are gene chip images acquired in the same area by the same gene chip scanner; The first de-degradation module includes a noise encoding unit, used for feature encoding of random noise based on the shared encoding network in the self-supervised de-degradation model for each iteration of training, to obtain noise encoding features of the random noise for different task networks; wherein, the task networks include a deblurring network corresponding to the deblurring task and a decontamination network corresponding to the decontamination task; a first processing unit is used for feature processing of the corresponding noise encoding features based on the deblurring network to obtain a bright-field estimated image and deblurred estimation data; a second processing unit is used for feature processing of the corresponding noise encoding features based on the decontamination network to obtain decontamination estimation data; The second de-degradation module includes: a third processing unit, used to perform de-degradation processing on the dark field degraded image using the decontamination estimation data to obtain a decontamination estimation image; and a fourth processing unit, used to perform deblurring processing on the decontamination estimation image using the deblurring estimation data to obtain a dark field estimation image. The first determining module is used to determine the comprehensive degradation loss based on at least one of the bright-field degraded image, the bright-field estimated image, the dark-field degraded image, and the dark-field estimated image; The model training module is used to perform self-supervised training on the self-supervised de-degeneration model based on the comprehensive degradation loss. The second determining module is used to determine the bright field estimated image output by the last iteration as the bright field de-degradation image and the dark field estimated image output by the last iteration as the dark field de-degradation image, with the goal of minimizing the comprehensive degradation loss.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 11.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 11.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 11.

Images