A document digitization processing auxiliary method and device for improving precision and efficiency

Abstract

The application discloses a kind of precision and efficiency promotion literature digitization processing auxiliary method and device, the method includes: the image data marked by user is marked result pre-processing;Image data contains several picture words;The image data after pre-processing is carried out OCR identification, and each picture word obtains corresponding recognition word;The error recognition possibility of each word is judged based on the pixel distribution of recognition word and picture word;In response to the judgment result of possible error recognition, corresponding picture word is obtained from image data Target screenshot word;Target pre-stored screenshot word is found from character library based on target screenshot word, and corresponding character of target pre-stored screenshot word is used to replace corresponding recognition word.The scheme designs self-correction mechanism in OCR conversion text process, the result of possible recognition error is judged by pixel distribution, and accurate character with higher accuracy is found by using screenshot word, which can significantly reduce error recognition word, improve proofreading efficiency and accuracy.

Description

Technical Field

[0001] This invention belongs to the field of document digitization technology, and in particular relates to an auxiliary method and device for improving the accuracy and efficiency of document digitization. Background Technology

[0002] Documents are an important carrier of knowledge. Traditional documents mostly exist in physical form such as paper. The retrieval and reading of physical documents are extremely inconvenient. When faced with a massive amount of data, researchers often have to spend a lot of time and energy to browse and search, which is inefficient.

[0003] With the rapid development of computer technology, internet technology, and artificial intelligence technology, digital processing offers a feasible solution to the aforementioned challenges. Through technologies such as image scanning and OCR recognition, paper documents can be quickly converted into digital images and further recognized into editable text formats.

[0004] Currently, the digitization of documents mainly includes preliminary data processing, annotation, OCR recognition, and review and correction. Preliminary data processing involves scanning physical documents to obtain image data. Annotation is a manual process of dividing the image data into columns and drawing bounding boxes; this prepares the data for subsequent OCR recognition. The more standardized and accurate the bounding boxes, the more accurate the OCR recognition will be. OCR recognition then converts the text in the image data into editable text based on the annotation results. The review and correction stage involves manual cleaning of the OCR results, including deleting erroneous, invalid, duplicate, and blank content, and correcting incorrect paragraphing and text errors.

[0005] However, existing document digitization technologies still have many shortcomings. For example, to ensure the accuracy of OCR recognition, the bounding box process requires a significant amount of manual effort. Furthermore, the inherent technical limitations of OCR recognition often lead to errors when dealing with easily misspelled words, handwritten fonts, and complex layouts, necessitating extensive manual proofreading. This not only reduces processing efficiency but also hinders accuracy improvement. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for improving the accuracy and efficiency of document digitization processing in response to the above-mentioned problems. This method reduces the workload of manual processing by adding an automated processing method to the manual-dominated digitization process, while improving the processing accuracy.

[0007] A method for improving the accuracy and efficiency of document digitization processing, the method comprising:

[0008] Preprocess the annotation results of user-annotated image data;

[0009] The image data includes several image words;

[0010] OCR recognition is performed on the preprocessed image data, and the corresponding recognized character is obtained for each character in the image.

[0011] The probability of misidentification of each character is determined based on the pixel distribution of the recognized characters and the characters in the image;

[0012] In response to a possible misidentification result, the corresponding image text is extracted from the image data to obtain the target screenshot text;

[0013] Based on the target screenshot character, search for the target pre-stored screenshot character from the character library;

[0014] In response to finding the target pre-stored screenshot character, the corresponding recognized character is replaced with the text corresponding to the target pre-stored screenshot character;

[0015] If no pre-stored screenshot character is found, mark the corresponding identification character.

[0016] In the above-mentioned methods for assisting in the digitization of documents to improve accuracy and efficiency, the annotation includes drag-and-drop boxes and label boxes;

[0017] Preprocessing of user-annotated image data includes:

[0018] The position of the drag box is identified and judged. If it does not conform to the established rules, the size of the corresponding drag box is adjusted to conform to the established rules.

[0019] In the above-mentioned methods to assist in the digitization of documents and improve accuracy and efficiency, the width of the first draw box on each page / column is used as the standard draw box.

