Image correction method, model and system based on deep learning

Abstract

The present invention discloses an image correction method, model and system based on deep learning, which belongs to the field of image processing. In view of the problem that the existing distorted image correction effect is poor and the cost is high, the present invention provides an image correction method based on deep learning, which includes constructing a model and using the model to correct the image, wherein the model construction includes splicing the processed distorted image and the corresponding super-resolution image to form a first training data set; and training the image correction model. The present invention introduces super-resolution images in the image correction model training process, so that the image clarity and detail expression are improved when the distorted image is corrected, so that it can adapt to complex and changeable interference environments, further improve the correction effect, and ensure that the final corrected image quality is higher and more reliable; at the same time, the introduction of super-resolution images enhances its generalization ability when processing different types of distorted images, so as to achieve the purpose of significantly improving the correction effect.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and more specifically, relates to an image correction method, model and system based on deep learning. Background Art

[0002] In modern industrial environments, there are many types of electrical equipment, and the electromagnetic environment is becoming increasingly complex, which has led to an increase in the number of image interference phenomena. Due to the randomness of these interference signals, traditional computer vision methods have difficulty in effectively correcting distorted images, resulting in a decrease in the image quality of the equipment, affecting product detection and identification. Existing technologies mainly use hardware shielding, filters and other means to reduce interference, but hardware shielding and filters are expensive and difficult to completely eliminate interference, making them ineffective in dealing with high-intensity interference and random interference. At the same time, traditional image correction methods are difficult to cope with complex and changing interference environments.

[0003] Corresponding improvements have also been made to the above problems, such as Chinese patent application number CN202211112047.8, published on November 10, 2023. The patent discloses an image processing method, which includes: obtaining a document image; inputting the document image into a classification model, and obtaining classification information of the document image through the classification model; processing the document image based on the classification information according to at least one of image enhancement processing based on an enhancement model and image correction processing based on a correction model to obtain a preprocessed image, and the preprocessed image carries classification information; performing color model processing on the preprocessed image based on the classification information to obtain a target image corresponding to the document image, and the color model processing is used to adjust at least one of the saturation, contrast and brightness of the preprocessed image. The shortcoming of this patent is that although model processing after classifying the document image greatly improves the work efficiency, the overall image processing effect still needs to be improved.

[0004] Another example is Chinese patent application number CN202310739205.0, published on November 21, 2023. The patent discloses a high-quality image generation method based on an improved generative adversarial network, including the following steps: obtaining a training data set; constructing an image generation network, improving the generator and discriminator network; training the improved GAN network, by alternately training the generator and the discriminator, so that the generator gradually learns to generate text image data close to the real image; generating high-quality text image data, using the trained generator, inputting a noise vector, and generating high-quality text image data. The shortcoming of this patent is that the data diversity in the training data set is insufficient, which affects its generalization ability, resulting in the final image processing effect is still not ideal. Summary of the invention

[0005] 1. Problems to be solved

[0006] In view of the problem that the existing distorted image correction has poor effect and high cost, the present invention provides an image correction method, model and system based on deep learning. The present invention introduces super-resolution images in the image correction model training process, so that the image clarity and detail performance are improved when correcting the distorted image, so that it can adapt to the complex and changeable interference environment, further improve the correction effect, and ensure that the final corrected image quality is higher and more reliable; at the same time, the introduction of super-resolution images enhances its generalization ability when processing different types of distorted images, so as to achieve the purpose of significantly improving the correction effect; and the correction of the distorted image through the model does not need to reduce interference through hardware shielding, filters and other means, thereby reducing costs.

[0007] 2. Technical solution

[0008] To solve the above problems, the present invention adopts the following technical solutions.

[0009] A deep learning-based image correction method includes constructing a model and correcting an image using the model; constructing the model includes the following steps:

[0010] S1. Obtaining a first training data set, specifically including:

[0011] S11: Acquire a first picture set, perform distortion processing and super-resolution processing on each first picture in the first picture set, and obtain a one-to-one corresponding distorted picture and a corresponding super-resolution picture;

[0012] S12: splicing the processed distorted image and the corresponding super-resolution image to form a first training data set;

[0013] S2. Training an image correction model: putting the first training data set into an image correction model for training, and finally obtaining an image correction model.

