A method and system for character recognition

Abstract

This invention relates to a method and system for character recognition. The method includes: performing optical character recognition on an image file to be recognized to obtain multiple text characters; determining multiple characters to be adjusted from the multiple text characters; selecting a predetermined number of recognition result characters with the highest confidence from the recognition result list of each character to be adjusted to form a recognition result set for each character to be adjusted; obtaining at least two neighboring characters for each character to be adjusted, and forming a character group to be adjusted for each character to be adjusted by combining each character to be adjusted and at least two neighboring characters; processing the recognition result set and character group to be adjusted for each character to be adjusted based on an association rule mining model to determine the text association degree; selecting the recognition result character with the highest confidence as the target text character for each character to be adjusted, and replacing the corresponding character to be adjusted with the target text character to obtain an adjusted text file.

Description

Technical Field

[0001] The present invention relates to the technical field of character recognition, and more specifically to a method and system for performing character recognition. Background Art

[0002] At present, the existing OCR (Optical Character Recognition) mainly reduces the difficulty of feature extraction algorithm and improves recognition accuracy by preprocessing the image to be recognized as follows: (1) Binarization: Since the amount of information contained in color images is too large, before recognizing the printed characters in the image, it is necessary to binarize the image so that the image only contains black foreground information and white background information, thereby improving the efficiency and accuracy of the recognition process; (2) Image denoising: Since the quality of the image to be recognized is limited by the input device, environment, and printing quality of the document, before recognizing the printed characters in the image to be recognized, it is necessary to perform denoising on the image to be recognized based on the characteristics of the noise, thereby improving the accuracy of the recognition process; (3) Tilt correction: Since the scanning and shooting processes involve manual operations, the image to be recognized input into the computer will have some tilt to a certain extent. Before recognizing the printed characters in the image to be recognized, it is necessary to perform image direction detection and correct the image direction based on the direction detection result; (4) Text feature extraction: In terms of recognition rate alone, text feature extraction is the core of OCR. What features to use and how to extract them directly affect the performance of character recognition. Therefore, in the early stage of OCR research, there were many research reports on feature extraction. Features can be considered as the bargaining chips for recognition. Simple distinctions can be divided into two categories: one is statistical features, such as the black / white dot ratio in the text area. When the text area is divided into multiple areas, the black / white dot ratio of each area becomes a numerical vector in space; the other is structural features, such as obtaining the number and position of the stroke endpoints and intersections of the character after the text image is thinned, or using the stroke segments as features and matching them with special comparison methods. At present, the recognition methods of online handwriting input software on the market are mostly based on this structural method. (5) Comparison database: After calculating and determining the features for the input text, whether using statistical or structural features, it is necessary to compare the database or feature database for comparison. The content of the database should contain all the characters in the character set that you want to recognize, so as to obtain the feature group obtained by the same feature extraction method as the input text.

[0003] At present, the recognition accuracy of English characters with clear picture standards on the market can reach more than 99%, but because there are many Chinese characters in daily use (the number of commonly used Chinese characters is more than 2,000), the recognition accuracy of Chinese characters with clear picture standards is about 95%. The clear picture standard refers to the image standard without tilt, cover or pollution, the text is clear and the image recognition rate meets the standard. However, in actual production and life, many scenes that need to recognize text have image tilt, light changes, image cover and pollution, and text font changes, which further reduce the recognition rate of Chinese characters. Even if the image is processed by methods such as image denoising, tilt correction and text feature extraction to reduce the decline in recognition rate caused by non-standard images, it is difficult to further improve the accuracy of Chinese character recognition after the recognition rate reaches a certain level due to the limitations of the collected images themselves. Summary of the invention

[0004] The technical solution of the present invention is to combine optical character recognition technology (OCR) with natural language recognition processing (NLP), to correct the recognized text by judging the content of the recognized text, determining the association between natural language characters, and judging whether the recognized characters logically conform to the grammatical rules of the natural language, so as to improve the recognition accuracy.

