Data processing method and system for bill identification and storage medium

Abstract

The invention relates to the technical field of bill data processing, in particular to a data processing method and system for bill recognition and a storage medium. The method comprises the following steps: carrying out multispectral image acquisition on a bill to obtain a bill multispectral image set; performing cross-spectrum registration on the bill multispectral image set to obtain an aligned bill image set; performing weight feature fusion on the aligned bill image set to obtain an enhanced bill feature map; carrying out topological transformation feature extraction on the enhanced bill feature map to obtain bill geometric feature mapping data; performing local semantic decoupling on the enhanced bill feature map based on bill geometric feature mapping data to obtain a bill topological feature point set; and performing information entropy analysis on the bill topological feature point set to obtain a bill information entropy gradient distribution diagram. According to the invention, the recognition accuracy of non-standard bills and special marks is enhanced, and the recognition effect of fuzzy or damaged bills is improved at the same time.

Description

Technical Field

[0001] The present invention relates to the technical field of bill data processing, and in particular to a data processing method, system and storage medium for bill identification. Background Art

[0002] In the process of bill recognition, the diversity and complexity of bills are a significant technical challenge. There are many types of bills, including but not limited to checks, invoices, receipts, etc., which have huge differences in format, font, paper quality, etc. These differences lead to the need for algorithms to have extremely high adaptability and robustness in the process of bill recognition. However, the existing bill recognition system often has a significant decrease in recognition accuracy when facing non-standard bills or special marks on bills, which is a problem that cannot be ignored in fields such as financial transactions and tax audits that require high accuracy. Secondly, bill recognition technology also faces challenges when dealing with blurred or damaged bills. In practical applications, bills become blurred due to long-term storage, improper storage conditions, or damage during transportation. In addition, some bills have the problem of blurring themselves due to poor printing quality. When dealing with these blurred or damaged bills, the existing bill recognition technology often cannot accurately identify the key information on the bill, resulting in inaccurate recognition results. Summary of the invention

[0003] Based on this, it is necessary for the present invention to provide a data processing method, system and storage medium for bill identification to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a data processing method for bill identification includes the following steps:

[0005] Step S1: collect multispectral images of bills to obtain a multispectral image set of bills; perform cross-spectral registration on the multispectral image set of bills to obtain an aligned bill image set; perform weighted feature fusion on the aligned bill image set to obtain an enhanced bill feature map;

[0006] Step S2: extracting topological transformation features of the enhanced bill feature map to obtain bill geometric feature mapping data; performing local semantic decoupling of the enhanced bill feature map based on the bill geometric feature mapping data to obtain a bill topological feature point set;

[0007] Step S3: performing information entropy analysis on the bill topological feature point set to obtain a bill information entropy gradient distribution map; constructing an adaptive coding mapping table based on the bill information entropy gradient distribution map to obtain a bill dynamic coding sequence;

[0008] Step S4: performing semantic relevance evaluation on the bill dynamic coding sequence to obtain semantic consistency reconstruction mapping data; performing semantic correction on the bill dynamic coding sequence according to the semantic consistency reconstruction mapping data to obtain high-confidence bill feature representation data;

[0009] Step S5: Perform prior knowledge graph matching on high-confidence bill feature representation data to obtain a bill candidate recognition result set; perform semantic reasoning on the bill candidate recognition result set to obtain bill semantic verification data; perform intelligent error correction on the bill candidate recognition result set based on the bill semantic verification data to obtain the final bill recognition output data.

[0010] The present invention significantly improves the adaptability and robustness to the diversity and complexity of bills through multispectral image acquisition and cross-spectral registration, and can effectively cope with bills of different formats, fonts and paper qualities. The recognition accuracy of non-standard bills and special marks is enhanced through weight feature fusion and topological transformation feature extraction, while the recognition effect of blurred or damaged bills is improved and high accuracy is maintained. The accuracy of information extraction is further improved through information entropy analysis and the construction of adaptive coding mapping tables, even when part of the information is lost or incomplete. The semantic consistency of the recognition results is enhanced through semantic relevance evaluation and semantic correction, and high-confidence bill feature representation data is obtained. The recognition accuracy is improved by combining the intelligent error correction mechanism of prior knowledge graph matching and semantic reasoning, especially in areas with high accuracy requirements such as financial transactions and tax audits.

[0011] Preferably, the present invention further provides a data processing system for bill identification, which is used to execute the data processing method for bill identification as described above, and the data processing system for bill identification comprises:

[0012] The image acquisition module is used to acquire multispectral images of bills to obtain a multispectral image set of bills; perform cross-spectral registration on the multispectral image set of bills to obtain an aligned bill image set; perform weighted feature fusion on the aligned bill image set to obtain an enhanced bill feature map;

[0013] A topological feature extraction module is used to extract topological transformation features of the enhanced bill feature map to obtain bill geometric feature mapping data; based on the bill geometric feature mapping data, the enhanced bill feature map is locally semantically decoupled to obtain a bill topological feature point set;

[0014] The information entropy coding module is used to perform information entropy analysis on the bill topological feature point set to obtain the bill information entropy gradient distribution map; construct an adaptive coding mapping table based on the bill information entropy gradient distribution map to obtain the bill dynamic coding sequence;

[0015] The semantic evaluation module is used to evaluate the semantic relevance of the bill dynamic coding sequence to obtain semantic consistency reconstruction mapping data; semantically correct the bill dynamic coding sequence according to the semantic consistency reconstruction mapping data to obtain high-confidence bill feature representation data;

[0016] The knowledge matching module is used to perform prior knowledge graph matching on high-confidence bill feature representation data to obtain a bill candidate recognition result set; perform semantic reasoning on the bill candidate recognition result set to obtain bill semantic verification data; perform intelligent error correction on the bill candidate recognition result set based on the bill semantic verification data to obtain the final bill recognition output data.

[0017] The present invention uses the multispectral image acquisition and cross-spectral registration capabilities of the image acquisition module to enable the system to capture more bill details and enhance image processing capabilities. The topological feature extraction module deeply mines bill features and provides rich geometric and topological feature information, while the information entropy coding module uses information entropy analysis and adaptive coding technology to achieve accurate coding of bill feature point sets, enhancing coding accuracy and reliability. The semantic evaluation module optimizes at the semantic level, and improves the semantic consistency and confidence of the recognition results through semantic relevance evaluation and correction. The intelligent error correction mechanism of the knowledge matching module, combined with prior knowledge graph matching and semantic reasoning, further improves recognition accuracy.

[0018] Preferably, a computer-readable storage medium stores a data processing method for bill identification as described above that can be loaded and executed by a processor. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Other features, objects and advantages of the present invention will become more apparent from reading the detailed description with reference to the following drawings:

[0020] Figure 1 A schematic flow chart of steps of a data processing method for bill identification according to an embodiment is shown.

[0021] Figure 2 A detailed flow chart of step S26 of an embodiment is shown.

[0022] Figure 3 A detailed flow chart of step S4 of an embodiment is shown. DETAILED DESCRIPTION

[0023] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are 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 technicians in this field without creative work are within the scope of protection of the present invention.

[0024] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0025] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0026] To achieve this, please refer to Figures 1 to 3 The present invention provides a data processing method for bill identification, comprising the following steps:

[0027] Step S1: collect multispectral images of bills to obtain a multispectral image set of bills; perform cross-spectral registration on the multispectral image set of bills to obtain an aligned bill image set; perform weighted feature fusion on the aligned bill image set to obtain an enhanced bill feature map;

[0028] Step S2: extracting topological transformation features of the enhanced bill feature map to obtain bill geometric feature mapping data; performing local semantic decoupling of the enhanced bill feature map based on the bill geometric feature mapping data to obtain a bill topological feature point set;

[0029] Step S3: performing information entropy analysis on the bill topological feature point set to obtain a bill information entropy gradient distribution map; constructing an adaptive coding mapping table based on the bill information entropy gradient distribution map to obtain a bill dynamic coding sequence;

[0030] Step S4: performing semantic relevance evaluation on the bill dynamic coding sequence to obtain semantic consistency reconstruction mapping data; performing semantic correction on the bill dynamic coding sequence according to the semantic consistency reconstruction mapping data to obtain high-confidence bill feature representation data;

[0031] Step S5: Perform prior knowledge graph matching on high-confidence bill feature representation data to obtain a bill candidate recognition result set; perform semantic reasoning on the bill candidate recognition result set to obtain bill semantic verification data; perform intelligent error correction on the bill candidate recognition result set based on the bill semantic verification data to obtain the final bill recognition output data.