[0020] The identification and judgment of the position of the pull box includes determining whether the horizontal position and width of the subsequent pull box are consistent with the standard pull box. If so, it is considered to conform to the established rules; otherwise, the corresponding pull box is adjusted so that its horizontal position and width are consistent with the standard pull box.

[0021] In the above-mentioned auxiliary methods for improving the accuracy and efficiency of document digitization, the box position recognition and judgment includes the judgment of the position of the box frame line and the judgment of the spacing between adjacent characters in the box.

[0022] The established rules include that no image text is cut off by the frame line of the frame and the spacing between adjacent image texts within the frame is less than a set spacing threshold.

[0023] Adjusting the size of the corresponding pull box includes adjusting the size of the pull box so that the truncated image text is enclosed within the pull box or excluded from the pull box; adjusting the size of the pull box so that image text that is relatively close to the edge of the pull box and has a spacing greater than a set spacing threshold is excluded from the pull box.

[0024] Among the aforementioned methods for improving the accuracy and efficiency of document digitization, the preprocessing of user-annotated image data also includes:

[0025] Find the beginning and end coordinates of the text in each drop box in turn, and determine whether the text in the corresponding drop box is an independent paragraph based on the beginning and / or end coordinates.

[0026] In response to the condition that the paragraph is not independent, add the paragraph in the following dropdown list as the same paragraph in the preceding dropdown list.

[0027] Among the aforementioned methods for improving the accuracy and efficiency of document digitization, the methods for determining whether a bounding box represents an independent paragraph include:

[0028] Based on the overall pixel distribution of each page / column of image data, determine the coordinate positions of the rightmost and leftmost characters, and accordingly determine a standard end-of-text horizontal coordinate and a standard beginning-of-text horizontal coordinate for each page / column.

[0029] Find the coordinates of the first and last characters of the text in each slider in turn. The coordinates of the first character in the slider are the coordinates of the first character in the slider, and the coordinates of the last character in the slider are the coordinates of the last character in the slider.

[0030] If the horizontal coordinate at the end of the text in the first dropdown box matches the corresponding standard horizontal coordinate at the end of the text, and the horizontal coordinate at the beginning of the text in the second dropdown box also matches the standard horizontal coordinate at the beginning of the text, then the second dropdown box is considered a non-independent paragraph.

[0031] In the above-mentioned document digitization processing auxiliary method for improving accuracy and efficiency, the text library pre-stores several error-correcting text pairs, each error-correcting text pair including a pre-stored screenshot character and its corresponding text.

[0032] Searching for target pre-stored screenshot characters from the text library based on the target screenshot characters includes:

[0033] Compare the pixel distribution of the target screenshot character with that of several pre-stored screenshot characters;

[0034] Extract the pre-stored screenshot characters with the highest pixel distribution similarity that is higher than the similarity threshold as the target pre-stored screenshot characters;

[0035] This method also includes:

[0036] In response to the failure to find the pre-stored screenshot character, the corresponding recognized character is marked and its corresponding target screenshot character is saved;

[0037] In response to the user's input operation based on the target screenshot text, the target screenshot text and the text provided by the user are stored as a new pair of error-correcting text in the text library.

[0038] In the above-mentioned document digitization processing auxiliary method for improving accuracy and efficiency, the first OCR recognition is performed on the pre-processed image data based on the first OCR module.

[0039] This method also includes:

[0040] The text that has undergone the first OCR recognition and error correction process will be handed over to the user for manual verification.

[0041] The text that has been verified and processed by the user will be used as the first text document;

[0042] The second OCR module performs a second OCR recognition on the preprocessed image data to generate a second text document.

[0043] Compare the recognized characters of the first text document and the second text document one by one. Based on the comparison results, highlight the recognized characters in the first text document that are inconsistent with the recognition results of the second text document.

[0044] The first OCR module and the second OCR module are different OCR modules, and it is preferred that the recognition accuracy of the second OCR module is higher than that of the first OCR module.

[0045] A document digitization processing auxiliary device for improving accuracy and efficiency includes an annotation module, a first OCR module and a verification module, an annotation result processing module located between the annotation module and the first OCR module, and an automatic error correction module located between the first OCR module and the verification module.

[0046] The annotation result processing module is used to preprocess the annotation results of the image data annotated by the user; the image data includes several image words.