[0014] Furthermore, the super-resolution processing step in step S11 includes:

[0015] Constructing a super-resolution network model, and performing super-resolution processing on each first image in the first image set based on the super-resolution network model to obtain a corresponding super-resolution image;

[0016] and / or:

[0017] Based on image processing technology and mathematical algorithms, super-resolution processing is performed on each first picture in the first picture set to obtain a corresponding super-resolution picture.

[0018] Furthermore, the construction of the super-resolution network model includes the following steps:

[0019] S111: Obtain a second training data set: Obtain a second picture set, perform super-resolution processing on each second picture in the second picture set to obtain a super-resolution picture corresponding to each second picture; splice the processed super-resolution picture with the corresponding second picture to form a second training data set;

[0020] S112: Training a super-resolution network model: putting the second training data set into a super-resolution network model for training, and finally obtaining a super-resolution network model.

[0021] Furthermore, the super-resolution network model and the image correction model both adopt a GAN network, and the generation network in the GAN network adopts a Unet structure, and the discrimination network in the GAN network adopts a CNN structure.

[0022] Furthermore, the distortion processing in step S11 includes the following steps:

[0023] Obtaining an initial picture set, where the initial picture set is a picture set generated by a device;

[0024] Screening the initial picture set to obtain a real distorted picture set and a normal picture set, wherein the real distorted picture set is a distorted picture set generated by the device;

[0025] Analyze the real distorted image set and find the corresponding initial distortion law;

[0026] According to the initial distortion rule, a periodic offset in a certain direction or certain directions is applied to each normal image in the normal image set to obtain an artificially generated theoretical distorted image set;

[0027] Correcting the theoretical distorted image set according to the real distorted image set to obtain a corrected theoretical distorted image set;

[0028] The initial distortion law is adjusted according to the corrected theoretical distortion image set to obtain the final actual distortion law;

[0029] According to the actual distortion law, a periodic offset in a certain direction or certain directions is applied to each normal picture in the normal picture set to obtain the final artificially generated distorted picture set.

[0030] Furthermore, obtaining the initial picture set also includes data expansion of the initial picture set: initial pictures in the initial picture set are processed into blocks one by one to obtain an expanded initial picture set.

[0031] Furthermore, the method for correcting an image using the model is: inputting a distorted image and / or a normal image, and outputting a normal and super-resolution image and / or a super-resolution image through the image correction model.

[0032] A deep learning-based image correction model, wherein the deep learning-based image correction model is used to execute the deep learning-based image correction method as described in any one of the above items.

[0033] An image correction system based on deep learning, comprising:

[0034] Model building module: used to build an image correction model; wherein the construction of the image correction model includes the following steps:

[0035] S1: Obtaining a first training data set, specifically including:

[0036] S11: Acquire a first picture set, perform distortion processing and super-resolution processing on each first picture in the first picture set, and obtain a one-to-one corresponding distorted picture and a corresponding super-resolution picture;

[0037] S12: splicing the processed distorted image and the corresponding super-resolution image to form a first training data set;

[0038] S2: Training the image correction model: Put the first training data set into the image correction model for training, and finally obtain the image correction model

[0039] Correction module: used to input the image into the image correction model for correction.

[0040] 3. Beneficial effects

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) The present invention corrects distorted images by constructing an image correction model, thereby improving efficiency and image quality. At the same time, super-resolution images are introduced in the image correction model training process. Super-resolution images provide more pixel information and details than original low-resolution images, so that when correcting distorted images, image clarity and detail performance are improved, thereby being able to adapt to complex and changeable interference environments, further improving the correction effect, and ensuring that the final corrected image quality is higher and more reliable. At the same time, the introduction of super-resolution images increases the diversity of image correction model data, which helps the image correction model learn more knowledge and skills about image processing, thereby enhancing its generalization ability when processing different types of distorted images, thereby achieving the purpose of significantly improving the correction effect.

[0043] (2) The super-resolution processing in the present invention can use a super-resolution network model to obtain a super-resolution image. This method can prevent image information from being lost and has stronger capabilities and higher flexibility in restoring image details, improving image quality, and processing complex scenes. It can also use traditional methods based on image processing technology and mathematical algorithms to obtain super-resolution images. This method usually has low computational complexity, high computational efficiency, and is easy to implement. Both methods can be selected and adjusted according to actual conditions and have high flexibility.