[0005] According to one aspect of the present invention, there is provided a method for character recognition, the method comprising:

[0006] Acquire an image file to be recognized that includes a plurality of text characters, perform image processing on the image file to be recognized, determine whether the image file after image processing meets character recognition requirements, and when the character recognition requirements are met, perform optical character recognition on the image file to be recognized to obtain a text file, wherein the text file includes a plurality of text characters, and each text character includes a plurality of recognition result characters;

[0007] Obtaining the confidence of all recognition result characters of each of the multiple text characters, and determining the text character whose maximum confidence among all recognition result characters of the multiple text characters is less than the confidence threshold as the character to be adjusted, so as to determine the multiple characters to be adjusted;

[0008] Sorting the multiple recognition result characters of each character to be adjusted in descending order of confidence to obtain a recognition result list of each character to be adjusted, and selecting a predetermined number of recognition result characters with the largest confidence from the recognition result list of each character to be adjusted to form a recognition result set of each character to be adjusted;

[0009] Acquire at least two adjacent characters of each to-be-adjusted text character, and form each to-be-adjusted text character and the at least two adjacent characters into a to-be-adjusted character group of each to-be-adjusted text character, and process the recognition result set of each to-be-adjusted character and the to-be-adjusted character group based on an association rule mining model to determine the text association degree;

[0010] The confidence of each recognition result character in the recognition result set is calculated through the confidence and text association of each character to be adjusted, and the recognition result character with the largest confidence is selected as the target text character of each character to be adjusted. The target text character is used to replace the corresponding character to be adjusted to obtain an adjusted text file.

[0011] Performing image processing on the image file to be identified includes:

[0012] Perform image noise reduction processing on the image file to be identified to obtain an image file after image noise reduction processing.

[0013] Performing image processing on the image file to be identified includes:

[0014] The image file subjected to the image noise reduction process is subjected to tilt correction to obtain an image file subjected to image processing.

[0015] Determining whether the image file after image processing meets the character recognition requirements includes:

[0016] Determine whether the image noise of the image file after image processing is lower than a noise threshold, and when the image noise of the image file after image processing is lower than the noise threshold, determine whether the image inclination of the image file after image processing is lower than an inclination threshold;

[0017] When the image inclination of the image file after image processing is lower than the inclination threshold, it is determined that the image file after image processing meets the character recognition requirement.

[0018] Before obtaining the image file to be recognized containing text characters, the following steps are also included:

[0019] Collect text corpus, where the content of the text corpus is consistent with the type of characters to be recognized;

[0020] The collected text corpus is analyzed using an association rule mining algorithm, and an association relationship between each text character and multiple adjacent text characters is established, thereby establishing an association rule mining model.

[0021] According to another aspect of the present invention, there is provided a system for performing character recognition, the system comprising:

[0022] A recognition device, which obtains an image file to be recognized that includes a plurality of text characters, performs image processing on the image file to be recognized, determines whether the image file after image processing meets character recognition requirements, and when the character recognition requirements are met, performs optical character recognition on the image file to be recognized to obtain a text file, wherein the text file includes a plurality of text characters, and each text character includes a plurality of recognition result characters;

[0023] An acquisition device is provided to acquire the confidence of all recognition result characters of each of the plurality of text characters, and to determine the text character whose maximum confidence among all the recognition result characters of the plurality of text characters is less than the confidence threshold as the character to be adjusted, so as to determine the plurality of characters to be adjusted;

[0024] A selection device is provided for sorting a plurality of recognition result characters of each character to be adjusted in descending order of confidence to obtain a recognition result list of each character to be adjusted, and selecting a predetermined number of recognition result characters with the largest confidence from the recognition result list of each character to be adjusted to form a recognition result set of each character to be adjusted;

[0025] A determination device is provided, which obtains at least two adjacent characters of each text character to be adjusted, and forms a character group to be adjusted of each text character to be adjusted with each text character to be adjusted and the at least two adjacent characters, and processes the recognition result set of each character to be adjusted and the character group to be adjusted based on an association rule mining model to determine the text association degree;

[0026] The replacement device calculates the confidence of each recognition result character in the recognition result set through the confidence and text association of each character to be adjusted, selects the recognition result character with the largest confidence as the target text character of each character to be adjusted, and replaces the corresponding character to be adjusted with the target text character to obtain an adjusted text file.