[0032] In this embodiment, first, a camera equipped with multiple filters is used to collect multispectral images of bills, and the camera can capture images in visible light, near infrared and infrared bands. The obtained bill multispectral image set includes image data of three bands. Using the image registration function in the OpenCV library, these image sets are cross-spectral registered to eliminate the geometric distortion caused by spectral differences, and the aligned bill image set is obtained. Using the fusion algorithm in the scikit-image library of Python, the weighted feature fusion of the aligned bill image set is performed according to the feature intensity of each band to obtain an enhanced bill feature map. Topological transformation is applied to the enhanced bill feature map, and morphological operations are performed using the imtophat function in the image processing toolbox of MATLAB to obtain bill geometric feature mapping data. Based on these geometric feature mapping data, the local semantic decoupling technology is further used to obtain the bill topological feature point set by setting semantic thresholds and clustering algorithms. The information entropy analysis of the bill topological feature point set is performed, and the information entropy of each feature point is calculated using the entropy library of Python to obtain the bill information entropy gradient distribution map. According to this distribution map, an adaptive coding mapping table is constructed to generate a bill dynamic coding sequence. The semantic relevance of the bill dynamic coding sequence is evaluated, and the association rule mining technology in the machine learning library scikit-learn is used to evaluate the relevance between the features in the sequence to obtain semantic consistency reconstruction mapping data. Based on this data, the bill dynamic coding sequence is semantically corrected, and the feature vectors are classified and optimized using the support vector machine (SVM) to obtain high-confidence bill feature representation data. Finally, the high-confidence bill feature representation data is matched with the prior knowledge graph, and the feature data is matched with the known bill ontology using the Protégé ontology reasoning tool to obtain the bill candidate recognition result set. Semantic reasoning is performed on these candidate result sets, and logical reasoning is performed using HermiT OWL Reasoner to obtain bill semantic verification data. Based on these verification data, intelligent error correction algorithms, such as Bloom filters, are used to correct the bill candidate recognition result set, and accurate bill recognition output data is finally obtained.

[0033] Preferably, step S1 comprises the following steps:

[0034] Step S11: photographing the original image of the bill to obtain the original image of the bill, and collecting the shooting device parameters of the original image of the bill to obtain a device image collection parameter set;

[0035] Specifically, a smart device with a high-resolution camera, such as a smartphone equipped with a 20-megapixel camera, can be used to take a photo of an A4-sized invoice. During the shooting process, make sure that the invoice is laid flat on a surface with even lighting to ensure that the image is clear and without shadows. After the shooting is completed, the device image acquisition parameter set is obtained from the camera of the smart device, which includes parameters such as ISO setting (e.g., ISO 200), exposure time (e.g., 1 / 500 second), focal length (e.g., 4 mm), aperture size (e.g., f / 2.0), etc. These parameters will be recorded in the device image acquisition parameter set.

[0036] Step S12: Select the spectral range of the spectral imaging of the bill according to the device image acquisition parameter set to obtain the device shooting spectral range data;

[0037] Specifically, a spectral range can be selected based on the ISO and exposure time parameters recorded in the device image acquisition parameter set, which can maximize the contrast and clarity of the image while reducing noise. Select a spectral range from 450 nanometers (blue) to 900 nanometers (near infrared). Within this range, several specific bands are set, such as 450-490 nanometers (blue), 500-550 nanometers (green), 570-590 nanometers (yellow), 600-650 nanometers (red) and 780-900 nanometers (near infrared). The selection of these bands will be based on the capabilities of the device and the specific features present on the bill, such as fluorescence or the absorption characteristics of specific inks. After the selection is completed, these parameters are input into the spectral imaging device to obtain the spectral range data captured by the device.

[0038] Step S13: performing multispectral image acquisition on the bill according to the spectral range data captured by the device to obtain a multispectral image set of the bill; wherein the multispectral image acquisition includes visible light band image acquisition, near infrared band image acquisition and infrared band image acquisition;

[0039] Specifically, a high-resolution camera equipped with an RGB filter can be used to shoot the bill at a resolution of 5 million pixels according to the spectral range data captured by the device. The camera is set to automatic white balance and automatic exposure mode, and the ISO is set to 100 to ensure the true restoration of color. The red, green, and blue channels are shot under daylight conditions (D65 light source) to obtain images in the visible light band. Next, the RGB filter is removed, replaced with a near-infrared filter (e.g., 800-1700 nanometers), and the camera's sensitivity is adjusted to ISO 400, and the exposure time is set to 1 / 100 second. Under the irradiation of a near-infrared light source, the bill is photographed to obtain an image in the near-infrared band. Finally, an infrared camera (capable of capturing infrared bands above 1700 nanometers) is used to shoot the bill under the irradiation of an infrared light source. The camera is set to ISO 800, and the exposure time is set to 1 / 200 second, and a multispectral image set including visible light, near-infrared, and infrared bands is finally obtained.

[0040] Step S14: extracting characteristic associated points from the bill multispectral image set to obtain a bill image characteristic reference point set;

[0041] Specifically, image processing software, such as MATLAB or OpenCV, can be used to detect feature points in each band of images. For example, the SIFT (Scale Invariant Feature Transform) algorithm is used to detect feature points such as corners and edges in the image. In images of different bands, corresponding feature point pairs are found. For example, for a corner point detected in the red channel image of the visible light band, the corresponding corner point is found in the near infrared band image and their coordinates are recorded. The matched feature point pairs are verified. This can be done by calculating the distance between the feature point descriptors and setting a threshold. Only when the distance between the descriptors is less than this threshold, the match is considered valid. After verification, the coordinates of all valid feature point pairs are recorded in a bill image feature reference point set. For example, 100 feature points are detected in the visible light band image, 80 feature points are detected in the near infrared band image, and 60 feature points are detected in the infrared band image. Through matching and verification, a reference point set containing 50 feature points is finally obtained, and these points have corresponding matching points in all bands.

[0042] Step S15: constructing a registration transformation matrix for the bill multispectral image set according to the bill image feature reference point set to obtain the bill multispectral image registration transformation matrix, and performing geometric transformation registration on the bill multispectral image set based on the bill multispectral image registration transformation matrix to obtain an aligned bill image set;

[0043] Specifically, the function findHomography() in the computer vision library OpenCV can be used to calculate the registration transformation matrix. Four non-collinear points are selected from the reference point set of the bill image feature, corresponding to the same point in the visible light, near infrared and infrared band image sets. The coordinates of these points are used as input to call the findHomography() function to obtain a 3x3 transformation matrix that can map the image of one band to the image of another band. After obtaining the registration transformation matrix, the warpPerspective() function in OpenCV is used to perform geometric transformation registration on the images of each band. For example, the transformation matrix is ​​applied to the image of the visible light band to align it with the image of the near infrared band. Set the interpolation method to INTER_LINEAR to reduce distortion during the registration process. The same method is also applied to align the image of the infrared band with the image of the visible light band. Finally, a set of aligned multispectral images is obtained, which are consistent in spatial position.

[0044] Step S16: Perform weighted feature fusion on the aligned bill image set to obtain an enhanced bill feature map.

[0045] Specifically, a weight can be assigned to each band of the image in the aligned bill image set, and this weight is based on the band's ability to identify bill features. For example, the near-infrared band is considered to be particularly sensitive to certain anti-counterfeiting features on bills, so it is given a higher weight. A weight of 0.4 is assigned to the visible light band, a weight of 0.4 is assigned to the near-infrared band, and a weight of 0.2 is assigned to the infrared band. Before fusion, the image of each band is preprocessed to enhance features and reduce noise. For example, Gaussian blur is applied to each image, and then a thresholding method is used to highlight important features. Use the addWeighted() function in OpenCV to fuse the images of the three bands. For example, for each pixel point, the contribution of the visible light band is 0.4 times the pixel value, the contribution of the near-infrared band is 0.4 times the pixel value, and the contribution of the infrared band is 0.2 times the pixel value. Add these weighted pixel values ​​to obtain the pixel value of the enhanced bill feature map, and finally obtain the enhanced bill feature map.

[0046] The present invention enhances the comprehensiveness of image acquisition by capturing original images and acquiring device parameters. By optimizing the spectral imaging selection, the optimal spectral range can be selected according to the device parameters, thereby improving the effectiveness of image acquisition. By improving the multi-spectral image acquisition capability, the visible light, near-infrared and infrared bands are covered, and invisible features on the bill can be captured, thereby enhancing the recognition capability. By improving the accuracy of feature extraction, the key feature points on the bill can be identified and utilized, providing an accurate reference for image registration and feature fusion. The accuracy of image registration is improved by constructing a registration transformation matrix and geometric transformation registration. The effect of aligning the image set is enhanced by weighted feature fusion, integrating the information in different spectral images, generating a richer bill feature map, and providing reliable data for feature extraction and recognition.

[0047] Preferably, step S2 comprises the following steps:

[0048] Step S21: performing geometric feature mapping transformation on the enhanced bill feature map to obtain a bill geometric feature mapping space;

[0049] Specifically, geometric feature extraction can be performed on the enhanced bill feature map using image processing software, such as MATLAB or OpenCV. Edge detection algorithms, such as the Canny edge detector, are applied to identify edges in the image. The parameters of the Canny algorithm are set to a high threshold of 80 and a low threshold of 40 to ensure that the detected edges are significant. Image registration techniques, such as homography based on feature points, are used to map the enhanced bill feature map into a unified geometric space. Four points that are obvious and non-collinear in the enhanced bill feature map are selected, and corresponding points are selected in the target geometric space. The findHomography() function in OpenCV is used to calculate the homography matrix, which maps points in the source image to points in the target space. Apply the homography transformation and use the warpPerspective() function in OpenCV to transform the enhanced bill feature map into a new geometric feature mapping space, and finally obtain an aligned and unified bill geometric feature mapping space.