[0047] The automatic error correction module includes an error identification module and an error correction module;

[0048] The error recognition module is used to determine the probability of misrecognition of each character based on the pixel distribution of the recognized character and the image character; the recognized character is obtained by the first OCR module through OCR recognition of the preprocessed image data;

[0049] The error correction module is used to extract the target screenshot character from the image data of the corresponding image character of the character that may be misidentified; search for the target pre-stored screenshot character from the text library based on the target screenshot character; in response to the target pre-stored screenshot character being found, replace the corresponding character with the text corresponding to the target pre-stored screenshot character; in response to the failure to find the pre-stored screenshot character, mark the corresponding character.

[0050] A document digitization processing auxiliary device to improve accuracy and efficiency also includes a second OCR module and a verification result processing module;

[0051] The second recognition module is used to perform a second OCR recognition on the preprocessed image data to generate a second text document;

[0052] The verification result processing module is used to take the text verified by the user as the first text document, compare the recognized characters of the first text document and the second text document one by one, and based on the comparison results, highlight the recognized characters in the first text document that are inconsistent with the recognition results of the second text document.

[0053] The advantages of this invention are:

[0054] A preprocessing step is added between the annotation module and the first OCR module to assist manual work, allowing for more flexibility in the manual bounding box process and reducing the burden of manual annotation. At the same time, it helps generate relatively standardized annotation results, reducing errors caused by human operation, thereby improving annotation accuracy and even the accuracy of subsequent OCR recognition, and reducing the additional costs caused by errors and rework.

[0055] Since high annotation accuracy can be guaranteed even without strict user annotation, this auxiliary processing step can improve annotation speed while ensuring annotation accuracy.

[0056] Designing a self-correction mechanism in the OCR to text conversion process, judging the possible recognition errors by pixel distribution, and using screenshots to find more accurate replacement text can effectively improve the accuracy of the OCR to text conversion process, significantly reduce the number of incorrectly recognized characters, thereby reducing the proofreading burden of subsequent manual proofreading and improving proofreading efficiency, while also improving proofreading accuracy.

[0057] A second OCR recognition process is introduced, which utilizes the differences in recognition characteristics between different OCR modules to perform a complementary fallback full check by comparing two texts, thereby further ensuring the accuracy of the document digitization results.

[0058] Through auxiliary processing and design in multiple stages, we have effectively reduced labor costs and improved project quality. Attached Figure Description

[0059] Figure 1 This is a flowchart of the document digitization processing auxiliary method for improving accuracy and efficiency in Embodiment 1 of the present invention;

[0060] Figure 2 This is a schematic diagram of the frame size adjustment process in the annotation result preprocessing process of the present invention (Embodiment 1).

[0061] Figure 3 This is a schematic diagram of the paragraph appending process in the annotation result preprocessing process of Embodiment 1 of the present invention;

[0062] Figure 4 This is a schematic diagram of the Chinese character library in Embodiment 1 of the present invention;

[0063] Figure 5 This is a schematic diagram of the automatic error correction process in the OCR to text conversion process in Embodiment 1 of the present invention;

[0064] Figure 6 This is a schematic diagram of the automatic error correction process when no replaceable text is found during the OCR to text conversion process in Embodiment 1 of the present invention;

[0065] Figure 7 This is a block diagram of the document digitization processing auxiliary device for improving accuracy and efficiency in Embodiment 2 of the present invention;

[0066] Figure 8 This is a schematic diagram of the processing of a case of non-standard bounding box in Embodiment 3 of the present invention and the preprocessing of the annotation results for this case;

[0067] Figure 9 This is a schematic diagram of the processing of another case of non-standard bounding box in Embodiment 3 of the present invention and the preprocessing of the annotation results for this case. Detailed Implementation

[0068] Example 1

[0069] like Figure 1 As shown, this invention provides an auxiliary method for improving the accuracy and efficiency of document digitization processing. While still primarily manual, it significantly reduces the workload of manual processing and improves processing accuracy through a series of auxiliary processing techniques. The specific method is as follows:

[0070] 1. Perform annotation preprocessing on the user-annotated image data.

