A distortion correction method, device, equipment and computer readable storage medium

Abstract

The application discloses a distortion correction method, device and equipment and a computer readable storage medium, and the method comprises the following steps: generating a distortion model according to randomly generated distortion parameters, inputting an original image dataset into the distortion model to obtain a distorted image and distortion data; constructing a distortion correction network, training the distortion correction network by using the distorted image and the distortion data, and obtaining a distortion correction model; inputting the distorted image into the distortion correction model to obtain a coarsely corrected image; inputting the coarsely corrected image and the original image into a registration network to obtain a plurality of registration results and perform evaluation, and obtaining a spatial transformation matrix of a group of coarsely corrected images-original images with the best evaluation result; inputting a to-be-corrected distorted image into the distortion correction model to obtain a coarsely corrected image, and applying the spatial transformation matrix to the coarsely corrected image of the to-be-corrected distorted image to obtain a corrected image. The distortion correction method has good universality and high calculation efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing technology and optical display, and particularly relates to a distortion correction method and device, equipment and a computer readable storage medium. BACKGROUND

[0002] With the continuous development of imaging technology, various types of lenses are applied in different scenes, such as standard lenses for shooting ordinary scenes, wide-angle lenses for shooting wide scenes such as buildings and landscapes, long-focus lenses for long-distance imaging, macro lenses for shooting small objects such as insects and flowers, and fisheye lenses with very wide angles. However, due to the physical properties and manufacturing defects of lenses and optical elements in optical systems, distortion problems inevitably occur in imaging.

[0003] There are currently many solutions to the problem of imaging distortion. Mainly including the following several kinds: correction method based on multi-angle image calculation camera parameter, correction method based on specified features (such as edge lines, feature points, etc.) in the image, correction method based on image registration, and correction method through machine learning to calculate the mapping relationship between the images before and after distortion, etc. For example, Li et al. generated a large number of distorted images and data by using a variety of distortion parameters that meet the uniform distribution and applying them to different distortion models, and trained the generated data in a convolutional neural network based on supervised learning, thereby obtaining a model that can correct multiple distortion types (Li X, Zhang B, Sander P V, Liao J. Blind Geometric Distortion Correction on Images Through Deep Learning [C] / / 2019IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, CA, USA: IEEE, 2019: 4850-4859 [2023-04-23].). The Chinese patent document with application number CN202011344480.5 discloses an end-to-end radial distortion correction method based on a convolutional neural network, which uses randomly sampled distortion parameters to apply to a radial distortion model, generates distorted images and data and puts them into an end-to-end convolutional neural network for training, thereby obtaining a model for correcting radial distortion. The Chinese patent document with application number CN202010970309.9 discloses a distortion two-dimensional code recognition method and device, storage medium and electronic device, which detects feature points for specific two-dimensional code image data, and uses thin plate spline transformation on the image after selecting specific feature points to achieve distortion correction of two-dimensional code images based on image registration.

[0004] The main shortcomings of the prior art include:

[0005] The method of calculating camera parameters for correction relies on multi-angle image information, and multi-angle data of the same object cannot be obtained in all cases. In addition, the correction method based on camera parameters also needs to be manually fine-tuned multiple times, so the correction process is relatively complex.

[0006] The correction method based on specific features of an image relies on features in the distorted image, and if there are no available features or the available features are not obvious in the distorted data, the correction effect will be poor.

[0007] Most correction methods only model and correct one fixed type of distortion, and the size of the distortion that can be corrected is limited. For other types of distortion or images with large distortion, the correction result will be seriously affected.

[0008] The correction level of the existing correction method is limited, and the closer to the edge of the image, the more likely the correction result will be problematic, and this phenomenon is widespread. SUMMARY

[0009] In view of the shortcomings of the prior art, the present application provides a distortion correction method with good versatility and high computational efficiency.