[0050] Step S22: Calculate the characteristic curvature tensor of the enhanced bill feature map based on the bill geometric feature mapping space to obtain the bill local geometric structure curvature field;

[0051] Specifically, the curvature tensor calculation tool in the image processing software can be used to analyze the local geometric structure of the bill. The curvature tensor is a second-order derivative matrix that describes the local curvature of the image and can be obtained by calculating the gradient of the image. First, the gradient of the enhanced bill feature map is calculated. The Sobel operator, which is an edge detection operator, is used to calculate the first-order derivative (gradient) of the image. The kernel size of the Sobel operator is set to 3x3, which is used to calculate the gradient of the image in the x and y directions. After the gradient is obtained, the second-order derivative, that is, the curvature tensor, is calculated. This can be done by applying the Sobel operator to the gradient image again, or using ready-made functions in the image processing software, such as MATLAB's imgradientxy() function, which can directly calculate the gradient and curvature of the image, and finally obtain the curvature field of the local geometric structure of the bill, which is a matrix describing the curvature of each point in the image. For example, if the calculation result of the curvature tensor shows that the curvature is large in a certain area (such as a corner mark on the bill), this indicates that the area has important geometric features, such as edges or corners. Through the above steps, the local geometric structure curvature field of the bill is finally obtained.

[0052] Step S23: performing geometric invariance mapping on the bill feature points in the enhanced bill feature map based on the bill local geometric structure curvature field to obtain the bill geometric feature invariant;

[0053] Specifically, the curvature field data can be analyzed to identify the local geometric structure of the bill feature points. Using MATLAB's image processing toolbox, the principal curvatures and principal directions of the curvature tensor are calculated for each feature point. The principal curvatures are selected as geometric invariant features because they remain unchanged under rotation and translation transformations. For each feature point, the maximum and minimum principal curvatures are calculated and these values ​​are recorded. These principal curvature values ​​are used to perform geometric invariance mapping on the feature points. For example, if the maximum principal curvature of a feature point is 0.05 and the minimum principal curvature is 0.02, these values ​​are mapped into a feature space where each point in this space corresponds to a unique curvature pair. In this way, a geometric feature invariant is constructed for each feature point in the enhanced bill feature map, which describes the local geometric structure of the point and is invariant to the geometric transformation of the image.

[0054] Step S24: performing topological invariance feature encoding on the bill geometric feature invariant to obtain bill geometric feature mapping data;

[0055] Specifically, topological invariant features, such as shape descriptors and angle information, can be selected, which remain unchanged under topological transformations of the image. For example, a shape context descriptor is used to describe the structure around each feature point. Use a coding scheme, such as scale-invariant feature transform (SIFT) or speeded up robust features (SURF), to encode these topological features. For example, the SIFT algorithm is used to extract key points and a descriptor vector is generated for each key point. For each feature point, the bill geometric feature invariant (such as the principal curvature) and the topological feature descriptor (such as the SIFT descriptor) are combined to form a comprehensive feature vector. This vector remains unchanged under topological transformations and can uniquely identify the feature points on the bill. These feature vectors are stored in a database, each vector is associated with the corresponding feature point, and finally a bill geometric feature mapping dataset is obtained.

[0056] Step S25: semantically clustering the bill geometric feature mapping data to obtain bill semantic cluster mapping data;

[0057] Specifically, feature vectors can be extracted from the bill geometric feature map dataset. This dataset contains geometric and topological feature descriptors for each feature point, for example, each feature vector contains principal curvature values, SIFT descriptors, etc. Select the K-means clustering algorithm. The K-means algorithm requires a preset number of clusters (K value), which is set according to the expected number of feature types on the bill. For example, if there are 5 different types of feature points (such as edges, corners, text areas, etc.) on the expected bill, set the K value to 5. Before clustering, the feature vectors are normalized to ensure that each feature has the same weight during the clustering process. Using the Min-Max normalization method, each feature is scaled to the range of [0,1]. Clustering is performed using the KMeans class in Python's machine learning library scikit-learn. Take the normalized feature vector as input, set n_clusters=5 (that is, the K value determined previously), and select an initialization algorithm (such as k-means++) to optimize the initial centroid selection for clustering. After clustering is completed, evaluate the quality of the clustering results. The Silhouette Coefficient is used to evaluate the clustering effect of each point to ensure that the clustering result is meaningful. The value of the Silhouette Coefficient ranges from -1 to 1, and the larger the value, the better the clustering effect. Finally, each feature point is assigned to its corresponding cluster, and the bill semantic clustering mapping data is generated. This data includes the centroid coordinates of each cluster and the cluster label of each point.

[0058] Step S26: Perform local semantic decomposition on the enhanced bill feature map according to the bill semantic clustering mapping data to obtain a bill semantic subspace, and reconstruct topological feature points on the bill semantic subspace to obtain a bill topological feature point set.

[0059] Specifically, please refer to the sub-steps of step S26 for the detailed implementation process of this embodiment.

[0060] The present invention enhances the ability to recognize geometric features of bills through geometric feature mapping transformation and characteristic curvature tensor calculation, and deeply understands the local geometric structure of bills. The recognition stability and accuracy of the system under different conditions are ensured through geometric invariance mapping and topological invariance feature encoding, while generating highly robust bill geometric feature mapping data. The system's ability to recognize and classify semantic information on bills is further improved through semantic clustering and local semantic decomposition, and the bill feature map is decomposed into subspaces with clear semantics, which facilitates analysis. By reconstructing the topological feature point set based on the semantic subspace, the key feature points on the bill are accurately located.

[0061] Preferably, step S26 comprises the following steps:

[0062] Step S261: Perform semantic association evaluation on the bill semantic clustering mapping data to obtain the bill semantic association weight matrix; specifically, key features can be extracted from the bill semantic clustering mapping data, such as the density, distribution range and shape features of the feature points in each cluster. These features will be used to evaluate the semantic association between different clusters. Use similarity measurement methods in the machine learning library scikit-learn, such as cosine similarity, to evaluate the semantic association between different clusters. Cosine similarity evaluates the similarity of two vectors by measuring the angle between them. For each pair of clusters, the cosine similarity between them is calculated and stored as a weight in the bill semantic association weight matrix. For example, if there are 5 clusters, a 5x5 matrix will be constructed, in which each element (i, j) represents the similarity weight between the i-th cluster and the j-th cluster. L2 normalization is selected because it can handle different scales of vectors. Finally, the semantic association weight matrix is ​​verified to ensure that the weight distribution is reasonable and can reflect the actual semantic association between different clusters. For example, the association weight between the text area and the digital area is expected to be high because they usually have similar semantic meanings in the bill.

[0063] Step S262: semantically deconstructing the enhanced bill feature graph based on the bill semantic association weight matrix to obtain bill multi-layer semantic deconstruction mapping data;

[0064] Specifically, community discovery algorithms in graph theory, such as the Louvain algorithm, can be used to hierarchically deconstruct the semantic association weight matrix of the bill. The Louvain algorithm can identify community structures, that is, tightly connected groups of nodes in the graph. When applying the Louvain algorithm, modularity optimization is selected as the quality indicator of community division. The higher the modularity value, the tighter the connection within the community and the sparser the connection with the outside. The semantic association weight matrix is ​​used as input and the Louvain algorithm is run to identify different semantic layers in the bill. The algorithm will output the community to which each node (that is, each cluster) belongs, and these communities represent different semantic layers in the bill. Based on the results of community discovery, multi-layer semantic deconstruction mapping data of the bill is constructed. These data include the semantic layer to which each cluster belongs, the feature description of the clusters within the layer, and the relationship between layers. Verify the rationality of the deconstruction results, and adjust the algorithm parameters, such as the resolution of community division, as needed to ensure that the deconstruction results can accurately reflect the semantic structure of the bill. Through the above steps, the multi-layer semantic deconstruction mapping data of the bill is finally obtained, which reveals the structure and relationship of different semantic layers in the bill.

[0065] Step S263: Perform semantic-geometric collaborative decoupling on the enhanced bill feature map according to the bill multi-layer semantic deconstruction mapping data and the bill local geometric structure curvature field to obtain bill semantic feature mapping data;

[0066] Specifically, the multi-layer semantic deconstruction mapping data of the bill can be fused with the local geometric structure curvature field. Using Python's NumPy library, the two data sets are aligned according to the coordinates of the feature points. Principal component analysis (PCA) is selected as a dimensionality reduction technique to achieve the collaborative decoupling of semantic and geometric information. PCA can reduce the dimension of the data while retaining the most important features. The PCA class in the scikit-learn library is used to perform dimensionality reduction on the fused data. The n_components parameter of PCA is set to 2, which means reducing the data to a two-dimensional space. After dimensionality reduction, the two-dimensional coordinates of each feature point are obtained, which contain both semantic and geometric information. These coordinates are analyzed to determine whether they can clearly distinguish different features on the bill, such as text areas, digital areas, etc. Finally, the PCA results are used as bill semantic feature mapping data, which contain the semantic and geometric information of the original enhanced bill feature map.

