Bill image recognition method and device, equipment and medium

Abstract

The invention relates to the technical field of image detection, can be applied to business system platforms of financial science and technology, medical health and the like, and discloses a bill image recognition method, device, equipment and medium, and the method comprises the steps: carrying out the quality enhancement of an original voucher image, and obtaining a voucher enhanced image; performing angle correction on the bill area according to the boundary position of the voucher enhanced image to obtain a standardized voucher image; extracting local texture features and global semantic features, and obtaining multi-modal fusion features according to the local texture features and the global semantic features; identifying a bill type according to the multi-modal fusion feature, and respectively extracting bill field information and a bill identification quality threshold; analyzing a compliance score according to the bill type and the bill field information, and obtaining an abnormal recognition level according to the compliance score and a bill recognition quality threshold value; and determining a bill recognition result according to the bill type, the bill field information, the compliance score and the anomaly recognition level. According to the invention, the bill image recognition accuracy can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image detection, and in particular to a bill image recognition method, device, equipment and medium. BACKGROUND

[0002] With the acceleration of the digital transformation of the insurance industry, intelligent claims settlement has become a core means to improve service efficiency. The current insurance claims materials are multi-source and heterogeneous, containing medical bills, accident photos, etc., and the image quality is uneven. Traditional optical character recognition technology needs to cope with the needs of multi-type material analysis and different quality image adaptation. The existing single model feature extraction method needs to consider both local details and global semantics, and also needs to combine business scenarios for accurate identification and verification to support standardized processing of high-value businesses. Therefore, in order to meet the efficient and accurate processing needs of intelligent claims settlement for claims images, the identification and verification methods need to be innovated and optimized to improve the accuracy of bill image recognition.

[0003] In the field of medical health, staff need to process a large number of medical bill images such as medical records, examination reports, and medical invoices. Traditional optical character recognition technology relies on a single model to extract features, or focuses on local text edges and other details while ignoring the overall layout semantics of the bill, resulting in many recognition errors of key information such as amounts and diagnosis results in medical bill images, and low accuracy of bill image recognition.

[0004] In the field of financial technology business, staff need to process multiple types of bill images related to vehicle insurance, health insurance, etc., such as vehicle repair bills, medical expense lists, and policyholder identification, etc. These images often have different resolutions, partial area occlusion, or mixed formats. Traditional processing methods use a single model to extract features, making it difficult to consider both local text details and global layout semantics of the bill, and not deeply combining business logic to associate and verify the identified fields, resulting in low accuracy of bill image recognition.

[0005] The existing technology has difficulty in simultaneously considering local texture and global semantics when processing insurance claims original document images, lacks cross-attention and gating fusion mechanisms to effectively interact features, and does not construct a composite scoring system combining image quality, local clarity, and business rules for compliance evaluation. Abnormal detection relies on manual rules, and does not set dynamic quality thresholds for different types of bills, resulting in insufficient robustness in identifying low-quality, unstructured documents, and ultimately low accuracy of bill image recognition. SUMMARY

[0006] The present application provides a bill image recognition method, device, equipment and medium to solve the problem of low accuracy of bill image recognition.

[0007] In a first aspect, a bill image recognition method is provided, comprising: obtain an original credential image of a target service, perform quality enhancement on the original credential image to obtain a credential enhanced image; identify a boundary position of a bill region in the credential enhanced image, perform angle correction on the bill region according to the boundary position to obtain a standardized credential image; extract local texture features and global semantic features of the standardized credential image respectively, perform feature fusion on the local texture features and the global semantic features to obtain multi-modal fusion features of the standardized credential image; identify a bill type corresponding to the standardized credential image according to the multi-modal fusion features, and extract bill field information and a bill recognition quality threshold of the standardized credential image respectively; analyze a compliance score of the standardized credential image according to the bill type and the bill field information, and compare the compliance score with the bill recognition quality threshold to obtain an abnormal recognition level of the standardized credential image; determine a bill recognition result of the standardized credential image according to the bill type, the bill field information, the compliance score, and the abnormal recognition level.

[0008] In a second aspect, a bill image recognition device is provided, comprising: an original credential image enhancement module configured to obtain an original credential image of a target service, and perform quality enhancement on the original credential image to obtain a credential enhanced image; a bill region angle correction module configured to identify a boundary position of a bill region in the credential enhanced image, and perform angle correction on the bill region according to the boundary position to obtain a standardized credential image; a multi-modal fusion feature acquisition module configured to extract local texture features and global semantic features of the standardized credential image respectively, and perform feature fusion on the local texture features and the global semantic features to obtain multi-modal fusion features of the standardized credential image; a bill type identification module configured to identify a bill type corresponding to the standardized credential image according to the multi-modal fusion features, and extract bill field information and a bill recognition quality threshold of the standardized credential image respectively; an abnormal recognition level acquisition module configured to analyze a compliance score of the standardized credential image according to the bill type and the bill field information, and compare the compliance score with the bill recognition quality threshold to obtain an abnormal recognition level of the standardized credential image; a bill recognition result determination module configured to determine a bill recognition result of the standardized credential image according to the bill type, the bill field information, the compliance score, and the abnormal recognition level.

[0009] In a third aspect, a computer device is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the bill image recognition method when executing the computer program.

[0010] In a fourth aspect, a computer readable storage medium is provided, which stores a computer program, and the computer program implements the steps of the bill image recognition method when executed by a processor.

