Image document risk analysis system based on deep learning

Abstract

The invention relates to the technical field of image document risk analysis, in particular to an image document risk analysis system based on deep learning, and the system comprises a capturing module which is used for recognizing and capturing a plurality of frames of to-be-analyzed images in an image document; the image processing module is used for performing image processing on each to-be-analyzed image to generate a supplementary analysis image set of each to-be-analyzed image; the image recognition module is used for performing image recognition on each to-be-analyzed image to determine feature information; the generation module is used for determining a plurality of keywords according to the feature information and determining a combination of a plurality of risk words according to each keyword; the matching analysis module is used for carrying out matching with a preset sensitive word library based on the keyword or a combination of a plurality of risk words to determine whether a sensitive keyword exists or not, and determining a sensitive image position corresponding to the image document based on the sensitive keyword; the image document risk analysis system is accurate in image document risk identification and high in identification precision and identification efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of image document risk analysis, and in particular to an image document risk analysis system based on deep learning. Background Art

[0002] Risk analysis aims to identify risks in image documents, filter sensitive images, maintain information security, and avoid the leakage of sensitive images. The basis of risk analysis is image capture of image documents, and the existing image capture method is to capture single-frame images.

[0003] At present, OCR technology is mostly used to perform image recognition on captured single-frame images, and then to identify risks in image documents. The principle of OCR technology is: if you want to recognize text from dynamic images, the usual practice is to first divide the dynamic image into single-frame images for capture, and then apply OCR technology to each frame for text recognition. Since this technology performs text recognition on each frame, for large dynamic images or dynamic images composed of several frames of the same or similar single-frame images, the use of OCR technology is time-consuming, inflexible, and inefficient.

[0004] Chinese Patent Publication No.: CN116012762A discloses a method and system for analyzing video images of traffic intersections for power equipment, the method comprising: obtaining monitoring video stream data of traffic intersections; extracting frames of monitoring video stream data of traffic intersections, preliminarily identifying the extracted images, and screening out images suspected of having power safety hazards; intelligently identifying and finely classifying images suspected of having power safety hazards and video image analysis models, identifying high-risk safety hazard images for warning; adopting a backbone-neck-head network structure, extracting feature information of each dimension of the image through the backbone network, enhancing the feature information of each dimension through the neck network, fusing shallow shape, texture features and deep semantic information, and based on the scale characteristics of the fused feature map, using the head network to dynamically adjust the size of the reference frame, extracting the scale information of the target, extracting and analyzing the image features from shallow to deep, and finally realizing the detection of the target.

[0005] It can be seen that the above-mentioned traffic intersection video image analysis method and system for power equipment have the following problems: the method of acquiring images is to capture a number of single-frame images, the number of captures is limited, and the capture efficiency is low, and images containing potential risks and corresponding potential risks may be missed, and the accuracy of potential risks cannot be guaranteed, and the time-series frame images cannot be combined for analysis, resulting in poor recognition accuracy of video images. Summary of the invention

[0006] To this end, the present invention provides an image document risk analysis system based on deep learning to overcome the problem in the prior art that the image document recognition process cannot perform combined analysis on time-series frame images, resulting in poor image document recognition accuracy.

[0007] To achieve the above objectives, the present invention provides an image document risk analysis system based on deep learning, comprising:

[0008] A capture module, which is used to identify and capture an image document, and capture a plurality of frames of images to be analyzed from the image document according to the recognition result of the image document, wherein the recognition result includes a static image and a dynamic image;

[0009] an image processing module connected to the capture module, and configured to perform image processing on each of the images to be analyzed based on the recognition result to generate a supplementary analysis image set for each of the images to be analyzed;

[0010] An image recognition module, connected to the image processing module, for performing image recognition on each of the images to be analyzed and the images in the supplementary analysis image set to determine feature information, wherein the feature information includes text information and graphic information;

[0011] A generation module, connected to the image recognition module, for determining a number of keywords based on the feature information, and determining a combination of a number of risk words based on each of the keywords;

[0012] A matching analysis module is connected to the generation module and is used to match the keyword or a combination of the risk words with a preset sensitive word library to determine whether a sensitive keyword exists, and to determine a sensitive image position corresponding to the image document based on the sensitive keyword.

