A remote intelligent invoice collection method based on financial invoice data model
By combining OCR technology and deep learning algorithms to process and identify invoice images, the problem of image quality affecting recognition accuracy in traditional methods is solved, and efficient and accurate invoice data processing and storage is achieved.
Patent Information
- Application Number
- CN202411930820.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-26
- Publication Date
- 2025-04-04
AI Technical Summary
The traditional remote intelligent invoice collection method affects the OCR recognition accuracy when the image quality is uneven, and requires manual classification, entry and storage, which is time-consuming and error-prone.
The remote intelligent invoice collection method based on the financial invoice data model is adopted, and combined with OCR technology, deep learning algorithms, image processing technology and feature extraction and matching technology, the identification, preprocessing, character cutting and feature extraction of invoice information is carried out to ensure image quality and improve recognition accuracy.
By improving the accuracy and robustness of image recognition, reducing manual intervention, improving the accuracy and consistency of invoice data, and improving overall processing efficiency.
Smart Images
Figure CN119763138A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of invoice identification, and more specifically, to a remote intelligent invoice collection method based on a financial invoice data model. Background Art
[0002] As financial and tax management are increasingly digitalized, a remote intelligent invoice collection method based on the financial invoice data model has emerged. This method mainly relies on cloud computing, big data and artificial intelligence technologies to achieve remote, intelligent and efficient processing of invoices. With the advancement of enterprise informatization, electronic invoices have gradually replaced paper invoices and become the mainstream form of invoices. Electronic invoices have the advantages of easy storage, fast transmission speed and convenient search, which provides a basis for remote intelligent collection of invoices. On this basis, a financial invoice data model is constructed to standardize and analyze invoice data to ensure the accuracy and consistency of the data.
[0003] The above disclosed technical solutions have at least the following technical problems: In practical applications, the image quality of traditional remote intelligent invoice collection methods varies. In particular, invoice images may be blurry, noisy, and have other problems, which will affect the accuracy of OCR recognition. In addition, traditional invoice data often requires manual classification, entry, and storage, which is time-consuming and error-prone. In view of the above problems, the present invention proposes a solution. Summary of the invention
[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a remote intelligent invoice collection method based on a financial invoice data model, which solves the problems mentioned in the above background technology by utilizing optical character recognition (OCR) technology in combination with deep learning algorithms, image processing technology, feature extraction and matching technology.
[0005] To achieve the above object, the present invention provides the following technical solutions: A remote intelligent invoice collection method based on a financial invoice data model comprises the following steps: recognizing invoice information based on OCR technology, preprocessing the recognized invoice text image data, sending the processed image to a remote control terminal, and obtaining image comparison data; performing character segmentation on the preprocessed image, and extracting features from the segmented characters, and obtaining the efficiency of current character extraction according to a preset feature extraction evaluation model; matching the extracted character features with pre-trained template features to determine the correct position and direction of each character; and establishing an invoice data model to classify and structure the various types of recognized invoices.
[0006] In a preferred embodiment, the acquisition of image comparison data is specifically as follows: a portion of the image data to be tested is divided into multiple groups of the same size, and the sharpness evaluation result of each group of image data is obtained by calculating the Laplace operator of the image; the obtained sharpness evaluation results are compared in turn to obtain the maximum sharpness evaluation result, that is, the maximum clarity of the image; the sharpness evaluation result data of each group of image data of a portion of image data are added and averaged to obtain the sharpness average value of the current image data, that is, the overall clarity of the image; the total number of pixels in the image data is obtained, and the noise area of the image data to be tested, that is, the number of noise pixels, is extracted by a local statistical method, and the number of noise pixels in the image is compared with the total number of pixels to obtain the noise ratio; a portion of the image data is divided into multiple equal areas, and a local window function is used to calculate the noise ratio. Calculate the contrast of each area of the image to obtain the maximum contrast and the minimum contrast; divide the absolute value of the difference between the maximum contrast and the minimum contrast by the sum of the maximum contrast and the minimum contrast to obtain the global contrast, and use the global contrast as the ideal contrast; enhance the contrast of the image through a gamma correction algorithm, and obtain the enhanced global contrast, and divide the enhanced global contrast by the global contrast before enhancement to obtain the contrast enhancement effect; input the above maximum clarity, the overall clarity of the image, the noise ratio, the number of total pixels, the ideal contrast and the contrast enhancement effect into a preset image comparison data formula to obtain the comparison data of the image to be tested; compare the image comparison data with a preset threshold value, and if it is lower than the preset value, reject the image data to the output end, and re-collect and output it.