[0010] The technical scheme of the present application is as follows:

[0011] A distortion correction method, comprising the following steps:

[0012] Define a plurality of distortion mathematical models, and generate a distortion model according to randomly generated distortion parameters, input the original image data set into the distortion model to obtain a distorted image and distortion data;

[0013] Construct a distortion correction network based on a convolutional neural network, and train the distortion correction network using the distorted image and the distortion data to obtain a distortion correction model;

[0014] Input the distorted image into the distortion correction model to obtain a coarse correction image;

[0015] Input the coarse correction image and the corresponding original image into a registration network to obtain a plurality of registration results;

[0016] Evaluate the registration results and select the best one coarse correction image-original image pair according to the evaluation results to obtain a spatial transformation matrix of the image pair;

[0017] The distortion correction model is inputted with the to-be-corrected distorted image to obtain a coarse correction image of the to-be-corrected distorted image, and the spatial transformation matrix is applied to the coarse correction image of the to-be-corrected distorted image to obtain a corrected image.

[0018] Preferably, the distortion mathematical model comprises at least one of six distortion models of barrel, pillow, rotation, cut, projection and wave.

[0019] Further, a set of random distortion parameters are generated by using a random sampling method satisfying uniform distribution, and a random sampling formula is as follows:

[0020]

[0021] wherein a and b are the starting point and the ending point of the interval respectively, and f(x) represents a probability density function of random value in the interval [a, b].

[0022] The randomly generated distortion parameters are brought into a specified distortion correction network to obtain a specific distortion model, and an original image set is inputted into the distortion model to obtain a distorted image and distortion data, wherein the distortion data records a flow field of the original image transformed to the distorted image.

[0023] The distortion correction network comprises a distortion parameter calculation network and a distortion type prediction network.

[0024] The distortion parameter calculation network comprises an encoder and a decoder, and the encoder and the decoder are respectively composed of three convolutional layers and five residual blocks, each residual block containing two convolutional layers; the down-sampling process of the encoder is realized by using a convolutional layer with a step of 2, and the up-sampling of the decoder is realized by using an inverse convolutional layer with a step of 2. The distortion parameter calculation network is used to output predicted distortion parameters.

[0025] The distortion type prediction network comprises an encoder and a classification network; the structure of the encoder is the same as that of the encoder in the distortion parameter calculation network; and the classification network is composed of two convolutional layers and one fully connected layer. The distortion type prediction network is used to output a predicted distortion type.

[0026] The distorted image is inputted into the distortion correction network, and after passing through the encoder, enters the decoder and the classification network respectively, the losses output by the distortion parameter calculation network and the distortion type prediction network are calculated respectively, and the network weights are constantly updated through back propagation, so as to complete the training of the distortion correction network.

[0027] During the training, the loss function of the distortion parameter calculation network is an endpoint error (EPE) to evaluate the distance between the predicted distorted image and the real image, and the endpoint error formula is as follows:

[0028]

[0029] Wherein: N is the total number of pixel points; u i And Respectively, the predicted distortion vector and the true vector. The smaller the value of EPE, the smaller the difference between the two vectors, that is, the more accurate the prediction.

[0030] During training, the loss function of the distortion type prediction network is the cross-entropy loss function, and its formula is as follows:

[0031]

[0032] Wherein, y i,j Indicates whether the true label of the i th sample belongs to the j th class; p i,j Indicates the probability that the model predicts that the i th sample belongs to the j th class.

[0033] The distortion image is input into the trained distortion correction model to obtain a rough correction image.

[0034] The rough correction image and the corresponding original image are input into the registration network, wherein the rough correction image is used as a floating image, and the original image is used as a reference image; the related parameters of the registration network are configured, and image registration operation is performed to obtain a plurality of registration results. Calculate the registration results of all images using evaluation indicators and record them.

[0035] Preferably, the registration results are evaluated, and the evaluation indicators include at least one of mean square error (MSE), structural similarity index measure (SSIM), dice coefficient (Dice Coefficient), and peak signal-to-noise ratio (PSNR).

[0036] Compare all registration result evaluation indicators, select a group of rough correction images-original images with the best evaluation indicators, obtain the spatial transformation matrix of the image pair, and sequentially perform spatial transformation on the to-be-corrected image sequence according to the spatial transformation matrix to obtain a final corrected image sequence.

