A BPO image review system based on large model

Abstract

The invention discloses a BPO picture audit system based on a large model, which relates to the technical field of graphic data reading, including: obtaining edge contours in a picture based on picture accuracy and extracting contour features; obtaining edge contours of the comparison picture by identifying pictures uploaded by users and retrieving comparison pictures with similar features through a search engine, comparing the high-quality contour feature set of the pictures uploaded by users with the edge contour features of the comparison pictures, extracting picture data based on the outer contour of the user-uploaded picture that is different from the distinguishing contour features of the comparison pictures, performing path simulation, and determining multiple paths from one point to multiple points in the picture data; determining text data contained in the distinguishing contour features based on the paths, obtaining feature points in the picture to be audited through a picture audit system, analyzing the matching degree between the feature data and the text, processing the text, determining the information contained in the corresponding text, and auditing the deformed text.

Description

Technical Field

[0001] The present invention relates to the technical field of graphic data reading, in particular to a BPO picture review system based on a large model. Background Art

[0002] BPO image review systems are often used to automatically and efficiently review large amounts of image content to ensure that it complies with platform or business regulations. The following are some common functions and features:

[0003] Pornography Detection: Detect whether images contain pornographic or inappropriate content.

[0004] Advertisement detection: Identify advertising content such as watermarks, QR codes, barcodes, etc. in images.

[0005] Sensitive information identification: Detect whether the image contains sensitive signs, places or objects.

[0006] Public figure recognition: Identify whether the image contains public figures, such as artists, athletes, etc.

[0007] Image quality detection: Evaluate the clarity and aesthetics of the image, and detect whether there are problems such as blur or loss of focus.

[0008] Customized audit: Customize audit labels and policies based on business needs, and flexibly adjust audit standards.

[0009] Announcement No. CN113344015A discloses a method and device for reviewing annotated images, which relates to the field of computer technology. A specific implementation of the method includes: counting the number of annotated frames for multiple annotated images of the same image data to determine the correct number of frames that the image data should be annotated with; reviewing the number of annotated frames for multiple annotated images based on the correct number of frames; matching the position and annotated text of each annotated frame of the annotated image that has passed the annotated frame number review with the reference frame; and determining the annotated image that has passed the annotated frame number review and whose positions and annotated texts have successfully matched as the annotated image that has passed the final review. This implementation does not rely on the subjective experience and attitude of the reviewer, avoids review errors such as misjudgment and omission, improves the accuracy of the annotated images after review, can quickly locate the annotation errors during the review process, can effectively judge the spaces and pixel differences within the text line, and does not require the annotator to locate the rejected data where the modification is needed.

[0010] Image review can reduce the number of illegal images on the platform and ensure the platform's network environment. The existing image review methods often focus on graphic images composed of color blocks, lines, shapes and other elements. Through special color blocks, lines, shapes and other feature recognition, the data information contained in the image is obtained to make corresponding reviews.

[0011] When extracting text data contained in an image, the text often produces ambiguity, resulting in illegal information from normal text combinations. In order to pass the image review system, some texts are deformed, resulting in the image review system being unable to recognize the existence of the text and identifying it as a graphic. However, the graphic generated by the text does not contain illegal features after the image review. For this reason, this application proposes a BPO image review system based on a large model to improve the deformed text contained in the image. Summary of the invention

[0012] One of the purposes of the present invention is to provide a BPO picture review system based on a large model to review the text content of pictures containing deformed text.

[0013] To achieve the above objectives, the present invention is implemented through the following technical solutions: A BPO picture review system based on a large model, comprising:

[0014] An image processing unit, which obtains the image uploaded by the user and performs noise reduction on the image, performs preprocessing on the noise-reduced image, and generates a plurality of preprocessed difference images;

[0015] The feature extraction unit obtains the edge contour in the image based on the image accuracy and extracts the contour features: the image is divided into multiple regions according to the image accuracy, the contour edges between the multiple regions are determined, and different regions are marked and the contours in each region are obtained to determine the contour edges;

[0016] The denoising unit performs filtering and denoising on the contour features, removes the contour feature noise, improves the contour feature image quality, and generates a high-quality contour feature set;

[0017] The contour comparison unit identifies the user-uploaded image and retrieves the comparison image with similar features from the search engine to obtain the edge contour of the comparison image. The comparison image is compared with the edge contour features of the comparison image through the high-quality contour feature set of the user-uploaded image to determine the comparison image that has the same contour features as the user-uploaded image and has different contour features.

