A high-performance lightweight recognition model deployment method

Abstract

The present invention discloses a high-performance lightweight recognition model deployment method. The present invention integrates OCR recognition technology, AI reimbursement software and edge box to realize a remote sporadic expense reimbursement business model based on a trusted environment, improves the reliability of data security and the scalability of the business system, and reduces the work cycle of business implementation and deployment. On the technical level, the present invention constructs a general OCR recognition model for vertical fields based on a hybrid classification model, a small-angle rotation model, and a table generation model. At the same time, it realizes one-click deployment of complex models on the edge box end. Through artificial intelligence technology, the present invention integrates hybrid sample classification, small-angle rotation, table line generation, and semantic intelligent matching models on the basis of traditional remote sporadic expense reimbursement, and quickly deploys to the edge terminal with one click, forming a new model of trusted remote sporadic expense reimbursement.

Description

Technical Field

[0001] The present invention relates to image recognition and natural language processing, wherein image recognition specifically involves a hybrid classification model, a small-angle rotation model, and a table generation model, and natural language processing involves intelligent matching of medical codes. Background Art

[0002] Existing off-site sporadic expense reimbursement is mostly handled manually, which is costly and time-consuming. Traditional cloud services or local server-side deployment models consume a lot of resources and lack data security. The intelligent reimbursement system based on edge boxes can effectively reduce manual workload and improve work efficiency. At the same time, through the intelligent conversion, identification, comparison and review of medical insurance sporadic reimbursement materials, and seamless connection with the medical insurance settlement system through interfaces or data export, the full digitalization and intelligence of medical insurance sporadic reimbursement is realized, greatly improving the work efficiency of medical insurance agencies. Summary of the invention

[0003] The technical problem to be solved by the present invention is that there are many types of medical reimbursement materials, and the text arrangement in the materials is complex. Without a multi-sample classification model and table lines, the structured processing is greatly affected.

[0004] In order to solve the above technical problems, the technical solution of the present invention is to disclose a high-performance lightweight recognition model deployment method, which is characterized by comprising the following steps:

[0005] Step 1: Mixed sample classification:

[0006] Automatically detect expense reimbursement type images in mixed samples and classify them;

[0007] Step 2: Small angle rotation:

[0008] After the image is straightened at a large angle, a small angle estimation is performed to obtain the rotation angle, and the image is rotated so that the text in the image remains horizontal;

[0009] Step 3: Generate table:

[0010] Automatically detect the table area in the expense reimbursement type image and map it to the Excel table;

[0011] Step 4: NLP semantic intelligent matching to realize intelligent reimbursement coding:

[0012] Extract NLP semantic features for reimbursement codes that need to be intelligent;

[0013] Based on the natural language information obtained by OCR recognition and the reimbursement function to be implemented input by the user, the NLP semantic matching calculation model and NLP semantic calculation algorithm are used to perform intelligent matching calculation to determine the reimbursement code that needs to be intelligently implemented;

[0014] Step 5: Deploy the model on the edge box with one click:

[0015] The models required for the mixed sample classification in step 1, the small-angle rotation in step 2, the table generation in step 3, and the semantic intelligent matching model in step 4 and the system environment required for the model operation are set in a container. The container can implement the method described in any one of the above steps 1 to 4 in the corresponding computer device system, and all the above models and fusion applications can be built and started with one click using the container, so as to realize one-click deployment on the edge computer.

[0016] Preferably, step 1 comprises the following steps:

[0017] Step 101, processing the input file into a picture type;

[0018] Step 102: Send the image to the image target detection model, and the image target detection model automatically detects the area belonging to the type of ID card, social security card, expense list, medical invoice, laboratory test report, and value-added tax invoice, and defines the area as the foreground area;

[0019] Step 103: Send the detected foreground areas one by one to the neural network classifier, output the probability that each foreground area belongs to the type of identity card, social security card, expense list, medical invoice, test report or value-added tax invoice, and take the maximum type probability as the type of the foreground area.

