Distorted character recognition method and system
By constructing and iteratively training a character recognition model with character segmentation, distortion correction, and recognition networks, the problem of accuracy in recognizing distorted characters in complex backgrounds was solved, achieving efficient distorted character recognition and reducing data annotation costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTH CENTRAL UNIVERSITY FOR NATIONALITIES
- Filing Date
- 2023-12-09
- Publication Date
- 2026-04-10
AI Technical Summary
Existing text recognition methods suffer from decreased accuracy when faced with complex backgrounds, multi-directional fonts, characters of varying sizes, and distorted characters. They are particularly susceptible to interference in practical application scenarios such as photovoltaic power plants.
A text recognition model is constructed that includes a character segmentation network, a distortion correction network, and a character recognition network. Through pre-training and iterative updates, the data annotation cost is reduced, and image enhancement technology is combined to segment, correct, and recognize distorted characters.
It improved the accuracy of recognizing distorted characters, reduced data annotation costs, and enhanced the model's recognition capabilities in complex scenarios.
Smart Images
Figure CN121838162A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of character recognition, and in particular to a twisted character recognition method and system. BACKGROUND
[0002] About 91% of the information that humans recognize and understand about the world comes from vision. Similarly, computer vision is the basis for machines to understand the world and is a hot topic in artificial intelligence research. Character recognition is also an important research direction in artificial intelligence. In our lives, characters are ubiquitous, and our clothing, food, shelter and transportation all depend on them. Character recognition technology enables batch character input operations, which input a large amount of files, information, forms, data, etc. into computers, successfully replacing human resources and reducing labor costs, thereby greatly improving people's work and life efficiency.
[0003] Scene character recognition technology originated from traditional character recognition, which initially supported industrial production and social activities, such as zip code recognition for postal delivery, to improve work efficiency and reduce labor costs. Deep learning entered the field of view of people in 2012, and has attracted great attention in the academic field. The detection and recognition technology of characters has developed rapidly, and the realization of natural scene character intelligent real-time detection and recognition is a direction with a wide range of application scenarios and a hot topic in the field of artificial intelligence research.
[0004] Most of the commonly used character recognition methods currently input the image to be recognized as a whole into a character recognition model, first use a convolutional neural network to extract image features, then use a sequence feature extraction network that can obtain context information, and finally obtain the final character information through transcription. This character recognition method has achieved a high level of recognition accuracy for printed body recognition, high-quality image recognition, etc.
[0005] However, in actual application scenarios, images often have very complex backgrounds, non-uniform fonts, fonts with multiple orientations, different sizes, non-uniform languages, and distorted character fonts, which can significantly reduce the accuracy of character recognition. For example, in images that need to be recognized in the daily operation of a photovoltaic power station, there are often barriers, icons, and lines that are similar to characters and can easily be incorrectly recognized as characters, resulting in a decrease in accuracy. Therefore, there is an urgent need to provide a solution to improve the above problems. SUMMARY
[0006] The purpose of this invention is to provide a method and system for distorted character recognition. During the training phase of the character recognition model, the character segmentation network can be trained without additional character-level annotation, which greatly reduces the cost of data annotation. Furthermore, by combining the character segmentation network, the distortion correction network, and the character recognition network, the problem of decreased accuracy caused by character distortion can be reduced, thereby improving the accuracy of the character recognition model for distorted characters.
[0007] In a first aspect, the present invention provides a method for recognizing distorted characters, comprising the following steps:
[0008] A character recognition model with a character segmentation network, a distortion correction network, and a character recognition network is constructed, and the character segmentation network in the character recognition model is pre-trained based on a public character segmentation set to obtain a pre-trained character recognition model; wherein, the distortion correction network has a localization grid, a grid generator, and a sampler;
[0009] Image information containing distorted characters but without character segmentation annotations is obtained. Based on the pre-trained character recognition model, the distorted characters in the image information are segmented, recombined, distorted, and recognized as characters. The training text information is then output. The parameters of the pre-trained character recognition model are iteratively updated based on the training text information and image information to obtain the trained character recognition model.
[0010] The system acquires image information to be recognized, and performs character recognition on the characters in the image based on the trained character recognition model after cutting, recombining, and distortion correction, and outputs the recognized character information.