[0018] An extraction unit extracts image data based on an outer contour of the image uploaded by the user that is different from the contour features of the comparison image, identifies the image data, determines the content of the image data, and performs path simulation based on the image data to determine multiple paths from one point to multiple points in the image data;

[0019] The path fitting model accesses the text library and corrects the path. Based on the path, the text data contained in the distinguishing contour features is determined, and the text is arranged and combined to determine the multiple meanings of the text contained in the picture uploaded by the user.

[0020] Determine the multiple meanings expressed by the deformable text, search for the meaning expressed by the text appearing on the platform, and determine the graphic data contained in the pictures uploaded by the user.

[0021] In one or more embodiments of the present invention, the user-uploaded pictures after noise reduction are preprocessed in a manner including: grayscale, normalization, binarization, contrast enhancement, and morphological operations to determine a plurality of distinguishing pictures;

[0022] When extracting contour features of the user-uploaded image, the contour features in the difference image are extracted simultaneously to determine the difference range of the contour features between the user-uploaded image and the difference image.

[0023] In one or more embodiments of the present invention, the gradient of each pixel in the image is calculated by an edge detection algorithm based on the picture accuracy, and the magnitude of the gradient represents the intensity of the brightness change, and the direction represents the direction of the change;

[0024] Find connected edge pixels, extract contours from user-uploaded images, mark contour features, and determine boundaries of different regions;

[0025] The image is divided into multiple regions based on contour features and each region is labeled.

[0026] In one or more embodiments of the present invention, the method for comparing the outlines of the user-uploaded image and the comparison image is as follows:

[0027] Identify user-uploaded images, search for similar images through search engines, and retrieve comparison images with the same edge features;

[0028] Extracting contour features of the comparison picture, and determining a processing state of the comparison picture based on a missing state of the contour features of the comparison picture;

[0029] The high-quality contour features of the pictures uploaded by the user are matched and compared with the contour features of the pictures to determine the distinguishing contour features.

[0030] In one or more embodiments of the present invention, distinguishing contour features are extracted, image data within the outer contour is extracted based on the distinguishing feature contour as the outer contour, the image data is identified, the content of the image data is determined, and the content contained in the image data is analyzed.

[0031] In one or more embodiments of the present invention, the path simulation method based on the outer contour internal image data is as follows:

[0032] Determine the outer contour shape: read the distinguishing contour features and extract the outer contour through the findContours function;

[0033] Determine multiple path points based on the outer contour: perform polygonal approximation on each contour through a polygonal approximation algorithm to determine the path points;

[0034] Connecting path points based on the fit of the path points to the outer contour shape: generating a simulation path based on the path points, and ensuring that the simulation path is within the outer contour shape;

[0035] Determine multiple simulation paths: connect the path points into complete paths;

[0036] Identify simulated paths and remove invalid paths based on text strokes.

[0037] In one or more embodiments of the present invention, the specific steps of determining the path points by polygonal approximation are as follows:

[0038] Calculate the perimeter of the contour: Use the arcLength function to calculate the perimeter of the contour;

[0039] Set the approximation accuracy: Set the approximation accuracy parameter epsilon to 1% to 5% of the circumference according to the circumference of the contour;

[0040] Perform polygon approximation: Use the approxPolyDP function to perform polygon approximation and generate simplified polygons;

[0041] Determine the simplified polygon corner points as path points and mark the path points.

[0042] In one or more embodiments of the present invention, coordinates are established based on the outer edge area to determine the coordinate position of the simulated path:

[0043] For a continuous curve in the simulation path, the curvature of the simulation path is calculated by the derivative of the curve:

[0044] The rectangular coordinate equation of the simulated path curve is: y=k(x) , the curvature of the simulated path curve at point M K for:

[0045] ;

[0046] For the curvature of the simulated path at discrete points, the three-point method is used to calculate the curvature of the simulated path:

[0047] Determine three consecutive points in the simulated path curve: P1(x1,y1), P2(x2,y2) and P3(x3,y3), simulate the path curvature K for:

[0048] ;

[0049] Calculate the curvature of the simulated path generated by the adjacent path points and determine the invalid path based on the change in curvature;

[0050] Determine the change of text stroke paths, remove invalid paths that do not conform to the text strokes and invalid paths within the curvature change range.