[0013] Furthermore, the capture module includes:

[0014] An identification unit, used for identifying the image document according to the similarity between any two frames of the image document;

[0015] A static capture unit, connected to the recognition unit, for statically capturing the static image to generate a single-frame static capture image;

[0016] A dynamic capture unit, connected to the recognition unit, for dynamically capturing the dynamic image to generate a plurality of frames of dynamic capture images;

[0017] A storage unit connected to the static capture unit and the dynamic capture unit, and used for classifying and storing the static capture images and the dynamic capture images;

[0018] Wherein, the image to be analyzed includes the static captured image and the dynamic captured image.

[0019] Furthermore, the recognition unit determines the recognition result of the single image document according to the similarity between any two frames of the single image document.

[0020] Furthermore, the dynamic capture unit determines a capture interval of the images to be analyzed of the dynamic image based on a preset duration, and determines whether to adjust the number of images to be analyzed based on a comparison result of a first similarity between adjacent images to be analyzed captured within the capture interval and a first preset similarity.

[0021] Furthermore, the image processing module determines whether to perform overlap processing on the time-series adjacent images to be analyzed according to the comparison result between the second similarity of the time-series adjacent images to be analyzed corresponding to the dynamic image and the second preset similarity and the time sequence interval of the images to be analyzed;

[0022] The overlap processing is to superimpose the image features of the images to be analyzed that are adjacent in time sequence to form an additional analysis image.

[0023] Furthermore, the image processing module is also used to perform image adjustment on the image to be analyzed and the additional analysis image to generate a plurality of supplementary analysis images of a single image to be analyzed or a single additional analysis image;

[0024] The image adjustment includes brightness adjustment, contrast adjustment, color adjustment, image noise reduction, deformation processing, image rotation, image sharpening and image blurring.

[0025] Furthermore, the image recognition module includes:

[0026] A partitioning unit, used to divide the image to be analyzed, the additional analysis image and the corresponding supplementary analysis image into a plurality of area blocks respectively;

[0027] an extraction and recognition unit, connected to the partitioning unit, for extracting the feature information from the area block based on image recognition;

[0028] The classification unit is connected to the extraction and recognition unit and is used to classify the feature information, and the classification result is text information or graphic information.

[0029] Furthermore, the generating module is preset with a graphic recognition unit, the graphic recognition unit recognizes keywords corresponding to the graphic information based on the graphic information, and forms a number of associative words based on the keywords;

[0030] Each character in the associative word is composed according to the synonyms, homophones or near-phonetic characters of each character in the keyword.

[0031] Furthermore, the generation module determines a combination of the plurality of risk words based on the plurality of keywords determined by the text information of the single image to be analyzed, the single additional analysis image or the single supplementary analysis image, the keywords corresponding to the graphic information and the associative words.

[0032] Furthermore, the generation module combines words between the graphic information according to the keywords and characters in the associative words corresponding to each graphic information of the single image to be analyzed, the single additional analysis image or the single supplementary analysis image to generate a number of graphic phrases, and combines words between the graphic and text information with a number of keywords determined by the text information of the corresponding image to generate a combination of the several risk words.

[0033] Compared with the prior art, the beneficial effect of the present invention lies in that the image document risk analysis system based on deep learning of the present invention selects different image capture methods through the recognition of image documents, and then performs image processing on the images to be analyzed to generate a supplementary analysis image set. It can capture images for analysis in a targeted manner according to the static and dynamic characteristics of the image documents, thereby improving the recognition accuracy of image documents. At the same time, the risk analysis system of the present invention converts images into keywords for sensitive keyword matching, and the recognition efficiency of keyword text is higher than that of picture recognition, thereby improving the risk identification efficiency of the image documents of the present invention.

[0034] Furthermore, the capture module of the present invention is equipped with an identification unit to classify image documents, so as to capture corresponding images of the image documents through a static capture unit or a dynamic capture unit, thereby improving the image capture efficiency of image documents, avoiding the capture of too many images in static documents to form redundant images, and avoiding the increased workload and low processing efficiency caused by the subsequent processing of redundant images, thereby improving the risk identification efficiency of image documents of the present invention.

[0035] Furthermore, the recognition unit of the present invention reduces the difficulty of image document classification and recognition and improves the recognition efficiency of image documents by analyzing the similarity between any two frames of images, and thus uses a static capture unit or a dynamic capture unit to capture the corresponding image document based on the recognition result, thereby improving the working efficiency of the capture module and further improving the working efficiency of the image document risk analysis system.