[0071] The image data is obtained from scanning and basic processing of physical documents, and contains several images containing text. Scanning and processing physical documents to obtain image data is currently the first step in the digitization process of physical documents; this solution does not propose improvements to this and therefore will not be elaborated upon here. The annotation process involves dividing the document into columns, drawing boxes, and labeling different sections. The purpose is to define content information by dividing the image into regions based on different text content, such as drawing boxes for titles and paragraphs. Generally, the entire text needs to be labeled, which is done manually and is consistent with existing technology, so it will not be elaborated upon here.

[0072] 2. The first OCR module is used to perform the first OCR recognition on the preprocessed image data. Each character in the image is recognized to obtain the corresponding recognized character. The goal of OCR recognition is to group the content that has been bounded in the annotation process and is labeled as part of the same paragraph. The recognized text content will also be grouped into the same paragraph.

[0073] 3. Determine the probability of misrecognition for each character based on the pixel distribution of the recognized character and the image characters. Specifically, the probability of misrecognition for each character is determined based on the similarity of pixel distribution. In this embodiment, it is preferred that a similarity of not less than 92% is considered a correct recognition. Conversely, if the similarity is less than 92%, it is considered a misrecognition, and the judgment result is "possibly misrecognized". When put into use, the percentage of similarity used to classify correct or incorrect recognition is determined according to specific circumstances, such as the characteristics of the OCR module built into the digitization tool used, and the specific calculation method of pixel distribution similarity.

[0074] 4. In response to the judgment result of possible misidentification, extract the target screenshot character from the image data; based on the target screenshot character, search for the target pre-stored screenshot character from the text library;

[0075] 5. In response to finding the target pre-stored screenshot character, replace the corresponding recognized character with the text corresponding to the target pre-stored screenshot character;

[0076] If no pre-stored screenshot character is found, mark the corresponding identification character.

[0077] Step 1, the preprocessing of the user-annotated image data includes:

[0078] The system identifies and judges the position of the drag-and-drop box, and adjusts its size to conform to the established rules if it does not meet them. Specifically, for example... Figure 2 As shown:

[0079] Use the width of the first drag box on each page / column as the standard drag box. Figure 2 In the text, the section "The development and improvement of computer network technology..." is the first pull box, and this pull box is used as the standard pull box.

[0080] The identification and judgment of the position of the pull box includes determining whether the horizontal position and width of the subsequent pull box are consistent with the standard pull box. If so, it is considered to conform to the established rules; otherwise, the corresponding pull box is adjusted so that its horizontal position and width are consistent with the standard pull box. Figure 2 The section titled "With the advent of the internet age..." is located in the second dropdown menu. Figure 2 On the left, the horizontal position and width are inconsistent with the standard draw frame. Therefore, this method adjusts the horizontal position and size according to the standard draw frame to obtain the desired result. Figure 2 The result is shown on the right. This adjustment is applied to the entire page / column, resulting in a more standardized bounding box. Users only need to ensure the first box is standard. Testing has shown that this optimization significantly improves the accuracy of recognizing text in edge-images. It achieves a significant improvement in annotation accuracy without substantially increasing the user's workload.

[0081] Furthermore, the preprocessing of the annotation results for the image data annotated by the user further includes:

[0082] According to the overall pixel distribution of each page / column of the image data, a standard end horizontal coordinate and a standard start horizontal coordinate are determined for each page / column. The horizontal coordinate of the rightmost character in each page / column is used as the standard end horizontal coordinate, and the horizontal coordinate of the leftmost character is used as the standard start horizontal coordinate. For example Figure 3 in the figure, the horizontal coordinates where characters such as '全' and '系' are located in the left column are determined as the standard end horizontal coordinates of the left column, and the horizontal coordinates where characters such as '不' and '木' are located in the right column are determined as the standard start horizontal coordinates of the right column.

[0083] The start coordinate and end coordinate of the text within each bounding box are searched for in sequence. Here, the start coordinate refers to the coordinate of the first character in the bounding box, and the end coordinate refers to the coordinate of the last character in the bounding box.

[0084] When the end horizontal coordinate of the previous bounding box conforms to the position of the corresponding standard end horizontal coordinate, and the start horizontal coordinate of the next bounding box also conforms to the standard start horizontal coordinate, it is considered that the next bounding box is a non-independent paragraph, and the subsequent bounding box is appended as the same paragraph as the previous bounding box. This situation generally occurs when paginating and columnating. Therefore, when put into use, this judgment can be made for the first and last bounding boxes of each page / column. It should be noted that conforming to the corresponding standard end horizontal coordinate position and standard start horizontal coordinate means being within the set error range from the standard end horizontal coordinate position and standard start horizontal coordinate, rather than being exactly the same.