[0011] The beneficial effects of the distorted character recognition method provided by this invention are as follows: during the training phase of the character recognition model, the character segmentation network can be trained without additional character-level annotation, which greatly reduces the cost of data annotation. Furthermore, by combining the character segmentation network, the distortion correction network, and the character recognition network, the problem of accuracy decline caused by character distortion and deformation can be reduced, thereby improving the accuracy of the character recognition model for distorted characters.
[0012] Optionally, the process of segmenting and recombining characters in the image information based on the pre-trained character recognition model includes: preprocessing the image information and dividing it into a training set and a test set; the character segmentation network in the pre-trained character recognition model performs single-character segmentation and sequential recombination on the distorted characters in the training set, and outputs a recombined training image.
[0013] Optionally, the process of correcting the distortion of characters in the image information based on the pre-trained character recognition model includes: based on the tilt angle of the distorted characters in the reconstructed training image, the positioning grid in the distortion correction network of the pre-trained character recognition model locates reference points on the reconstructed training image; the grid generator generates a sampling network based on the positioning reference points; and the sampler performs horizontal correction on the distorted characters in the reconstructed training image based on the sampling network to obtain a corrected training image.
[0014] Optionally, the process of recognizing the distorted characters in the image information based on the pre-trained character recognition model includes: extracting image features and sequence features from the corrected training image based on the character recognition network in the pre-trained character recognition model, and outputting training text information based on the image features and the sequence features.
[0015] Optionally, the process of iteratively updating the parameters of the pre-trained text recognition model based on training text information and image information includes: training the pre-trained text recognition model based on the training set data to obtain a preliminary text recognition model, until the recognition accuracy of the preliminary text recognition model is greater than or equal to a preset accuracy threshold when tested based on the test set data, to obtain the trained text recognition model.
[0016] Optionally, the process of training the pre-trained character recognition model based on the training set includes: generating pseudo-labels with empty initial values; updating the values of the pseudo-labels based on the number of correctly recognized characters in the training text information; updating the parameters of the character segmentation network in the pre-trained character recognition model based on the values of the pseudo-labels; and updating the parameters of the distortion correction network and the character recognition network in the pre-trained character recognition model based on the character labels corresponding to the distorted characters in the training set and the reconstructed training image, thereby obtaining a primary character recognition model.
[0017] In a second aspect, the present invention provides a distorted character recognition system for implementing any of the above-mentioned optional distorted character recognition methods, including:
[0018] The model building module is used to build a text recognition model with a character segmentation network, a distortion correction network, and a character recognition network;
[0019] The pre-training module is used to pre-train the character segmentation network in the character recognition model to obtain the pre-trained character recognition model.
[0020] The model training module is used to acquire image information containing distorted characters and train the pre-trained character recognition model based on the image information to obtain the trained character recognition model.
[0021] The text recognition module is used to acquire image information to be recognized, and to perform text recognition after cutting, recombining and distorting the characters in the image to be recognized based on the trained text recognition model, and output the recognized text information.
[0022] Optionally, it also includes an image submission module for inputting the image information to be recognized into the text recognition module.
[0023] Optionally, it also includes an information display module, used to display a reconstructed image after the characters in the image to be recognized are cut and reconstructed based on the trained character recognition model, and to display the text information recognized based on the trained character recognition model. Attached Figure Description
[0024] Figure 1 This is a flowchart illustrating the steps of a distorted character recognition method according to an embodiment of the present invention;
[0025] Figure 2 This is a schematic diagram of the structure of the distortion correction network constructed in the embodiments of the present invention;
[0026] Figure 3 This is a schematic diagram of the character recognition network constructed in an embodiment of the present invention;
[0027] Figure 4 This is a schematic diagram of the structure of a distorted character recognition system according to an embodiment of the present invention.
[0028] Explanation of reference numerals in the attached figures:
[0029] 100. Model building module; 200. Pre-training module; 300. Model training module; 400. Text recognition module; 500. Image submission module; 600. Information display module; 700. Error collection module; 800. Model optimization module. Detailed Implementation
[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.
[0031] See Figure 1This invention provides a method for recognizing distorted characters, comprising the following steps:
[0032] S1. Model Construction: Construct a character recognition model with a character segmentation network, a distortion correction network, and a character recognition network;
[0033] S2. Model pre-training: The character segmentation network in the character recognition model is pre-trained based on the public character segmentation set to obtain the pre-trained character recognition model.