[0007] In a preferred embodiment, the preprocessed image is segmented into characters, and features of the segmented characters are extracted according to a preset feature extraction evaluation model, specifically as follows: the position of the characters in the image data is found by calculating the horizontal and vertical projections of the text area, the image data is segmented, and individual characters are extracted; feature extraction is performed on the segmented characters based on a machine learning algorithm, and the extracted features include the shape, texture, and direction of the characters; a corresponding feature extraction method is selected according to the feature type; if it is a shape feature, contour extraction and geometric feature calculation methods are used; if it is a texture feature, gray-level co-occurrence matrix and Fourier transform method are used.
[0008] In a preferred embodiment, the efficiency of obtaining the current character extraction is specifically as follows: obtaining the total number of characters cut from the image, and calculating the number of correctly cut characters and the number of incorrectly cut characters by comparing them with the marked real characters; in the process of character cutting, recording the number of character fragments generated by all cutting, that is, the total number of cuts, and at the same time, based on morphology, setting a constraint condition to extract character fragments that are too small and incomplete to obtain the number of cut fragments; inputting the above-mentioned total number of characters, the number of correctly cut characters, the number of incorrectly cut characters, the total number of cuts and the number of cut fragments into a preset character cutting accuracy formula to obtain the character cutting accuracy; using a feature extraction algorithm to extract all useful features from image data and text data, counting the total number of all extracted features, and recording the feature extraction process The start and end time of the feature extraction process are used to obtain the feature extraction time; the average extraction time of each feature extraction is obtained by obtaining the extraction time and averaging them, and the average extraction time is used as the ideal time; by evaluating the contribution of each feature in character recognition, it is determined which features are effective and the number of effective features is obtained; the features that are highly correlated with other features and have no additional contribution to the recognition results are eliminated through feature selection methods, and the number of features that have no additional contribution to the recognition results, that is, the number of redundant features, is recorded; the feature extraction effectiveness is calculated based on the preset feature extraction effectiveness coefficient formula according to the total number of extracted features, the number of effective features, the number of redundant features, the ideal time and the feature extraction time; the feature extraction effectiveness is obtained by multiplying the feature extraction effectiveness by the character cutting accuracy.
[0009] In a preferred embodiment, the contribution of each feature in character recognition is evaluated as follows: the extracted features are converted into feature vectors, each feature vector corresponds to a character and a character category, and the feature vectors are standardized; the original performance of the trained model on the test set is evaluated using a test data set, and the evaluation indicators include recognition accuracy, character error rate, and word error rate; the data order of each feature column is disrupted one by one, and the performance changes of the model are re-evaluated, and the performance degradation caused by each feature to the model is calculated, thereby obtaining the feature contribution; according to the comparison between the performance degradation ratio and the preset value, if the performance degradation ratio is higher than the preset value, the feature corresponding to the performance degradation ratio is a valid feature.
[0010] In a preferred embodiment, the extracted character features are matched with pre-trained template features to determine the correct position and direction of each character as follows: the extracted character features are matched with pre-trained template features to determine the correct position and direction of each character; according to the matching results, the characters are corrected, if the character recognition is wrong, the wrong characters are replaced with correct characters, if the character position and direction are incorrect, they are adjusted; the corrected text is further processed, including re-correcting the recognized text, typesetting the recognized text, and classifying the recognized text; the re-correction of the recognized text includes correcting typos and removing extra spaces; the typesetting of the recognized text includes adjusting the font, font size, and line spacing; the classification of the recognized text includes classifying the text into Chinese, English, and numbers.