[0067] Step S264: performing a topological semantic convergence transformation on the bill multi-layer semantic deconstruction mapping data based on the bill semantic feature mapping data to obtain a bill semantic subspace;

[0068] Specifically, the U-Net network can be used for semantic segmentation and feature extraction of images. U-Net can capture the contextual information of the image and perform effective feature fusion. Use deep learning frameworks such as TensorFlow or PyTorch to train the U-Net model. Take the bill semantic feature mapping data as input and the labeled bill image as the label for supervised learning. Set the model's loss function to cross entropy loss, and use the Adam optimizer for parameter optimization. After training, use the U-Net model to process the bill semantic feature mapping data to obtain the topological semantic convergence transformation result of the bill multi-layer semantic deconstruction mapping data. The output of the U-Net model is the semantic subspace of the bill, which is a high-dimensional space in which each dimension represents a specific semantic feature in the bill. Through thresholding, this high-dimensional space is converted into a binary image in which the value of each pixel represents the probability that the point belongs to a specific semantic category.

[0069] Step S265: Reconstruct feature points of the bill semantic subspace according to the bill geometric feature invariant to obtain a bill topological feature point set.

[0070] Specifically, the invariants of the geometric features of the bill, such as the principal curvature values, can be obtained from the previous steps. These invariants are extracted from the curvature field of the local geometric structure of the bill and remain unchanged under geometric transformations. For example, the principal curvature of each key point on the bill has been calculated and these values ​​are stored in a database. The Iterative Closest Point (ICP) algorithm is selected to reconstruct the feature points. The ICP algorithm is a point cloud matching technology that can accurately reconstruct feature points while keeping the geometric features unchanged. The geometric feature invariants of the bill are matched with the feature points in the semantic subspace. For example, there is a feature point represented as a high curvature area in the semantic subspace, and points with similar principal curvature values ​​will be found and matched. Using the ICP implementation in the OpenCV library, the matched feature points are taken as input, the number of iterations of the algorithm is set to 10, and the distance threshold is set to 0.01 (determined according to the resolution of the bill image and the density of the feature points). The algorithm will iteratively adjust the position of the feature points until the best match is found. Through the ICP algorithm, we get a set of reconstructed feature points that geometrically match the original bill and have corresponding semantic information in the semantic subspace. For example, if a feature point geometrically corresponds to the edge of the bill and is identified as part of the text area in the semantic subspace, then this point is an important member of the bill topological feature point set. Finally, verify the accuracy and completeness of the reconstructed feature point set. Use visualization tools to check whether the reconstructed feature points match the actual structure of the bill. Through the above steps, we finally get a set of bill topological feature points that combine the geometric and semantic information of the bill.

[0071] The present invention enhances the ability to understand the semantic content of bills and improves the semantic parsing accuracy of feature maps through semantic association evaluation and semantic hierarchical deconstruction. By combining multi-layer semantic deconstruction mapping data and local geometric structure curvature field, semantic-geometric collaborative decoupling is achieved, thereby obtaining more accurate bill semantic feature mapping data. The construction of semantic subspace is optimized through topological semantic convergence transformation, so that the system can accurately identify and reconstruct the topological feature point set of the bill. The reliability of recognition results is improved through semantic association evaluation and feature point reconstruction, and the possibility of misrecognition is reduced.

[0072] Preferably, step S3 comprises the following steps:

[0073] Step S31: topologically segmenting the enhanced bill feature graph according to the bill topological feature point set to obtain bill feature block mapping data;

[0074] Specifically, you can choose to use the watershed algorithm in the image processing library OpenCV for topological segmentation. The watershed algorithm is a segmentation method based on image gradients. It simulates the process of water flowing into different areas and is suitable for segmentation based on feature points. The enhanced bill feature map is input into the watershed algorithm, and the bill topological feature point set is used as markers. The algorithm will segment the image into different areas based on these markers. For example, 10 key points are marked on the bill, and the watershed algorithm will segment the bill into more than 10 areas based on these points. After the segmentation is completed, the bill feature block mapping data is obtained, which describes the boundaries and internal features of each segmented area. Each area is assigned a unique identifier and is associated with the corresponding topological feature point. Check the segmentation results through visualization tools to ensure that each area is consistent with the actual structure of the bill and that the key feature points are correctly contained in the corresponding area.

[0075] Step S32: performing information entropy estimation on the bill feature block mapping data based on the enhanced bill feature map to obtain a bill local entropy value matrix;

[0076] Specifically, the local entropy of the image can be used to estimate the information entropy of each region. Local entropy can describe the texture complexity of the image region and is very useful for capturing the variability within the region. For each segmented region, its local entropy value is calculated. The Gray-Level Co-occurrence Matrix (GLCM) is used to calculate the local entropy. The GLCM describes the statistical characteristics of pixel pairs in the image, from which the entropy value is calculated. For example, GLCM in 13 different directions is calculated for each region, and the average value is taken as the local entropy value of the region. An entropy value is assigned to each segmented region, and these values ​​are stored in a matrix, namely the bill local entropy value matrix. The rows of this matrix represent different regions, and the columns represent different entropy value calculation parameters (such as entropy values ​​in different directions). Through the above steps, the bill local entropy value matrix is ​​finally obtained.

[0077] Step S33: performing statistics on the bill local entropy value matrix to obtain a bill information entropy distribution histogram, and performing gradient mapping according to the bill information entropy distribution histogram to obtain a bill information entropy gradient distribution map;

[0078] Specifically, Python's NumPy library can be used to perform statistical analysis on the local entropy matrix of the bill. Calculate the frequency of each entropy value and generate an information entropy distribution histogram. For example, it is found that there are 30% of the areas in the entropy range of 0 to 0.5, and 70% of the areas in the range of 0.5 to 1.0. Use the Matplotlib library to visualize this histogram, where the x-axis represents the range of entropy values ​​and the y-axis represents the percentage of the area within each entropy value range. Use gradient mapping technology. Select a gradient mapping method based on the peak of the histogram, for example, map the highest peak of the histogram to red, the second highest peak to yellow, and the lowest peak to blue. Use image processing software such as GIMP or Photoshop to color the histogram according to the gradient mapping rules to generate a bill information entropy gradient distribution map. This distribution map intuitively shows the size of the information entropy of different areas on the bill in the form of color changes.

[0079] Step S34: extracting feature points from the bill information entropy gradient distribution map according to a preset information entropy threshold to obtain a set of bill key entropy change points;

[0080] Specifically, an information entropy threshold can be set according to the information entropy gradient distribution map of the bill and business requirements. For example, 0.6 is selected as the threshold, which means that areas with entropy values ​​greater than 0.6 will be focused on because these areas contain more complex information. The threshold function threshold in the OpenCV library is used to process the information entropy gradient distribution map of the bill. The threshold is set to 0.6, and a suitable threshold type is selected, such as THRESH_BINARY, so that areas above the threshold are marked as white (value 255) and areas below the threshold are marked as black (value 0). After processing, a binary image is obtained, in which white areas represent high information entropy areas. The findContours function is further used to identify the boundaries of these white areas and extract the key entropy change point set. The extracted key entropy change point set is verified to ensure that they correspond to key feature points on the bill, such as the edges of the text area, the corners of the digital area, etc. Finally, the extracted key entropy change point set is output as a list containing the coordinates of each key point and the corresponding entropy value.

[0081] Step S35: generating adaptive coding rules based on the bill key entropy change point set to obtain bill coding mapping rules;

[0082] Specifically, please refer to the sub-steps of step S35 for the detailed implementation process of this embodiment.

[0083] Step S36: extracting feature representations of the bill topological feature point set based on the enhanced bill feature graph to obtain a bill feature representation data set;

[0084] Step S37: Convert the bill feature representation data set into code according to the bill code mapping rule to obtain a bill dynamic code sequence.

[0085] Specifically, the bill coding mapping rules can be prepared. These rules define how to map the features in the bill feature characterization dataset to the coding sequence. The bill feature characterization dataset contains information such as multispectral image acquisition, geometric features, and semantic features of the bill, which are used to generate the dynamic coding sequence of the bill. The Huffman coding algorithm is selected for coding conversion. Huffman coding is a data compression algorithm that constructs the optimal prefix code according to the frequency of feature occurrence and is suitable for generating compact coding sequences. The heapq library of Python is used to build the Huffman tree. According to the feature frequency in the bill coding mapping rule, a priority queue is constructed, and the Huffman tree is gradually constructed until all features are included in the tree. Once the Huffman tree is constructed, the tree is traversed to assign a unique code to each feature in the bill feature characterization dataset. The length of the code is set to be dynamically adjusted according to the importance and frequency of the feature, and frequently occurring features will be assigned shorter codes. Through Huffman coding, the bill dynamic coding sequence is finally obtained, which is a compact binary sequence that contains all the feature information of the bill, and each feature has a unique code. Verify the accuracy and effectiveness of the bill dynamic coding sequence. Check whether the coding sequence can accurately represent the original bill feature representation data set and whether it can effectively compress the data. Finally, output the generated bill dynamic coding sequence as a file or database record. Through the above steps, the bill dynamic coding sequence is finally obtained.

[0086] The present invention can obtain fine bill feature block mapping data through topological segmentation. Information entropy estimation provides quantitative analysis for the complexity of bill feature block, and introduces a new measurement method. By statistically analyzing the local entropy value matrix and generating the information entropy distribution histogram, the distribution of the global features of the bill can be mastered. The key entropy change point set of the bill is effectively identified through the application of gradient mapping and preset thresholds. Based on these point sets, adaptive coding rules are generated, which enhances the flexibility and adaptability of the coding process, making the coding results more consistent with the actual characteristics of the bill. Comprehensive feature representation extraction provides rich information for coding conversion, and the generation of dynamic coding sequences provides a new and dynamic feature representation method for bill identification and verification.