[0034] S3. Model Training: Obtain image information containing distorted characters, and perform character segmentation, recombination, distortion correction, and character recognition on the characters in the image information based on the pre-trained character recognition model to output training text information. Iteratively update the parameters of the pre-trained character recognition model based on the training text information and image information to obtain the trained character recognition model.
[0035] S4. Text Recognition: Obtain the image information to be recognized, and perform text recognition on the characters in the image to be recognized based on the trained text recognition model after cutting, recombining and distorting correction, and output the recognized text information.
[0036] In some embodiments, when performing step S1, see [reference needed]. Figure 2 As shown, the constructed warp correction network has a localization grid, a grid generator, and a sampler. The localization grid is used to locate a set of reference points on the image, and the grid generator generates a sampling network. Finally, the sampler generates a corrected image through the sampling network. Since both the localization grid and the grid generator are differentiable, the warp correction network can be trained through backpropagation using errors during training in step S3.
[0037] In some embodiments, when performing step S1, see [reference needed]. Figure 3 As shown, the constructed character recognition network is a network model based on an attention mechanism, which includes an encoder module and a decoder module. The encoder module can extract sequence features from the corrected image output by the distortion correction network and input the sequence features into the decoder module. The decoder module outputs the text sequence sequentially.
[0038] In some embodiments, when performing step S2, the publicly available character segmentation set used contains text information and the specific location of each character in the image. That is, the publicly available character segmentation set is a manually labeled dataset. The character segmentation network is pre-trained using this publicly available character segmentation set, enabling it to possess basic character segmentation capabilities. Specifically, after pre-training the character segmentation network using the publicly available character segmentation set, the pre-trained network can use a convolutional neural network with residual connections to predict the character positions in a given image, segment the characters in the given image, and output them sequentially. This removes useless background from the image, resulting in a reconstructed image containing only text information.
[0039] In some embodiments, before performing step S2 to pre-train the character segmentation network in the character recognition model based on the public character segmentation set, the public character segmentation set is pre-processed to improve data quality, such as resolution enhancement and contrast enhancement, and then integrated with the original segmentation set to obtain a larger dataset, thereby improving the pre-training effect of the character segmentation network.
[0040] In practice, character-level annotation is too costly, training data for character segmentation networks is difficult to obtain, and training solely with publicly available character segmentation sets leads to low accuracy in real-world applications. Therefore, pre-training the character segmentation network using the publicly available character segmentation set (as in step S2) followed by training it with image information containing distorted characters (as in step S3) significantly reduces training costs and improves the overall training performance of the model.
[0041] In some embodiments, the image information containing distorted characters used in step S3 includes text information and text labels, but does not include the position information of the text in the image. Thus, this image information is used to train the pre-trained text recognition model, and the training of the character segmentation model can be achieved without additional character-level annotation, which can reduce the cost of data annotation.
[0042] In some embodiments, image information containing distorted characters is sometimes of low quality, mainly due to factors such as dim lighting and blurry images caused by camera shake. Therefore, image enhancement is needed to facilitate the training of the character recognition model, reduce errors caused by blurry images, and improve the applicability of the trained character recognition model.
[0043] Specifically, when enhancing image information, image quality is improved by increasing the resolution and contrast of the image to be recognized, so that the image information can be more easily recognized by the text recognition model.
[0044] In some embodiments, super-resolution (SR) technology is used to enhance the resolution of image information.
[0045] However, with the development of deep learning and the continuous improvement of computing power in various hardware, more and better super-resolution technologies are emerging. Therefore, in some embodiments, supervised learning-based deep learning methods can also be used to implement super-resolution technology. Compared with difference-based and reconstruction-based super-resolution image reconstruction algorithms, supervised learning-based deep learning methods are more effective.
[0046] Specifically, when using supervised learning-based deep learning methods to enhance the resolution of image information, the high-resolution image is first downsampled into a low-resolution image, and then the low-resolution image is input into a neural network to train a super-resolution image, thereby completing the training of the super-resolution model.