[0011] In a preferred embodiment, the invoice data model is established to classify and structure the various types of identified invoices as follows: for each invoice image, the OCR model needs to identify the text information therein, including the invoice number, amount, and date; a labeled invoice image data set is collected, and the images in the data set contain clear invoice information; the invoice data model is constructed by combining the deep learning model of CNN and RNN, wherein CNN is used to extract image features and RNN is used to process sequence data; the invoice data model is trained with the OCR data set, and the OCR model is verified to evaluate the accuracy of the model in text recognition; the various types of identified invoices are classified and a database is constructed using the trained invoice data model.
[0012] The technical effects and advantages of the remote intelligent invoice collection method based on the financial invoice data model of the present invention are as follows: 1. The present invention combines OCR technology with a deep learning model to accurately identify text information in invoice images. Image processing technologies such as image sharpness assessment, noise removal, and contrast enhancement are used to ensure that the input image meets the optimal quality standards, thereby improving the accuracy of image recognition. At the same time, by optimizing the cutting and feature extraction of characters and combining machine learning algorithms to select the most effective features, the accuracy of character recognition is further improved.
[0013] 2. The present invention adopts a number of image quality assessment technologies, such as image clarity, contrast, and noise ratio calculation. Through these methods, the system can automatically identify and remove noise areas in the image, thereby ensuring the quality of the input image. Technologies such as image sharpness assessment and gamma correction effectively improve the detail performance of the image and avoid recognition errors caused by poor image quality. This enables the system to adapt to image inputs of different qualities in practical applications and improves robustness.
[0014] 3. The present invention uses an evaluation model to dynamically adjust the efficiency of character recognition. By evaluating the contribution of each feature, it is possible to identify in real time which features play a key role in improving recognition accuracy. By calculating indicators of multiple dimensions such as recognition efficiency and accuracy, the system can optimize the model according to actual conditions so that the performance of each recognition is always maintained at the best level. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 The present invention is a schematic structural diagram of a method for remote intelligent collection of invoices based on a financial invoice data model. DETAILED DESCRIPTION
[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0017] Embodiment 1, Figure 1 The present invention provides a remote intelligent invoice collection method based on a financial invoice data model, comprising the following steps: S1, invoice information is recognized based on OCR technology, the recognized invoice text image data is pre-processed, the processed image is sent to the remote control end, and the image comparison data is obtained.
[0018] In this embodiment, the invoice information is recognized based on the OCR technology, the recognized invoice text image data is pre-processed, and the processed image is sent to the remote control terminal, as follows: Collect annotated invoice datasets. The data includes the image of each invoice and its corresponding category label. The data labels are discrete category labels, including but not limited to VAT invoices and ordinary invoices. Preprocess the recognized invoice text image data, including denoising, scaling, normalization and binarization; The processed image is sent to the authorized remote control terminal using the AES encryption algorithm.
[0019] In this embodiment, the comparison data of the acquired image is specifically as follows: A set of image data to be tested is divided into multiple groups of the same size, and the sharpness evaluation result of each group of image data is obtained by calculating the Laplacian operator of the image; The obtained sharpness evaluation results are compared in sequence to obtain the maximum sharpness evaluation result, that is, the maximum clarity of the image; The sharpness evaluation result data of each group of image data of a piece of image data are added and averaged to obtain the average sharpness of the current image data, that is, the overall clarity of the image; The total number of pixels in the image data is obtained, and the noise area of the image data to be tested, that is, the number of noise pixels, is extracted by a local statistical method, and the number of noise pixels in the image is compared with the total number of pixels to obtain the noise ratio; Divide an image data into multiple equal regions, calculate the contrast of each region of the image based on the local window function, and obtain the maximum contrast and minimum contrast; The absolute value of the difference between the maximum contrast and the minimum contrast is divided by the sum of the maximum contrast and the minimum contrast to obtain the global contrast, which is taken as the ideal contrast. The contrast of the image is enhanced by a gamma correction algorithm, and the enhanced global contrast is obtained. The enhanced global contrast is divided by the global contrast before enhancement to obtain a contrast enhancement effect. Input the above maximum definition, overall definition of the image, noise ratio, total number of pixels, ideal contrast and contrast enhancement effect into a preset image comparison data formula to obtain comparison data of the image to be tested; The image comparison data is compared with the preset threshold. If it is lower than the preset value, the image data is rejected to the output end for re-collection and output.