[0036] Furthermore, the dynamic capture unit of the present invention adjusts the number of images to be analyzed by determining the capture interval and comparing the first similarity with the first preset similarity based on the capture interval, thereby avoiding the capture of too many images to form redundant images and increasing the workload of subsequent processing of redundant images, and avoiding possible omissions of sensitive images, thereby ensuring the comprehensiveness and accuracy of subsequent extraction of sensitive keywords, recognition accuracy and risk analysis.

[0037] Furthermore, the image processing module of the present invention determines whether the dynamic image needs to be overlapped by identifying the comparison results of the second similarity of several temporally adjacent images to be analyzed corresponding to the dynamic image and the second preset similarity and the temporal interval of the images to be analyzed. For the images to be analyzed that need to be overlapped, new additional analysis images are formed by overlapping, so as to fully identify the images that need to be risk analyzed and avoid omission of risk points in the image document, thereby improving the accuracy of the risk identification of the image document of the present invention. Moreover, after the overlap processing, the total number of images for risk analysis is reduced, further improving the efficiency of the risk identification of the image document of the present invention.

[0038] Furthermore, the image processing module of the present invention can make the image clear and expose hidden image information by performing image adjustment on the image to be analyzed and the added analysis image, thereby avoiding the outflow of hidden sensitive images and sensitive information due to factors such as poor image quality, deformation, and color influence, and helps the subsequent image recognition module and generation module to determine the recognition of feature information and sensitive keywords, thereby improving the accuracy and efficiency of feature information generation and sensitive keyword recognition, thereby improving the risk identification efficiency and recognition accuracy of the image documents of the present invention, and realizing safe and effective supervision of image documents.

[0039] Furthermore, the image recognition module of the present invention divides the image to be analyzed, the additional analysis image and the corresponding supplementary analysis image into regions through a partitioning unit, so as to extract feature information from the image by region through an extraction and recognition unit and classify the feature information through a classification unit, thereby improving the extraction efficiency of feature information and the working efficiency of the image recognition module, helping the generation module to more efficiently identify sensitive keywords and sensitive images, thereby improving the risk identification efficiency of the present invention for image documents.

[0040] Furthermore, the generation module of the present invention is provided with a graphic recognition unit, which realizes information conversion and information extraction of the graphic information by recognizing keywords of the graphic information, and generates a combination of several risk words by generating associative words for the keywords, so as to match the combination of the keywords and the several risk words with a preset sensitive word library through a matching analysis module, thereby avoiding the omission of sensitive information due to the use of associative words of keywords in the graphic information, improving the accuracy and comprehensiveness of subsequent determination and identification of sensitive keywords, and improving the accuracy of risk identification.

[0041] Furthermore, the generation module of the present invention determines a combination of the several risk words through the several keywords determined by the text information, the keywords corresponding to the graphic information and the associative words, thereby avoiding the omission of the combination of the several risk words determined by the keywords generated corresponding to the text information and graphic information and the associative words, thereby improving the accuracy and comprehensiveness of subsequent determination and identification of sensitive keywords, and thereby improving the accuracy of risk identification.

[0042] Furthermore, the generation module of the present invention generates a combination of the several risk words by combining words between the graphic phrases and the text information of the corresponding images, thereby avoiding the omission of sensitive keywords and sensitive images due to the omission of the combination of several risk words generated by the word combination, improving the accuracy and comprehensiveness of the combination of several risk words generated by the generation module, and improving the risk identification accuracy of the image document of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a structural block diagram of an image document risk analysis system based on deep learning according to an embodiment of the present invention;

[0044] Figure 2 It is a schematic diagram of the structure of a capture module according to an embodiment of the present invention;

[0045] Figure 3 Schematic diagram of the structure of the image recognition module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0047] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0048] It should be noted that, in the description of the present invention, terms such as "up", "down", "left", "right", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the drawings. This is merely for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present invention.

[0049] In addition, it should be noted that in the description of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0050] See also Figure 1 As shown, it is a structural block diagram of an image document risk analysis system based on deep learning according to an embodiment of the present invention. An embodiment of the present invention provides an image document risk analysis system based on deep learning, including:

[0051] A capture module, which is used to identify and capture an image document, and capture a plurality of frames of images to be analyzed from the image document according to the recognition result of the image document, wherein the recognition result includes a static image and a dynamic image;

[0052] an image processing module connected to the capture module, and configured to perform image processing on each of the images to be analyzed based on the recognition result to generate a supplementary analysis image set for each of the images to be analyzed;

[0053] An image recognition module, connected to the image processing module, for performing image recognition on each of the images to be analyzed and the images in the supplementary analysis image set to determine feature information, wherein the feature information includes text information and graphic information;

[0054] A generation module, connected to the image recognition module, for determining a number of keywords based on the feature information, and determining a combination of a number of risk words based on each of the keywords;

[0055] A matching analysis module is connected to the generation module and is used to match the keyword or a combination of the risk words with a preset sensitive word library to determine whether a sensitive keyword exists, and to determine a sensitive image position corresponding to the image document based on the sensitive keyword.