[0047] The formula for downsampling a high-resolution image into a low-resolution image is shown below:
[0048] I x =D(I y ;δ)
[0049] I x =(I y )↓ s
[0050]
[0051] In the formula, I x I represents the low-resolution image output. y The input high-resolution image is represented by D(), the downsampling function is represented by s, the noise is represented by n, the blur kernel is represented by k, and the other parameters in the downsampling function are represented by δ; ↓ represents downsampling.
[0052] In some embodiments, image information is enhanced using both direct and indirect contrast enhancement methods.
[0053] Specifically, histogram stretching and histogram equalization are the most widely used indirect contrast enhancement methods. Histogram stretching adjusts the histogram by stretching the contrast, thereby increasing the difference in gray levels between the foreground and background to enhance contrast. Histogram equalization, on the other hand, adjusts the gray values through a cumulative function to enhance contrast.
[0054] Specifically, the grayscale statistical histogram of image information is a discrete function, as shown in the following formula:
[0055] p s(S k ) = N k / N(0≤s k ≤1)
[0056] In the formula, N represents the total number of pixels in the image information, and p s (s k ) represents the probability of the k-th gray level appearing in the image information, denoted by N. k s is the independent variable. k The k-th gray level value of the image information is represented by p. s (s k The curve obtained by the function is the histogram of image information.
[0057] Specifically, by processing image information using the histogram equalization formula, the histogram can be transformed into a uniformly distributed form. This enhances the dynamic range of pixel grayscale values in the image, thereby improving the overall contrast of the image. The histogram equalization formula is shown below:
[0058]
[0059] In the formula, N j This indicates the number of pixels at the current gray level, where N represents the total number of pixels in the image information, and L represents the total number of gray levels in the image.
[0060] In fact, when enhancing image information, sharpness and noise reduction can also be achieved through dehazing and noise reduction algorithms.
[0061] In fact, in step S3, the character segmentation network can segment and reassemble the image information sequentially to obtain a reassembled image. The reassembled image contains only the text information of distorted characters, thereby removing interference information in the image information and preserving the contextual information of the text sequence as much as possible. The distortion correction network can correct the distorted characters in the reassembled image, correcting distorted or angled characters into horizontal characters, which is beneficial for the character recognition network to recognize them better. The input of the character recognition network is the output of the distortion correction network, and the encoder and decoder extract the image features and contextual information of the corrected image information to recognize the text information in the image.
[0062] In some embodiments, the process of performing step S3, which involves segmenting and recombining characters in the image information based on a pre-trained character recognition model, includes the following steps:
[0063] After preprocessing, the image information is divided into training and testing sets.
[0064] The character segmentation network in the pre-trained text recognition model sequentially segments and reassembles the distorted characters in the training set, outputting a reassembled training image.
[0065] Specifically, after preprocessing the image information, including enhancing resolution and contrast, it is divided into a training set and a test set according to a preset ratio. The training set is used to train the text recognition model as a whole, while the test set is used to test the recognition accuracy of the trained model. In practice, the preset ratio for distinguishing between the training and test sets can be adaptively adjusted based on the size of the image information.
[0066] Specifically, in the process of the character segmentation network in the pre-trained text recognition model sequentially segmenting and recombining the distorted characters in the training set, the character segmentation network presets the position of the distorted characters in the image information based on the convolutional neural network with residual connections, and after segmenting the distorted characters, outputs and recombines them sequentially into a recombined training image containing only the text information of the distorted characters.
[0067] In some embodiments, the process of performing step S3 to correct the distortion of characters in the image information based on the pre-trained character recognition model includes: based on the tilt angle of the distorted characters in the reconstructed training image, the distortion correction network in the pre-trained character recognition model performs horizontal correction on the distorted characters in the reconstructed training image to obtain a corrected training image.
[0068] Specifically, when the distortion correction network performs horizontal correction on the distorted characters in the reconstructed training image, it locates a set of reference points for each distorted character in the reconstructed training image, generates a sampling network corresponding to each distorted character based on the reference points, and uses the corresponding sampling network to correct each distorted character separately. The corrected distorted characters are then reconstructed sequentially to obtain the corrected training image.
[0069] In some embodiments, the process of performing character recognition on image information based on the pre-trained character recognition model in step S3 includes: extracting image features and sequence features from the correction training image based on the character recognition network in the pre-trained character recognition model, and outputting training text information based on the image features and the sequence features.
