A graph signature recognition processing method and system based on multi-source feature fusion and AI

Abstract

The application provides a kind of based on multi-source feature fusion and AI's graph signature recognition processing method and system, belongs to pattern recognition technical field, specifically includes: feature library construction module is responsible for extracting the multidimensional feature of frame template, is converted into standardization template feature vector, is stored in template feature library;Vector extraction module is responsible for extracting feature and converting into standardization candidate feature vector;Target frame screening module is responsible for using AI algorithm and graph signature recognition processing method, the candidate feature vector is compared with the template feature vector, similarity is calculated and target frame is screened out;Frame recognition module extracts the drawing vector size of target frame and the actual physical size of template, integrates the identification result of map specification, the calculation result of actual physical size of target frame, outputs identification report, improves the identification reliability.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, and in particular relates to an image recognition processing method and system based on multi-source feature fusion and AI. Background Technology

[0002] In fields such as engineering design, architectural construction, and machinery manufacturing, DWG format drawings are the core carrier for conveying design intent and guiding construction production. The drawing label (including the drawing frame, drawing size, scale, etc.) serves as the basic identifier of the drawing, and its recognition and parameter extraction are crucial steps in the digital management and automated interpretation of drawings. With the integrated development of computer-aided design (CAD) technology and artificial intelligence technology, drawing label recognition technology is gradually evolving from manual recognition to automated recognition. Currently, various drawing label recognition software and technical solutions have emerged.

[0003] Currently, drawing frame recognition technologies in the industry mainly fall into two categories: one is based on system-built-in fixed standard drawing frame blocks, requiring designers to use these blocks without breaking them down. The system identifies the drawing frame based on the block name, which is very unfriendly to historical drawings, as the drawing frames on historical drawings must be manually replaced with these standard drawing frame blocks for recognition. Furthermore, adding a new drawing frame requires tedious setup, creating the drawing frame as a block and setting certain parameters before it can be released to designers. The other type is based on rectangular outline recognition technology, which blindly checks closed shapes in drawings and identifies them as drawing frames. However, this often results in rectangular elements in the drawing being identified as drawing frames, and the blind test results require manual setting of the drawing size and scale, as the system cannot automatically recognize these. Both types of technologies have significant limitations and cannot meet the diverse needs of practical engineering applications.

[0004] Therefore, by combining multi-source feature fusion, reliable identification and processing of image tags can be achieved. However, in actual use, there may be too much interference risk from certain features, which may lead to deviations in the identification results. Therefore, how to identify and process feature vectors using multiple features based on the identification deviation risk of different feature vectors, thereby improving the reliability of the identification and processing, has become an urgent technical problem to be solved.

[0005] In summary, to address the aforementioned technical issues, this application provides a method and system for image signature recognition based on multi-source feature fusion and AI. Summary of the Invention

[0006] To achieve the objectives of this invention, the following technical solution is adopted: Specifically, in the first aspect, this application provides a signature recognition and processing system based on multi-source feature fusion and AI, specifically including: The feature library construction module is responsible for format validation and preprocessing of the imported frame templates, extracting multi-dimensional features of the frame templates, converting them into standardized template feature vectors, and storing them in the template feature library. The vector extraction module is responsible for locating candidate frame regions of the DWG drawing to be identified through contour detection and closed shape recognition, extracting features and converting them into standardized candidate feature vectors. The target frame filtering module is responsible for determining the standardized template feature vector based on multi-dimensional features for similarity analysis and recognition processing in the standardized template feature vector based on the recognition deviation data of the frame under different standardized template feature vectors, and using it as the target type feature vector. Based on the update results of feature analysis targets under different target type feature vectors, the module determines the image recognition processing method under the target type feature vector. Using AI algorithms and image recognition processing methods, the module compares the candidate feature vector with the template feature vector, calculates the similarity, and filters out the target frame. The drawing frame recognition module extracts the drawing vector dimensions and the actual physical dimensions of the template of the target drawing frame, integrates the recognition results of the drawing sheet specifications and the calculation results of the actual physical dimensions of the target drawing frame, and outputs a recognition report.

[0007] Furthermore, the multidimensional features include geometric features, line features, and structural features.

[0008] Furthermore, the target frame is a frame with a similarity coefficient greater than or equal to a similarity coefficient threshold.