[0011] In the scheme implemented by the bill image recognition method, device, equipment and medium, the original voucher image of a target service can be obtained by a client, the original voucher image is quality-enhanced to obtain a voucher enhanced image, the boundary position of a bill region in the voucher enhanced image is recognized, the bill region is angle-corrected according to the boundary position to obtain a standardized voucher image, the local texture feature and the global semantic feature of the standardized voucher image are extracted respectively, the local texture feature and the global semantic feature are fused to obtain a multi-modal fusion feature of the standardized voucher image, the bill type corresponding to the standardized voucher image is recognized according to the multi-modal fusion feature, and the bill field information and the bill recognition quality threshold of the standardized voucher image are extracted respectively, the compliance score of the standardized voucher image is analyzed according to the bill type and the bill field information, the compliance score is compared with the bill recognition quality threshold to obtain the abnormal recognition level of the standardized voucher image, and the bill recognition result of the standardized voucher image is determined according to the bill type, the bill field information, the compliance score and the abnormal recognition level. In the present application, the original voucher image is quality-enhanced, the image quality is improved by matching a differential strategy according to the image quality score, the bill region boundary is accurately recognized and the angle is corrected to obtain a standardized voucher image, then the local texture and the global semantic feature are extracted and deeply fused to obtain a multi-modal fusion feature, then the bill type is recognized based on the feature, the field information and the quality threshold are extracted, the compliance score is calculated in combination with the bill type and the field information, the abnormal level is determined by comparing the score with the threshold, and finally the recognition result is determined by comprehensively determining the multi-dimensional information. The problem of low bill image recognition accuracy can be solved. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0013] Figure 1is a schematic diagram of an application environment of a bill image recognition method in an embodiment of the present application; Figure 2 is a schematic diagram of a flow of a bill image recognition method in an embodiment of the present application; Figure 3 is Figure 2 is a schematic diagram of a specific implementation flow of step S4 in the embodiment; Figure 4 is Figure 2 is a schematic diagram of a specific implementation flow of step S5 in the embodiment; Figure 5 is a schematic diagram of a structure of a bill image recognition device in an embodiment of the present application; Figure 6 is a schematic diagram of a structure of a computer device in an embodiment of the present application; Figure 7 is another schematic diagram of a structure of a computer device in an embodiment of the present application. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0015] The bill image recognition method provided by the embodiments of the present application can be applied in, for example, Figure 1In an application environment of the application, a client communicates with a server through a network. The server can obtain an original credential image of a target service through the client, perform quality enhancement on the original credential image to obtain a credential enhanced image, identify a boundary position of a bill region in the credential enhanced image, perform angle correction on the bill region according to the boundary position to obtain a standardized credential image, extract local texture features and global semantic features of the standardized credential image respectively, perform feature fusion on the local texture features and the global semantic features to obtain multi-modal fusion features of the standardized credential image, identify a bill type corresponding to the standardized credential image according to the multi-modal fusion features, and extract bill field information and a bill recognition quality threshold of the standardized credential image respectively. The server analyzes a compliance score of the standardized credential image according to the bill type and the bill field information, compares the compliance score with the bill recognition quality threshold to obtain an abnormal recognition level of the standardized credential image, and determines a bill recognition result of the standardized credential image according to the bill type, the bill field information, the compliance score and the abnormal recognition level. In the application, quality enhancement is performed on an original credential image, a differential strategy is matched according to an image quality score to improve image quality, a bill region boundary is accurately identified and an angle is corrected to obtain a standardized credential image, then local texture features and global semantic features are extracted and deeply fused to obtain multi-modal fusion features, a bill type is identified based on the features, field information and a quality threshold are extracted, a compliance score is calculated according to the bill type and the field information, the score is compared with the threshold to determine an abnormal level, and finally a recognition result is determined by comprehensively considering multi-dimensional information, so that the problem of low accuracy of bill image recognition can be solved. The client can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices. The server can be implemented by an independent server or a server cluster composed of multiple servers. The application will be described in detail below through specific embodiments.

[0016] Please refer to Figure 2 as shown in the figure, Figure 2 A flowchart of a bill image recognition method provided by an embodiment of the application is shown in the figure, which includes the following steps: S1, obtaining an original credential image of a target service, performing quality enhancement on the original credential image to obtain a credential enhanced image.

[0017] In the embodiment of the application, the target service refers to an insurance service, and the original credential image refers to a multi-source heterogeneous material image used for claim settlement audit in an insurance claim settlement scenario.

[0018] In detail, the original voucher image of the target service is obtained by receiving a multi-source heterogeneous claim settlement material image in JPG / PNG / PDF format corresponding to the target service in an insurance claim settlement scenario.

[0019] In the embodiment of the present application, the voucher enhanced image refers to the voucher image obtained by calculating the image quality score through NIQE (an image quality evaluation algorithm) and matching the corresponding enhancement strategy for the original voucher image obtained in the insurance claim settlement scenario.

[0020] In the embodiment of the present application, the quality enhancement of the original voucher image to obtain the voucher enhanced image comprises: calculating the quality score of the original voucher image by using a preset image quality evaluation algorithm; when the quality score is greater than or equal to a preset score threshold, performing illumination enhancement processing on the original voucher image to obtain a first initial enhanced image; when the quality score is less than the preset score threshold, performing histogram equalization processing on the original voucher image to obtain a second initial enhanced image; performing definition enhancement processing on the first initial enhanced image or the second initial enhanced image by using a preset super-resolution processing network to obtain the voucher enhanced image.

[0021] In detail, the preset image quality evaluation algorithm refers to the NIQE algorithm; the quality score is the image quality quantization result calculated by the algorithm for the original voucher image in the insurance claim settlement scenario, and the score range is 0-10, and the lower the score, the higher the image quality. Running the NIQE algorithm on the obtained original voucher image in the insurance claim settlement to obtain the corresponding quality score, which is used to automatically match the image enhancement strategy.

[0022] Specifically, the preset score threshold refers to the NIQE algorithm quality score limit for judging whether the original voucher image in the insurance claim settlement needs to enable a specific enhancement strategy, and specifically is 7.5; the first initial enhanced image refers to the image obtained after performing illumination enhancement processing on the original voucher image when the NIQE quality score of the original voucher image is greater than or equal to 7.5. When the score is greater than or equal to 7.5, the original voucher image is processed by using the Retinex (an image enhancement technology) enhancement combined with non-local mean denoising to perform light enhancement and denoising, and the generated image after processing is the first initial enhanced image.

[0023] Further, the second initial enhanced image refers to the image obtained after performing histogram equalization processing on the original voucher image in the insurance claim settlement when the quality score calculated by the NIQE algorithm is less than the preset score threshold 7.5. When the score is less than 7.5, only the original voucher image is processed by histogram equalization, and the generated image after processing is the second initial enhanced image.