[0070] Specifically, since the characters in the calibration training images are already a definite character sequence, text recognition can be converted into sequence recognition. After the text recognition network extracts the sequence features from the calibration training images, it predicts the text sequence in sequence, and can output the training text information recognized by the text recognition network.
[0071] In some embodiments, the process of iteratively updating the parameters of the pre-trained text recognition model based on training text information and image information in step S3 includes: training the pre-trained text recognition model based on training set data to obtain a preliminary text recognition model, until the recognition accuracy of the preliminary text recognition model is greater than or equal to a preset accuracy threshold when tested based on test set data, thus obtaining the trained text recognition model.
[0072] In practice, training the pre-trained character segmentation network using training set data that does not contain character segmentation annotations includes: generating pseudo-labels with empty initial values; updating the values of the pseudo-labels based on the number of correctly recognized characters in the training text information; updating the parameters of the character segmentation network in the pre-trained text recognition model based on the values of the pseudo-labels; and updating the parameters of the distortion correction network and the character recognition network in the pre-trained text recognition model based on the character labels corresponding to the reconstructed training images and the distorted characters in the training set, thus obtaining a primary text recognition model.
[0073] Specifically, when the training set image is input for the first time, the character segmentation network is used to determine the position of the character, and the character is segmented and input into the distortion correction network and the character recognition network for corresponding processing. After the recognition result is output, the recognition result is judged based on the text recognition labels of the training set. If the recognition result is correct, it is considered that the character segmentation network has correctly segmented the character, and the value of the pseudo-label is updated. Subsequently, the correctly predicted characters are occluded, and the occluded characters are input into the text recognition model again for iterative recognition, and the value of the pseudo-label is continuously updated until all characters are recognized or the preset number of iterations is reached. Then, the loss value is calculated based on the value of the pseudo-label, thereby updating the parameters of the character segmentation network.
[0074] For example, when masking correctly predicted characters, methods such as solid color masking, mosaic masking, and masking with real scene background images can be used, as long as it does not affect the segmentation and recognition of other characters.
[0075] In fact, updating the pseudo-label value is recording the number of characters correctly identified by the character recognition model in the image. For example, if there are five characters in the training set image label, and after multiple character recognition steps by the character recognition model, the pseudo-label value is updated to four, it means that the character segmentation network correctly predicted the positions of four characters.
[0076] In some embodiments, when testing the recognition accuracy of the preliminary text recognition model based on test set data, the test set is input into the preliminary text recognition model, and after the recognition result is output, the recognition result is compared with the text recognition labels of the test set and the recognition accuracy is calculated. When the recognition accuracy is less than a preset accuracy threshold, a training set is added to train the preliminary text recognition model for a new round. When the recognition accuracy of the preliminary text recognition model based on the test set data is greater than or equal to the preset accuracy threshold, the trained text recognition model is obtained.
[0077] See Figure 4 This invention also provides a distorted character recognition system that implements the distorted character recognition method in any of the above embodiments, comprising:
[0078] The model building module 100 is used to build a character recognition model with a character segmentation network, a distortion correction network, and a character recognition network;
[0079] The pre-training module 200 is used to pre-train the character segmentation network in the character recognition model to obtain the pre-trained character recognition model.
[0080] The model training module 300 is used to acquire image information containing distorted characters and train the pre-trained character recognition model based on the image information to obtain the trained character recognition model.
[0081] The text recognition module 400 is used to acquire image information to be recognized, and to perform text recognition after cutting, recombining and distorting the characters in the image to be recognized based on the trained text recognition model, and output the recognized text information.
[0082] In some embodiments, the image information to be recognized obtained by the text recognition module may be a single image, a folder containing images, or a compressed file.
[0083] In some embodiments, see Figure 4 The distorted text recognition system also includes an image submission module 500, which is used to interact with users who need text recognition, to receive image information submitted by users that needs text recognition, and to input the received image information into the text recognition module.
[0084] In some embodiments, see Figure 4 The distorted character recognition system also includes an information display module 600, which is used to display the reconstructed image after the characters in the image to be recognized are cut and reconstructed based on the trained character recognition model, and to display the text information recognized based on the trained character recognition model.