[0037] The application also provides a distortion correction device, comprising:

[0038] A distortion image generation unit defines a plurality of distortion mathematical models, generates a distortion model according to randomly generated distortion parameters, inputs an original image dataset into the distortion model, and obtains a distortion image and distortion data;

[0039] The distortion coarse correction unit constructs a distortion correction network based on a convolutional neural network, trains the distortion correction network using the distortion image and distortion data, and obtains a distortion correction model; and inputs the distortion image into the distortion correction model to obtain a coarse correction image;

[0040] The registration unit inputs the coarse correction image and the corresponding original image into a registration network to obtain a plurality of registration results;

[0041] The correction quality evaluation unit evaluates the registration results, selects a group of coarse correction images and original images with the best evaluation result, and obtains a spatial transformation matrix of the image pair;

[0042] The image correction unit inputs the to-be-corrected distortion image into the distortion correction model to obtain a coarse correction image of the to-be-corrected distortion image, applies the spatial transformation matrix to the coarse correction image of the to-be-corrected distortion image, and obtains a corrected image.

[0043] The present application also provides a computer device comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the above-mentioned distortion correction method.

[0044] The present application also provides a computer-readable storage medium having a program stored thereon, wherein the program is executed by a processor to implement the above-mentioned distortion correction method.

[0045] Compared with the prior art, the present application has at least the following beneficial effects:

[0046] The present application divides the distortion correction of the image into two processes of coarse correction and fine correction, and when processing the distortion problem in a fixed imaging system, only one calculation is required, which can be applied to other distortion images, has good universality and high calculation efficiency.

[0047] The present application is the first image distortion correction method combining a deep learning correction network and an image registration algorithm, which utilizes deep learning to correct multiple types of distortion and combines the image registration algorithm to further improve the distortion correction effect.

[0048] The present application does not depend on fixed type features in the distortion image, and automatically extracts the required features through a convolutional neural network, so that it can correct distortion images of multiple styles. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 A flowchart of a distortion correction method provided by the embodiment of the present application is shown;

[0050] Figure 2 A model structure diagram of a distortion correction method provided by the embodiment of the present application is shown.

[0051] Figure 3 This is a schematic diagram of an image registration algorithm provided in an embodiment of the present invention;

[0052] Figure 4 The diagram shows the correction effect of the present invention and the prior art on barrel distortion data; where (a)-(e) are the input distorted image, the correction effect of reference 1, the correction effect of reference 2, the correction effect of the present invention, and the true value, respectively.

[0053] Figure 5 The diagram shows the correction effect of the present invention and the prior art on pincushion distortion data; where (a)-(e) are the input distorted image, the correction effect of reference 1, the correction effect of reference 2, the correction effect of the present invention, and the true value, respectively. Detailed Implementation

[0054] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not limit it in any way.

[0055] like Figure 1 As shown, the distortion correction method of the present invention includes the following steps:

[0056] (a) Define multiple distortion mathematical models and randomly generate distortion parameters;

[0057] (b) Generate a distortion model based on the distortion parameters and input it into the original image dataset to obtain distorted images and data;

[0058] (c) Construct a distortion correction network and train it with distorted images and data to obtain a distortion correction model;

[0059] (d) Put the distorted image obtained in step (b) into the distortion correction model to obtain the coarse correction image;

[0060] (e) Put the coarsely corrected image and the original image from step (a) into the registration model to obtain several registration results;

[0061] (f) Evaluate all registration results using an evaluation function;

[0062] (g) Preserve the spatial transformation matrix of the image pair with the best registration results;

[0063] (h) Input the distorted image to be corrected into the trained distortion correction model to obtain a coarse corrected image of the distorted image to be corrected. Apply the spatial transformation matrix to the coarse corrected image to obtain the corrected image.

[0064] Step (a) includes:

[0065] (a1) defining a plurality of mathematical models of distortion, including: six kinds of distortion models of barrel, pillow, rotation, cut, projection and wave, for performing distortion transformation on the input original image;

[0066] (a2) generating a set of random distortion parameters using a random sampling method satisfying uniform distribution, the formula is as follows:

[0067]

[0068] Wherein: a, b are the starting point and the end point of the interval respectively, and f(x) represents the probability density function of random value in the interval [a, b].

[0069] Step (b) includes:

[0070] (b1) bringing the randomly generated distortion parameters into the specified distortion model to obtain a specific distortion model;

[0071] (b2) putting the original image data set into the generated distortion model to obtain the distorted image and the distortion data, wherein the distortion data records the flow field of the original image transformed to the distorted image.