[0024] Further, the preset super-resolution processing network refers to a lightweight ESRGAN network. It is determined whether the image to be processed is a first initial enhanced image or a second initial enhanced image, and then the lightweight ESRGAN network is called to perform super-resolution magnification processing on the images with a resolution lower than 600 dpi in the two types of initial enhanced images, so that the resolution is uniformly adjusted to 1024*768 to realize clarity enhancement.

[0025] Exemplarily, in a medical scene, the Retinex enhancement + non-local mean denoising is adopted on the medical invoice original image (NIQE quality score 8.2>=7.5) in JPG format to obtain a first initial enhanced image, and the histogram equalization is adopted on the medical record original image (NIQE quality score 6.8<7.5) to obtain a second initial enhanced image, and then the lightweight ESRGAN network is called to perform super-resolution magnification to 1024*768 on the initial enhanced images with a resolution lower than 600 dpi, so as to complete clarity enhancement.

[0026] Exemplarily, in a financial scene, the Retinex enhancement + non-local mean denoising is adopted on the bank transfer voucher original image (NIQE quality score 8.0>=7.5) in PDF format to obtain a first initial enhanced image, and the histogram equalization is adopted on the customer asset proof original image (NIQE quality score 7.0<7.5) to obtain a second initial enhanced image, and then the lightweight ESRGAN network is called to perform super-resolution magnification to 1024*768 on the initial enhanced images with a resolution lower than 600 dpi, so as to complete clarity enhancement.

[0027] S2, identify the boundary position of the bill region in the voucher enhanced image, and perform angle correction on the bill region according to the boundary position to obtain a standardized voucher image.

[0028] In the embodiment of the application, the bill region refers to the image region corresponding to various types of vouchers that need to be intelligently identified and verified in an insurance claim settlement scene; and the boundary position refers to the edge coordinate information for clearly defining the specific range of the bill region in the voucher enhanced image detected by the improved YOLOv8 model (You Only Look Once version 8, single-stage target detection algorithm version 8).

[0029] In detail, after the preprocessing and quality enhancement of the original claim settlement image are completed, the improved YOLOv8 model is used to detect the obtained voucher enhanced image, so as to identify the boundary position of the bill region in the voucher enhanced image.

[0030] In the embodiment of the present application, the standardized certificate image refers to a normalized certificate image that meets the requirements of subsequent multi-modal feature fusion modeling and business logic verification in terms of clarity, angle, and image quality.

[0031] In detail, the spatial transformation network (STN) is used to adjust the angle of the bill area in the photo and control the error to be no more than ±0.3°, so as to complete the angle correction.

[0032] S3, local texture features and global semantic features of the standardized certificate image are extracted respectively, the local texture features and the global semantic features are fused, and multi-modal fusion features of the standardized certificate image are obtained.

[0033] In the embodiment of the present application, the local texture features of the standardized certificate image refer to image detail information extracted by the ResNet18 backbone network (CNN branch); and the global semantic features refer to image overall level information captured by the Swin-Tiny architecture (Transformer branch).

[0034] In the embodiment of the present application, the local texture features and the global semantic features of the standardized certificate image are extracted respectively, including: performing layer-by-layer texture analysis on the standardized certificate image to obtain initial local features; performing dimension compression operation on the initial local features to obtain low-dimensional local features; extracting texture feature information in the low-dimensional local features, and determining the local texture features of the standardized certificate image according to the texture feature information; performing associated recognition on the certificate field in the standardized certificate image to obtain initial global features; performing semantic normalization processing on the initial global features to obtain the global semantic features of the standardized certificate image.

[0035] In detail, the initial local features refer to image local detail information such as text edges and seals extracted from the standardized certificate image. The layer-by-layer texture analysis on the standardized certificate image by the ResNet18 backbone network of the CNN branch is implemented and completed, and finally the feature map with a size of 1 / 8 of the input image is output.

[0036] Specifically, the low-dimensional local features refer to features with lower dimensions and retaining core local texture information after dimension reduction of the initial local features extracted by the CNN branch ResNet18 backbone network. The dimension compression operation is performed on the initial local features output by the CNN branch, so as to obtain the low-dimensional local features.

[0037] Further, the texture feature information refers to key information reflecting the local details of the standardized voucher image, which is screened from the low-dimensional local features. After dimension compression of the extracted initial local features to obtain low-dimensional local features, the texture feature information such as the edge of the text and the seal is extracted from the low-dimensional local features, and then the local texture features of the standardized voucher image are determined according to the texture feature information.

[0038] Further, the voucher field refers to key information items related to the business in the standardized voucher image in the insurance claim scenario; the initial global feature refers to information capable of reflecting the overall semantic logic of the standardized voucher image; the Swin-Tiny architecture of the Transformer branch is used to process the standardized voucher image, which can associate and identify the voucher field in the image, and further capture the global semantic information such as the layout of the document and the field association.

[0039] In addition, the initial global feature is regularized and optimized at the semantic level by the semantic processing capability of the architecture itself, so that the feature is more suitable for the demand for global semantic understanding in the insurance claim business scenario, and finally the global semantic feature capable of accurately reflecting the overall semantic logic of the standardized voucher image is formed.

[0040] Exemplarily, in the medical scenario, for the medical invoice standardized voucher image submitted by the patient, the drug name edge and the seal are extracted as texture information after layer-by-layer texture analysis by the CNN branch ResNet18 backbone network, and the local texture features are determined after dimension compression of the initial local features; at the same time, the amount and the diagnosis date are associated and identified by the Transformer branch Swin-Tiny architecture, and the global semantic feature reflecting the overall semantic logic of the invoice is obtained after semantic standardization processing of the initial global feature.

[0041] Exemplarily, in the financial scenario, for the bank deposit receipt standardized voucher image submitted by the customer, the bank seal edge and the handwritten deposit amount handwriting are extracted as texture information after layer-by-layer texture analysis by the CNN branch ResNet18 backbone network, and the local texture features are determined after dimension compression of the initial local features; at the same time, the depositor's name, the deposit amount, and the due date are associated and identified by the Transformer branch Swin-Tiny architecture, and the global semantic feature reflecting the overall semantic logic of the bank deposit receipt is obtained after semantic standardization processing of the initial global feature.