[0044] (3) The super-resolution network model and image correction model in the present invention both use GAN networks. The generator and discriminator in the GAN network are continuously trained against each other and the competitive learning mechanism helps to continuously optimize the model and improve the quality of the generated image and the correction effect. The structure of the GAN network can be flexibly adjusted according to specific task requirements and has strong adaptability. The generated network adopts the Unet structure, and the feature splicing between different levels can realize the sharing of underlying information, improve the ability to restore details and enhance the robustness of the model, so that it can perform better in different scenarios.

[0045] (4) When distorting an image, the present invention obtains an initial distortion law by analyzing a set of real distorted images, and then artificially distorts the normal image. The artificially distorted image is then corrected using the set of real distorted images. The initial distortion law is adjusted using the corrected distorted image to obtain a final distortion law. The normal image is distorted using the final distortion law to quickly obtain a theoretical distorted image that is highly similar to the real distorted image, thereby increasing the authenticity of the data. At the same time, in order to ensure the data volume, the initial image set is expanded, which not only increases the data volume but also makes the data set more detailed, thereby ensuring that the correction capability of the image correction model is greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Generate sample images for distorted images;

[0047] Figure 2 Generate sample images for super-resolution images;

[0048] Figure 3 Splice sample graphs for the first training data set;

[0049] Figure 4 Correct sample images for images;

[0050] Figure 5 Schematic diagram of the construction of the image correction model. DETAILED DESCRIPTION

[0051] The present invention is further described below in conjunction with specific embodiments and drawings.

[0052] Example 1

[0053] A method for image correction based on deep learning includes constructing a model and using the model to correct an image. Specifically, the model constructed in this application is as follows: Figure 5 As shown, the following steps are included:

[0054] S1: Obtaining a first training data set, specifically including:

[0055] S11: obtaining a first picture set, performing distortion processing and super-resolution processing on each first picture in the first picture set, and obtaining a one-to-one corresponding distorted picture and a corresponding super-resolution picture; it is worth noting in this step that each first picture in the first picture set here is a normal picture, that is, a non-distorted picture;

[0056] S12: The processed distorted image and the corresponding super-resolution image are spliced ​​to form a first training data set such as Figure 3 As shown; in this step, the distorted image can be used as the left half of the spliced ​​image, and the super-resolution image can be used as the right half of the spliced ​​image;

[0057] S2: training the image correction model: putting the first training data set into the image correction model for training, setting appropriate training hyperparameter values ​​and selecting a suitable optimizer, and finally obtaining the image correction model.

[0058] The present invention corrects the distorted image by constructing an image correction model, thereby improving efficiency and image quality. At the same time, a super-resolution image is introduced in the image correction model training process. The super-resolution image provides more pixel information and details than the original low-resolution image, so that the image clarity and detail performance are improved when the distorted image is corrected, so that it can adapt to complex and changeable interference environments, further improve the correction effect, and ensure that the final corrected image quality is higher and more reliable. At the same time, the introduction of the super-resolution image increases the diversity of the image correction model data, which helps the image correction model learn more knowledge and skills about image processing, thereby enhancing its generalization ability when processing different types of distorted images, and achieving the purpose of significantly improving the correction effect. In addition, the super-resolution image provides more precise image information, which helps the model learn a more accurate image reconstruction method during the training process, can reduce errors and distortions in the correction process, and improve correction accuracy.

[0059] The applicant also verified the effect and used the image correction model of the present application to perform anti-distortion processing on industrial equipment images, and compared the image quality before and after correction. The results showed that the clarity and details of the corrected images were significantly improved, and its PSNR and SSIM indicators achieved the expected results.

[0060] It is worth noting that the training data set of the existing image correction model usually uses a data set of spliced ​​normal images and distorted images. This data set mainly focuses on the geometric correction of the image, and has limited effect on improving the image quality (such as clarity and detail recovery); and the types of distortions in the training data are not rich enough, and the generalization ability of the model may be limited, resulting in limited correction effect on the image. The present application is mainly based on the fact that in modern industrial environments, there are many types of electrical equipment, the electromagnetic environment is increasingly complex, and the phenomenon of interference to the image is also increasing. Due to the randomness of these interference signals, the distortion of the image is also random. Traditional computer vision methods are difficult to achieve effective correction of distorted images, resulting in a decrease in the image quality of the equipment, affecting the detection and identification of the product, and the existing image correction model is difficult to achieve a good correction effect on randomly generated distorted images. Therefore, the image correction method based on deep learning in this application is used, which makes it have a good correction effect on distorted images caused by high-intensity interference and random interference.