[0072] Step (c) includes:

[0073] (c1) constructing a distortion correction neural network model, which includes two branches of distortion parameter calculation network and distortion type prediction network;

[0074] (c2) putting the distorted image and data into the model for training.

[0075] The distortion parameter calculation network is composed of an encoder and a decoder, and the encoder and the decoder are respectively composed of three convolutional layers and five residual blocks, each residual block containing two convolutional layers; and the down-sampling process of the encoder is realized by using convolutional layer with a step of 2, and the up-sampling of the decoder is realized by using deconvolutional layer with a step of 2, and the network outputs the predicted distortion parameters.

[0076] The distortion type prediction network is composed of an encoder and a classification network, and the structure of the encoder is the same as that in the distortion parameter calculation network, and the classification part is composed of two convolutional layers and one fully connected layer, and the network outputs the predicted distortion type.

[0077] Step (c2) includes:

[0078] (c2-1) putting the distorted image as input into the double-branch network, and after passing through the encoder, entering the decoder and the classification prediction network respectively.

[0079] (c2-2) Calculate the loss of the double-branch network output respectively, and constantly update the network weight through back propagation, so as to complete the training of the network model. Wherein, two loss functions are included, the first one is the loss function of the distortion type prediction network, and the endpoint error (EPE) is selected to evaluate the distance between the predicted distortion image and the real image, and the formula is as follows:

[0080]

[0081] Wherein: N is the total number of pixel points; u i And The predicted distortion vector and the real vector respectively; the smaller the value of EPE, the smaller the difference between the two vectors, that is, the more accurate the prediction.

[0082] The second loss function is to evaluate the loss of the predicted distortion type and the real distortion type, and the cross-entropy loss is selected as the loss function, and the formula is as follows:

[0083]

[0084] Wherein, y i,j Indicates whether the real label of the i-th sample belongs to the j-th class; p i,j Indicates the probability that the model predicts that the i-th sample belongs to the j-th class.

[0085] Step (d) includes: putting the distortion image into the trained model for prediction to obtain the corrected result, which is recorded as the coarse correction image.

[0086] Step (e) includes:

[0087] (e1) Put the coarse correction image and the original image as input into the registration network, wherein the coarse correction image is taken as the floating image and the original image is taken as the reference image;

[0088] (e2) Configure the parameters related to the registration network, including: the number of down-sampling, the number of iterations, the interpolation method, the optimization function, the transformation type, etc., and perform image registration operation, so as to realize further correction of the coarse correction image.

[0089] Step (f) includes: calculating the evaluation index for all registration results to obtain the registration results of all images and record them;

[0090] Wherein, the evaluation indexes are respectively selected as a mean square error (MSE), a structural similarity index measure (SSIM), a DICE coefficient and a peak signal-to-noise ratio (PSNR).

[0091] Step (g) comprises: comparing the evaluation indexes of all registration results, and screening a transformation matrix of a group with the best evaluation index.

[0092] Step (h) comprises: according to the screened transformation matrix, sequentially performing spatial transformation on the pre-correction image sequence to obtain a final correction image sequence.

[0093] The application further provides a distortion correction device, comprising:

[0094] A distortion image generation unit comprises a plurality of distortion mathematical models and randomly generates distortion parameters, generates a distortion model according to the distortion parameters and inputs an original image data set to obtain a distortion image and data, and is used for training a distortion correction network;

[0095] A distortion correction network unit is used for loading a trained distortion correction network and performing coarse correction on input image data;

[0096] A registration algorithm unit is used for loading a coarse correction result and an original image as a floating image and a reference image respectively, performing registration on the images according to preset image registration parameters, saving the registered images and used transformation matrices;

[0097] A correction quality evaluation unit is used for loading a registered image and an original image sequence, evaluating all target images by specifying evaluation indexes, and screening a group with the best evaluation result, and retaining a corresponding transformation matrix;

[0098] An image correction unit is used for loading a transformation matrix, performing transformation on a coarse correction image of a distortion image according to the transformation matrix, and obtaining a corrected image.

[0099] The application further provides an electronic device comprising a memory and a processor, wherein the memory stores executable codes, and the processor is used for implementing the above-mentioned distortion correction method when executing the executable codes.