[0087] Preferably, step S35 includes the following steps:

[0088] Step S351: perform information entropy complexity evaluation on the bill key entropy change point set to obtain bill information complexity spectrum mapping data;

[0089] Specifically, the bill key entropy change point set contains key feature points with higher information entropy values ​​on the bill. The mutual information method in Python's machine learning library scikit-learn can be used to evaluate the information entropy complexity of these key point sets. Mutual information is a measure of the amount of shared information between two variables. For each pair of points in the key entropy change point set, calculate the mutual information value between them. Set the calculation parameters of the mutual information, for example, select a suitable distance threshold (such as 5 pixels) and angle threshold (such as 45 degrees) to determine which point pairs will be used for the calculation of mutual information. Generate bill information complexity spectrum mapping data based on the calculated mutual information value. This spectrum mapping data can be a matrix in which each row represents a key point and each column represents the mutual information value of the point with all other points. Verify the information complexity spectrum mapping data to ensure that they can accurately reflect the information entropy complexity between key points on the bill.

[0090] Step S352: performing nonlinear information association dimensionality reduction on the information features in the enhanced bill feature graph according to the bill information complexity spectrum mapping data and the bill key entropy change point set to obtain bill information association dimensionality reduction mapping data;

[0091] Specifically, you can choose to use the t-SNE (t-distributed Stochastic Neighbor Embedding) algorithm for nonlinear information association dimensionality reduction. Use the t-SNE implementation in Python's scikit-learn library and take the bill information complexity spectrum mapping data as input. Set the perplexity parameter of t-SNE to 30 (indicating the number of neighboring points), the learning_rate parameter to 200, and the number of iterations to 1000. The t-SNE algorithm reduces the high-dimensional information complexity spectrum mapping data to a two-dimensional or three-dimensional space to obtain bill information association dimensionality reduction mapping data. These data can be used to visualize the information correlation between key points on the bill. Check the dimensionality reduction mapping results through visualization tools to ensure that the information correlation between key points is effectively retained. Through the above steps, the bill information association dimensionality reduction mapping data is finally obtained.

[0092] Step S353: Decomposing the information feature dimensional space of the bill key entropy change point set based on the bill information associated dimensionality reduction mapping data to obtain the bill information feature deconstruction subspace;

[0093] Specifically, the bill information association dimensionality reduction mapping data can be analyzed, which reduces the high-dimensional information association to two-dimensional or three-dimensional space through the t-SNE algorithm. Choose to use principal component analysis (PCA) for dimensional space decomposition. PCA is a statistical method that can transform data into a new coordinate system through an orthogonal transformation so that the first largest variance of any projection of the data is the largest in the first aspect of the new coordinate, the second largest variance is the largest in the second aspect, and so on. Use the PCA class in Python's scikit-learn library to perform dimensional space decomposition on the dimensionality reduction mapping data. Set the n_components parameter of PCA to 2, which means decomposing the data into a two-dimensional space. After PCA processing, the bill information feature deconstruction subspace is obtained, which is a two-dimensional space containing two main components, and each dimension represents the direction of the largest variance in the data. Check the results of PCA decomposition through visualization tools to ensure that the key entropy change point set is reasonably represented in the new subspace and the key features are retained.

[0094] Step S354: constructing information coding basis functions according to the deconstruction subspace of the bill information feature to obtain the bill coding basis functions;

[0095] Specifically, the feature deconstruction subspace of bill information can be analyzed to determine the main distribution area of ​​key features. The radial basis function (RBF) is selected as the encoding basis function because RBF can map the input data to a high-dimensional space through nonlinear mapping. The RBF class in Python's scikit-learn library is used to construct a radial basis function network. The n_components parameter of RBF is set to 10, which means that 10 basis functions will be constructed to cover the main features of the data. The data of the feature deconstruction subspace of bill information is used as input to train the RBF network. The length_scale parameter of RBF is set to 1.0, which controls the width of the basis function, and the optimization mode of the length_scale parameter is set to optimized, so that the algorithm automatically adjusts this parameter to obtain the best fit. After the training is completed, the bill encoding basis functions are obtained, which can map the key entropy change point set of the bill to a new feature space, where each basis function corresponds to a specific feature or pattern in the data.

[0096] Step S355: performing information feature mapping transformation on the bill key entropy change point set through the bill coding basis function to obtain bill feature mapping transformation data;

[0097] Specifically, support vector machines (SVM) can be used for mapping transformation. SVM performs well in processing nonlinear data and performing effective classification. The SVM class in Python's scikit-learn library is used to map the key entropy change point set of the bill. The encoding basis function is used as the kernel function of the SVM, and the kernel function type is set to RBF, where the gamma parameter is set to 0.1 and the C parameter is set to 1.0. These parameters control the decision boundary of the SVM and the penalty strength of the error term. Through the SVM mapping transformation, the bill feature mapping transformation data is obtained. These data map the original key entropy change point set to a new high-dimensional space, where each point is represented by the response value of the encoding basis function.

[0098] Step S356: constructing the code transformation domain of the bill key entropy change point set based on the bill feature mapping transformation data to obtain the bill code transformation domain mapping data;

[0099] Specifically, the bill feature map transformation data can be analyzed first, which represents the position of the key entropy change point set in the new mapping space. The autoencoder neural network is selected to construct the encoding transformation domain because the autoencoder can learn the effective representation of the data and capture the intrinsic structure of the data. Use a deep learning framework such as TensorFlow or PyTorch to build an autoencoder model. The autoencoder consists of an encoder and a decoder. The encoder compresses the input data into a low-dimensional representation, while the decoder restores the representation to the original data. Take the bill feature map transformation data as input and train the autoencoder model. Set the hidden layer size of the autoencoder to 128, use the ReLU activation function, and use the mean square error (MSE) as the loss function. After the training is completed, the encoder part of the autoencoder is used to process the bill feature map transformation data to obtain the bill encoding transformation domain mapping data. These data represent the low-dimensional representation of the key entropy change point set in the encoding transformation domain.

[0100] Step S357: Generate encoding rules based on the bill feature mapping transformation matrix and the bill coding transformation domain mapping data to obtain the bill coding mapping rules.

[0101] Specifically, the bill feature mapping transformation matrix and the bill encoding transformation domain mapping data can be prepared. The bill feature mapping transformation matrix contains the feature vectors of the key entropy change point set after SVM mapping transformation, while the bill encoding transformation domain mapping data contains the low-dimensional representation of the autoencoder model. The decision tree algorithm is selected to generate encoding rules, and the DecisionTreeClassifier class in Python's scikit-learn library is used to build a decision tree model. The max_depth parameter of the decision tree is set to 5 to control the depth of the tree and prevent overfitting, and the min_samples_split parameter is set to 10 to determine the minimum number of samples required to split internal nodes. The bill feature mapping transformation matrix is ​​used as the feature input, and the bill encoding transformation domain mapping data is used as the label to train the decision tree model. The model will learn how to predict the category or attribute in the encoding transformation domain mapping data based on the information in the feature mapping transformation matrix. After training, the decision tree model will generate a series of decision rules that define how to map the data in the feature mapping transformation matrix to the encoding transformation domain mapping data. These rules are extracted to form the bill encoding mapping rules. The performance of the decision tree model is verified by cross-validation to ensure the accuracy and generalization ability of the encoding mapping rules. Finally, the generated bill coding mapping rules are output as a rule set, which can be used to guide the automatic encoding process of the bill key entropy change point set and to realize the rapid retrieval and matching of features in the bill recognition system.

[0102] The present invention can deeply understand the complexity of bill information through information entropy complexity evaluation and obtain accurate information complexity spectrum mapping data. By using these data to perform nonlinear information association dimensionality reduction, key features can be effectively extracted from high-dimensional data, simplifying the data processing process while retaining key information. By performing dimensional space decomposition on the key entropy change point set, a bill information feature deconstruction subspace is constructed to more accurately locate and analyze the key features of the bill. By constructing an information coding basis function, a standardized coding method is provided for bill features, thereby enhancing the consistency and reliability of coding. By constructing a coding transform domain, a unified coding framework is provided for the bill key entropy change point set, simplifying the coding process and improving coding efficiency. The flexibility and adaptability of coding are improved by generating adaptive coding rules, thereby improving the accuracy and robustness of the system in recognizing bill features, especially in the face of complex or noisy interference.

[0103] Preferably, step S4 comprises the following steps:

[0104] Step S41: extracting semantic association features from the bill dynamic code sequence to obtain a bill semantic association feature vector;

[0105] Specifically, you can choose to use the Bag of Words (BoW) model to extract semantic association features from the dynamic coding sequence of the bill. The BoW model can convert the coding sequence into a set of vectors of fixed length, each of which represents the distribution of a semantic concept. Extract all semantic concepts from the dynamic coding sequence of the bill and build a vocabulary. For example, if the bill contains text, numbers, and symbols, use these elements as entries in the vocabulary. For the dynamic coding sequence of each bill, calculate the frequency of occurrence of each entry in the vocabulary and generate a feature vector. The length of this vector is equal to the size of the vocabulary, and the value of each element represents the frequency of the corresponding semantic concept. Through the above steps, the semantic association feature vector of the bill is finally obtained.