[0027] The recognition device performs image processing on the image file to be recognized, including:

[0028] The recognition device performs image noise reduction processing on the image file to be recognized to obtain an image file after image noise reduction processing.

[0029] The recognition device performs image processing on the image file to be recognized, including:

[0030] The recognition device performs tilt correction on the image file after the image noise reduction process to obtain an image file after image processing.

[0031] The recognition device determines whether the image file after image processing meets the character recognition requirements including:

[0032] The recognition device determines whether the image noise of the image file after image processing is lower than a noise threshold, and when the image noise of the image file after image processing is lower than the noise threshold, determines whether the image inclination of the image file after image processing is lower than an inclination threshold;

[0033] When the image inclination of the image file after image processing is lower than the inclination threshold, the recognition device determines that the image file after image processing meets the character recognition requirement.

[0034] It also includes a preprocessing device for collecting text corpus, wherein the content of the text corpus is consistent with the type of characters to be recognized;

[0035] The collected text corpus is analyzed using an association rule mining algorithm, and an association relationship between each text character and multiple adjacent text characters is established, thereby establishing an association rule mining model.

[0036] The beneficial effect of the present invention is that due to the shortcomings of the image itself and the fact that there are many commonly used Chinese characters during optical character recognition, the accuracy of Chinese recognition is not high. The commonly used technical means is to correct the deviation of the image through a series of calculations to improve the accuracy of the image. However, due to the shortcomings of the image itself and the fact that there are many Chinese characters, it is very difficult to continue to improve the accuracy after the recognition accuracy reaches a certain level, and the benefits of the improvement are very low.

[0037] The method of the present invention combines optical character technology with natural language analysis technology, and performs semantic analysis on the recognized text content. When the recognized content does not conform to the logical law of natural language, it is corrected through an algorithm, and finally the recognized text content conforms to the language law, thereby improving the recognition accuracy.

[0038] Most optical recognition technologies use methods such as image noise reduction, tilt correction, and feature extraction to collect image content. The effect improved by this method is limited by the content of the image itself, and it is easy to reach the limit, and therefore it is very difficult to further improve. There are many commonly used Chinese characters, and there are many classification categories when using machine learning algorithms, which further reduces the recognition accuracy. The technical solution proposed in the present invention takes a different approach. On the basis of the original optical character recognition, it analyzes the semantics and content of the extracted text, and corrects the text by making the extracted text conform to the logical rules of natural language, thereby improving the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] A more complete understanding of exemplary embodiments of the present invention may be obtained by referring to the following drawings:

[0040] Figure 1A flowchart of a method for performing character recognition according to an embodiment of the present invention;

[0041] Figure 2 A flowchart of a method for performing character recognition according to another embodiment of the present invention;

[0042] Figure 3 Schematic diagram of the structure of a system for performing character recognition according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] Figure 1 The present invention is a method for character recognition according to an embodiment of the present invention. The first three steps are the commonly used optical character recognition (OCR) process. The following steps are the process of comprehensively judging and correcting the text characters whose confidence level is lower than the threshold in the OCR recognition results by using NLP recognition technology.

[0044] Start by obtaining a picture containing text. Perform image processing on the picture containing text. Image processing includes: image noise reduction, tilt correction, feature extraction, etc. Then perform text recognition on the image processed picture to obtain a recognition result. For example, use an OCR recognition model to perform text recognition on the image processed picture.

[0045] Determine whether the confidence of the text recognition in the recognition result reaches a threshold. When the threshold is reached, the recognition result can be output. For example, when the confidence of the text recognition of all the texts in the recognition result reaches the threshold, the recognition result is output.

[0046] If it is determined that the confidence of the text recognition in the recognition result does not reach the threshold, the adjacent character recognition results are extracted. Then the association rule mining model is determined, wherein the association rule mining algorithm is determined by obtaining the expected data. Then the text association is calculated, and the recognition result is judged by the confidence and text association. When the recognition result meets the requirements, the recognition result is output.