[0042] In the embodiment of the application, the multi-modal fusion feature refers to a feature having both local detail representation capability and global semantic association capability, which is obtained based on deep fusion of the local texture feature and the global semantic feature.

[0043] In the embodiment of the present application, the local texture feature and the global semantic feature are fused to obtain the multi-modal fusion feature of the standardized credential image, which comprises: The local texture feature is mapped into a query vector, and the global semantic feature is respectively mapped into a key vector and a value vector; The feature attention weight of the query vector and the key vector is calculated; The value vector is weighted and summed using the feature attention weight to obtain the initial fusion feature of the standardized credential image; The first feature weight coefficient corresponding to the local texture feature and the second feature weight coefficient corresponding to the global semantic feature are generated by using a preset two-way gating unit; The local weighted feature of the local texture feature is determined according to the first feature weight coefficient, and the global weighted feature of the global semantic feature is determined according to the second feature weight coefficient; The initial fusion feature, the local weighted feature and the global weighted feature are added element by element to obtain the multi-modal fusion feature of the standardized credential image.

[0044] In detail, the query vector refers to the vector obtained by performing feature projection on the local texture feature extracted from the standardized credential image by the CNN branch composed of ResNet18 backbone network; the key vector and the value vector refer to two types of vectors obtained by performing feature projection on the global semantic feature captured from the standardized credential image by the Transformer branch composed of Swin-Tiny architecture. The mapping processing is performed on the local texture feature extracted by the CNN branch to obtain the query vector, and the mapping processing is performed on the global semantic feature captured by the Transformer branch to obtain the key vector and the value vector.

[0045] Specifically, the feature attention weight refers to the coefficient for measuring the importance of different parts of the global semantic feature to the local texture feature, which is obtained by calculating the matching degree of the query vector and the key vector by vector dot product and normalized by the Softmax function; the initial fusion feature refers to the feature generated by focusing on the key area and preliminarily fusing the local texture information and the global semantic information by weighting and summing the value vector using the above feature attention weight. The similarity calculation and feature weighting operation in the cross-modal attention fusion are completed, the vector dot product operation is performed on the projected query vector and the key vector to quantify the matching degree between them, the operation result is input into the Softmax function for normalization processing to obtain the feature attention weight which can reflect the importance of the feature, and finally the value vector is weighted and summed using the weight, so that the initial fusion feature which fuses the local and global key information is obtained.

[0046] Further, the preset two-path gating unit is a two-path gating structure preset for dynamically controlling the weight proportion of the local texture feature and the global semantic feature in the fusion process; the first feature weight coefficient refers to a coefficient generated after the local texture feature extracted by the CNN branch is processed by one of the two-path gating units, and used to measure the importance of the local texture feature in the final fusion feature; the second feature weight coefficient refers to a coefficient generated after the global semantic feature captured by the Transformer branch is processed by the other of the two-path gating units, and used to measure the importance of the global semantic feature in the final fusion feature. The gating fusion mechanism is completed, that is, the local texture feature output by the CNN branch and the global semantic feature output by the Transformer branch are independently processed by the two-path gating units preset, and the two-path gating units will be according to the effectiveness of the feature and the relevance to the insurance claim image recognition task.

[0047] Further, the local weighted feature refers to a feature obtained after the first feature weight coefficient generated by the preset two-path gating unit and the local texture feature extracted by the CNN branch are operated, which reflects the corresponding importance of the local texture feature in the final fusion; the global weighted feature refers to a feature obtained after the second feature weight coefficient generated by the preset two-path gating unit and the global semantic feature captured by the Transformer branch are operated, which reflects the corresponding importance of the global semantic feature in the final fusion. After obtaining the first feature weight coefficient and the second feature weight coefficient generated by the two-path gating units, the first feature weight coefficient is operated with the local texture feature output by the CNN branch to obtain the local weighted feature, and the second feature weight coefficient is operated with the global semantic feature output by the Transformer branch to obtain the global weighted feature.

[0048] In addition, the key region information in the initial fusion feature, the local detail information in the local weighted feature, and the global semantic information in the global weighted feature are integrated to obtain a multi-modal fusion feature of the standardized voucher image, which has local detail representation capability, global semantic association capability, and key region focusing capability.

[0049] Exemplarily, in a medical scene, the local texture features such as drug name edges and doctor signatures extracted by the ResNet18 backbone network are mapped into query vectors, and the global semantic features such as the overall layout of the invoice, the association between the diagnosis and treatment items and the amount field captured by the Swin-Tiny architecture are respectively mapped into key vectors and value vectors, the matching degree is calculated through vector dot product, the feature attention weight is obtained through Softmax normalization, the initial fusion feature is obtained by weighting and summing the value vectors with the weight, and the weight coefficients corresponding to the local and global semantic features are generated by means of two gating units, the local and global weighted features are obtained by respectively operating the corresponding features, and finally the three types of features are added element by element to obtain the multi-modal fusion feature of the medical invoice.

[0050] Exemplarily, in a financial scene, the local texture features such as the deposit amount digital edge and the bank seal extracted by the ResNet18 backbone network are mapped into query vectors, and the global semantic features such as the overall layout of the deposit receipt, the association between the deposit period and the interest rate field captured by the Swin-Tiny architecture are respectively mapped into key vectors and value vectors, the matching degree is calculated through vector dot product, the feature attention weight is obtained through Softmax normalization, the initial fusion feature is obtained by weighting and summing the value vectors with the weight, and the weight coefficients corresponding to the corresponding features are generated by means of two gating units, the local and global weighted features are obtained by respectively operating the corresponding features, and finally the three types of features are added element by element to obtain the multi-modal fusion feature of the bank deposit receipt.

[0051] S4, identifying the type of the standardized voucher image according to the multi-modal fusion feature, and respectively extracting the field information of the standardized voucher image and the quality threshold of the bill recognition.

[0052] In the embodiment of the application, the type of the bill refers to the types of multiple source heterogeneous claim materials involved in the insurance claim scene, such as medical invoices, driver's licenses, etc.

[0053] In detail, after obtaining the multi-modal fusion feature, the type of the standardized voucher image is predicted by relying on the classification task in multi-task joint reasoning and using the multi-modal fusion feature.