[0051] In one or more embodiments of the present invention, the path fitting model obtains the simulated path and fits the text in the text library, and the steps are as follows:

[0052] Use cubic Bezier curve to fit the path. The Bezier curve defines the simulation path through endpoints and control points. The control points are determined by calculating the bending position of the simulation path and the ratio of its two ends.

[0053] Feature point extraction: Extract feature points from the fitted simulation path to describe the shape and structure of the simulation path;

[0054] Use optical character recognition (OCR) technology to extract text information from the fitted path and convert it into digital text;

[0055] Path optimization: Correct and optimize the path based on the recognized text data, and increase the text data range corresponding to the simulated path.

[0056] In one or more embodiments of the present invention, the simulation path is corrected by the text data, the strokes corresponding to the simulation path are defined, the simulation path is corrected according to the strokes corresponding to the simulation path, and the text library is fitted again by the corrected simulation path;

[0057] By combining the fitted characters, homophones and similar characters are combined to extend the semantics, and determine the multiple meanings corresponding to the deformed characters;

[0058] Review the various meanings of morphed text and analyze whether the morphed text contained in the pictures uploaded by users complies with the platform review rules.

[0059] Through the above technical solution, the present invention has the following beneficial effects:

[0060] 1. Obtain feature points in the image to be reviewed through the image review system, determine the data contained in the image through the feature points, extract the complete image, determine the feature data in the image, analyze the degree of match between the feature data and the text, process the text, determine the information contained in the corresponding text, and review the deformed text.

[0061] 2. Search for comparison pictures corresponding to the pictures uploaded by the user through the search engine to determine the distinguishing contour features between the pictures uploaded by the user and the comparison pictures, and then distinguish the contour features to identify the deformed text in the pictures uploaded by the user, thereby improving the review effect of the picture review system on deformed text.

[0062] 3. Perform path simulation through the outer contour of the deformed text to determine the various simulation paths contained in the outer contour area, determine the text data corresponding to the simulation path based on the simulation path, determine the multiple meanings of the text corresponding to the outer contour based on the text data, and review the deformed text based on the multiple meanings.

[0063] 4. Use the path fitting model to correct the outer contour of the deformed text, ensure the accuracy of the simulation path when fitting the text, ensure the accuracy of the deformed text recognition, determine the meaning expressed by the simulation path, reduce the volume of permutations and combinations required in the fitting process, and achieve the purpose of rapid recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a system schematic diagram of the present invention. DETAILED DESCRIPTION

[0065] The following will disclose multiple embodiments of the present invention with the accompanying drawings. For the purpose of clear description, many practical details will be described together in the following description. However, it should be understood that these practical details should not be used to limit the present invention. In other words, in some embodiments of the present invention, these practical details are not necessary. And if it is possible to implement, the features of different embodiments can be applied interchangeably.

[0066] Unless otherwise defined, all words (including technical and scientific terms) used herein have their usual meanings, which can be understood by those familiar with this field. Furthermore, the definitions of the above words in commonly used dictionaries should be interpreted in the content of this specification as consistent with the meanings of the relevant fields of the present invention. Unless otherwise clearly defined, these words will not be interpreted as idealized or overly formal meanings.

[0067] See also Figure 1 The present invention provides a BPO picture review system based on a large model, which can review pictures, especially pictures with deformed text in them, read graphic data, and determine text information.

[0068] In one embodiment, the image review system includes:

[0069] An image processing unit, which obtains the image uploaded by the user and performs noise reduction on the image, performs preprocessing on the noise-reduced image, and generates a plurality of preprocessed difference images;

[0070] The feature extraction unit obtains the edge contour in the image based on the image accuracy and extracts the contour features: the image is divided into multiple regions according to the image accuracy, the contour edges between the multiple regions are determined, and different regions are marked and the contours in each region are obtained to determine the contour edges;

[0071] The denoising unit performs filtering and denoising on the contour features, removes the contour feature noise, improves the contour feature image quality, and generates a high-quality contour feature set;

[0072] The contour comparison unit identifies the user-uploaded image and retrieves the comparison image with similar features from the search engine to obtain the edge contour of the comparison image. The comparison image is compared with the edge contour features of the comparison image through the high-quality contour feature set of the user-uploaded image to determine the comparison image that has the same contour features as the user-uploaded image and has different contour features.