[0056] In a specific embodiment, the image document risk analysis system is used to perform risk analysis on image documents stored in a computer system or in a cloud storage to identify image documents containing sensitive information.

[0057] The image document risk analysis system of the present invention is based on deep learning. By identifying the image document to select different image capture methods, and then performing image processing on the image to be analyzed to generate a supplementary analysis image set, it can capture images for analysis in a targeted manner according to the static and dynamic characteristics of the image document, thereby improving the recognition accuracy of the image document. At the same time, the risk analysis system of the present invention converts images into keywords for sensitive keyword matching. The recognition efficiency of keyword text is higher than that of picture recognition, thereby improving the risk identification efficiency of the image document of the present invention.

[0058] See also Figure 2 As shown, it is a schematic diagram of the structure of the capture module according to an embodiment of the present invention;

[0059] Specifically, the capture module includes:

[0060] An identification unit, used for identifying the image document according to the similarity between any two frames of the image document;

[0061] A static capture unit, connected to the recognition unit, for statically capturing the static image to generate a single-frame static capture image;

[0062] A dynamic capture unit, connected to the recognition unit, for dynamically capturing the dynamic image to generate a plurality of frames of dynamic capture images;

[0063] A storage unit connected to the static capture unit and the dynamic capture unit, and used for classifying and storing the static capture images and the dynamic capture images;

[0064] Wherein, the image to be analyzed includes the static captured image and the dynamic captured image.

[0065] It can be understood that the capture module may also include a unit for information transmission, a unit for sending and receiving instructions, and a unit for storing instructions, or the above-mentioned functions can be implemented by the units of the capture module in the present invention, or there are other methods or system components for implementing the above-mentioned functions not mentioned in the present invention, all of which fall within the scope of protection of the present invention and will not be repeated here.

[0066] In implementation, the image capture performed by the static capture unit and the dynamic capture unit is to obtain a single frame image of each frame image played in time sequence of the image document.

[0067] The capture module of the present invention is provided with an identification unit to classify image documents, so that corresponding images of the image documents are captured by a static capture unit or a dynamic capture unit, which can improve the image capture efficiency of image documents, avoid capturing too many images in static documents to form redundant images, and avoid the increase in workload and low processing efficiency caused by the subsequent processing of redundant images, thereby improving the risk identification efficiency of image documents of the present invention.

[0068] Specifically, the recognition unit determines the recognition result of the single image document according to the similarity between any two frames of the single image document.

[0069] It can be understood that the recognition unit determines whether the image document is the static image or the dynamic image based on the comparison between the similarity between any two frames of images and the preset static similarity.

[0070] It is understandable that the preset static similarity ranges from 90% to 99%, which changes as the image of the image document is affected by external factors including the loss of some pixels. In implementation, the similarity calculation and comparison of the two images can adopt any method in the existing technology, which will not be repeated here.

[0071] In this embodiment, if the similarity between any two frames of the single image document is greater than the preset static similarity, the recognition unit determines that the recognition result of the single image document is a static image.

[0072] If the similarity between any two frames of the single image document is less than or equal to the preset static similarity, the recognition unit determines that the recognition result of the single image document is a dynamic image.

[0073] In a specific embodiment, when the image document is too large, in order to improve recognition efficiency, the recognition unit scans and recognizes the image document at a fast forward speed or reduces the overall pixels of the image document for scanning and recognition. If the image in the image document changes locally or as a whole, the recognition unit recognizes the image document as a dynamic image. In other embodiments, the recognition unit can also recognize the image document in other ways.

[0074] In another specific embodiment, the recognition unit is connected to a display device, which can be displayed to an operator; or / and, the recognition unit can select two frames of images to be recognized in the image document by means including fast forwarding and pulling the image document, thereby determining whether the image document is a static image or a dynamic image as the recognition result, and pass the recognition result to a capture module for subsequent processing; in other embodiments, the device connected to the recognition unit includes but is not limited to the display device, and the recognition unit can also complete the recognition of the image document in other ways or processes.