[0020] Preferably, step 2 comprises the following steps:

[0021] Step 201: Detect text lines in the image using a segmentation-based text detection model to obtain position information of all text lines;

[0022] Step 202: filter and discard curved text lines;

[0023] Step 203: simulate and estimate the slope, aspect ratio, and height of the remaining text lines, and then further filter out some outlier text lines in combination with a clustering algorithm;

[0024] Step 204: Estimate the tilt angle of the entire image based on the slope of the remaining text lines, and perform a preliminary small-angle rotation and straightening on the image;

[0025] Step 205: Use the FLD line detection algorithm to perform line detection on the image to obtain position information of horizontal and vertical lines in the image;

[0026] Step 206, calculating the slope and length information of the line segment, and filtering out some short lines, lines close to the image edge, and interfering lines by combining the clustering algorithm and the text line position information;

[0027] Step 207: Estimate the tilt angle of the entire image based on the slope of the remaining horizontal or vertical lines, and perform a second small-angle rotation to the image to straighten it.

[0028] Preferably, step 3 is implemented based on a table generation module, a text detection generation module and table text area mapping, wherein:

[0029] The table generation module uses the deep learning semantic segmentation method to classify the table line area as the foreground and the non-table line area as the background, and then obtains each table unit based on the minimum connected domain method;

[0030] The text detection generation module automatically detects text areas and then identifies the text content in each area;

[0031] The table text area mapping module automatically matches the generated table cells with the text areas detected by the text detection and recognition model.

[0032] Preferably, the matching rule adopted by the table text area mapping module is: for each generated table cell, the overlap value between all detected text areas and the cell is calculated, and the text content corresponding to the text area with the highest overlap is identified as the text content in the table, which is organized and output to an Excel table.

[0033] Preferably, in step 5, the model operating environment includes the deep learning framework required by the model and the corresponding database for storage.

[0034] Preferably, in step 5, all required models, databases, deep learning frameworks and other required environments are written into a container through a computer script, and a startup command for the container is set.

[0035] Preferably, in step 5, an operating environment required by the container is installed on an edge computer device and a computer-readable storage medium.

[0036] Preferably, the computer device is characterized in that the system can run the container where the model is located and support the system environment required by the container to implement the method described in any one of steps 1 to 4.

[0037] The present invention integrates OCR recognition technology, AI reimbursement software and edge boxes to realize a remote sporadic expense reimbursement business model based on a trusted environment, improves the reliability of data security and the scalability of the business system, and reduces the work cycle of business implementation and deployment. On the technical level, the present invention constructs a general OCR recognition model for vertical fields based on a hybrid classification model, a small-angle rotation model, and a table generation model. At the same time, it realizes one-click deployment of complex models on the edge box end. Through artificial intelligence technology, the present invention integrates hybrid sample classification, small-angle rotation, table line generation, and semantic intelligent matching models on the basis of traditional remote sporadic expense reimbursement, and quickly deploys it to the edge terminal with one click, forming a new model of trusted remote sporadic expense reimbursement. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 The types of foreground areas obtained in the embodiments of the present invention are illustrated. DETAILED DESCRIPTION

[0039] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0040] A high-performance lightweight recognition model deployment method disclosed in an embodiment of the present invention includes the following steps:

[0041]

Step 1

[0042] This step automatically detects expense reimbursement type images in the mixed sample and classifies them. It mainly includes the following steps:

[0043] Step 101: Process the input file into image types, including jpg, jpeg, png and other formats;

[0044] Step 102: Send the image to the image object detection model, and the model automatically detects seven types of areas including identity card, social security card, expense list, medical invoice, laboratory test report, and value-added tax invoice, which are collectively referred to as foreground areas;

[0045] Step 103: Figure 1 As shown, the detected foreground areas (each image in the mixed sample may include multiple foreground areas) are sent to the neural network classifier one by one, and the probability that each foreground area belongs to the above seven types is output, and the maximum type probability is taken as the type of the foreground area.