[0085] In some embodiments, see Figure 4The distorted character recognition system also includes an error collection module 700, which can collect images with recognition errors for subsequent optimization of the character recognition model.
[0086] In some embodiments, see Figure 4 The distorted character recognition system also includes a model optimization module 800, which can further optimize and train the character recognition model based on the images with recognition errors collected by the error collection module.
[0087] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.
Claims
1. A method for recognizing distorted characters, characterized in that, Includes the following steps: A character recognition model with a character segmentation network, a distortion correction network, and a character recognition network is constructed, and the character segmentation network in the character recognition model is pre-trained based on a public character segmentation set to obtain a pre-trained character recognition model; wherein, the distortion correction network has a localization grid, a grid generator, and a sampler, and the character recognition network is a network based on an attention mechanism; Image information containing distorted characters but without character segmentation annotations is obtained. Based on the pre-trained character recognition model, the distorted characters in the image information are segmented, recombined, distorted, and recognized as characters. The training text information is then output. The parameters of the pre-trained character recognition model are iteratively updated based on the training text information and image information to obtain the trained character recognition model. The system acquires image information to be recognized, and performs character recognition on the characters in the image based on the trained character recognition model after cutting, recombining, and distortion correction, and outputs the recognized character information.
2. The distorted character recognition method according to claim 1, characterized in that, The process of segmenting and recombining characters in the image information based on the pre-trained character recognition model includes: The image information is preprocessed and then divided into a training set and a test set; The character segmentation network in the pre-trained character recognition model performs single-character segmentation on the distorted characters in the training set and reassembles them sequentially, outputting a reassembled training image.
3. The distorted character recognition method according to claim 2, characterized in that, The process of correcting the distortion of distorted characters in the image information based on the pre-trained character recognition model includes: Based on the tilt angle of the distorted characters in the reconstructed training image, the positioning grid in the distortion correction network of the pre-trained character recognition model locates reference points on the reconstructed training image. The grid generator generates a sampling network based on the positioning reference points, and the sampler performs horizontal correction on the distorted characters in the reconstructed training image based on the sampling network to obtain a corrected training image.
4. The distorted character recognition method according to claim 3, characterized in that, The process of recognizing distorted characters in the image information based on the pre-trained character recognition model includes: Based on the character recognition network in the pre-trained character recognition model, image features and sequence features are extracted from the corrected training images, and training text information is output based on the image features and the sequence features.
5. The distorted character recognition method according to claim 4, characterized in that, The process of iteratively updating the parameters of the pre-trained character recognition model based on training text and image information includes: The pre-trained character recognition model is trained based on the training set data to obtain a preliminary character recognition model. The trained character recognition model is obtained when the recognition accuracy of the preliminary character recognition model is greater than or equal to a preset accuracy threshold based on the test set data.
6. The distorted character recognition method according to claim 5, characterized in that, The process of training the pre-trained text recognition model based on the training set includes: Generate pseudo-labels with empty initial values, and update the values of the pseudo-labels based on the number of correctly recognized characters in the training text information; The parameters of the character segmentation network in the pre-trained text recognition model are updated based on the values of the pseudo-labels. Based on the recombined training image and the character labels corresponding to the distorted characters in the training set, the parameters of the distortion correction network and the character recognition network in the pre-trained character recognition model are updated to obtain the primary character recognition model.
7. A distorted character recognition system, used to implement the distorted character recognition method as described in any one of claims 1 to 6, characterized in that, include: The model building module is used to build a text recognition model with a character segmentation network, a distortion correction network, and a character recognition network; The pre-training module is used to pre-train the character segmentation network in the character recognition model to obtain the pre-trained character recognition model. The model training module is used to acquire image information containing distorted characters and train the pre-trained character recognition model based on the image information to obtain the trained character recognition model. The text recognition module is used to acquire image information to be recognized, and to perform text recognition after cutting, recombining and distorting the characters in the image to be recognized based on the trained text recognition model, and output the recognized text information.
8. A distorted character recognition system according to claim 7, characterized in that, It also includes an image submission module, used to input the image information to be recognized into the text recognition module.
9. A distorted character recognition system according to claim 7, characterized in that, It also includes an information display module, which is used to display the reconstructed image after the characters in the image to be recognized are cut and reconstructed based on the trained character recognition model, and to display the text information recognized based on the trained character recognition model.