[0009] Furthermore, the method for determining the identification result of the map sheet size is as follows: The drawing scale is calculated using a formula. Based on the drawing vector size of the target drawing frame and the drawing scale, the actual physical size of the target drawing frame is converted and compared with the standard drawing size library to determine the drawing size specification.

[0010] Secondly, this application provides a method for image recognition processing based on multi-source feature fusion and AI, applied to the aforementioned image recognition processing system based on multi-source feature fusion and AI, specifically including: S1 uses the identification data of the image tag to determine the identification deviation data of the image frame under different standardized template feature vectors. Using the identification deviation data of the image frame under different standardized template feature vectors, S1 determines the standardized template feature vector in the standardized template feature vector that is processed for similarity analysis based on multi-dimensional features and uses it as the target type feature vector. S2 determines the identification strategy for feature analysis targets in the target identification based on the target type feature vector and the degree of correlation between the identification deviation data of the frame under the target type feature vector and the identification deviation data of the model; S3 determines the image recognition processing method under the target type feature vector based on the update results of the feature analysis target under different target type feature vectors; S4 uses the image recognition processing method to determine the matching analysis features of the target under the target type feature vector, and uses the matching analysis features under different target type feature vectors and the fluctuation range of similarity under different matching analysis features to determine the updated target of the target type feature vector.

[0011] Furthermore, the recognition deviation data of the frame under the standardized template feature vector is determined based on the number of recognition deviations of the frame under the standardized template feature vector.

[0012] Furthermore, the method for determining the target type feature vector is as follows: S11 uses the recognition deviation data of the frame under different standardized template feature vectors to determine the number of recognition deviations of the frame under the standardized template feature vectors; S12 determines the recognition deviation probability of the frame under the standardized template feature vector based on the number of recognition deviations of the frame under the standardized template feature vector; S13 determines the target type feature vector in the standardized template feature vector based on the recognition deviation probability of the frame under different standardized template feature vectors.

[0013] The beneficial effects of this invention are as follows: Based on the recognition deviation data of the bounding boxes under different standardized template feature vectors, the overall recognition deviation risk of the AI ​​model is determined. Using the overall recognition deviation risk of the AI ​​model, the target type feature vector under the standardized template feature vector is determined, thereby laying the foundation for improving the overall recognition reliability while reducing the difficulty of recognition processing.

[0014] By utilizing the matching analysis feature data under different target type feature vectors and the fluctuation range of similarity coefficients under the matching analysis features, the accuracy of tag recognition and analysis processing under different target type feature vectors combined with features of different dimensions is determined. This accuracy in tag recognition and analysis processing under different target type feature vectors combined with features of different dimensions determines the update target of the target type feature vector, thereby further realizing the update and recognition processing of the target type feature vector and ensuring the reliability of tag recognition and processing.

[0015] Other features and advantages will be set forth in the following description, and the objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] The above and other features and advantages of the present invention will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0018] Figure 1 This is a framework diagram of a tag recognition and processing system based on multi-source feature fusion and AI; Figure 2 This is a flowchart of a signature recognition processing method based on multi-source feature fusion and AI; Figure 3 This is a flowchart illustrating the method for determining the feature vector of the target type; Figure 4 This is a flowchart illustrating the method for determining the identification strategy of the target through feature analysis in image-based target recognition. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0020] Example 1 Specifically, such as Figure 1 As shown, this application provides a signature recognition processing system based on multi-source feature fusion and AI, specifically including: The feature library construction module is responsible for format validation and preprocessing of the imported frame templates, extracting multi-dimensional features of the frame templates, converting them into standardized template feature vectors, and storing them in the template feature library. The vector extraction module is responsible for locating candidate frame regions of the DWG drawing to be identified through contour detection and closed shape recognition, extracting features and converting them into standardized candidate feature vectors. The target frame filtering module is responsible for determining the standardized template feature vector based on multi-dimensional features for similarity analysis and recognition processing in the standardized template feature vector based on the recognition deviation data of the frame under different standardized template feature vectors, and using it as the target type feature vector. Based on the update results of feature analysis targets under different target type feature vectors, the module determines the image recognition processing method under the target type feature vector. Using AI algorithms and image recognition processing methods, the module compares the candidate feature vector with the template feature vector, calculates the similarity, and filters out the target frames. The drawing frame recognition module extracts the drawing vector dimensions and the actual physical dimensions of the template of the target drawing frame, integrates the recognition results of the drawing sheet specifications and the calculation results of the actual physical dimensions of the target drawing frame, and outputs a recognition report.