[0073] An extraction unit extracts image data based on an outer contour of the image uploaded by the user that is different from the contour features of the comparison image, identifies the image data, determines the content of the image data, and performs path simulation based on the image data to determine multiple paths from one point to multiple points in the image data;

[0074] The path fitting model accesses the text library and corrects the path. Based on the path, the text data contained in the distinguishing contour features is determined, and the text is arranged and combined to determine the multiple meanings of the text contained in the picture uploaded by the user.

[0075] Determine the multiple meanings expressed by the deformable text, search for the meaning expressed by the text appearing on the platform, and determine the graphic data contained in the pictures uploaded by the user.

[0076] In this embodiment, the text data contained in the picture uploaded by the user is uploaded in the form of similar text or homophones. Since there is no difficulty in recognizing text, it is only necessary to determine the expression meaning of the similar text or homophones, that is, the temperature data contained in the uploaded picture can be reviewed. For the picture review system, the deformed text is difficult to recognize, and the user can make associations based on the deformed text. After the uploaded picture is reviewed by the picture review system, there are still illegal text data.

[0077] The text path is simulated through the text outline, and the strokes contained in multiple text outlines are simulated. The text in the text library is retrieved through these strokes, and the text path is fitted. The various text data contained in the text outline are determined, and the text arrangement and combination determine the various meanings contained in the uploaded image, so as to review the deformable text.

[0078] In one embodiment, the user-uploaded image after noise reduction is preprocessed in a manner including: grayscale, normalization, binarization, contrast enhancement, and morphological operations to determine a plurality of distinguishing images;

[0079] When extracting contour features of the user-uploaded image, the contour features in the difference image are extracted simultaneously to determine the difference range of the contour features between the user-uploaded image and the difference image.

[0080] In this embodiment, there are differences between the difference pictures after preprocessing and the pictures uploaded by the user. Compared with the pictures uploaded by the user, multiple difference pictures lack contour features. By determining the contour features contained in multiple difference pictures and the pictures uploaded by the user, the common contour features contained in the difference pictures and the pictures uploaded by the user can be determined.

[0081] When searching for similar images based on images uploaded by users through a search engine, the processing status of the similar images can be determined by determining changes in contour features of the similar images.

[0082] In one embodiment, the gradient of each pixel in the image is calculated by an edge detection algorithm based on the picture accuracy, the magnitude of the gradient represents the intensity of the brightness change, and the direction represents the direction of the change;

[0083] Find connected edge pixels, extract contours from user-uploaded images, mark contour features, and determine boundaries of different regions;

[0084] The image is divided into multiple regions based on contour features and each region is labeled.

[0085] In this embodiment, when marking the area, it is ensured that the changes in each area are clear and there is no overlap. The contour features in the picture uploaded by the user are determined through the accuracy of the picture, and the area is split according to the contour features. The contour is divided into different areas. When there is text data in the picture, part of the picture will be covered. Therefore, the occlusion range of the text can be determined by the contour features.

[0086] By differentiating the precision between the multiple regions, the connection between the multiple regions can be determined, thereby determining multiple text data with unrelated outer contours.

[0087] In one embodiment, the method for comparing the outlines of the user-uploaded image and the comparison image is as follows:

[0088] Identify user-uploaded images, search for similar images through search engines, and retrieve comparison images with the same edge features;

[0089] Extracting contour features of the comparison picture, and determining a processing state of the comparison picture based on a missing state of the contour features of the comparison picture;

[0090] The high-quality contour features of the pictures uploaded by the user are matched and compared with the contour features of the pictures to determine the distinguishing contour features.

[0091] In this embodiment, by searching the pictures uploaded by the user, comparison pictures similar to the pictures uploaded by the user are determined. The pictures uploaded by the user are searched by the search engine to determine whether the search engine has pictures that are consistent with them and original pictures that are not covered by text. After obtaining the original pictures that are not covered by text, the position of the text and the outer contour of the text can be determined.

[0092] Among them, when the high-quality contour features of the user-uploaded picture completely match the contour features of the comparison picture, the text data in the user-uploaded picture cannot be obtained. Therefore, when performing picture contour feature matching, the text contour in the user-uploaded picture can be determined based on the distinguishing contour features.