[0085] Since the steps of the preprocessing of the annotation results are carried out after the user's annotation, whether there is columnating can be directly known based on the user's bounding boxes, and the implementation is relatively simple, so it will not be elaborated here.

[0086] For traditional tools, when bounding the main text, every time a box is drawn, the tool will automatically annotate the paragraph. For example, when drawing the first box, it annotates paragraph 1, and the second box annotates paragraph 2. Sometimes, the text content of two adjacent boxes is in the same paragraph. During the manual bounding and annotation process, there is an option for paragraph appending, and the user can append the next box as a continuation of the previous box's paragraph. However, in actual annotation, it is very easy to forget to append after drawing the box. The above design of this solution automatically appends paragraphs based on format recognition for this situation, which can well solve the problem of forgetting to append during the manual annotation process, reduce incorrect annotation, and improve the OCR recognition effect.

[0087] For example Figure 3As shown, one page of a certain document is displayed in two columns. The left side shows the content in the left column, and the right side shows some content in the right column. The content within the right box is the same paragraph as the left box. During the process of dragging the box, the user forgot to append it, resulting in the system tool automatically marking it as "Paragraph 2" and obtaining Figure 3 the result above. After the above judgment, the content of the two boxes is the same paragraph, so the annotation result preprocessing of paragraph appending is performed on it, and finally Figure 3 the result below can be obtained.

[0088] Furthermore, as Figure 4 shown, a number of error-correction word pairs are pre-stored in the text library. Each pair of error-correction word pairs includes a pre-stored screenshot word and its corresponding word. Step 3 specifically includes:

[0089] As Figure 5 shown, in a certain image data, the picture word is recognized as "find". After comparing the pixel distribution, it is considered that the recognition may be incorrect. Subsequently, is intercepted from the image data to obtain the target screenshot word . Compare the pixel distribution of the target screenshot word with the pixel distributions of a number of pre-stored screenshot words pre-stored in the text library. Finally, the pre-stored screenshot word is found in the text library to meet the condition that the similarity degree of pixel distribution with the target screenshot word is the highest and higher than the degree threshold. The found pre-stored screenshot word is used as the target pre-stored screenshot word, and then the word "I" corresponding to the target pre-stored screenshot word is used to replace the recognition result "find".

[0090] When no pre-stored screenshot word is found, the corresponding recognized word is marked and its corresponding target screenshot word is saved, and the association relationship between the recognized word and the target screenshot word is established for the user to perform the warehousing operation on it later. As Figure 6 shown, there is currently no pre-stored screenshot word in the text library with a pixel distribution similarity degree higher than the degree threshold to , so the recognition result "valley" is highlighted.

[0091] When the user receives the text after OCR recognition and error-correction processing, the user will perform manual verification and cleaning on the text content, including deleting spaces, correcting errors, etc. During this process, under the improvement of this solution, the user can correct and warehouse the marked recognized words. For example, when selecting "valley" for warehousing, the system will retrieve the corresponding target screenshot word , and then use the word "rong" input by the user as the corresponding word of the target screenshot word . The two are stored in the word library as a new pair of error-correction word pairs. The next time OCR faces Furthermore, it can automatically correct errors when it identifies them incorrectly.

[0092] Furthermore, this method also includes:

[0093] The text that has been verified and processed by the user will be used as the first text document;

[0094] The second OCR module performs a second OCR recognition on the preprocessed image data to generate a second text document.

[0095] The system compares the recognized characters in the first and second text documents one by one. Based on the comparison results, it highlights the recognized characters in the first text document that are inconsistent with the recognition results in the second text document.

[0096] The first OCR module and the second OCR module are different OCR modules. It is preferred to use the second OCR module because its recognition accuracy is higher than that of the first OCR module. The first OCR module is a built-in module of the digital processing tool system, while the second OCR module is an OCR module added to this solution to improve processing accuracy.