[0061] In one embodiment, the super-resolution processing step in step S11 includes: constructing a super-resolution network model, and performing super-resolution processing on each first image in the first image set based on the super-resolution network model to obtain a corresponding super-resolution image such as Figure 2 As shown. In this embodiment, a super-resolution network model is selected for super-resolution processing, and the model can automatically learn complex nonlinear mapping relationships from a large amount of data, so that when processing super-resolution tasks, it can more accurately restore the high-frequency details and texture information of the image, and there is no easy information loss like the traditional method; and the model can extract local and global features in the image, including edges, textures, shapes, etc. These feature information is crucial for super-resolution processing and helps to generate more realistic and natural high-resolution images; More importantly, the model usually adopts an end-to-end training method, that is, directly from low-resolution image input to high-resolution image output Overall optimization, this method simplifies the image processing process and avoids the complex pre-processing and post-processing steps in the traditional method; Therefore, the use of super-resolution models for super-resolution processing has stronger capabilities and higher flexibility in restoring image details, improving image quality, and processing complex scenes. For the super-resolution network model, its input is a picture degraded from a high-definition picture (such as camera blur, sensor noise, picture compression, ringing, and pictures caused by overshoot artifacts); the output is a super-resolution picture.

[0062] Specifically, the construction of the super-resolution network model includes the following steps:

[0063] S111: Obtain a second training data set: Obtain a second picture set, perform super-resolution processing on each second picture in the second picture set to obtain a super-resolution picture corresponding to each second picture; splice the processed super-resolution picture with the corresponding second picture to form a second training data set;

[0064] S112: training a super-resolution network model: placing the second training data set into a super-resolution network model for training, setting appropriate training hyperparameter values ​​and selecting an appropriate optimizer, and introducing image evaluation indicators such as PSNR and SSIM, and finally obtaining a super-resolution network model.

[0065] In one embodiment, the super-resolution processing step in step S11 includes performing super-resolution processing on each first image in the first image set based on image processing technology and mathematical algorithms to obtain a corresponding super-resolution image. It is worth noting that the methods based on image processing technology and mathematical algorithms include traditional methods such as interpolation methods, sparse representation-based methods, local embedding-based methods, and example-based methods. The use of these traditional methods for super-resolution processing has high computational efficiency because these methods usually have low computational complexity and are easy to understand and implement; and these methods can generate high-quality super-resolution images when the image noise is small and the resolution degradation model is known.

[0066] In one embodiment, it is worth noting that since image distortion is generally caused by signal interference, signal interference will affect the image quality of the entire device. Since signal interference has a certain randomness, only artificial distortion simulation of the image can be performed to achieve a similar effect; therefore, Figure 1 As shown, the distortion processing in step S11 includes the following steps:

[0067] Obtaining an initial picture set, where the initial picture set is a picture set generated by a device;

[0068] Screening the initial picture set to obtain a real distorted picture set and a normal picture set, wherein the real distorted picture set is a distorted picture set generated by the device; and the normal picture set is a normal picture set generated by the device;

[0069] Analyze the real distorted image set and find the corresponding initial distortion law;

[0070] According to the initial distortion rule, a periodic offset in a certain direction or certain directions is applied to each normal image in the normal image set to obtain an artificially generated theoretical distorted image set;

[0071] Correcting the theoretical distorted image set according to the real distorted image set to obtain a corrected theoretical distorted image set;

[0072] The initial distortion law is adjusted according to the corrected theoretical distortion image set to obtain the final actual distortion law;

[0073] According to the actual distortion law, a periodic offset in a certain direction or certain directions is applied to each normal picture in the normal picture set to obtain the final artificially generated distorted picture set.