[0106] Step S42: performing semantic consistency entropy value evaluation on the enhanced bill feature graph according to the bill semantic association feature vector to obtain bill semantic entropy value mapping data;

[0107] Specifically, conditional entropy can be used to evaluate semantic consistency because conditional entropy can measure the uncertainty of another random variable under the condition of a given random variable. For each pair of semantic concepts in the semantic association feature vector of the bill, the conditional entropy between them is calculated. The mutual_info_classif function in the scikit-learn library of Python is used to calculate the conditional entropy, and discrete_features = True is set to process discrete features. For example, if there are two semantic concepts "date" and "amount", the conditional entropy between the two concepts will be calculated to evaluate their co-occurrence consistency in all bills. According to the calculation results of conditional entropy, the bill semantic entropy value mapping data is generated. This dataset contains the entropy value between each pair of semantic concepts, reflecting the semantic consistency between them. For example, if the conditional entropy between "date" and "amount" is low, it indicates that the two concepts have a strong co-occurrence relationship in the bill. The semantic entropy value mapping data is analyzed to verify the accuracy of the entropy value evaluation. It is expected that the entropy value between related semantic concepts is low, while the entropy value between unrelated semantic concepts is high. For example, if the conditional entropy between "payer" and "payee" is high, it indicates that the information of these two concepts in some bills is inconsistent or incomplete. Finally, the bill semantic entropy mapping data is applied to the enhanced bill feature map to highlight those areas with high or low semantic consistency. For example, different colors or marks are used on the enhanced bill feature map to represent semantic areas with different entropy values, thereby providing an intuitive visual reference for further analysis and recognition. Through the above steps, the bill semantic entropy mapping data is finally obtained.

[0108] Step S43: Based on the bill semantic entropy value mapping data, the bill dynamic coding sequence is semantically deconstructed to obtain bill semantic association deconstruction mapping data;

[0109] Specifically, a clustering algorithm can be used to deconstruct the semantic association of the dynamic coding sequence of the bill. Select the DBSCAN algorithm, which does not require the number of clusters to be specified in advance and can handle areas of different densities. Use the DBSCAN class in Python's scikit-learn library to perform clustering. Set the eps parameter to 0.5 and the min_samples parameter to 10. These parameters control the neighborhood size and the minimum number of samples for each cluster in the clustering process. The DBSCAN algorithm will divide the points in the dynamic coding sequence of the bill into different clusters based on the similarity in the semantic entropy mapping data. Each cluster represents a group of coding sequences with strong semantic associations, records the cluster label of each point, and generates the bill semantic association deconstruction mapping data.

[0110] Step S44: reconstructing the semantic consistency features of the bill dynamic coding sequence according to the bill semantic association deconstruction mapping data to obtain semantic consistency reconstruction mapping data;

[0111] Specifically, non-negative matrix factorization (NMF) can be used to reconstruct semantically consistent features. NMF is a data dimensionality reduction technique that decomposes a non-negative matrix into the product of two non-negative matrices and is suitable for processing image and text data. The decomposition is performed using the NMF class in Python's scikit-learn library. The n_components parameter is set to 5, which means that the bill dynamic coding sequence is decomposed into a linear combination of 5 basic features. After NMF decomposition, the reconstructed form of the bill dynamic coding sequence is obtained, and each sequence is represented as the weighted sum of 5 basic features. These weight sums reflect the semantic consistency features of the original sequence, which are used as semantic consistency reconstruction mapping data. The effect of NMF decomposition is verified by comparing the original bill dynamic coding sequence and the reconstructed sequence.

[0112] Step S45: Performing a semantic relevance confidence assessment on the bill dynamic coding sequence according to the semantic consistency reconstruction mapping data to obtain a bill semantic confidence assessment index;

[0113] Specifically, the semantic consistency reconstruction mapping data contains the reconstructed form of the bill dynamic coding sequence, and each sequence is represented as the weighted sum of basic features. Support vector machine (SVM) can be used to evaluate the confidence of semantic relevance. Use the SVC class in Python's scikit-learn library to train the SVM model. The semantic consistency reconstruction mapping data is used as the feature input, and the true label of the bill (if any) is used as the target output. Set the kernel parameter to linear and the C parameter to 1.0. These parameters control the decision boundary of the SVM and the penalty strength of the error term. Use the trained SVM model to predict the bill dynamic coding sequence and obtain the confidence score of each prediction. Set the probability parameter to True so that the SVM outputs the confidence of the prediction. Record the confidence score of each bill dynamic coding sequence. These scores reflect the degree of certainty of the model that each sequence belongs to a specific category, which is used as the bill semantic confidence evaluation indicator.

[0114] Step S46: Based on the bill semantic association feature vector and the bill semantic confidence evaluation index, semantic correction is performed on the bill dynamic coding sequence to obtain high-confidence bill feature representation data.

[0115] Specifically, a bill semantic association feature vector and a bill semantic confidence evaluation index can be prepared. Semantic correction is performed using a rule-based approach. This approach allows the bill dynamic coding sequence to be adjusted based on the confidence score and the semantic association feature vector. A series of rules are formulated. For example, if the confidence score of a bill dynamic coding sequence is lower than a certain threshold (e.g., 0.7), its semantic association feature vector will be checked and adjusted according to the rules. For example, if the frequency of a semantic concept in a sequence is abnormally high and the confidence score is low, the weight of the concept is reduced. The bill dynamic coding sequence is corrected according to the formulated rules. For each sequence, the weight in the semantic association feature vector is adjusted to improve the confidence score. The corrected bill dynamic coding sequence will have higher semantic consistency and confidence, and it will be used as high-confidence bill feature representation data. The effect of the semantic correction is verified by comparing the confidence scores before and after correction.

[0116] The present invention enhances the system's ability to capture the semantic information of bills by extracting semantically associated feature vectors from the bill's dynamic coding sequence, thereby improving the accuracy of recognition. By evaluating the semantic consistency entropy value of the enhanced bill feature graph, the system can quantify and evaluate the semantic consistency of the bill features, providing an important reference for subsequent processing. In addition, the system can also deeply analyze the semantic structure of the bill based on the semantic entropy mapping data, and reconstruct the bill's dynamic coding sequence through semantic association deconstruction mapping data, further enhancing the semantic consistency of the bill features. The semantic association confidence assessment provides a quantitative indicator for the reliability of bill features, and the generation of high-confidence bill feature representation data improves the system's recognition ability in complex or ambiguous situations. These measures not only enhance the robustness of the system, making it stable in the face of noise and abnormal data, but also improve the system's generalization ability, optimize the decision-making process, and ultimately enhance the user experience, providing more reliable and satisfactory recognition results.

[0117] Preferably, step S5 comprises the following steps:

[0118] Step S51: constructing a semantic ontology for the bill to obtain a bill semantic ontology knowledge base;

[0119] Specifically, a bill semantic ontology knowledge base can be designed, which contains the key concepts in the bill field and the relationships between them. For example, define concepts such as "date", "amount", "payer", and "payee", and describe the semantic relationships between them. Use Protégé as an ontology construction tool. In Protégé, a new ontology is created, and the above concepts are defined as classes (Class). Define the properties (Property) between these classes, such as "belong to", "pay", etc., and set their domains (Domain) and ranges (Range). Add specific instances (Individual) to the ontology, such as specific bill samples, and specify classes and property values ​​for these instances. Use Protégé's reasoning function to verify the consistency of the ontology and ensure that all defined relationships and rules are reasonable. Finally, save the constructed bill semantic ontology knowledge base as an OWL file.

[0120] Step S52: normalizing the high-confidence bill feature representation data to obtain standard bill feature representation data, and annotating the standard bill feature representation data with semantic graph mapping based on the bill semantic ontology knowledge base to obtain a bill candidate recognition result set;

[0121] Specifically, the data representing the high-confidence bill features can be standardized. Using the Z-score standardization method, each feature value is converted to a distribution with a mean of 0 and a standard deviation of 1. For example, if a feature value is "amount", the mean and standard deviation of all "amount" feature values ​​are calculated and the formula is used: Standardize, where x is the original value, μ is the mean, and σ is the standard deviation. Use the StandardScaler class in Python's scikit-learn library to perform the standardization operation. Use the bill semantic ontology knowledge base to annotate the standardized bill feature representation data with a semantic graph map. Find the corresponding ontology concept for each standardized feature value and create nodes and edges in the semantic graph. Merge the annotated bill feature representation data set into a candidate recognition result set. This set contains all standardized and annotated bill features and their locations in the semantic graph. Check the results of the semantic graph mapping annotation through visualization tools to ensure that each feature is correctly mapped to the corresponding ontology concept. Finally, output the bill candidate recognition result set into a structured data format, such as JSON or XML.