[0054] In the embodiment of the application, the field information of the bill is a detection task modeled by multi-modal feature fusion, and the multi-modal fusion feature is formed by cross-modal attention fusion and gating fusion of the local texture features extracted by the CNN branch and the global semantic features captured by the Transformer branch, and the key field coordinates and text content related to the insurance claim in the standardized voucher image are obtained by regression; the quality threshold of the bill recognition is a differentiated dynamic threshold determined based on a dynamic quality evaluation model for different types of bills.

[0055] In the embodiment of the application, reference is made to Figure 3As shown, the extracting the bill field information and the bill recognition quality threshold of the standardized voucher image respectively comprises: S31, identifying a voucher information region of the standardized voucher image, and extracting bill field information of the standardized voucher image in the voucher information region; S32, performing multi-dimensional analysis on three-dimensional quality indicators of the standardized voucher image to obtain quality score data; S33, determining the bill recognition quality threshold of the standardized voucher image according to a differentiated judgment limit of the quality score data.

[0056] In detail, the voucher information region refers to a specific region in the standardized voucher image containing key fields related to insurance claims. The coordinates and text content of the key fields in the voucher information region are obtained by using the above multi-modal fusion feature regression to complete the determination of the bill field information.

[0057] Specifically, the three-dimensional quality indicators refer to three dimensions for evaluating the quality of the standardized voucher image, which are respectively an image quality basic score accounting for 40%, an image quality basic score calculated by a NIQE algorithm to evaluate the image blur degree and noise level, accounting for 30%, a local clarity score calculated by a gradient to measure the clarity of the local region of the image, and a business compliance score calculated according to business rules, accounting for 30%; the quality score data refers to a composite score obtained by weighting the above three-dimensional quality indicators according to the corresponding weights. The image quality basic score is obtained by calculating the blur degree and noise level of the standardized voucher image by the NIQE algorithm, the local clarity score is obtained by calculating the clarity of the local region of the image by the gradient, and the business compliance score is obtained by evaluating the image according to the business rules, and finally the scores of the three dimensions are weighted according to the weights of 40%, 30% and 30% to obtain the quality score data.

[0058] Further, the differentiated judgment limit refers to different quality score standards set for different bill types in the insurance claims scenario to distinguish whether the bill needs manual review. After obtaining the quality score data by the dynamic quality evaluation model, different quality score standards are set for different bill types according to the differences in the review requirements of different bill types in the insurance claims business.

[0059] S5, analyzing the compliance score of the standardized voucher image according to the bill type and the bill field information, comparing the compliance score with the bill recognition quality threshold, and obtaining an abnormality recognition level of the standardized voucher image.

[0060] In the embodiment of the application, the compliance score is one of the dimensions for calculating the composite score in the dynamic quality evaluation model, accounting for 30% of the composite score.

[0061] In detail, when the image resolution is lower than 300 dpi and the file format is not JPG / PNG, the score will be deducted, and a differentiated dynamic threshold will be set for different bill types, and the compliance score needs to be weighted with the image quality basic score accounting for 40% and the local sharpness score accounting for 30% to form the final quality score.

[0062] In the embodiment of the application, the abnormality recognition level is obtained by comparing the compliance score accounting for 30% of the final quality score and determined according to the business rules with the differentiated bill recognition quality threshold set for different bill types in the business logic consistency verification stage.

[0063] In the embodiment of the application, as shown in Figure 4 The comparison of the compliance score and the bill recognition quality threshold to obtain the abnormality recognition level of the standardized voucher image comprises: S41, the numerical comparison of the compliance score and the bill recognition quality threshold is performed to obtain a numerical comparison result; S42, according to the correspondence between the score interval and the abnormality level in the preset business rules, the comparison result is matched in level to obtain a level matching result; S43, the abnormality recognition level of the standardized voucher image is determined according to the level matching result.

[0064] In detail, the numerical comparison result is whether the final quality score of the standardized voucher image is lower than the bill recognition quality threshold corresponding to the bill type after the numerical comparison of the final quality score of the standardized voucher image and the bill recognition quality threshold corresponding to the bill type. The image quality basic score is calculated by the NIQE algorithm, the local sharpness score is calculated by the gradient, the compliance score is determined according to the business rules, and then the three dimensions are weighted according to the corresponding weight to obtain the final quality score. Then, the differentiated bill recognition quality threshold set for the current bill type is called, and the final quality score is numerically compared with the threshold, so that the numerical comparison result of whether the final quality score is lower than the threshold is obtained.

[0065] Specifically, the preset business rule includes two aspects, one is the rule for calculating compliance score in dynamic quality assessment, and the other is the rule related to the Bayesian network constructed based on the knowledge graph in causal reasoning verification; the score interval is the final quality score composed of the image quality basic score accounting for 40%, the local sharpness score accounting for 30%, and the compliance score accounting for 30% in the dynamic quality assessment model; the abnormal level is a level division corresponding to the score interval and the business abnormal situation, including the related level in the dynamic quality assessment when the final quality score is lower than the threshold value of the corresponding bill type, and the levels such as the first level alarm and the second level alarm detected in the causal reasoning verification; the level matching result is the comparison of the final quality score and the score interval corresponding to the bill type, combined with the abnormal situation detected in the causal reasoning verification. The final quality score containing the compliance score is calculated through the dynamic quality assessment model, the score interval preset for the current bill type is called, the final quality score is compared with the score interval, and whether there is a business abnormality such as time line and price is detected through the causal reasoning verification. Then, according to the corresponding relationship between the score interval and the abnormal level, the comparison result and the detected business abnormality are matched in level, and the level matching result containing the specific abnormal level and the manual review suggestion is obtained.

[0066] Further, in the business logic consistency verification stage, the final quality score composed of the image quality basic score accounting for 40%, the local sharpness score accounting for 30%, and the compliance score accounting for 30% is calculated through the dynamic quality assessment model, and then the differentiated bill identification quality threshold preset for the current bill type is called, the final quality score is compared with the threshold value to obtain a numerical comparison result, and the causal reasoning verification based on the Bayesian network constructed based on the insurance knowledge graph is performed. Then, according to the corresponding relationship between the score interval and the abnormal level, the numerical comparison result and the business abnormality detection situation are combined to form a level matching result, and finally the abnormal identification level of the standardized voucher image is determined according to the level matching result.