[0093] In one embodiment, distinguishing contour features are extracted, and image data within the outer contour is extracted based on the distinguishing feature contour as an outer contour, the image data is identified, the content of the image data is determined, and the content contained in the image data is analyzed.

[0094] In this embodiment, when it is impossible to determine whether the image data within the outer contour is text or a picture, in order to ensure the accuracy of the detection, the image data within the outer contour is identified and the content contained in the image data is determined. First, the image data contained in the outer contour is searched through a search engine, and a retrieval image containing similar image data is retrieved to obtain the retrieval image content, and determine whether the retrieval image corresponds to text.

[0095] In one embodiment, the path simulation method based on the image data inside the outer contour is as follows:

[0096] Determine the outer contour shape: read the distinguishing contour features and extract the outer contour through the findContours function;

[0097] Determine multiple path points based on the outer contour: perform polygonal approximation on each contour through a polygonal approximation algorithm to determine the path points;

[0098] Connecting path points based on the fit of the path points to the outer contour shape: generating a simulation path based on the path points, and ensuring that the simulation path is within the outer contour shape;

[0099] Determine multiple simulation paths: connect the path points into complete paths;

[0100] Identify simulated paths and remove invalid paths based on text strokes.

[0101] In this embodiment, by determining the path points and simulating the path, multiple simulated paths can be generated within the outer contour area, multiple structures passing through the deformed characters can be determined, and the text data corresponding to the deformed characters can be determined based on the multiple structures of the deformed characters, thereby achieving the purpose of recognizing the deformed characters.

[0102] In one embodiment, the specific steps of determining the path points by polygonal approximation are as follows:

[0103] Calculate the perimeter of the contour: Use the arcLength function to calculate the perimeter of the contour;

[0104] Set the approximation accuracy: Set the approximation accuracy parameter epsilon to 1% to 5% of the circumference according to the circumference of the contour;

[0105] Perform polygon approximation: Use the approxPolyDP function to perform polygon approximation and generate simplified polygons;

[0106] Determine the simplified polygon corner points as path points and mark the path points.

[0107] In this embodiment, the path points are determined by polygonal approximation, and multiple path points within the outer contour can be obtained. The path is simulated by the path points generated by polygonal approximation. However, since the path points generated by polygonal approximation of deformable text may be repeated, it is necessary to screen the simulated paths to simplify the volume of the simulated paths and improve the review efficiency of the image review system.

[0108] Exemplary, the code to implement polygonal approximation is as follows:

[0109] approx_contours = []

[0110] for contour in contours:

[0111] # Calculate the perimeter of the contour

[0112] perimeter = cv2.arcLength(contour, True)

[0113] # Set the approximation accuracy

[0114] epsilon = 0.01 * perimeter

[0115] # Perform polygonal approximation

[0116] approx = cv2.approxPolyDP(contour, epsilon, True)

[0117] # Save the approximated polygon

[0118] approx_contours.append(approx).

[0119] In one embodiment, coordinates are established based on the outer edge area to determine the simulated path coordinate position:

[0120] For a continuous curve in the simulation path, the curvature of the simulation path is calculated by the derivative of the curve:

[0121] The rectangular coordinate equation of the simulated path curve is: y=k(x) , the curvature of the simulated path curve at point M K for:

[0122] ;

[0123] For the curvature of the simulated path at discrete points, the three-point method is used to calculate the curvature of the simulated path:

[0124] Determine three consecutive points in the simulated path curve: P1(x1,y1), P2(x2,y2) and P3(x3,y3), simulate the path curvature K for:

[0125] ;

[0126] Calculate the curvature of the simulated path generated by the adjacent path points and determine the invalid path based on the change in curvature;

[0127] Determine the change of text stroke paths, remove invalid paths that do not conform to the text strokes and invalid paths within the curvature change range.

[0128] In this embodiment, the curvature variation range can be set according to the audit intensity, so that the audit intensity can be increased or reduced, and adjusted according to usage. Since consistent text stroke paths are all fixed, when the path point is simulated within the outer contour, a simulated path that does not conform to the stroke path will be generated. Removing invalid paths can increase the audit speed.

[0129] In one embodiment, the path fitting model obtains the simulated path and fits the text in the text library in the following steps:

[0130] Use cubic Bezier curve to fit the path. The Bezier curve defines the simulation path through endpoints and control points. The control points are determined by calculating the bending position of the simulation path and the ratio of its two ends.