[0100] The application further provides a computer readable storage medium, wherein a computer program is stored on the computer readable storage medium, and the computer program is executed by a processor to implement the above-mentioned distortion correction method.

[0101] Embodiments

[0102] AsFigure 1 As shown, the present application corrects any imaging distortion data. First, the original image data is transformed according to the distortion model to obtain the distorted image and the parameters used for distortion, then the generated data is input into the distortion correction network for training to obtain the trained distortion correction model; then the model is used to correct any distortion image sequence to obtain a coarse correction image sequence; the coarse correction image sequence and the corresponding original image sequence are put into the image registration algorithm for further fine correction; after fine correction, all results are evaluated using the specified evaluation index, and the transformation matrix of the best result is retained; finally, the transformation matrix is applied to other images in the distortion image sequence to obtain the final corrected image sequence.

[0103] As shown, the original image is first input into the distortion model, which contains six different types of distortion, to obtain the distorted image and distortion data, then it is put into the double-branch distortion correction network for training to obtain the distortion correction network model; then the distortion image sequence is put into the correction network model for coarse correction to obtain the coarse correction image sequence, which is put into the registration algorithm with the original image sequence to obtain the registration result, and the evaluation index is used to evaluate the registration result, and the transformation matrix of the best image pair is selected and retained; finally, the transformation matrix is applied to all coarse correction image sequences to obtain the final corrected image sequence. Figure 2 As shown, the flowchart of the image registration algorithm in the present application is shown, wherein the main parameters include the number of downsampling, the number of iterations, the interpolation method, the optimization function, the transformation type, etc. The main process of the registration algorithm is as follows: first, the coarse correction result obtained by the distortion correction network and the original image are respectively taken as the floating image and the fixed image, then the images are pyramid multi-scale sampled according to the set registration parameters, then the floating image is interpolated and put into the loss function to evaluate the difference between the reference image and the floating image, the optimization function is used to optimize the difference, and the result of the loss function is fed back to the transformation matrix. Finally, when the loss function between the fixed image and the floating image does not meet the set value or the number of iterations does not reach the set value, the image is continuously optimized and the spatial transformation matrix is adjusted, otherwise the final registration result is directly output.

[0104] Figure 3 As shown, two sets of barrel distortion data are selected, and the distortion correction methods of the present application and references 1-4 are compared, and from the results it can be seen that the distortion correction method of the present application is closest to the true value.

[0105] As shown, two sets of barrel distortion data are selected, and the distortion correction methods of the present application and references 1-4 are compared, and from the results it can be seen that the distortion correction method of the present application is closest to the true value. Figure 4 As shown, two sets of barrel distortion data are selected, and the distortion correction methods of the present application and references 1-4 are compared, and from the results it can be seen that the distortion correction method of the present application is closest to the true value.

[0106] Figure 5 ​​As shown, two groups of pillow-shaped distortion data are selected, and the distortion correction comparison test is carried out by using the method of the present application and the methods of references 1-4 (reference 1: Li X, Zhang B, Sander P V, Liao J. Blind Geometric Distortion Correction on Images Through Deep Learning [C] / / 2019 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, CA, USA: IEEE, 2019: 4850-4859; reference 2: Klein, Stefan, Marius Staring, Keelin Murphy, Max A. Viergever, and Josien PW Pluim. "Elastix: a toolbox for intensity-based medical image registration." IEEE transactions on medical imaging 29, no. 1 (2009): 196-205; reference 3: Rosten, Edward, and Rohan Loveland. "Camera distortion self-calibration using the plumb-line constraint and minimal Hough entropy." Machine Vision and Applications 22 (2011): 77-85; reference 4: Liao, Kang, Chunyu Lin, Yao Zhao, and Moncef Gabbouj. "DR-GAN: Automatic radial distortion rectification using conditional GAN in real-time." IEEE Transactions on Circuits and Systems for Video Technology 30, no. 3 (2019): 725-733.), and the results show that the distortion correction method of the present application is closest to the true value.

[0107] The correction effects of the method of the present application and references 1-4 on the distortion data are shown in Table 1.