[0075] All of the above identification methods fall within the protection scope of the present invention and will not be described in detail here.

[0076] The recognition unit of the present invention reduces the difficulty of image document classification and recognition and improves the recognition efficiency of image documents by analyzing the similarity between any two frames of images, thereby using a static capture unit or a dynamic capture unit image document to capture the corresponding image based on the recognition result, thereby improving the working efficiency of the capture module and further improving the working efficiency of the image document risk analysis system.

[0077] Specifically, the dynamic capture unit determines the capture interval of the images to be analyzed of the dynamic image based on a preset duration, and determines whether to adjust the number of images to be analyzed based on a comparison result of a first similarity between adjacent images to be analyzed captured in the capture interval and a first preset similarity.

[0078] It can be understood that the image to be analyzed is a capture of a single frame image of the dynamic image; the preset duration is determined by the capture unit according to the duration of the dynamic image, and the capture interval is determined with reference to the visual persistence time and visual association time of the human eye.

[0079] The first preset similarity is determined according to a specific application scenario and must be able to be used as a comparison standard to determine whether to adjust the number of images to be analyzed. In implementation, the calculation and comparison of the first similarity of adjacent images to be analyzed can adopt any method in the prior art, which will not be repeated here.

[0080] The visual persistence time is usually 0.1s to 0.4s, and the visual association time is usually 0.1s to 2s. In the present embodiment, the capture interval is 0.4s, that is, at each moment, the human eye recognizes a frame image of an image document acquired with a 0.4s exposure and a 0.4s interval.

[0081] In this embodiment, the image similarity between the captured adjacent images to be analyzed is calculated and recorded as the first similarity. If the first similarity between the captured adjacent images to be analyzed is greater than or equal to the first preset similarity, it is determined that there is no need to adjust the number of images to be analyzed.

[0082] If the first similarity between the captured adjacent images to be analyzed is less than the first preset similarity, an additional image located in the middle of the capture interval is captured as the image to be analyzed.

[0083] The dynamic capture unit of the present invention adjusts the number of images to be analyzed by determining the capture interval and comparing the first similarity with the first preset similarity based on the capture interval, thereby avoiding the capture of too many images to form redundant images and increasing the workload of subsequent processing of redundant images, and avoiding possible omissions of sensitive images, thereby ensuring the comprehensiveness and accuracy of subsequent extraction of sensitive keywords, recognition accuracy and risk analysis.

[0084] Specifically, the image processing module determines whether to perform overlap processing on the time-series adjacent images to be analyzed according to the comparison result of the second similarity of the time-series adjacent images to be analyzed corresponding to the dynamic image and the second preset similarity and the time sequence interval of the images to be analyzed;

[0085] The overlap processing is to superimpose the image features of the images to be analyzed that are adjacent in time sequence to form an additional analysis image.

[0086] After the determination and adjustment of the number of images to be analyzed are completed, the image similarity between the images to be analyzed that are adjacent in time sequence is calculated and recorded as a second similarity;

[0087] If the second similarity is less than the second preset similarity, the image processing module does not perform the overlap processing on the images to be analyzed that are adjacent in time sequence;

[0088] If the second similarity is greater than or equal to the second preset similarity, the second similarity satisfies the first condition for the overlap processing, and it is necessary to perform overlap processing on the time-series adjacent images to be analyzed according to the time interval.

[0089] The second preset similarity is determined according to a specific application scenario and must be able to be used as a comparison standard to determine whether to perform overlap processing on the temporally adjacent images to be analyzed. In implementation, the calculation of the second similarity can adopt any method in the prior art, which will not be repeated here.

[0090] In this embodiment, the second preset similarity is smaller than the first preset similarity.

[0091] It can be understood that the timing interval is determined by the time that can produce visual association. Preferably, in this embodiment, the timing interval is set to 2s.

[0092] On the premise of satisfying the first condition, if the time interval is greater than or equal to 2s, the image processing module does not perform the overlap processing on the time-adjacent images to be analyzed; if the time interval is less than 2s, the time interval satisfies the overlap processing condition, and it is determined that the time-adjacent images to be analyzed need to be overlapped.

[0093] Specifically, only when the second similarity and the time interval both meet the overlap processing condition, the image processing module performs overlap processing on the time-series adjacent images to be analyzed.

[0094] It can be understood that the overlap processing is to overlap the images to be analyzed that are adjacent in time sequence, retain the different image information in two or more images to be overlapped, and generate an overlapped image.