[0067] S6, determining a bill identification result of the standardized voucher image according to the bill type, the bill field information, the compliance score and the abnormal identification level.

[0068] In the embodiment of the application, the bill identification result refers to the comprehensive judgment result finally generated by integrating the bill type, the bill field information, the compliance score and the abnormal identification level, including the specific category of the bill, the detailed information of the key information, the score result of whether the business rule is met, the level of abnormal situation, and the suggestion of whether manual review is needed. The result can be directly used for automatic audit in the insurance claim process, and at the same time provides clear guidance for subsequent manual review. Moreover, the related data is also used for continuous optimization and iteration of the model.

[0069] In the embodiment of the present application, the bill recognition result of the standardized voucher image is determined according to the bill type, the bill field information, the compliance score, and the abnormality identification level, and includes: The bill field information is analyzed according to the bill type to obtain a field correlation analysis result; A compliance judgment result of the standardized voucher image is determined according to the compliance score; The field correlation analysis result and the compliance judgment result are analyzed according to the abnormality processing rule corresponding to the abnormality identification level to obtain an abnormality trace analysis result; The bill recognition result of the standardized voucher image is determined according to the abnormality trace analysis result.

[0070] In detail, the field correlation analysis result is a determination result that reflects whether the fields meet the insurance business logic, which is obtained by associating the bill type determined through multi-task joint inference, and the correlation verification of the bill key field coordinates and text content obtained by the detection task. The Bayesian network constructed based on the insurance knowledge graph is combined with the business rules corresponding to the bill type to perform causal reasoning verification such as timeline verification and price anomaly verification on the bill field information, thereby completing the field correlation analysis and obtaining the corresponding result.

[0071] Specifically, the compliance judgment result is a determination result that reflects whether the standardized voucher image meets the insurance claim audit requirements and whether it needs to trigger manual review, which is generated by matching the corresponding bill type difference dynamic threshold after combining the compliance score of the standardized voucher image. The compliance score is obtained by weighted calculation based on the image quality basic score, the local clarity score, and the business compliance score. The judgment of whether the standardized voucher image is compliant and whether it needs manual review is completed by comparing the different dynamic thresholds of different bill types, and then the compliance judgment result is obtained.

[0072] Further, the abnormality processing rule is designed based on the Bayesian network constructed based on the insurance knowledge graph, and is used for detecting the trigger logic of the abnormality in the field correlation analysis result. The abnormality trace analysis result is a result of locating the root cause of the abnormality and specifying the specific data basis corresponding to the abnormality by analyzing the field correlation analysis result and the compliance judgment result through the abnormality processing rule. The specific reasons and corresponding data support for the abnormality are located by matching and analyzing the logical abnormality in the field correlation analysis result and the non-compliance item in the compliance judgment result, thereby completing the abnormality trace analysis and obtaining the abnormality trace analysis result.

[0073] Furthermore, by integrating the root causes of anomalies identified in the anomaly tracing analysis results with specific data, and combining them with the previously determined invoice type, invoice field information, compliance score, and anomaly identification level, a comprehensive judgment is formed that includes the specific category of the invoice, key information details, compliance determination, anomaly situation and root cause, and whether manual review is required. This determines the invoice recognition result of the standardized voucher image. This result can be used for automated review of insurance claims, and the relevant data will also be fed back to the continuous optimization stage of the model to update the model weights.

[0074] As can be seen, the above scheme enhances the quality of the original voucher image by matching a differentiated strategy based on the image quality score, accurately identifies the boundaries of the voucher region and corrects the angle, and obtains a standardized voucher image. Then, local texture and global semantic features are extracted and deeply fused to obtain multimodal fusion features. Next, based on these features, the voucher type is identified, field information and quality thresholds are extracted, compliance scores are calculated by combining the voucher type and field information, the score is compared with the threshold to determine the anomaly level, and finally, the recognition result is determined by comprehensively considering multi-dimensional information. This can solve the problem of low accuracy in voucher image recognition.

[0075] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0076] In one embodiment, a ticket image recognition device is provided, which corresponds one-to-one with the ticket image recognition method described in the above embodiments. For example... Figure 5 As shown, the document image recognition device 100 includes an original document image enhancement module 101, a document region angle correction module 102, a multimodal fusion feature acquisition module 103, a document type recognition module 104, an anomaly recognition level acquisition module 105, and a document recognition result determination module 106. Detailed descriptions of each functional module are as follows: The original voucher image enhancement module 101 is used to acquire the original voucher image of the target business, enhance the quality of the original voucher image, and obtain an enhanced voucher image. The document area angle correction module 102 is used to identify the boundary position of the document area in the document enhancement image, and perform angle correction on the document area according to the boundary position to obtain a standardized document image; The multimodal fusion feature acquisition module 103 is used to extract the local texture features and global semantic features of the standardized voucher image respectively, and fuse the local texture features and the global semantic features to obtain the multimodal fusion features of the standardized voucher image; The bill type identification module 104 is configured to identify a bill type corresponding to the standardized certificate image according to the multi-modal fusion feature, and extract bill field information and a bill recognition quality threshold of the standardized certificate image respectively. The abnormality identification level acquisition module 105 is configured to analyze a compliance score of the standardized certificate image according to the bill type and the bill field information, and compare the compliance score with the bill recognition quality threshold to obtain an abnormality identification level of the standardized certificate image. The bill recognition result determination module 106 is configured to determine a bill recognition result of the standardized certificate image according to the bill type, the bill field information, the compliance score and the abnormality identification level.

[0077] In an embodiment, the original certificate image enhancement module 101, when performing quality enhancement on the original certificate image to obtain a certificate enhanced image, is configured to: calculate a quality score of the original certificate image by using a preset image quality evaluation algorithm; when the quality score is greater than or equal to a preset score threshold, perform illumination enhancement processing on the original certificate image to obtain a first initial enhanced image; when the quality score is less than the preset score threshold, perform histogram equalization processing on the original certificate image to obtain a second initial enhanced image; perform definition enhancement processing on the first initial enhanced image or the second initial enhanced image by using a preset super-resolution processing network to obtain a certificate enhanced image.