[0108] Table 1

[0109] Method Peak signal-to-noise ratio (PSNR) Structural similarity (SSIM) Reference 1 13.645 0.307 Reference 2 15.006 0.372 Reference 3 13.558 0.347 Reference 4 14.801 0.383 The present invention 21.118 0.736

[0110] The above-described embodiments have explained the technical solutions and beneficial effects of the present application in detail. It should be understood that the above-described is only a specific embodiment of the present application and is not used to limit the present application. Any modification, supplement and equivalent replacement, etc. made within the principle range of the present application shall be included in the protection scope of the present application.

Claims

1. A distortion correction method, characterized in that, Includes the following steps: Define several distortion mathematical models, and generate distortion models based on randomly generated distortion parameters. Input the original image dataset into the distortion models to obtain distorted images and distortion data. A distortion correction network is constructed based on a convolutional neural network, and the distortion correction network is trained using the distorted images and distortion data to obtain a distortion correction model. The distorted image is input into the distortion correction model to obtain a coarsely corrected image; The coarsely corrected image and the corresponding original image are input into the registration network to obtain several registration results; The registration results are evaluated, and the coarsely corrected image-original image with the best evaluation result is selected to obtain the spatial transformation matrix of the image pair. The distorted image to be corrected is input into the distortion correction model to obtain a coarse corrected image. The spatial transformation matrix is ​​then applied to the coarse corrected image to obtain the corrected image.

2. The distortion correction method according to claim 1, characterized in that, The distortion mathematical model includes at least one of six distortion models: barrel, pincushion, rotation, tangent, projection, and waveform.

3. The distortion correction method according to claim 1, characterized in that, A set of random distortion parameters is generated using a random sampling method that satisfies a uniform distribution; the random sampling formula is as follows: Where: a and b are the starting and ending points of the interval, respectively, and f(x) represents the probability density function of randomly taking values ​​within the interval [a,b].

4. The distortion correction method according to claim 1, characterized in that, The distortion correction network includes a distortion parameter calculation network and a distortion type prediction network; The distortion parameter calculation network includes an encoder and a decoder; the encoder and decoder are each composed of three convolutional layers and five residual blocks, and each residual block contains two convolutional layers; the downsampling process of the encoder is implemented using a convolutional layer with a stride of 2, and the upsampling of the decoder is implemented using a deconvolutional layer with a stride of 2. The distortion type prediction network includes an encoder and a classification network; the structure of the encoder is the same as that of the encoder in the distortion parameter calculation network. The classification network consists of two convolutional layers and one fully connected layer.

5. The distortion correction method according to claim 1, characterized in that, During training, the loss function of the network, calculated using distortion parameters, is the endpoint error, and its formula is as follows: Where: N is the total number of pixels; u i and These are the predicted distortion vector and the true vector, respectively. The loss function of the distortion type prediction network is the cross-entropy loss function, and its formula is as follows: Among them, y i,j Indicates whether the true label of the i-th sample belongs to the j-th category; p i,j This represents the probability that the model predicts the i-th sample belongs to the j-th category.

6. The distortion correction method according to claim 1, characterized in that, The registration results are evaluated using at least one of the following metrics: mean square error, structural similarity, DICE coefficient, and peak signal-to-noise ratio.

7. A distortion correction device, characterized in that, include: The distortion image generation unit defines several distortion mathematical models and generates distortion models based on randomly generated distortion parameters. The original image dataset is input into the distortion model to obtain distorted images and distortion data. The distortion coarse correction unit constructs a distortion correction network based on a convolutional neural network, and trains the distortion correction network using the distorted image and distortion data to obtain a distortion correction model; the distorted image is then input into the distortion correction model to obtain a coarsely corrected image; The registration unit inputs the coarsely corrected image and the corresponding original image into the registration network to obtain several registration results; The calibration quality evaluation unit evaluates the registration results, selects the coarsely calibrated image-original image with the best evaluation results, and obtains the spatial transformation matrix of the image pair. The image correction unit inputs the distorted image to be corrected into the distortion correction model to obtain a coarsely corrected image of the distorted image to be corrected, and applies the spatial transformation matrix to the distorted image to be corrected to obtain the corrected image.

8. A computer device, characterized in that, The device includes a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the distortion correction method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the distortion correction method as described in any one of claims 1-6.

Images