[0046]

Step 2

[0047] The image after large-angle correction is estimated at a small angle (within 45 degrees) to obtain the rotation angle. The image is rotated so that the text in the image remains horizontal. The main steps include:

[0048] Step 201: Detect text lines in the image using a segmentation-based text detection model to obtain position information of all text lines;

[0049] Step 202: filter and discard curved text lines;

[0050] Step 203: simulate and estimate the slope, aspect ratio, height and other information of the remaining text lines, and then further filter out some outlier text lines in combination with the clustering algorithm;

[0051] Step 204: Estimate the tilt angle of the entire image based on the slope of the remaining text lines, and perform a preliminary small-angle rotation and straightening on the image;

[0052] Step 205: Use the FLD line detection algorithm to perform line detection on the image to obtain the position information of the horizontal and vertical lines in the image. In this step, the line detection is performed after the initial small-angle rotation, which can significantly improve the line detection effect;

[0053] Step 206: Calculate the slope, length and other information of the line segment, and filter out some short lines, lines close to the image edge and interfering lines by combining the clustering algorithm and the text line position information;

[0054] Step 207: Estimate the tilt angle of the entire image based on the slope of the remaining horizontal or vertical lines, and perform a second small-angle rotation to the image to straighten it.

[0055]

Step 3

[0056] This step automatically detects the table area in the expense reimbursement type image and maps it to an Excel table, making it easier to analyze the data structure. This step is mainly divided into three modules, namely, table generation module, text detection generation module, and table text area mapping.

[0057] The table generation module is different from the traditional image processing straight line detection and then table synthesis. The present invention uses a deep learning semantic segmentation method to classify the table line area as foreground and the non-table line area as background, and then obtains each table unit based on the minimum connected domain method, and the result is more accurate.

[0058] The text detection generation module automatically detects text areas and then identifies the text content in each area.

[0059] The table text area mapping module automatically matches the generated table cells with the text areas detected by the text detection and recognition model. The matching rule is that for each generated table cell, the overlap value between all detected text areas and the cell is calculated, and the text content corresponding to the text area with the highest overlap is used as the text content in the table, which is then organized and output to an Excel table.

[0060]

Step 4

[0061] NLP semantic feature extraction will be performed on the reimbursement codes that need to be intelligent.

[0062] Based on the natural language information obtained by OCR recognition and the reimbursement function to be implemented input by the user, the NLP semantic matching calculation model and NLP semantic calculation algorithm are used to perform intelligent matching calculations to determine the reimbursement code that needs to be intelligently implemented.

[0063] Reimbursement function is performed based on the results of NLP semantic intelligent matching.

[0064] [Step 5] One-click deployment of the model on the edge box

[0065] The model required for mixed sample classification, small angle rotation, table line generation, and semantic intelligent matching model and the system environment required for model operation are set in a container. The container is characterized in that the method described in any one of the above steps 1 to 4 can be implemented in the corresponding computer device system.

[0066] The model running environment includes the deep learning framework required by the above models and the corresponding database for storage.

[0067] All the required models, databases, deep learning frameworks and other required environments can be written into a container through computer scripts, and the startup command of the container can be set.

[0068] The operating environment required by the container is installed on an edge computer device (including a processor, a memory) and a computer-readable storage medium. The computer device is characterized in that the system can run the container where the model is located and support the system environment required by the container to implement any of the methods described in steps 1 to 4 above.

[0069] Use containers to build and start all the above models and integrated applications with one click, and achieve one-click deployment on edge computers.