[0097] After the initial ORC identification and automatic error correction, manual inspection and adjustment are performed. The final result is obtained after manual processing and inspection. Traditionally, this step is the final result. This embodiment adds a final full inspection process after these steps, using cross-comparison based on the different recognition characteristics of different OCR modules to further improve the accuracy of the result.

[0098] Example 2

[0099] This embodiment provides an auxiliary device for improving the accuracy and efficiency of document digitization processing, used to implement the method described in Embodiment 1. Figure 7 As shown, it includes an annotation module 1, a first OCR module 2, and a verification module 3. Specifically, an annotation result processing module 4 is designed between the annotation module 1 and the first OCR module 2, an automatic error correction module 5 is designed between the first OCR module 2 and the verification module 3, a verification result processing module 7 is designed after the verification module 3, and a second OCR module 6 is also provided after the annotation result processing module 4.

[0100] The annotation module 1, first OCR module 2, and verification module 3 are modules inherent to the digitization tool itself. This solution is characterized by the addition of an annotation result processing module 4, automatic error correction module 5, second OCR module 6, and verification result processing module 7, which significantly improves the efficiency and accuracy of users digitizing documents using the tool. It is important to note that this solution focuses on automating several processing steps; whether or not to add manual checks and verifications at each stage, or the number of manual quality checks, is not within the scope of this solution.

[0101] The annotation result processing module 4 is used to preprocess the annotation results of the image data annotated by the user. The image data contains several image words.

[0102] The automatic error correction module 5 includes an error identification module 51 and an error correction module 52;

[0103] The error recognition module 51 is used to determine the probability of error recognition for each character based on the pixel distribution of the recognized character and the image character; the recognized character is obtained by the first OCR module 2 through OCR recognition of the preprocessed image data;

[0104] The error correction module 52 is used to extract the target screenshot character from the image number of the corresponding image character of the character that may be misidentified; search for the target pre-stored screenshot character from the text library based on the target screenshot character; in response to the target pre-stored screenshot character being found, replace the corresponding character with the text corresponding to the target pre-stored screenshot character; in response to the pre-stored screenshot character not being found, mark the corresponding character.

[0105] The second recognition module is used to perform a second OCR recognition on the image data preprocessed by the annotation result processing module 4 to generate a second text document.

[0106] The verification result processing module 7 is used to take the text verified by the user as the first text document, compare the recognized characters of the first text document and the second text document one by one, and based on the comparison results, highlight the recognized characters in the first text document that are inconsistent with the recognition results of the second text document.

[0107] Through the above semi-automation improvements, auxiliary processing and preprocessing modules are added in multiple stages to realize automatic adjustment, correction and verification of annotation and text, effectively improving the accuracy of annotation results and text conversion results, thereby improving the overall level of data quality, while reducing the annotation burden of staff and improving the efficiency of digital work processing.

[0108] Example 3

[0109] This embodiment is similar to Embodiment 1, except that, in this embodiment, as... Figure 8 and Figure 9 As shown, the preprocessing of the annotation results for adjusting the bounding box size of user-annotated image data includes:

[0110] The position of the drag box is identified and determined, including the position of the drag box frame and the spacing between adjacent characters in the drag box.

[0111] Determine whether it conforms to the established rules. The established rules include that no picture characters are truncated by the bounding box lines of the drawn box, and the spacing between adjacent picture characters within the box is less than the set spacing threshold. When there is any situation of "picture characters being truncated by the bounding box lines" or "the spacing between adjacent picture characters within the box being greater than or equal to the set spacing threshold", it is considered not to conform to the established rules.

[0112] In response to not conforming to the established rules, adjust the size of the corresponding drawn box. First, adjust the size of the drawn box so that no picture characters are truncated by the drawn box, including adjusting the size of the drawn box to enclose the truncated picture characters within the drawn box or exclude them outside the drawn box. As Figure 8 shown, "of computer network technology" is truncated by the bounding box line. After adjustment, there is no truncation situation.

[0113] Then adjust according to the character spacing, and adjust the size of the drawn box so that the picture characters with a spacing greater than the set spacing threshold from adjacent picture characters and located relatively close to the bounding box edge are excluded outside the drawn box. As Figure 9 shown, on the left is a column of content, and on the right is another column of content. The spacing between "dai" and "mu" in the figure is significantly larger than the spacing between adjacent characters in a normal paragraph, and "mu" is relatively closer to the bounding box edge than "dai", so it is excluded outside the drawn box. After adjustment, the content on the right part is excluded outside the drawn box. If there is no situation where the picture is truncated by the bounding box line, directly enter the second-step adjustment.