[0020] The image comparison data calculation formula is as follows:
[0021] In the formula, is the image comparison data, is the maximum clarity, is the overall clarity of the image, is the total number of pixels, is the noise ratio, is the ideal contrast ratio, is the contrast enhancement effect, , and is the weight coefficient and satisfies .
[0022] It should be noted that the image comparison data The larger the value, the better the image has achieved in terms of clarity, noise suppression, and contrast enhancement. In the preprocessing process, it can better retain the useful information in the image while reducing noise and interference. This is crucial for subsequent image analysis, recognition, and processing tasks; image comparison data The smaller the value, the more obvious problems the image may have in terms of clarity, noise ratio, and contrast. This will cause the overall image quality to deteriorate, becoming blurry, noisy, and difficult to observe. It also means that a lot of useful information is lost during the preprocessing process, while more noise and interference are retained. This will have an adverse effect on subsequent image analysis, recognition, and processing tasks.
[0023] S2, performing character segmentation on the preprocessed image, and extracting features from the segmented characters, and obtaining the efficiency of the current character extraction according to a preset feature extraction evaluation model.
[0024] In this embodiment, the preprocessed image is segmented into characters, and features of the segmented characters are extracted according to a preset feature extraction evaluation model, as follows: By calculating the projection of the text area in the horizontal and vertical directions, the position of the characters in the image data is found, the image data is segmented, and individual characters are extracted; The segmented characters are subjected to feature extraction based on a machine learning algorithm. The extracted features include the shape, texture, and direction of the characters. Select the corresponding feature extraction method according to the feature type; If it is a shape feature, contour extraction and geometric feature calculation methods are used; If it is a texture feature, the gray level co-occurrence matrix and Fourier transform method are used.
[0025] In this embodiment, the efficiency of obtaining the current character extraction is specifically as follows: Obtain the total number of characters cut from the image, and calculate the number of correctly cut characters and the number of incorrectly cut characters by comparing them with the real characters marked; During the character cutting process, the number of character fragments generated by all cutting is recorded, that is, the total number of cuttings. At the same time, based on morphology, a constraint condition is set to extract the character fragments that are too small and incomplete to obtain the number of cutting fragments; The total number of characters, the number of correctly cut characters, the number of incorrectly cut characters, the total number of cuts and the number of cut fragments are input into a preset character cutting accuracy formula to obtain the character cutting accuracy; Use feature extraction algorithms to extract all useful features from image data and text data, count the total number of all extracted features, and record the start and end time of the feature extraction process to obtain the feature extraction time; By obtaining the extraction time of each feature extraction and averaging them, the average extraction time of each feature extraction is obtained, and the average extraction time is taken as the ideal time; By evaluating the contribution of each feature in character recognition, determine which features are effective and obtain the number of effective features; The features that are highly correlated with other features and have no additional contribution to the recognition results are eliminated through feature selection methods, and the number of features that have no additional contribution to the recognition results, that is, the number of redundant features, is recorded; The feature extraction validity is calculated based on the total number of extracted features, the number of valid features, the number of redundant features, the ideal time and the feature extraction time based on the preset feature extraction validity coefficient formula; The efficiency of current character extraction is obtained by multiplying the feature extraction effectiveness and the character segmentation accuracy.
[0026] The calculation formula of the character cutting accuracy is as follows:
[0027] The calculation formula for the effectiveness of feature extraction is as follows:
[0028] The calculation formula of the efficiency of the current character extraction is as follows:
[0029] In the formula, is the character cutting accuracy, is the number of characters to be cut correctly, is the total number of characters cut, is the number of characters that were miscut, is the number of cut pieces, is the total number of cuts, , and is the weight coefficient, is the feature extraction effectiveness coefficient, is the effective feature number, is the total number of features, is the number of redundant features, is the feature extraction time, It’s the ideal time. , and is the weight coefficient, is the efficiency of current character extraction.