[0074] In one embodiment, obtaining the initial picture set also includes data expansion of the initial picture set: the initial pictures in the initial picture set are processed one by one in blocks to obtain the expanded initial picture set. It is worth noting that due to signal interference, it is time-consuming and difficult to collect normal picture samples in the detection scene, resulting in too few detection samples; therefore, data expansion of the initial picture set can solve the problem of too little data in a certain scene and make the data set more detailed.

[0075] In one embodiment, the super-resolution network model and the image correction model both use a GAN network, and the generation network in the GAN network uses a Unet structure, and the discriminant network in the GAN network uses a CNN structure; the loss of the generation network consists of two parts: the adversarial loss and the L1 pixel loss between the network-generated correction image and the real correction image, and the loss of the identification network is achieved by generating adversarial loss. The GAN network is a generative adversarial network. The core idea of ​​the adversarial network is to generate images by learning the pixel probability distribution of images in the Ground Truth through adversarial training of the discriminator and the generator. In this embodiment, the generation network adopts a Unet structure, and the feature splicing between different levels can realize the sharing of underlying information. The loss function adopts a combination of adversarial loss and cycle consistency loss, so that the generated image has the high and low frequency information of the Ground Truth image, improves the detail recovery ability and enhances the robustness of the model, so that it can perform better performance in different scenarios; and a multi-scale generation network is used to process images at different scales and fuse high and low-order features to improve the details and clarity of the image.

[0076] It is worth noting that the present application introduces residual blocks and skip connections in the network, which can effectively avoid the problem of gradient dissipation and enhance the network training effect; it also introduces a self-attention mechanism in the network, which can capture long-distance dependencies in the image and improve the quality of image generation. In addition, when training the image correction model, the pre-trained model will be used for transfer learning, which can accelerate the training process, achieve good results on small sample data sets, and improve training efficiency.

[0077] In one embodiment, the method for correcting an image using the model is: inputting a distorted image and / or a normal image, and outputting a normal and super-resolution image and / or a super-resolution image through the image correction model. Figure 4As shown in the figure, when the input is a distorted image, the distorted image is output as a normal and super-resolved image after passing through the image correction model; when the input is a normal image, the normal image is output as a normal and super-resolved image after passing through the image correction model. That is, no matter what type of image is input to the image correction model, the image correction model will perform super-resolution and distortion correction operations on the input image to achieve an output image with higher quality.

[0078] Example 2

[0079] A deep learning-based image correction model, wherein the deep learning-based image correction model is used to execute the deep learning-based image correction method as described in the above-mentioned embodiment 1.

[0080] The image correction model in this embodiment introduces super-resolution images during training. During the model training process, these super-resolution images can be used as target outputs to guide the model to learn how to restore and enhance the details of the image. When the model corrects the distorted image, it can draw on the detail recovery ability learned from the super-resolution image, so that the corrected image is more outstanding in clarity and detail performance; at the same time, the data diversity of the model is increased, thereby enhancing its generalization ability when processing different types of distorted images. Therefore, when the distorted image enters the image correction model described in this application, a higher quality and more reliable corrected image will be output.

[0081] Example 3

[0082] An image correction system based on deep learning, comprising:

[0083] Model building module: used to build an image correction model; wherein the construction of the image correction model includes the following steps:

[0084] S1: Obtaining a first training data set, specifically including:

[0085] S11: Acquire a first picture set, perform distortion processing and super-resolution processing on each first picture in the first picture set, and obtain a one-to-one corresponding distorted picture and a corresponding super-resolution picture;

[0086] S12: splicing the processed distorted image and the corresponding super-resolution image to form a first training data set;

[0087] S2: Training the image correction model: Put the first training data set into the image correction model for training, and finally obtain the image correction model

[0088] Correction module: used to input the image into the image correction model for correction.

[0089] The image correction system of the present application uses an image correction model to correct distorted images, so that there is no need to reduce interference through hardware shielding, filters, etc. to obtain higher image operations, saving costs; and the distorted image is corrected by the image correction model described in the present application, and the corrected image quality is higher and more reliable.

[0090] The examples described in the present invention are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various modifications and improvements made to the technical solutions of the present invention by engineers and technicians in this field should all fall within the protection scope of the present invention.