Claims

1. A high-performance lightweight recognition model deployment method, characterized in that: The following steps are involved: Step 1: Mixed sample classification: Automatically detect expense reimbursement type images in mixed samples and classify them; Step 2: Small angle rotation: After the image is straightened at a large angle, a small angle estimation is performed to obtain the rotation angle, and the image is rotated so that the text in the image remains horizontal; Step 3: Generate table: Automatically detect the table area in the expense reimbursement type image and map it to the Excel table; Step 4: NLP semantic intelligent matching to realize intelligent reimbursement coding: Extract NLP semantic features for reimbursement codes that need to be intelligent; Based on the natural language information obtained by OCR recognition and the reimbursement function to be implemented input by the user, the NLP semantic matching calculation model and NLP semantic calculation algorithm are used to perform intelligent matching calculation to determine the reimbursement code that needs to be intelligently implemented; Step 5: Deploy the model on the edge box with one click: The models required for the mixed sample classification in step 1, the small-angle rotation in step 2, the table generation in step 3, and the semantic intelligent matching model in step 4, as well as the system environment required for the model operation are set in a container. The container can implement the method described in any one of steps 1 to 4 in the corresponding computer device system, and all the above models and fusion applications can be built and started with one click using the container, so as to achieve one-click deployment on the edge computer.

2. A high-performance lightweight recognition model deployment method as claimed in claim 1, characterized in that: The step 1 comprises the following steps: Step 101, processing the input file into a picture type; Step 102: Send the image to the image target detection model, and the image target detection model automatically detects the area belonging to the type of ID card, social security card, expense list, medical invoice, laboratory test report, and value-added tax invoice, and defines the area as the foreground area; Step 103: Send the detected foreground areas one by one to the neural network classifier, output the probability that each foreground area belongs to the type of identity card, social security card, expense list, medical invoice, test report or value-added tax invoice, and take the maximum type probability as the type of the foreground area.

3. A high-performance lightweight recognition model deployment method as described in claim 1, characterized in that: The step 2 comprises the following steps: Step 201: Detect text lines in the image using a segmentation-based text detection model to obtain position information of all text lines; Step 202: filter and discard curved text lines; Step 203: simulate and estimate the slope, aspect ratio, and height of the remaining text lines, and then further filter out some outlier text lines in combination with a clustering algorithm; Step 204: Estimate the tilt angle of the entire image based on the slope of the remaining text lines, and perform a preliminary small-angle rotation and straightening on the image; Step 205: Use the FLD line detection algorithm to perform line detection on the image to obtain position information of horizontal and vertical lines in the image; Step 206: Calculate the slope and length information of the line segment, and filter out some short lines, lines close to the edge of the image, and interfering lines by combining the clustering algorithm and the text line position information; Step 207: Estimate the tilt angle of the entire image based on the slope of the remaining horizontal or vertical lines, and perform a second small-angle rotation to the image to straighten it.

4. A high-performance lightweight recognition model deployment method as claimed in claim 1, characterized in that: The step 3 is implemented based on a table generation module, a text detection generation module and table text area mapping, wherein: The table generation module uses the deep learning semantic segmentation method to classify the table line area as the foreground and the non-table line area as the background, and then obtains each table unit based on the minimum connected domain method; The text detection generation module automatically detects text areas and then identifies the text content in each area; The table text area mapping module automatically matches the generated table cells with the text areas detected by the text detection and recognition model.

5. A high-performance lightweight recognition model deployment method as described in claim 4, characterized in that: The matching rule adopted by the table text area mapping module is: for each generated table cell, the overlap value between all detected text areas and the cell is calculated, and the text content corresponding to the text area with the highest overlap is identified as the text content in the table, which is sorted and output to an Excel table.

6. A high-performance lightweight recognition model deployment method as claimed in claim 1, characterized in that: In step 5, the model running environment includes the deep learning framework required by the model and the corresponding database for storage.

7. A high-performance lightweight recognition model deployment method as claimed in claim 1, characterized in that: In step 5, all required models, databases, deep learning frameworks and other required environments are written into a container through a computer script, and the startup command of the container is set.

8. A high-performance lightweight recognition model deployment method as claimed in claim 1, characterized in that: In step 5, an operating environment required by the container is installed on an edge computer device and a computer-readable storage medium.

9. A high-performance lightweight recognition model deployment method as claimed in claim 8, characterized in that: The computer device is characterized in that the system can run the container where the model is located and support the system environment required by the container to implement the method described in any one of steps 1 to 4.

Images