[0122] Step S53: constructing semantic reasoning logic for the bill candidate recognition result set to obtain a bill semantic reasoning logic rule set;

[0123] Specifically, the bill candidate recognition result set obtained in step S52 can be analyzed, and these result sets contain the characteristic data of the bill and the corresponding semantic annotations. Use the OWL ontology language to construct a semantic reasoning logic rule set. For example, define a rule: if the attribute 'amount' exists in the bill, the attributes 'payer' and 'payee' must exist. This rule can be expressed in OWL as: Amount->(Payer and Payee). Use the HermiT reasoning engine, which is a high-performance OWL 2RL reasoning engine that can handle complex ontology reasoning tasks. Load the designed ontology and reasoning logic rules into the HermiT reasoning engine. Use the OWL API to programmatically implement this process, and import the ontology file (.owl) and the rule file (.rdf) into HermiT. Run the HermiT reasoning engine and execute the reasoning logic rule set. The engine will automatically derive all conclusions that meet the rules and generate new ontology facts. After the reasoning is completed, the bill semantic reasoning logic rule set is obtained, which contains all derived ontology facts and logic rules. Verify the correctness of the reasoning results and ensure that the reasoning logic rule set can accurately reflect the semantic structure and relationship of the bill.

[0124] Step S54: performing semantic relevance evaluation on the bill candidate recognition result set based on the bill semantic reasoning logic rule set to obtain bill semantic relevance evaluation data;

[0125] Specifically, the inference logic rule set contains defined semantic rules, and the candidate recognition result set contains the bill feature data to be evaluated. The inference logic rule set can be applied to the candidate recognition result set to identify data patterns that meet these rules. For example, if the rule set contains the rule "If the bill contains the 'amount' attribute, it must contain the 'payer' attribute", find bills that meet this pattern in the candidate recognition result set. Use the Apriori algorithm to evaluate semantic relevance. The Apriori algorithm is a commonly used association rule mining technology that can discover frequent patterns and association rules between data items. Use the apriori function in the mlxtend library of Python to perform association rule mining. Convert the data in the candidate recognition result set into a format suitable for the Apriori algorithm, for example, convert the data into a transaction list form, where each transaction contains a set of items. Set the min_support parameter of the Apriori algorithm to 0.01, indicating that the focus is on the item set that appears in at least 1% of the transactions; set the min_confidence parameter to 0.5, indicating that the focus is on the rules with a confidence of at least 50%. Through the Apriori algorithm, we obtain the bill semantic association evaluation data containing frequent item sets and strong association rules. These data reveal the potential correlation between bill features, for example, which features often appear together, or the appearance of a feature often indicates the existence of another feature. Finally, the semantic association evaluation data is generated. These data include statistics such as frequent item sets, support, confidence, and lift.

[0126] Step S55: performing knowledge graph constraint verification on the bill semantic association evaluation data based on the bill semantic ontology knowledge base to obtain the bill semantic constraint verification result;

[0127] Specifically, the bill semantic association evaluation data can be prepared, which contains frequent item sets and strong association rules. Use OWL Reasoner to perform knowledge graph constraint verification. OWL Reasoner can process OWL ontologies and check whether the data conforms to the constraints defined in the ontology. Load the bill semantic ontology knowledge base and semantic association evaluation data into OWLReasoner. Using the OWL API, import the ontology file (.owl) and the evaluation data file (such as RDF format) into Reasoner. Run OWL Reasoner to perform constraint verification on the semantic association evaluation data. Reasoner will check whether each data item conforms to the class, attribute, and relationship constraints defined in the ontology. Reasoner generates bill semantic constraint verification results, which indicate which data items conform to the ontology constraints and which do not, and point out the specific constraints that are violated.

[0128] Step S56: Perform semantic reasoning path planning on the bill candidate recognition result set according to the bill semantic constraint verification result to obtain the bill semantic reasoning path;

[0129] Specifically, the shortest path algorithm in graph theory can be used to plan the semantic reasoning path. For example, the Dijkstra algorithm is used to find the optimal reasoning path that meets the ontology constraints. The bill candidate recognition result set and the semantic constraint verification results are constructed into a graph model, in which nodes represent data items, edges represent the relationship between data items, and the weight of the edge represents the relationship strength or confidence. Use Python's networkx library to implement the Dijkstra algorithm and perform path planning on the graph model. Set the starting point of the algorithm to a specific data item in the candidate recognition result set and the end point to the reasoning target. The Dijkstra algorithm generates a bill semantic reasoning path, which is the shortest path from the starting point to the end point, which meets the ontology constraints and maximizes the relationship strength or confidence on the path. Verify the validity of the semantic reasoning path, and optimize the path planning parameters as needed, such as adjusting the edge weight calculation method or selecting different path planning algorithms. Through the above steps, the bill semantic reasoning path is finally obtained.

[0130] Step S57: performing semantic reasoning on the bill candidate recognition result set based on the bill semantic reasoning path to obtain bill semantic verification data;

[0131] Specifically, bill semantic reasoning paths can be prepared, which define the logical steps from input data to the reasoning target. Use Rule Engine for semantic reasoning. The rule engine is able to reason and make decisions on data based on a predefined set of rules. A series of reasoning rules are defined in the rule engine, which are based on the bill semantic ontology knowledge base and semantic reasoning paths. For example, the rule includes "if there is a payer and a payee, there is a transaction." The bill candidate identification result set is input into the rule engine, and the defined reasoning rules are executed. The engine will reason on the data according to the rules and paths to generate new semantic verification data. After the reasoning is completed, the bill semantic verification data is finally obtained, which contains the reasoning results and intermediate derivation steps.

[0132] Step S58: Intelligently correct the bill candidate recognition result set according to the bill semantic verification data to obtain the final bill recognition output data.

[0133] Specifically, the bill semantic verification data can be prepared, which includes the reasoning results and intermediate derivation steps. Use machine learning models for intelligent error correction. The model can identify and correct errors or inconsistencies in the data. Use historical bill data to train an error correction model, such as a support vector machine (SVM) or a neural network. The correct bill data is used as a positive sample, and the incorrect bill data is used as a negative sample. The bill candidate recognition result set is input into the trained error correction model, and the model will identify potential errors and make correction suggestions. According to the correction suggestions of the model, the candidate recognition result set is adjusted to generate the final bill recognition output data.

[0134] The present invention provides background information by constructing a bill semantic ontology knowledge base, making the bill recognition process more accurate and efficient. By standardizing the high-confidence bill feature representation data and annotating the semantic graph mapping, not only the consistency and comparability of the bill features are improved, but also the semantic information of the features is enhanced. The constructed semantic reasoning logic rule set further improves the logic and accuracy of the recognition process, and the semantic relevance evaluation and knowledge graph constraint verification based on these rules ensure the reliability of the recognition results. The accuracy of bill recognition is significantly improved through semantic reasoning path planning and semantic verification and intelligent error correction mechanism, especially when dealing with complex or ambiguous bills. In summary, the present invention can adapt to different bill types and changes, and can maintain stable recognition performance even when faced with incomplete or damaged bill information.

[0135] Preferably, the present invention further provides a data processing system for bill identification, which is used to execute the data processing method for bill identification as described above, and the data processing system for bill identification comprises:

[0136] The image acquisition module is used to acquire multispectral images of bills to obtain a multispectral image set of bills; perform cross-spectral registration on the multispectral image set of bills to obtain an aligned bill image set; perform weighted feature fusion on the aligned bill image set to obtain an enhanced bill feature map;

[0137] A topological feature extraction module is used to extract topological transformation features of the enhanced bill feature map to obtain bill geometric feature mapping data; based on the bill geometric feature mapping data, the enhanced bill feature map is locally semantically decoupled to obtain a bill topological feature point set;

[0138] The information entropy coding module is used to perform information entropy analysis on the bill topological feature point set to obtain the bill information entropy gradient distribution map; construct an adaptive coding mapping table based on the bill information entropy gradient distribution map to obtain the bill dynamic coding sequence;

[0139] The semantic evaluation module is used to evaluate the semantic relevance of the bill dynamic coding sequence to obtain semantic consistency reconstruction mapping data; semantically correct the bill dynamic coding sequence according to the semantic consistency reconstruction mapping data to obtain high-confidence bill feature representation data;

[0140] The knowledge matching module is used to perform prior knowledge graph matching on high-confidence bill feature representation data to obtain a bill candidate recognition result set; perform semantic reasoning on the bill candidate recognition result set to obtain bill semantic verification data; perform intelligent error correction on the bill candidate recognition result set based on the bill semantic verification data to obtain the final bill recognition output data.

[0141] Preferably, a computer-readable storage medium stores a data processing method for bill identification as described above that can be loaded and executed by a processor.

[0142] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0143] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A data processing method for bill identification, characterized in that: The following steps are involved: Step S1: collect multispectral images of bills to obtain a multispectral image set of bills; Perform cross-spectral registration on the multispectral image set of bills to obtain an aligned bill image set; Perform weighted feature fusion on the aligned bill image set to obtain an enhanced bill feature map; Step S2: extracting the topological transformation features of the enhanced bill feature map to obtain bill geometric feature mapping data; Based on the bill geometric feature mapping data, the enhanced bill feature map is locally semantically decoupled to obtain the bill topological feature point set; Step S3: performing information entropy analysis on the bill topological feature point set to obtain a bill information entropy gradient distribution map; constructing an adaptive coding mapping table based on the bill information entropy gradient distribution map to obtain a bill dynamic coding sequence; Step S4: performing semantic relevance evaluation on the bill dynamic coding sequence to obtain semantic consistency reconstruction mapping data; performing semantic correction on the bill dynamic coding sequence according to the semantic consistency reconstruction mapping data to obtain high-confidence bill feature representation data; Step S5: Perform prior knowledge graph matching on high-confidence bill feature representation data to obtain a bill candidate recognition result set; perform semantic reasoning on the bill candidate recognition result set to obtain bill semantic verification data; perform intelligent error correction on the bill candidate recognition result set based on the bill semantic verification data to obtain the final bill recognition output data.