[0047] Before recognition, it also includes collecting common text corpora, the content of which is consistent with the type of characters to be recognized. For example, if the text content to be recognized is tax information, then the common text corpora collected are tax document corpora. The collected text corpora are analyzed using the association rule mining algorithm apripri. Each character has a relationship with the characters before and after it. An association rule mining model is established.

[0048] Subsequently, the optical character method is called. You can use a self-developed OCR tool or an API interface released on the market. This method uses a self-developed OCR interface to return the confidence of each recognized text. When the confidence of the recognized text is higher than the threshold, it is considered that the recognition reliability of the character is high, and the character recognition is judged to be accurate. When the confidence of the recognized text is lower than a certain threshold, it is judged that the recognition reliability of the character is low. The recognition results of the characters before and after the character are extracted, and the top n confidences are the results. The recognition results are added to the association rule mining model, and the text association is returned. The confidence of the recognized text is calculated through the confidence of the character and the text association, and the result with high confidence is selected as the recognition result output.

[0049] Figure 2 Flow chart of a method 200 for character recognition based on content correction according to another embodiment of the present invention. The method 200 comprises:

[0050] Step 201, obtain an image file to be recognized that contains multiple text characters, perform image processing on the image file to be recognized, determine whether the image file after image processing meets the character recognition requirements, and when the character recognition requirements are met, perform optical character recognition on the image file to be recognized to obtain a text file, wherein the text file includes multiple text characters, and each text character includes multiple recognition result characters.

[0051] Step 202, obtaining the confidence of all recognition result characters of each of the multiple characters, determining the character with the maximum confidence less than the confidence threshold among all the recognition result characters of the multiple characters as the character to be adjusted, so as to determine multiple characters to be adjusted.

[0052] Step 203, sorting the multiple recognition result characters of each character to be adjusted in descending order of confidence to obtain a recognition result list of each character to be adjusted, and selecting a predetermined number of recognition result characters with the largest confidence from the recognition result list of each character to be adjusted to form a recognition result set for each character to be adjusted.

[0053] Step 204, obtaining at least two adjacent characters of each text character to be adjusted, and forming each text character to be adjusted and at least two adjacent characters into a character group to be adjusted for each text character to be adjusted, and processing the recognition result set of each character to be adjusted and the character group to be adjusted based on the association rule mining model to determine the text association degree.

[0054] Step 205, calculate the confidence of each recognition result character in the recognition result set through the confidence and text association of each character to be adjusted, select the recognition result character with the largest confidence as the target text character of each character to be adjusted, and use the target text character to replace the corresponding character to be adjusted to obtain an adjusted text file.

[0055] Performing image processing on the image file to be identified includes: performing image noise reduction processing on the image file to be identified to obtain an image file after image noise reduction processing. Performing image processing on the image file to be identified includes: performing tilt correction on the image file after image noise reduction processing to obtain an image file after image processing. Determining whether the image file after image processing meets the character recognition requirements includes: determining whether the image noise of the image file after image processing is lower than the noise threshold, and when the image noise of the image file after image processing is lower than the noise threshold, determining whether the image tilt of the image file after image processing is lower than the tilt threshold; when the image tilt of the image file after image processing is lower than the tilt threshold, determining that the image file after image processing meets the character recognition requirements.

[0056] Before obtaining the image file to be recognized containing text characters, the method also includes: collecting text corpus, wherein the content of the text corpus is consistent with the type of characters to be recognized; using an association rule mining algorithm to analyze the collected text corpus, establishing an association relationship between each text character and multiple adjacent text characters, thereby establishing an association rule mining model.

[0057] Figure 3 Schematic diagram of the structure of a system for character recognition based on content correction according to an embodiment of the present invention. The system 300 comprises: a recognition device 301, an acquisition device 302, a selection device 303, a determination device 304, a replacement device 305 and a pre-processing device 306.

[0058] The recognition device 301 obtains an image file to be recognized that contains multiple text characters, performs image processing on the image file to be recognized, determines whether the image file after image processing meets the character recognition requirements, and when the character recognition requirements are met, performs optical character recognition on the image file to be recognized to obtain a text file, wherein the text file includes multiple text characters, and each text character includes multiple recognition result characters.