[0030] It should be noted that the higher the character cutting accuracy, the more accurate the character cutting process is, and the characters in the text can be correctly identified and cut, with fewer miscutting and cutting fragments. The lower the character cutting accuracy, the more errors there are in the cutting results (such as miscutting or cutting fragments). Low character cutting accuracy will cause subsequent character recognition systems to face more challenges, because inaccurate cutting will make it difficult for the recognition system to accurately distinguish characters, thereby reducing recognition accuracy.
[0031] In addition, the higher the feature extraction effectiveness coefficient is, the more efficient and effective the feature extraction process is, and a large number of valid features can be extracted while minimizing redundant features. The effective features can better characterize the distinguishing information of characters and are significantly helpful for subsequent character recognition. At the same time, a shorter feature extraction time means that the system is efficient and can complete feature extraction in a shorter time, providing enough time for the recognition process to process more samples. The lower the feature extraction effectiveness coefficient is, the more redundant features may be extracted, or fewer valid features may be extracted, or even lead to excessively long feature extraction time, which not only reduces the accuracy of character recognition, but also may increase processing time and affect the real-time performance of the system.
[0032] Finally, if the current character extraction efficiency is higher, it means that the character segmentation is very accurate, each character can be correctly extracted, and the feature extraction is also very effective, which can effectively extract the key information to distinguish the characters. This will directly improve the accuracy of the entire recognition system and reduce misrecognition and missed recognition; if the current character extraction efficiency is lower, it means that the character segmentation is inaccurate, which may cause the characters to be miscut or fragmented, and the extracted features may not be able to effectively distinguish the characters or have redundancy, resulting in reduced accuracy of the recognition results; the current character extraction efficiency directly determines the accuracy, efficiency and robustness of the character recognition system. High character segmentation accuracy and feature extraction validity coefficient values help improve system performance, otherwise it may lead to system performance degradation.
[0033] The contribution of each feature in character recognition is evaluated as follows: The extracted features are converted into feature vectors, each feature vector corresponds to a character and a character category, and the feature vectors are standardized; Use the test data set to evaluate the original performance of the trained model on the test set. The evaluation indicators include recognition accuracy, character error rate, and word error rate. The data order of each feature column is disrupted one by one, and the performance change of the model is re-evaluated. The performance degradation caused by each feature to the model is calculated to obtain the feature contribution. According to the comparison between the performance degradation ratio and the preset value, if the performance degradation ratio is higher than the preset value, the feature corresponding to the performance degradation ratio is a valid feature.
[0034] S3, matches the extracted character features with the pre-trained template features to determine the correct position and orientation of each character.
[0035] In this embodiment, the extracted character features are matched with the pre-trained template features to determine the correct position and direction of each character as follows: Match the extracted character features with the pre-trained template features to determine the correct position and orientation of each character; According to the matching results, the characters are corrected. If the character recognition is wrong, the wrong character is replaced with the correct character. If the character position and direction are incorrect, they are adjusted; Further processing the corrected text, including re-correcting the recognized text, typesetting the recognized text, and classifying the recognized text; The re-correction of the recognized text includes correcting typos and removing extra spaces; The typesetting of the recognized text includes adjusting the font, font size, and line spacing; The classifying of the recognized text includes classifying the text into Chinese, English, and numbers.
[0036] S4, establish an invoice data model to classify and structure the various types of invoices identified.
[0037] In this embodiment, the invoice data model is established to classify and structure the identified invoices as follows: For each invoice image, the OCR model needs to recognize the text information, including the invoice number, amount, and date; Collect annotated invoice image datasets, where the images contain clear invoice information; The invoice data model is constructed by combining the deep learning model of CNN and RNN, where CNN is used to extract image features and RNN is used to process sequence data; Train the invoice data model using the OCR dataset, validate the OCR model, and evaluate the model’s accuracy in text recognition; The various types of invoices identified are classified and a database is constructed using the trained invoice data model.