[0095] In this embodiment, the overlap processing is to use any image to be overlapped as a base image, and to add the distinguishing image features of other images to be overlapped with it in the base image according to their corresponding positions in the original image, so as to form an added analysis image.

[0096] The overlap processing allows accurate image capture of the image to be analyzed and the additional analysis image, thereby facilitating the retention and identification of the sensitive keywords, thereby improving the accuracy and effectiveness of risk identification and the security of the image document risk analysis system.

[0097] The image processing module of the present invention determines whether the dynamic image needs to be overlapped by identifying the comparison results of the second similarity of several temporally adjacent images to be analyzed corresponding to the dynamic image and the second preset similarity and the temporal interval of the images to be analyzed. For the images to be analyzed that need to be overlapped, new additional analysis images are formed by overlapping, so as to fully identify the images that need to be risk analyzed and avoid the omission of risk points in the image document, thereby improving the accuracy of the risk identification of the image document of the present invention. Moreover, after the overlap processing, the total number of images for risk analysis is reduced, further improving the efficiency of the risk identification of the image document of the present invention.

[0098] In another specific embodiment, the deep learning-based image document risk analysis system also includes a model unit for performing deep learning on the overlap processing process to generate a training model to improve the efficiency and accuracy of image overlap of image documents.

[0099] Specifically, the image processing module is further used to perform image adjustment on the image to be analyzed and the additional analysis image to generate a plurality of supplementary analysis images of a single image to be analyzed or a single additional analysis image;

[0100] The image adjustment includes brightness adjustment, contrast adjustment, color adjustment, image noise reduction, deformation processing, image rotation, image sharpening and image blurring.

[0101] In this embodiment, the deformation processing includes stretching, compressing, adjusting the image ratio, etc. at a set position or in a set direction of the image. The image characteristics can be changed through image adjustment.

[0102] It is understandable that the image adjustment can make the image clear, easy to determine the characteristic information, and expose the hidden image information, avoiding the leakage of hidden sensitive images and sensitive information due to factors such as poor image quality, deformation, and color influence, which helps to improve the accuracy and efficiency of image recognition and keyword generation. Improve the accuracy of risk identification.

[0103] The image processing module of the present invention can make the image clear and expose hidden image information by performing image adjustment on the image to be analyzed and the added analysis image, thereby avoiding the outflow of hidden sensitive images and sensitive information due to factors such as poor image quality, deformation, and color influence, and helps the subsequent image recognition module and generation module to determine the recognition of feature information and sensitive keywords, improve the accuracy and efficiency of feature information generation and sensitive keyword recognition, thereby improving the risk identification efficiency and recognition accuracy of the present invention for image documents, and realizing safe and effective supervision of image documents.

[0104] See also Figure 3 As shown, it is a schematic diagram of the structure of the image recognition module according to an embodiment of the present invention;

[0105] Specifically, the image recognition module includes:

[0106] A partitioning unit, used to divide the image to be analyzed, the additional analysis image and the corresponding supplementary analysis image into a plurality of area blocks respectively;

[0107] an extraction and recognition unit, connected to the partitioning unit, for extracting the feature information from the area block based on image recognition;

[0108] The classification unit is connected to the extraction and recognition unit and is used to classify the feature information, and the classification result is text information or graphic information.

[0109] In this embodiment, the partitioning unit extracts the information by identifying the outer contours of the information in the image to be analyzed, the additional analysis image and the corresponding supplementary analysis image, and uses the extracted information as information points to divide the information area according to the outer contours of several information points.

[0110] The extraction and identification unit analyzes the information points extracted by the partitioning unit, and extracts the feature information in the information points according to the outer contour of the feature information.

[0111] The classification unit recognizes the feature information as graphics or text and uses it as text information or graphics information.

[0112] In this embodiment, the technology used by the classification unit to identify the feature information is automatic recognition technology, including text recognition technology, which compares the feature information with the text features in the text library to identify the text information, and identifies the feature information that cannot be identified as text as graphic information; the information extraction in the form of outer contours adopts any one of the existing technologies including intelligent cutout technology, which will not be described in detail here. In other embodiments, the classification unit can also use other technologies in the existing technology to identify the feature information, which will not be described in detail here.

[0113] The image recognition module of the present invention divides the image to be analyzed, the additional analysis image and the corresponding supplementary analysis image into regions through a partitioning unit, so as to extract feature information from the image by region through an extraction and recognition unit and classify the feature information through a classification unit, thereby improving the extraction efficiency of feature information and the working efficiency of the image recognition module, helping the generation module to more efficiently identify sensitive keywords and sensitive images, thereby improving the risk identification efficiency of the present invention for image documents.