[0131] Feature point extraction: Extract feature points from the fitted simulation path to describe the shape and structure of the simulation path;

[0132] Use optical character recognition (OCR) technology to extract text information from the fitted path and convert it into digital text;

[0133] Path optimization: Correct and optimize the path based on the recognized text data, and increase the text data range corresponding to the simulated path.

[0134] In this embodiment, the text range corresponding to the simulation path is determined by fitting the path, and the number of simulation paths is greater than the number of strokes of the fitted text. Therefore, when fitting, the simulation path within the outer contour range will fit multiple text data, and the text data needs to be further extracted and identified to determine the meaning of the text contained in the outer contour range after combination.

[0135] In one embodiment, the simulation path is corrected using text data, the strokes corresponding to the simulation path are defined, the simulation path is corrected according to the strokes corresponding to the simulation path, and the text library is fitted again using the corrected simulation path.

[0136] In this embodiment, since the simulation path is limited by the range of the outer contour, and the outer contour is generated after the text is deformed, a simulation path often has multiple corresponding strokes. By defining the simulation path and determining the strokes of the simulation path, the accuracy of the text fitted by the simulation path is ensured.

[0137] By combining the fitted characters, homophones and similar characters are combined to extend the semantics, and determine the multiple meanings corresponding to the deformed characters;

[0138] Review the various meanings of morphed text and analyze whether the morphed text contained in the pictures uploaded by users complies with the platform review rules.

[0139] In summary, the technical solutions disclosed in the above embodiments of the present invention have at least the following advantages:

[0140] 1. Obtain feature points in the image to be reviewed through the image review system, determine the data contained in the image through the feature points, extract the complete image, determine the feature data in the image, analyze the degree of match between the feature data and the text, process the text, determine the information contained in the corresponding text, and review the deformed text.

[0141] 2. Search for comparison pictures corresponding to the pictures uploaded by the user through the search engine to determine the distinguishing contour features between the pictures uploaded by the user and the comparison pictures, and then distinguish the contour features to identify the deformed text in the pictures uploaded by the user, thereby improving the review effect of the picture review system on deformed text.

[0142] 3. Perform path simulation through the outer contour of the deformed text to determine the various simulation paths contained in the outer contour area, determine the text data corresponding to the simulation path based on the simulation path, determine the multiple meanings of the text corresponding to the outer contour based on the text data, and review the deformed text based on the multiple meanings.

[0143] 4. Use the path fitting model to correct the outer contour of the deformed text, ensure the accuracy of the simulation path when fitting the text, ensure the accuracy of the deformed text recognition, determine the meaning expressed by the simulation path, reduce the volume of permutations and combinations required in the fitting process, and achieve the purpose of rapid recognition.

[0144] Although the present invention is disclosed in conjunction with the above embodiments, it is not intended to limit the present invention. Anyone familiar with this technology can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the definition of the attached claims.

Claims

1. A BPO picture review system based on a large model, characterized in that: include: An image processing unit, which obtains the image uploaded by the user and performs noise reduction on the image, performs preprocessing on the noise-reduced image, and generates a plurality of preprocessed difference images; The feature extraction unit obtains the edge contour in the image based on the image accuracy and extracts the contour features: the image is divided into multiple regions according to the image accuracy, the contour edges between the multiple regions are determined, and different regions are marked and the contours in each region are obtained to determine the contour edges; The denoising unit performs filtering and denoising on the contour features, removes the contour feature noise, improves the contour feature image quality, and generates a high-quality contour feature set; The contour comparison unit identifies the user-uploaded image and retrieves the comparison image with similar features from the search engine to obtain the edge contour of the comparison image. The comparison image is compared with the edge contour features of the comparison image through the high-quality contour feature set of the user-uploaded image to determine the comparison image that has the same contour features as the user-uploaded image and has different contour features. An extraction unit extracts image data based on an outer contour of the image uploaded by the user that is different from the contour features of the comparison image, identifies the image data, determines the content of the image data, and performs path simulation based on the image data to determine multiple paths from one point to multiple points in the image data; The gradient of each pixel in the image is calculated based on the image accuracy through the edge detection algorithm. The magnitude of the gradient represents the intensity of the brightness change, and the direction represents the direction of the change. Find connected edge pixels, extract contours from user-uploaded images, mark contour features, and determine boundaries of different regions; Divide the image into multiple regions based on contour features and mark each region; The path fitting model accesses the text library and corrects the path. Based on the path, the text data contained in the distinguishing contour features is determined, and the text is arranged and combined to determine the multiple meanings of the text contained in the picture uploaded by the user. Based on the multiple meanings of the text contained in the user-uploaded picture, the meaning expressed by the text appearing on the search platform is determined to determine the graphic data contained in the user-uploaded picture; Establish coordinates based on the outer contour area and determine the coordinate position of the simulation path: For a continuous curve in the simulation path, the curvature of the simulation path is calculated by the derivative of the curve: The rectangular coordinate equation of the simulated path curve is: y=k(x) , the curvature of the simulated path curve at point M K for: ； For the curvature of the simulated path at discrete points, the three-point method is used to calculate the curvature of the simulated path; Calculate the curvature of the simulated path generated by the adjacent path points and determine the invalid path based on the change in curvature; Determine the change of text stroke paths, remove invalid paths that do not conform to the text strokes and invalid paths within the curvature change range.