[0078] In an embodiment, the multi-modal fusion feature acquisition module 103, when performing extraction of the local texture feature and the global semantic feature of the standardized certificate image respectively, is configured to: perform layer-by-layer texture analysis on the standardized certificate image to obtain an initial local feature; perform dimension compression operation on the initial local feature to obtain a low-dimensional local feature; extract texture feature information in the low-dimensional local feature, and determine the local texture feature of the standardized certificate image according to the texture feature information; perform associated recognition on a certificate field in the standardized certificate image to obtain an initial global feature; perform semantic normalization processing on the initial global feature to obtain the global semantic feature of the standardized certificate image.

[0079] In an embodiment, the multi-modal fusion feature acquisition module 103, when performing feature fusion of the local texture feature and the global semantic feature to obtain the multi-modal fusion feature of the standardized certificate image, is further configured to: map the local texture feature into a query vector, and map the global semantic feature into a key vector and a value vector respectively; calculate a feature attention weight of the query vector and the key vector; perform weighted summation on the value vector by using the feature attention weight to obtain an initial fusion feature of the standardized certificate image; generate a first feature weight coefficient corresponding to the local texture feature and a second feature weight coefficient corresponding to the global semantic feature by using a preset two-way gating unit; determine a local weighted feature of the local texture feature according to the first feature weight coefficient, and determine a global weighted feature of the global semantic feature according to the second feature weight coefficient; perform element-by-element addition on the initial fusion feature, the local weighted feature, and the global weighted feature to obtain the multi-modal fusion feature of the standardized certificate image.

[0080] In an embodiment, the bill type identification module 104, when performing extraction of the bill field information of the standardized certificate image and the bill identification quality threshold respectively, is configured to: identify a certificate information region of the standardized certificate image, and extract the bill field information of the standardized certificate image in the certificate information region; perform multi-dimensional analysis on a three-dimensional quality indicator of the standardized certificate image to obtain quality score data; determine the bill identification quality threshold of the standardized certificate image according to a differentiated judgment limit of the quality score data.

[0081] In an embodiment, the abnormality identification level acquisition module 105, when performing comparison of the compliance score and the bill identification quality threshold to obtain the abnormality identification level of the standardized certificate image, is configured to: perform numerical comparison of the compliance score and the bill identification quality threshold to obtain a numerical comparison result; perform level matching on the comparison result according to a preset corresponding relationship between score intervals and abnormality levels in a business rule to obtain a level matching result; determine the abnormality identification level of the standardized certificate image according to the level matching result.

[0082] In an embodiment, the bill identification result determination module 106, when determining the bill identification result of the standardized voucher image according to the bill type, the bill field information, the compliance score and the abnormality identification level, is configured to: perform field correlation analysis on the bill field information according to the bill type to obtain a field correlation analysis result; determine a compliance judgment result of the standardized voucher image according to the compliance score; perform abnormality trace analysis on the field correlation analysis result and the compliance judgment result according to an abnormality processing rule corresponding to the abnormality identification level to obtain an abnormality trace analysis result; determine the bill identification result of the standardized voucher image according to the abnormality trace analysis result.

[0083] The present application provides a bill image identification device, which enhances the quality of original voucher images, matches differentiated strategies to improve image quality according to image quality scores, accurately identifies bill area boundaries and corrects angles to obtain standardized voucher images, then extracts local texture and global semantic features and deeply fuses them to obtain multi-modal fusion features, then identifies bill types, extracts field information and quality thresholds based on the features, calculates compliance scores in combination with bill types and field information, compares scores and thresholds to determine abnormality levels, and finally determines identification results by comprehensively considering multi-dimensional information, which can solve the problem of low accuracy of bill image identification.

[0084] The specific limitations of the bill image identification device can be referred to the limitations of the bill image identification method in the above, which will not be repeated here. Each module in the above bill image identification device can be realized by software, hardware and their combinations in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each module.

[0085] In one embodiment, a computer device is provided, which can be a server, and its internal structure diagram can be as shown in Figure 6As shown in the figure. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes non-volatile and / or volatile storage media, internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with the external client through the network connection. The computer program is executed by the processor to realize the function or step of the server side of the bill image recognition method.

[0086] In one embodiment, a computer device is provided, which can be a client, and its internal structure diagram can be as shown in the figure. Figure 7 As shown in the figure. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes non-volatile storage media, internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with the external server through the network connection. The computer program is executed by the processor to realize the function or step of the client side of the bill image recognition method.

[0087] In one embodiment, a computer device is provided, including a memory, a processor and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to realize the following steps: Obtain an original voucher image of a target service, perform quality enhancement on the original voucher image to obtain a voucher enhanced image; Identify the boundary position of the bill region in the voucher enhanced image, and perform angle correction on the bill region according to the boundary position to obtain a standardized voucher image; Respectively extract local texture features and global semantic features of the standardized voucher image, and perform feature fusion on the local texture features and the global semantic features to obtain multi-modal fusion features of the standardized voucher image; According to the multi-modal fusion features, identify the bill type corresponding to the standardized voucher image, and respectively extract the bill field information and the bill recognition quality threshold of the standardized voucher image; According to the bill type and the bill field information, analyze the compliance score of the standardized voucher image, and compare the compliance score with the bill recognition quality threshold to obtain the abnormal recognition level of the standardized voucher image; According to the bill type, the bill field information, the compliance score and the abnormal identification level, a bill identification result of the standardized voucher image is determined.

[0088] In one embodiment, a computer readable storage medium is provided, and a computer program is stored on the computer readable storage medium, and the computer program is executed by a processor to implement the following steps: An original voucher image of a target service is acquired, and quality enhancement is performed on the original voucher image to obtain a voucher enhanced image. A boundary position of a bill region in the voucher enhanced image is identified, and angle correction is performed on the bill region according to the boundary position to obtain a standardized voucher image. Local texture features and global semantic features of the standardized voucher image are extracted respectively, the local texture features and the global semantic features are fused to obtain multi-modal fusion features of the standardized voucher image. A bill type corresponding to the standardized voucher image is identified according to the multi-modal fusion features, and bill field information and a bill identification quality threshold of the standardized voucher image are extracted respectively. A compliance score of the standardized voucher image is analyzed according to the bill type and the bill field information, and the compliance score is compared with the bill identification quality threshold to obtain an abnormal identification level of the standardized voucher image. A bill identification result of the standardized voucher image is determined according to the bill type, the bill field information, the compliance score and the abnormal identification level.