[0059] The acquisition device 302 acquires the confidence of all recognition result characters of each of the multiple characters, and determines the character whose maximum confidence among all recognition result characters of the multiple characters is less than the confidence threshold as the character to be adjusted, so as to determine multiple characters to be adjusted.

[0060] The selection device 303 sorts the multiple recognition result characters of each character to be adjusted in descending order of confidence to obtain a recognition result list of each character to be adjusted, and selects a predetermined number of recognition result characters with the largest confidence from the recognition result list of each character to be adjusted to form a recognition result set of each character to be adjusted.

[0061] The determination device 304 obtains at least two adjacent characters of each text character to be adjusted, and forms a character group to be adjusted for each text character to be adjusted with each text character to be adjusted and the at least two adjacent characters, and processes the recognition result set of each character to be adjusted and the character group to be adjusted based on the association rule mining model to determine the text association degree.

[0062] The replacement device 305 calculates the confidence of each recognition result character in the recognition result set through the confidence and text association of each character to be adjusted, selects the recognition result character with the largest confidence as the target text character of each character to be adjusted, and replaces the corresponding character to be adjusted with the target text character to obtain an adjusted text file.

[0063] The preprocessing device 306 is used to collect text corpus, wherein the content of the text corpus is consistent with the type of the character to be recognized; use the association rule mining algorithm to analyze the collected text corpus, establish an association relationship between each text character and multiple adjacent text characters, and thus establish an association rule mining model.

[0064] The recognition device 301 performs image processing on the image file to be recognized, including: the recognition device 301 performs image noise reduction processing on the image file to be recognized, so as to obtain an image file after image noise reduction processing. The recognition device 301 performs image processing on the image file to be recognized, including: the recognition device 301 performs tilt correction on the image file after image noise reduction processing, so as to obtain an image file after image processing. The recognition device 301 determines whether the image file after image processing meets the character recognition requirements, including: the recognition device 301 determines whether the image noise of the image file after image processing is lower than the noise threshold, and when the image noise of the image file after image processing is lower than the noise threshold, determines whether the image inclination of the image file after image processing is lower than the inclination threshold; when the image inclination of the image file after image processing is lower than the inclination threshold, the recognition device 301 determines that the image file after image processing meets the character recognition requirements.

Claims

1. A method for character recognition, the method comprising: Acquire an image file to be recognized that includes a plurality of text characters, perform image processing on the image file to be recognized, determine whether the image file after image processing meets character recognition requirements, and when the character recognition requirements are met, perform optical character recognition on the image file to be recognized to obtain a text file, wherein the text file includes a plurality of text characters, and each text character includes a plurality of recognition result characters; Before obtaining the image file to be recognized containing text characters, the following steps are also included: Collect text corpus, where the content of the text corpus is consistent with the type of characters to be recognized; Use association rule mining algorithms to analyze the collected text corpus, establish association relationships between each character and multiple adjacent characters, and thus establish an association rule mining model; Obtaining the confidence of all recognition result characters of each of the multiple text characters, and determining the text character whose maximum confidence among all recognition result characters of the multiple text characters is less than the confidence threshold as the character to be adjusted, so as to determine the multiple characters to be adjusted; Sorting the multiple recognition result characters of each character to be adjusted in descending order of confidence to obtain a recognition result list of each character to be adjusted, and selecting a predetermined number of recognition result characters with the largest confidence from the recognition result list of each character to be adjusted to form a recognition result set of each character to be adjusted; Acquire at least two adjacent characters of each to-be-adjusted text character, and form each to-be-adjusted text character and the at least two adjacent characters into a to-be-adjusted character group of each to-be-adjusted text character, and process the recognition result set of each to-be-adjusted character and the to-be-adjusted character group based on an association rule mining model to determine the text association degree; Calculate the confidence of each recognition result character in the recognition result set by the confidence and text association of each character to be adjusted, select the recognition result character with the largest confidence as the target text character of each character to be adjusted, and replace the corresponding character to be adjusted with the target text character to obtain an adjusted text file; It also includes correcting the recognized text by combining optical character recognition technology with natural language recognition processing, judging the content of the recognized text, determining the relationship between natural language characters, and judging whether the recognized characters logically conform to the grammatical rules of natural language, or correcting the text by causing the extracted text to conform to the logical rules of natural language.