[0114] Specifically, the generating module is preset with a graphic recognition unit, which recognizes keywords corresponding to the graphic information based on the graphic information, and forms a number of associative words based on the keywords;

[0115] Each character in the associative word is composed according to the synonyms, homophones or near-phonetic characters of each character in the keyword.

[0116] In a specific embodiment, the keywords corresponding to the graphic information are "rabbit" and "animal", and the associated words of "rabbit" are "bald man", "spitting seeds", "drawings", etc.

[0117] The generation module of the present invention is provided with a graphic recognition unit, which realizes information conversion and information extraction of the graphic information by recognizing keywords of the graphic information, and generates a combination of several risk words by generating associative words for the keywords, so as to match the combination of the keywords and the several risk words with a preset sensitive word library through a matching analysis module, thereby avoiding the omission of sensitive information due to the use of associative words of keywords in the graphic information, improving the accuracy and comprehensiveness of subsequent determination and identification of sensitive keywords, and improving the accuracy of risk identification.

[0118] In another specific embodiment, the deep learning-based image document risk analysis system also includes a keyword generation model unit, which is used to perform deep learning on the recognition of the keywords and generate a training model to improve the accuracy and efficiency of the generation of the graphic information keywords.

[0119] Specifically, the generation module determines a combination of the risk words based on the keywords determined by the text information of the single image to be analyzed, the single additional analysis image or the single supplementary analysis image, the keywords corresponding to the graphic information and the associative words.

[0120] In a specific embodiment, the keywords determined by the text information include "flower", "money", and "banknotes"; the keywords corresponding to the graphic information include "rabbit" and "animal"; the associated words of "rabbit" include "bald", "spitting seeds", and "drawings"; the associated words of "flower" include "painting", "words", and "slippery"; then the combination of the several risk words includes "painting", "bald and slippery", etc.

[0121] The generation module of the present invention determines a combination of the several risk words through the several keywords determined by the text information, the keywords corresponding to the graphic information and the associative words, thereby avoiding the omission of the combination of the several risk words determined by the keywords generated corresponding to the text information and graphic information and the associative words, thereby improving the accuracy and comprehensiveness of subsequent determination and identification of sensitive keywords, and further improving the accuracy of risk identification.

[0122] Specifically, the generation module combines the words between the graphic information according to the keywords and the characters in the associative words corresponding to each graphic information of the single image to be analyzed, the single additional analysis image or the single supplementary analysis image to generate a number of graphic phrases, and combines the words between the graphic and text information with a number of keywords determined by the text information of the corresponding image to generate a combination of the several risk words.

[0123] It can be understood that the sensitive keyword is a combination of the keyword or the risk word and a word that matches the preset sensitive word library.

[0124] The matching analysis module identifies an image containing several of the sensitive keywords as the sensitive image, and determines the location of the sensitive image.

[0125] The sensitive keywords, the sensitive images and the positions of the sensitive images correspond one to one.

[0126] In a specific embodiment, the keywords corresponding to one of the graphic information include "flying" and "superman", and the associated words of "superman" include "superpower", "money power", "strong", etc. The keywords corresponding to another graphic information include "torch", "bright", and "firecrackers", then the phrases between the graphics include "explosive superpower", and the corresponding text information of the image is "panda", and the word combination between the graphic and text information is "panda superpower" and "explosive panda", etc., and the combination of the several risk words is "panda superpower" and "explosive panda".

[0127] The generation module of the present invention generates the combination of the several risk words by combining the words between the graphic phrases and the text information of the corresponding images, thereby avoiding the omission of the combination of several risk words caused by the word combination, resulting in the omission of sensitive keywords and sensitive images, improving the accuracy and comprehensiveness of the combination of several risk words generated by the generation module, and improving the risk identification accuracy of the image document of the present invention.

[0128] In a specific embodiment, the deep learning-based image document risk analysis system also includes a display module for displaying the sensitive keywords, the sensitive images and the sensitive image locations to the operator so that the operator can process the displayed information; the display method includes message reminders, picture displays, sensitive image table generation and sound reminders.

[0129] In another specific embodiment, the deep learning-based image document risk analysis system also includes a processing module for deleting, filtering and storing the sensitive images to prevent the outflow of the sensitive images.

[0130] Any matters not described in the present invention are applicable to the prior art.