Claims

1. An image correction method based on deep learning, characterized in that: The method includes constructing a model and using the model to correct an image; the model construction includes the following steps: S1: Get the first training data set: S11: Obtain a first picture set, perform distortion processing and super-resolution processing on each first picture in the first picture set, and obtain a one-to-one corresponding distorted picture and a corresponding super-resolution picture; the distortion processing in step S11 includes the following steps: Obtaining an initial picture set, where the initial picture set is a picture set generated by a device; Screening the initial picture set to obtain a real distorted picture set and a normal picture set, wherein the real distorted picture set is a distorted picture set generated by the device; Analyze the real distorted image set and find the corresponding initial distortion law; According to the initial distortion rule, a periodic offset in a certain direction or certain directions is applied to each normal image in the normal image set to obtain an artificially generated theoretical distorted image set; Correcting the theoretical distorted image set according to the real distorted image set to obtain a corrected theoretical distorted image set; The initial distortion law is adjusted according to the corrected theoretical distortion image set to obtain the final actual distortion law; According to the actual distortion law, a periodic offset in a certain direction or certain directions is applied to each normal image in the normal image set to obtain the final distorted image set artificially generated; S12: splicing the processed distorted image and the corresponding super-resolution image to form a first training data set; S2: training an image correction model: putting the first training data set into the image correction model for training, and finally obtaining an image correction model.

2. The image correction method based on deep learning according to claim 1, characterized in that: The super-resolution processing steps in step S11 include: Constructing a super-resolution network model, and performing super-resolution processing on each first image in the first image set based on the super-resolution network model to obtain a corresponding super-resolution image; and / or: Based on image processing technology and mathematical algorithms, super-resolution processing is performed on each first picture in the first picture set to obtain a corresponding super-resolution picture.

3. The image correction method based on deep learning according to claim 2, characterized in that: The construction of the super-resolution network model includes the following steps: S111: Obtain a second training data set: Obtain a second picture set, perform super-resolution processing on each second picture in the second picture set to obtain a super-resolution picture corresponding to each second picture; splice the processed super-resolution picture with the corresponding second picture to form a second training data set; S112: Training a super-resolution network model: putting the second training data set into a super-resolution network model for training, and finally obtaining a super-resolution network model.

4. The image correction method based on deep learning according to claim 2 or 3, characterized in that: The super-resolution network model and the image correction model both adopt a GAN network, and the generation network in the GAN network adopts a Unet structure, and the discrimination network in the GAN network adopts a CNN structure.

5. The image correction method based on deep learning according to claim 1, characterized in that: The process of obtaining the initial picture set also includes data expansion of the initial picture set: initial pictures in the initial picture set are processed into blocks one by one to obtain an expanded initial picture set.

6. The image correction method based on deep learning according to claim 1, characterized in that: The method for correcting an image using the model is as follows: inputting a distorted image and / or a normal image, and outputting a normal and super-resolution image and / or a super-resolution image through an image correction model.

7. An image correction system based on deep learning, characterized in that: include: Model building module: used to build an image correction model; wherein the construction of the image correction model includes the following steps: S1: Get the first training data set: S11: Obtain a first picture set, perform distortion processing and super-resolution processing on each first picture in the first picture set, and obtain a one-to-one corresponding distorted picture and a corresponding super-resolution picture; the distortion processing in step S11 includes the following steps: Obtaining an initial picture set, where the initial picture set is a picture set generated by a device; Screening the initial picture set to obtain a real distorted picture set and a normal picture set, wherein the real distorted picture set is a distorted picture set generated by the device; Analyze the real distorted image set and find the corresponding initial distortion law; According to the initial distortion rule, a periodic offset in a certain direction or certain directions is applied to each normal image in the normal image set to obtain an artificially generated theoretical distorted image set; Correcting the theoretical distorted image set according to the real distorted image set to obtain a corrected theoretical distorted image set; The initial distortion law is adjusted according to the corrected theoretical distortion image set to obtain the final actual distortion law; According to the actual distortion law, a periodic offset in a certain direction or certain directions is applied to each normal image in the normal image set to obtain the final distorted image set artificially generated; S12: splicing the processed distorted image and the corresponding super-resolution image to form a first training data set; S2: Training the image correction model: Put the first training data set into the image correction model for training, and finally obtain the image correction model Correction module: used to input the image into the image correction model for correction.

Images