2. According to the method of claim 1, performing image processing on the image file to be identified comprises: Perform image noise reduction processing on the image file to be identified to obtain an image file after image noise reduction processing.

3. According to the method of claim 2, performing image processing on the image file to be identified comprises: The image file subjected to the image noise reduction process is subjected to tilt correction to obtain an image file subjected to image processing.

4. According to the method of claim 1, determining whether the image file after image processing meets the character recognition requirements comprises: Determine whether the image noise of the image file after image processing is lower than a noise threshold, and when the image noise of the image file after image processing is lower than the noise threshold, determine whether the image inclination of the image file after image processing is lower than an inclination threshold; When the image inclination of the image file after image processing is lower than the inclination threshold, it is determined that the image file after image processing meets the character recognition requirement.

5. A system for character recognition, the system comprising: A recognition device, which obtains an image file to be recognized that includes a plurality of text characters, performs image processing on the image file to be recognized, determines whether the image file after image processing meets character recognition requirements, and when the character recognition requirements are met, performs optical character recognition on the image file to be recognized to obtain a text file, wherein the text file includes a plurality of text characters, and each text character includes a plurality of recognition result characters; A preprocessing device, used for collecting text corpus, wherein the content of the text corpus is consistent with the type of characters to be recognized; An association rule mining algorithm is used to analyze the collected text corpus, and an association relationship is established between each text character and a plurality of adjacent text characters, thereby establishing an association rule mining model; an acquisition device is used to acquire the confidence of all recognition result characters of each text character in the plurality of text characters, and a text character whose maximum confidence among all recognition result characters in the plurality of text characters is less than a confidence threshold is determined as a character to be adjusted, so as to determine a plurality of characters to be adjusted; A selection device is provided for sorting a plurality of recognition result characters of each character to be adjusted in descending order of confidence to obtain a recognition result list of each character to be adjusted, and selecting a predetermined number of recognition result characters with the largest confidence from the recognition result list of each character to be adjusted to form a recognition result set of each character to be adjusted; A determination device is provided, which obtains at least two adjacent characters of each text character to be adjusted, and forms a character group to be adjusted of each text character to be adjusted with each text character to be adjusted and the at least two adjacent characters, and processes the recognition result set of each character to be adjusted and the character group to be adjusted based on an association rule mining model to determine the text association degree; A replacement device calculates the confidence of each recognition result character in the recognition result set through the confidence and text association of each character to be adjusted, selects the recognition result character with the largest confidence as the target text character of each character to be adjusted, and replaces the corresponding character to be adjusted with the target text character to obtain an adjusted text file; It also includes correcting the recognized text by combining optical character recognition technology with natural language recognition processing, judging the content of the recognized text, determining the relationship between natural language characters, and judging whether the recognized characters logically conform to the grammatical rules of natural language, or correcting the text by causing the extracted text to conform to the logical rules of natural language.

6. According to the system of claim 5, the recognition device performs image processing on the image file to be recognized, comprising: The recognition device performs image noise reduction processing on the image file to be recognized to obtain an image file after image noise reduction processing.

7. According to the system of claim 6, the recognition device performs image processing on the image file to be recognized, comprising: The recognition device performs tilt correction on the image file after the image noise reduction process to obtain an image file after image processing.

8. According to the system of claim 5, the recognition device determines whether the image file after image processing meets the character recognition requirement, comprising: The recognition device determines whether the image noise of the image file after image processing is lower than a noise threshold, and when the image noise of the image file after image processing is lower than the noise threshold, determines whether the image inclination of the image file after image processing is lower than an inclination threshold; When the image inclination of the image file after image processing is lower than the inclination threshold, the recognition device determines that the image file after image processing meets the character recognition requirement.

Images