2. A BPO picture review system based on a large model according to claim 1, characterized in that: The user-uploaded images after noise reduction are preprocessed in the following ways: grayscale, normalization, binarization, contrast enhancement, and morphological operations to determine multiple distinguishing images; When extracting contour features of the user-uploaded image, the contour features in the difference image are extracted simultaneously to determine the difference range of the contour features between the user-uploaded image and the difference image.

3. A BPO picture review system based on a large model according to claim 2, characterized in that: The method for comparing the outlines of user-uploaded images and comparison images is as follows: Identify user-uploaded images, search for similar images through search engines, and retrieve comparison images with the same edge features; Extracting contour features of the comparison picture, and determining a processing state of the comparison picture based on a missing state of the contour features of the comparison picture; The high-quality contour features of the pictures uploaded by the user are matched and compared with the contour features of the pictures to determine the distinguishing contour features.

4. A BPO picture review system based on a large model according to claim 3, characterized in that: Extract distinguishing contour features, extract image data within the outer contour based on the distinguishing feature contour as the outer contour, identify the image data, determine the content of the image data, and analyze the content contained in the image data.

5. A BPO picture review system based on a large model according to claim 4, characterized in that: The path simulation method based on the internal image data of the outer contour is as follows: Determine the outer contour shape: read the distinguishing contour features and extract the outer contour through the findContours function; Determine multiple path points based on the outer contour: perform polygonal approximation on each contour through a polygonal approximation algorithm to determine the path points; Connecting path points based on the fit of the path points to the outer contour shape: generating a simulation path based on the path points, and ensuring that the simulation path is within the outer contour shape; Determine multiple simulation paths: connect the path points into complete paths; Identify simulated paths and remove invalid paths based on text strokes.

6. A BPO picture review system based on a large model according to claim 5, characterized in that: The specific steps for determining path points through polygonal approximation are as follows: Calculate the perimeter of the contour: Use the arcLength function to calculate the perimeter of the contour; Set the approximation accuracy: Set the approximation accuracy parameter epsilon according to the perimeter of the contour; Perform polygon approximation: Use the approxPolyDP function to perform polygon approximation and generate simplified polygons; Determine the simplified polygon corner points as path points and mark the path points.

7. A BPO picture review system based on a large model according to claim 6, characterized in that: The path fitting model obtains the simulation path and fits the text in the text library. The steps are as follows: Use cubic Bezier curve to fit the path. The Bezier curve defines the simulation path through endpoints and control points. The control points are determined by calculating the bending position of the simulation path and the ratio of its two ends. Feature point extraction: Extract feature points from the fitted simulation path to describe the shape and structure of the simulation path; Use optical character recognition technology to extract text information from the fitted path and convert it into digital text; Path optimization: Correct and optimize the path based on the recognized text data, and increase the text data range corresponding to the simulated path.

8. A BPO picture review system based on a large model according to claim 7, characterized in that: Correct the simulation path through the text data, define the strokes corresponding to the simulation path, correct the simulation path according to the strokes corresponding to the simulation path, and perform text library fitting again through the corrected simulation path; By combining the fitted characters, homophones and similar characters are combined to extend the semantics, and determine the multiple meanings corresponding to the deformed characters; Review the various meanings of morphed text and analyze whether the morphed text contained in the pictures uploaded by users complies with the platform review rules.

Images