[0038] The calculation formula of the invoice data model is as follows:
[0039] In the formula, is the cross entropy loss value, is the number of categories, is the actual category label, is the class probability predicted by the model.
[0040] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.
[0041] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product.
[0042] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.
[0043] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.
[0044] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
[0045] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A remote intelligent invoice collection method based on a financial invoice data model, characterized in that: The steps include: Recognize invoice information based on OCR technology, pre-process the recognized invoice text image data, send the processed image to the remote control terminal, and obtain image comparison data; Perform character segmentation on the preprocessed image and extract features from the segmented characters. Obtain the efficiency of current character extraction based on a preset feature extraction evaluation model. Match the extracted character features with the pre-trained template features to determine the correct position and orientation of each character; Establish an invoice data model to classify and structure the various types of identified invoices.
2. According to claim 1, a method for remote intelligent collection of invoices based on a financial invoice data model is characterized in that: The comparison data of the obtained image is specifically as follows: A set of image data to be tested is divided into multiple groups of the same size, and the sharpness evaluation result of each group of image data is obtained by calculating the Laplacian operator of the image; The obtained sharpness evaluation results are compared in sequence to obtain the maximum sharpness evaluation result, that is, the maximum clarity of the image; The sharpness evaluation result data of each group of image data of a piece of image data are added and averaged to obtain the average sharpness of the current image data, that is, the overall clarity of the image; The total number of pixels in the image data is obtained, and the noise area of the image data to be tested, that is, the number of noise pixels, is extracted by a local statistical method, and the number of noise pixels in the image is compared with the total number of pixels to obtain the noise ratio; Divide an image data into multiple equal regions, calculate the contrast of each region of the image based on the local window function, and obtain the maximum contrast and minimum contrast; The absolute value of the difference between the maximum contrast and the minimum contrast is divided by the sum of the maximum contrast and the minimum contrast to obtain the global contrast, which is taken as the ideal contrast. The contrast of the image is enhanced by a gamma correction algorithm, and the enhanced global contrast is obtained. The enhanced global contrast is divided by the global contrast before enhancement to obtain a contrast enhancement effect. Input the above maximum definition, overall definition of the image, noise ratio, total number of pixels, ideal contrast and contrast enhancement effect into a preset image comparison data formula to obtain comparison data of the image to be tested; The image comparison data is compared with the preset threshold. If it is lower than the preset value, the image data is rejected to the output end for re-collection and output.
3. The method for remote intelligent collection of invoices based on a financial invoice data model according to claim 2, characterized in that: The preprocessed image is segmented into characters, and features of the segmented characters are extracted according to a preset feature extraction evaluation model, as follows: By calculating the projection of the text area in the horizontal and vertical directions, the position of the characters in the image data is found, the image data is segmented, and individual characters are extracted; The segmented characters are subjected to feature extraction based on a machine learning algorithm. The extracted features include the shape, texture, and direction of the characters. Select the corresponding feature extraction method according to the feature type; If it is a shape feature, contour extraction and geometric feature calculation methods are used; If it is a texture feature, the gray level co-occurrence matrix and Fourier transform method are used.