[0131] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims

1. A deep learning-based image document risk analysis system, characterized in that: include: A capture module, which is used to identify and capture an image document, and capture a plurality of frames of images to be analyzed from the image document according to the recognition result of the image document, wherein the recognition result includes a static image and a dynamic image; an image processing module connected to the capture module, and configured to perform image processing on each of the images to be analyzed based on the recognition result to generate a supplementary analysis image set for each of the images to be analyzed; An image recognition module, connected to the image processing module, for performing image recognition on each of the images to be analyzed and the images in the supplementary analysis image set to determine feature information, wherein the feature information includes text information and graphic information; A generation module, connected to the image recognition module, for determining a number of keywords based on the feature information, and determining a combination of a number of risk words based on each of the keywords; A matching analysis module is connected to the generation module and is used to match the keyword or a combination of the risk words with a preset sensitive word library to determine whether a sensitive keyword exists, and to determine a sensitive image position corresponding to the image document based on the sensitive keyword.

2. The image document risk analysis system based on deep learning according to claim 1 is characterized in that: The capture module comprises: An identification unit, used for identifying the image document according to the similarity between any two frames of the image document; A static capture unit, connected to the recognition unit, for statically capturing the static image to generate a single-frame static capture image; A dynamic capture unit, connected to the recognition unit, for dynamically capturing the dynamic image to generate a plurality of frames of dynamic capture images; A storage unit connected to the static capture unit and the dynamic capture unit, and used for classifying and storing the static capture images and the dynamic capture images; Wherein, the image to be analyzed includes the static captured image and the dynamic captured image.

3. The image document risk analysis system based on deep learning according to claim 2 is characterized in that: The recognition unit determines the recognition result of the single image document according to the similarity between any two frames of the single image document.

4. The image document risk analysis system based on deep learning according to claim 2 is characterized in that: The dynamic capture unit determines a capture interval of the images to be analyzed of the dynamic image based on a preset duration, and determines whether to adjust the number of images to be analyzed based on a comparison result of a first similarity between adjacent images to be analyzed captured in the capture interval and a first preset similarity.

5. The image document risk analysis system based on deep learning according to claim 4 is characterized in that: The image processing module determines whether to perform overlap processing on the images to be analyzed that are adjacent in time sequence and correspond to the dynamic image according to the comparison result of the second similarity of the images to be analyzed that are adjacent in time sequence and the second preset similarity and the time sequence interval of the images to be analyzed; The overlap processing is to superimpose the image features of the images to be analyzed that are adjacent in time sequence to form an additional analysis image.

6. The image document risk analysis system based on deep learning according to claim 5 is characterized in that: The image processing module is further used to perform image adjustment on the image to be analyzed and the additional analysis image to generate a plurality of supplementary analysis images of a single image to be analyzed or a single additional analysis image; The image adjustment includes brightness adjustment, contrast adjustment, color adjustment, image noise reduction, deformation processing, image rotation, image sharpening and image blurring.

7. The image document risk analysis system based on deep learning according to claim 5, characterized in that: The image recognition module comprises: A partitioning unit, used to divide the image to be analyzed, the additional analysis image and the corresponding supplementary analysis image into a plurality of area blocks respectively; an extraction and recognition unit, connected to the partitioning unit, for extracting the feature information from the area block based on image recognition; The classification unit is connected to the extraction and recognition unit and is used to classify the feature information, and the classification result is text information or graphic information.

8. The image document risk analysis system based on deep learning according to claim 5, characterized in that: The generating module is pre-set with a graphic recognition unit, the graphic recognition unit recognizes keywords corresponding to the graphic information based on the graphic information, and forms a plurality of associative words based on the keywords; Each character in the associative word is composed according to the synonyms, homophones or near-phonetic characters of each character in the keyword.

9. The image document risk analysis system based on deep learning according to claim 8, characterized in that: The generating module determines a combination of the plurality of risk words based on the plurality of keywords determined by the text information of the single image to be analyzed, the single additional analysis image or the single supplementary analysis image, the keywords corresponding to the graphic information and the associative words.

10. The image document risk analysis system based on deep learning according to claim 9, characterized in that: The generation module combines words between graphic information according to the keywords and characters in the associative words corresponding to each graphic information of a single image to be analyzed, a single image to be added for analysis, or a single image to be supplemented for analysis to generate a plurality of graphic phrases, and combines words between graphic and text information with a plurality of keywords determined by the text information of the corresponding image to generate a combination of the plurality of risk words.

Images