2. The data processing method for bill identification according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: photographing the original image of the bill to obtain the original image of the bill, and collecting the shooting device parameters of the original image of the bill to obtain a device image collection parameter set; Step S12: Select the spectral range of the spectral imaging of the bill according to the device image acquisition parameter set to obtain the device shooting spectral range data; Step S13: performing multispectral image acquisition on the bill according to the spectral range data captured by the device to obtain a multispectral image set of the bill; wherein the multispectral image acquisition includes visible light band image acquisition, near infrared band image acquisition and infrared band image acquisition; Step S14: extracting characteristic associated points from the bill multispectral image set to obtain a bill image characteristic reference point set; Step S15: constructing a registration transformation matrix for the bill multispectral image set according to the bill image feature reference point set to obtain the bill multispectral image registration transformation matrix, and performing geometric transformation registration on the bill multispectral image set based on the bill multispectral image registration transformation matrix to obtain an aligned bill image set; Step S16: Perform weighted feature fusion on the aligned bill image set to obtain an enhanced bill feature map.

3. The data processing method for bill identification according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: performing geometric feature mapping transformation on the enhanced bill feature map to obtain a bill geometric feature mapping space; Step S22: Calculate the characteristic curvature tensor of the enhanced bill feature map based on the bill geometric feature mapping space to obtain the bill local geometric structure curvature field; Step S23: performing geometric invariance mapping on the bill feature points in the enhanced bill feature map based on the bill local geometric structure curvature field to obtain the bill geometric feature invariant; Step S24: performing topological invariance feature encoding on the bill geometric feature invariant to obtain bill geometric feature mapping data; Step S25: semantically clustering the bill geometric feature mapping data to obtain bill semantic cluster mapping data; Step S26: Perform local semantic decomposition on the enhanced bill feature map according to the bill semantic clustering mapping data to obtain a bill semantic subspace, and reconstruct topological feature points on the bill semantic subspace to obtain a bill topological feature point set.

4. The data processing method for bill identification according to claim 3, characterized in that: Step S26 includes the following steps: Step S261: Perform semantic association evaluation on the bill semantic cluster mapping data to obtain a bill semantic association weight matrix; Step S262: semantically deconstructing the enhanced bill feature graph based on the bill semantic association weight matrix to obtain bill multi-layer semantic deconstruction mapping data; Step S263: Perform semantic-geometric collaborative decoupling on the enhanced bill feature map according to the bill multi-layer semantic deconstruction mapping data and the bill local geometric structure curvature field to obtain bill semantic feature mapping data; Step S264: performing a topological semantic convergence transformation on the bill multi-layer semantic deconstruction mapping data based on the bill semantic feature mapping data to obtain a bill semantic subspace; Step S265: Reconstruct feature points of the bill semantic subspace according to the bill geometric feature invariant to obtain a bill topological feature point set.

5. The data processing method for bill identification according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: topologically segmenting the enhanced bill feature graph according to the bill topological feature point set to obtain bill feature block mapping data; Step S32: performing information entropy estimation on the bill feature block mapping data based on the enhanced bill feature map to obtain a bill local entropy value matrix; Step S33: performing statistics on the bill local entropy value matrix to obtain a bill information entropy distribution histogram, and performing gradient mapping according to the bill information entropy distribution histogram to obtain a bill information entropy gradient distribution map; Step S34: extracting feature points from the bill information entropy gradient distribution map according to a preset information entropy threshold to obtain a set of bill key entropy change points; Step S35: generating adaptive coding rules based on the bill key entropy change point set to obtain bill coding mapping rules; Step S36: extracting feature representations of the bill topological feature point set based on the enhanced bill feature graph to obtain a bill feature representation data set; Step S37: Convert the bill feature representation data set into code according to the bill code mapping rule to obtain a bill dynamic code sequence.

6. The data processing method for bill identification according to claim 5, characterized in that: Step S35 includes the following steps: Step S351: perform information entropy complexity evaluation on the bill key entropy change point set to obtain bill information complexity spectrum mapping data; Step S352: performing nonlinear information association dimensionality reduction on the information features in the enhanced bill feature graph according to the bill information complexity spectrum mapping data and the bill key entropy change point set to obtain bill information association dimensionality reduction mapping data; Step S353: Decomposing the information feature dimensional space of the bill key entropy change point set based on the bill information associated dimensionality reduction mapping data to obtain the bill information feature deconstruction subspace; Step S354: constructing information coding basis functions according to the deconstruction subspace of the bill information feature to obtain the bill coding basis functions; Step S355: performing information feature mapping transformation on the bill key entropy change point set through the bill coding basis function to obtain bill feature mapping transformation data; Step S356: constructing the code transformation domain of the bill key entropy change point set based on the bill feature mapping transformation data to obtain the bill code transformation domain mapping data; Step S357: Generate encoding rules based on the bill feature mapping transformation matrix and the bill coding transformation domain mapping data to obtain the bill coding mapping rules.

7. The data processing method for bill identification according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: extracting semantic association features from the bill dynamic code sequence to obtain a bill semantic association feature vector; Step S42: performing semantic consistency entropy value evaluation on the enhanced bill feature graph according to the bill semantic association feature vector to obtain bill semantic entropy value mapping data; Step S43: Based on the bill semantic entropy value mapping data, the bill dynamic coding sequence is semantically deconstructed to obtain bill semantic association deconstruction mapping data; Step S44: reconstructing the semantic consistency features of the bill dynamic coding sequence according to the bill semantic association deconstruction mapping data to obtain semantic consistency reconstruction mapping data; Step S45: Performing a semantic relevance confidence assessment on the bill dynamic coding sequence according to the semantic consistency reconstruction mapping data to obtain a bill semantic confidence assessment index; Step S46: Based on the bill semantic association feature vector and the bill semantic confidence evaluation index, semantic correction is performed on the bill dynamic coding sequence to obtain high-confidence bill feature representation data.

8. The data processing method for bill identification according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: constructing a semantic ontology for the bill to obtain a bill semantic ontology knowledge base; Step S52: normalizing the high-confidence bill feature representation data to obtain standard bill feature representation data, and annotating the standard bill feature representation data with semantic graph mapping based on the bill semantic ontology knowledge base to obtain a bill candidate recognition result set; Step S53: constructing semantic reasoning logic for the bill candidate recognition result set to obtain a bill semantic reasoning logic rule set; Step S54: performing semantic relevance evaluation on the bill candidate recognition result set based on the bill semantic reasoning logic rule set to obtain bill semantic relevance evaluation data; Step S55: performing knowledge graph constraint verification on the bill semantic association evaluation data based on the bill semantic ontology knowledge base to obtain the bill semantic constraint verification result; Step S56: Perform semantic reasoning path planning on the bill candidate recognition result set according to the bill semantic constraint verification result to obtain the bill semantic reasoning path; Step S57: performing semantic reasoning on the bill candidate recognition result set based on the bill semantic reasoning path to obtain bill semantic verification data; Step S58: Intelligently correct the bill candidate recognition result set according to the bill semantic verification data to obtain the final bill recognition output data.

9. A data processing system for bill identification, characterized in that: Used to execute the data processing method for bill identification as claimed in claim 1, the data processing system for bill identification comprises: The image acquisition module is used to acquire multispectral images of bills to obtain a multispectral image set of bills; perform cross-spectral registration on the multispectral image set of bills to obtain an aligned bill image set; perform weighted feature fusion on the aligned bill image set to obtain an enhanced bill feature map; A topological feature extraction module is used to extract topological transformation features of the enhanced bill feature map to obtain bill geometric feature mapping data; based on the bill geometric feature mapping data, the enhanced bill feature map is locally semantically decoupled to obtain a bill topological feature point set; The information entropy coding module is used to perform information entropy analysis on the bill topological feature point set to obtain the bill information entropy gradient distribution map; construct an adaptive coding mapping table based on the bill information entropy gradient distribution map to obtain the bill dynamic coding sequence; The semantic evaluation module is used to evaluate the semantic relevance of the bill dynamic coding sequence to obtain semantic consistency reconstruction mapping data; semantically correct the bill dynamic coding sequence according to the semantic consistency reconstruction mapping data to obtain high-confidence bill feature representation data; The knowledge matching module is used to perform prior knowledge graph matching on high-confidence bill feature representation data to obtain a bill candidate recognition result set; perform semantic reasoning on the bill candidate recognition result set to obtain bill semantic verification data; perform intelligent error correction on the bill candidate recognition result set based on the bill semantic verification data to obtain the final bill recognition output data.

10. A computer-readable storage medium, characterized in that: The data processing method for bill identification according to any one of claims 1 to 8 is stored and can be loaded and executed by a processor.

Images