[0114] Specifically, for the situation where the bounding box line truncates picture characters, it can be set to default to enclose them within the box or exclude them outside the box. In this embodiment, it is preferably determined whether to pull the picture characters into the box or exclude them from the box according to the pixel distribution ratio of the picture characters inside and outside the box. Specifically, if the pixel proportion inside the box is more than that outside the box, adjust the size of the drawn box to enclose the relevant picture characters within the box, otherwise exclude them outside the box.

[0115] This embodiment has higher flexibility compared to Embodiment 1. It can not only adjust the horizontal size but also adjust the vertical size. When the user is drawing the box, they can be more casual not only in the horizontal direction but also in the vertical direction. And there is no strict requirement for the accuracy of the user's box drawing, and the user does not need to spend energy to ensure the specification of the first box.

[0116] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Those skilled in the art of the present invention can make various modifications or supplements to the described specific embodiments or use similar methods to replace them, but will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0117] Although this paper frequently uses terms such as annotation module 1, first OCR module 2, verification module 3, annotation result processing module 4, automatic error correction module 5, error identification module 51, error correction module 52, second OCR module 6, and verification result processing module 7, the possibility of using other terms is not excluded. The use of these terms is merely for the convenience of describing and explaining the essence of this invention; interpreting them as any additional limitation would contradict the spirit of this invention.

Claims

1. A document digitization processing assistance method for improving precision and efficiency, characterized by, The method comprises: The image data labeled by the user is preprocessed; The image data contains a plurality of picture characters; The preprocessed image data is subjected to OCR recognition, and each picture character obtains a corresponding recognized character; Based on the pixel distribution of the recognized character and the picture character, the possibility of error recognition of each character is determined, and the recognized character whose pixel distribution is lower than a preset value is determined as a possible error recognition; In response to the judgment result of the possible error recognition, the corresponding picture character is extracted from the image data to obtain a target screenshot character; Based on the target screenshot character, a target pre-stored screenshot character is searched from a character library, and a plurality of error correction character pairs are pre-stored in the character library, each error correction character pair includes a pre-stored screenshot character and a corresponding character, and the method for searching the target pre-stored screenshot character comprises: Comparing the pixel distribution of the target screenshot character with a plurality of pre-stored screenshot characters; Extracting the pre-stored screenshot character with the highest pixel distribution similarity and higher than the degree threshold as the target pre-stored screenshot character; In response to finding the target pre-stored screenshot character, the corresponding character of the target pre-stored screenshot character is used to replace the corresponding recognized character; In response to not finding the pre-stored screenshot character, the corresponding recognized character is marked, the corresponding target screenshot character is saved, and an association between the recognized character and the target screenshot character is established; In response to the user's operation of storing the target screenshot character based on the target screenshot character, the target screenshot character and the character provided by the user are stored in the character library as a new error correction character pair.

2. The precision and efficiency enhanced document digitization processing assistance method of claim 1, wherein, The labeling includes box drawing and frame labeling; The preprocessed image data labeled by the user comprises: The position of the box is identified and judged, and in response to not meeting the predetermined rules, the size of the corresponding box is adjusted to meet the predetermined rules.

3. The precision and efficiency enhanced document digitization processing assistance method of claim 2, wherein, The width of the first box of each page / column is used as a standard box; The identification and judgment of the box position includes judging whether the horizontal position and width of the subsequent box are consistent with the standard box, if yes, it is considered to meet the predetermined rules, otherwise, the horizontal position and width of the corresponding box are adjusted to be consistent with the standard box.

4. The precision and efficiency enhanced document digitization processing assistance method of claim 2, wherein, The identification and judgment of the box position includes the judgment of the box line position and the judgment of the distance between adjacent picture characters in the box; The predetermined rules include that no picture character is cut off by the box line and the distance between adjacent picture characters in the box is less than a set distance threshold; The size of the corresponding box is adjusted, including adjusting the size of the box to make the cut-off picture character be enclosed in the box or excluded from the box; The size of the corresponding box is adjusted to make the picture character close to the box edge and having a distance greater than the set distance threshold be excluded from the box.