4. The method for remote intelligent collection of invoices based on a financial invoice data model according to claim 3 is characterized in that: The efficiency of obtaining the current character extraction is as follows: Obtain the total number of characters cut from the image, and calculate the number of correctly cut characters and the number of incorrectly cut characters by comparing them with the real characters marked; During the character cutting process, the number of character fragments generated by all cutting is recorded, that is, the total number of cuttings. At the same time, based on morphology, a constraint condition is set to extract the character fragments that are too small and incomplete to obtain the number of cutting fragments; The total number of characters, the number of correctly cut characters, the number of incorrectly cut characters, the total number of cuts and the number of cut fragments are input into a preset character cutting accuracy formula to obtain the character cutting accuracy; Use feature extraction algorithms to extract all useful features from image data and text data, count the total number of all extracted features, and record the start and end time of the feature extraction process to obtain the feature extraction time; By obtaining the extraction time of each feature extraction and averaging them, the average extraction time of each feature extraction is obtained, and the average extraction time is taken as the ideal time; By evaluating the contribution of each feature in character recognition, determine which features are effective and obtain the number of effective features; The features that are highly correlated with other features and have no additional contribution to the recognition results are eliminated through feature selection methods, and the number of features that have no additional contribution to the recognition results, that is, the number of redundant features, is recorded; The feature extraction effectiveness is calculated based on the preset feature extraction effectiveness coefficient formula according to the total number of extracted features, the number of effective features, the number of redundant features, the ideal time and the feature extraction time; The efficiency of current character extraction is obtained by multiplying the feature extraction effectiveness and the character segmentation accuracy.
5. The method for remote intelligent collection of invoices based on a financial invoice data model according to claim 4, characterized in that: The contribution of each feature in character recognition is evaluated as follows: The extracted features are converted into feature vectors, each feature vector corresponds to a character and a character category, and the feature vectors are standardized; Use the test data set to evaluate the original performance of the trained model on the test set. The evaluation indicators include recognition accuracy, character error rate, and word error rate. The data order of each feature column is disrupted one by one, and the performance change of the model is re-evaluated. The performance degradation caused by each feature to the model is calculated to obtain the feature contribution. According to the comparison between the performance degradation ratio and the preset value, if the performance degradation ratio is higher than the preset value, the feature corresponding to the performance degradation ratio is a valid feature.
6. The method for remote intelligent collection of invoices based on a financial invoice data model according to claim 5, characterized in that: The extracted character features are matched with the pre-trained template features to determine the correct position and direction of each character as follows: Match the extracted character features with the pre-trained template features to determine the correct position and orientation of each character; According to the matching results, the characters are corrected. If the character recognition is wrong, the wrong character is replaced with the correct character. If the character position and direction are incorrect, they are adjusted; Further processing the corrected text, including re-correcting the recognized text, typesetting the recognized text, and classifying the recognized text; The re-correction of the recognized text includes correcting typos and removing extra spaces; The typesetting of the recognized text includes adjusting the font, font size, and line spacing; The classifying of the recognized text includes classifying the text into Chinese, English, and numbers.
7. The method for remote intelligent collection of invoices based on a financial invoice data model according to claim 6, characterized in that: The invoice data model is established to classify and structure the identified invoices as follows: For each invoice image, the OCR model needs to recognize the text information, including the invoice number, amount, and date; Collect annotated invoice image datasets, where the images contain clear invoice information; The invoice data model is constructed by combining the deep learning model of CNN and RNN, where CNN is used to extract image features and RNN is used to process sequence data; Train the invoice data model using the OCR dataset, validate the OCR model, and evaluate the model’s accuracy in text recognition; The various types of invoices identified are classified and a database is constructed using the trained invoice data model.
8. The method for remote intelligent collection of invoices based on a financial invoice data model according to claim 7, characterized in that: The image comparison data calculation formula is as follows: In the formula, is the image comparison data, is the maximum clarity, is the overall clarity of the image, is the total number of pixels, is the noise ratio, is the ideal contrast ratio, is the contrast enhancement effect, , and is the weight coefficient and satisfies .
9. The method for remote intelligent collection of invoices based on a financial invoice data model according to claim 8, characterized in that: The calculation formula of the character cutting accuracy is as follows: The calculation formula for the effectiveness of feature extraction is as follows: The calculation formula of the efficiency of the current character extraction is as follows: In the formula, is the character cutting accuracy, is the number of characters to be cut correctly, is the total number of characters cut, is the number of characters that were miscut, is the number of cut pieces, is the total number of cuts, , and is the weight coefficient, is the feature extraction effectiveness coefficient, is the effective feature number, is the total number of features, is the number of redundant features, is the feature extraction time, It’s the ideal time. , and is the weight coefficient, is the efficiency of current character extraction.