[0089] It should be noted that the functions or steps of the computer readable storage medium or the computer device described above can be referred to the related descriptions of the server side and the client side in the foregoing method embodiments. To avoid repetition, they will not be described here.

[0090] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0091] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is exemplified, and in actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0092] It should be noted that if non-company software tools or components appear in the embodiments of the present application, they are only used for example introduction and do not represent actual use.

[0093] The above-mentioned embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method of identifying a document image, characterized by, The method comprises the following steps: obtaining an original image of a target business credential, performing quality enhancement on the original image to obtain a credential enhanced image; identifying the boundary position of the bill region in the credential enhanced image, performing angle correction on the bill region according to the boundary position to obtain a standardized credential image; extracting local texture features and global semantic features of the standardized credential image respectively, and performing feature fusion on the local texture features and the global semantic features to obtain multi-modal fusion features of the standardized credential image; identifying the bill type corresponding to the standardized credential image according to the multi-modal fusion features, and extracting bill field information and bill recognition quality threshold of the standardized credential image respectively; analyzing the compliance score of the standardized credential image according to the bill type and the bill field information, and comparing the compliance score with the bill recognition quality threshold to obtain the abnormal recognition level of the standardized credential image; determining the bill recognition result of the standardized credential image according to the bill type, the bill field information, the compliance score and the abnormal recognition level.

2. The method of claim 1, wherein calculating the quality score of the original credential image by using a preset image quality evaluation algorithm; when the quality score is greater than or equal to a preset score threshold, performing illumination enhancement processing on the original credential image to obtain a first initial enhanced image; when the quality score is less than the preset score threshold, performing histogram equalization processing on the original credential image to obtain a second initial enhanced image; performing clarity enhancement processing on the first initial enhanced image or the second initial enhanced image by using a preset super-resolution processing network to obtain a credential enhanced image.

3. The method of claim 1, wherein The method comprises the following steps: performing layer-by-layer texture analysis on the standardized credential image to obtain initial local features; performing dimension compression operation on the initial local features to obtain low-dimensional local features; extracting texture feature information in the low-dimensional local features, and determining the local texture features of the standardized credential image according to the texture feature information; performing associated recognition on the credential field in the standardized credential image to obtain initial global features; performing semantic normalization processing on the initial global features to obtain the global semantic features of the standardized credential image.

4. The method of claim 1, wherein The method comprises the following steps: mapping the local texture features into query vectors, and mapping the global semantic features into key vectors and value vectors respectively; calculating the feature attention weight of the query vector and the key vector; performing weighted summation on the value vector by using the feature attention weight to obtain the initial fusion features of the standardized credential image; generating the first feature weight coefficient corresponding to the local texture features and the second feature weight coefficient corresponding to the global semantic features by using a preset two-way gating unit; determine a local weighted feature of the local texture feature according to the first feature weight coefficient, and determine a global weighted feature of the global semantic feature according to the second feature weight coefficient; perform element-by-element addition on the initial fusion feature, the local weighted feature and the global weighted feature to obtain a multi-modal fusion feature of the standardized voucher image.

5. The method of claim 1, wherein The extracting the bill field information and the bill recognition quality threshold of the standardized voucher image respectively includes: identifying a voucher information region of the standardized voucher image, and extracting the bill field information of the standardized voucher image in the voucher information region; performing multi-dimensional analysis on the three-dimensional quality indicators of the standardized voucher image to obtain quality score data; determining the bill recognition quality threshold of the standardized voucher image according to the differentiated judgment limit of the quality score data.

6. The method of claim 1, wherein The comparing the compliance score with the bill recognition quality threshold to obtain the abnormal recognition level of the standardized voucher image includes: numerical comparison of the compliance score and the bill recognition quality threshold to obtain a numerical comparison result; performing level matching on the comparison result according to the corresponding relationship between the score interval and the abnormal level in the preset business rule to obtain a level matching result; determining the abnormal recognition level of the standardized voucher image according to the level matching result.

7. The method of claim 1, wherein The determining the bill recognition result of the standardized voucher image according to the bill type, the bill field information, the compliance score and the abnormal recognition level includes: performing field association analysis on the bill field information according to the bill type to obtain a field association analysis result; determining a compliance judgment result of the standardized voucher image according to the compliance score; performing abnormal trace analysis on the field association analysis result and the compliance judgment result according to the abnormal processing rule corresponding to the abnormal recognition level to obtain an abnormal trace analysis result; determining the bill recognition result of the standardized voucher image according to the abnormal trace analysis result.

8. A bill image recognition apparatus characterized by comprising: It includes: an original voucher image enhancement module, configured to obtain an original voucher image of a target business, perform quality enhancement on the original voucher image, and obtain a voucher enhanced image; a bill region angle correction module, configured to identify a boundary position of a bill region in the voucher enhanced image, perform angle correction on the bill region according to the boundary position, and obtain a standardized voucher image; a multi-modal fusion feature acquisition module, configured to extract a local texture feature and a global semantic feature of the standardized voucher image respectively, perform feature fusion on the local texture feature and the global semantic feature, and obtain a multi-modal fusion feature of the standardized voucher image; a bill type identification module, configured to identify a bill type corresponding to the standardized voucher image according to the multi-modal fusion feature, and extract bill field information and a bill recognition quality threshold of the standardized voucher image respectively; an abnormality identification level obtaining module, configured to analyze a compliance score of the standardized receipt image according to the receipt type and the receipt field information, and compare the compliance score with the receipt identification quality threshold to obtain an abnormality identification level of the standardized receipt image; a receipt identification result determining module, configured to determine a receipt identification result of the standardized receipt image according to the receipt type, the receipt field information, the compliance score and the abnormality identification level.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the receipt image identification method in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the receipt image identification method in any one of claims 1 to 7.

Images