5. The precision and efficiency improving document digitizing processing assistance method according to claim 3 or 4, characterized by, The preprocessed image data labeled by the user further comprises: The text start coordinate and the text end coordinate of the text in each box are sequentially searched, and whether the text in the corresponding box is an independent paragraph is judged according to the text start coordinate and / or the text end coordinate; In response to being a non-independent paragraph, the subsequent box is added to the same paragraph as the previous box.

6. The precision and efficiency improved document digitizing processing assistance method according to claim 5, characterized by, The method for judging whether the box is an independent paragraph comprises: The coordinates of the rightmost character and the leftmost character are determined according to the overall pixel distribution of each page / column of the image data, and a standard text end horizontal coordinate and a standard text start horizontal coordinate are determined for each page / column. The first and last coordinates of the text in each pull box are found in sequence, the first coordinate of the text is the coordinate of the first character in the pull box, and the last coordinate of the text is the coordinate of the last character in the pull box; If the last horizontal coordinate of the current pull box meets the corresponding standard last horizontal coordinate position, and the first horizontal coordinate of the next pull box also meets the standard first horizontal coordinate position, the next pull box is considered as a non-independent paragraph.

7. The precision and efficiency enhanced document digitization processing assistance method of claim 1, wherein, The first OCR module is used to perform the first OCR recognition on the preprocessed image data; The method further comprises: The text after the first OCR recognition and the error correction processing is subjected to manual checking and processing by a user; The text after the user checking and processing is taken as a first text document; The second OCR module is used to perform the second OCR recognition on the preprocessed image data to generate a second text document; The recognized characters of the first text document and the second text document are compared one by one, and based on the comparison result, the recognized characters of the first text document that are inconsistent with the recognized result of the second text document are highlighted. The first OCR module and the second OCR module are different OCR modules.

8. A document digitization processing auxiliary device for improving precision and efficiency, comprising a marking module (1), a first OCR module (2) and a proofreading module (3), characterized in that, Further comprising a labeling result processing module (4) located between the labeling module (1) and the first OCR module (2), and an automatic error correction module (5) located between the first OCR module (2) and the checking module (3); The labeling result processing module (4) is used to perform labeling result preprocessing on the image data labeled by the user; the image data contains a plurality of picture characters; The automatic error correction module (5) comprises an error identification module (51) and an error correction module (52); The error identification module (51) is used to judge the possibility of error recognition of each character based on the pixel distribution of the recognized character and the picture character, and the recognized result with a pixel distribution similarity lower than a preset value between the recognized character and the picture character is judged as a possible error recognition; the recognized character is obtained by the first OCR module (2) performing OCR recognition on the preprocessed image data; The error correction module (52) is used to obtain a target screenshot character by cutting the corresponding picture character of the possible error recognition from the image data; The target pre-stored screenshot character is searched from a character library based on the target screenshot character; A plurality of error correction character pairs are pre-stored in the character library, each error correction character pair comprises a pre-stored screenshot character and a corresponding character, and the method for searching the target pre-stored screenshot character comprises: Comparing the pixel distribution of the target screenshot character with a plurality of pre-stored screenshot characters; Extracting the pre-stored screenshot character with the highest pixel distribution similarity and higher than a degree threshold as the target pre-stored screenshot character; In response to finding the target pre-stored screenshot character, the corresponding character of the target pre-stored screenshot character is used to replace the corresponding recognized character; In response to not finding the pre-stored screenshot character, the corresponding recognized character is marked, the corresponding target screenshot character is saved, and an association relationship between the recognized character and the target screenshot character is established; In response to the user's operation of storing the target screenshot character, the target screenshot character and the character provided by the user are taken as a new error correction character pair and stored in the character library.

9. The precision and efficiency enhanced document digitization processing aid of claim 8, wherein, Further comprising a second OCR module (6) and a checking result processing module (7); The second identification module is used to perform the second OCR recognition on the preprocessed image data to generate a second text document; The verification result processing module (7) is used to take the text verified by the user as the first text document, compare the recognition characters of the first text document and the second text document one by one, and highlight the recognition characters of the first text document that are inconsistent with the recognition results of the second text document based on the comparison results.

Images