[0021] Specifically, the construction method of the AI ​​model for similarity evaluation based on the AI ​​model is as follows: The similarity evaluation AI model constructs input features based on the standardized candidate feature vectors and standardized template feature vectors of the candidate bounding boxes, and uses the actual similarity labels between the candidate and template bounding boxes as the output target. It employs a Siamese neural network architecture for model construction. The Siamese neural network includes two structurally identical and parameter-shared feature extraction sub-networks, used to extract deep feature representations of the candidate and template feature vectors, respectively. Each feature extraction sub-network consists of multiple fully connected layers, with ReLU activation functions introduced to introduce non-linear mapping between the fully connected layers. Batch normalization and Dropout layers are also set between the fully connected layers to improve the model's generalization ability and prevent overfitting. After the two feature extraction sub-networks output the deep feature representations of the candidate and template feature vectors, respectively, a distance metric layer calculates the cosine similarity between the two deep feature representations, which serves as the similarity coefficient between the candidate and template bounding boxes. The training data is constructed based on the labeled data in the historical frame recognition process. The candidate frame feature vector is paired with the corresponding template feature vector as input samples, and the actual similarity of manually labeled data is used as label samples. The contrastive loss function is used to iteratively optimize the model parameters. The contrastive loss function brings the feature representation distance closer in similar frame pairs and pushes the feature representation distance further apart in dissimilar frame pairs, so that the model learns the semantic similarity mapping relationship between frame features, ensuring that the similarity evaluation results can accurately reflect the true matching degree between frames.

[0022] The update mechanism of the similarity assessment AI model is as follows: During the image recognition process, when the recognition deviation data indicates a discrepancy between the similarity assessment result under a standardized template feature vector and the actual matching situation, the candidate feature vector, template feature vector, and corrected actual similarity label in the deviation case are added as new training samples to the training dataset. The model is then incrementally trained to update the parameters of the feature extraction sub-network. During the update process, if the similarity assessment deviation probability of a standardized template feature vector continuously exceeds a preset deviation probability threshold in multiple consecutive recognition processes, the sampling weight of samples related to that feature vector in the training dataset is increased. This allows the model to pay more attention to the feature patterns under that feature vector in subsequent training, thereby achieving dynamic adaptive correction of the similarity assessment deviation under different template feature vectors. The significance of this update mechanism lies in ensuring that the similarity assessment AI model can adapt to the impact of diversified image frame templates and evolving drawing styles through continuous deviation data feedback and model iteration, thus maintaining the accuracy and reliability of similarity assessment in the long term.

[0023] The similarity assessment AI model is invoked in the image recognition process as follows: After the standardized candidate feature vectors for the candidate image frame region are extracted, the candidate feature vectors and the standardized template feature vectors from the template feature library are input into the two feature extraction sub-networks of the similarity assessment AI model. The model outputs the similarity coefficients between the candidate image frame and each template image frame. The template image frame with the highest similarity coefficient is taken as the best matching template. It is then determined whether the maximum similarity coefficient reaches or exceeds a preset similarity coefficient threshold. If it does, the candidate image frame is identified as the target image frame; otherwise, it is determined that there is no valid image frame in the candidate region. In the similarity confirmation stage of the matching analysis features, for matching analysis features whose matching weight values ​​meet the requirements, the feature values ​​of the candidate image frame and the template image frame in that feature dimension are input into the corresponding feature channel of the similarity assessment AI model. The model outputs the similarity coefficient in that feature dimension and determines whether the similarity coefficient is within the fluctuation range of the matching analysis feature. Only when the similarity coefficient is within the fluctuation range and meets the preset similarity coefficient threshold is the recognition result finally confirmed as accurate. This multi-dimensional feature dual verification mechanism improves the overall reliability of the image recognition result.

[0024] Specifically, the implementation methods of each module of the system are as follows: The specific implementation method of the feature library construction module is as follows: In the format verification and preprocessing stage, the format of the imported drawing frame template file is verified to determine whether the file format meets the preset format requirements. These preset format requirements include that the file type is DWG or DXF, the file size is within a preset size range, and the file contains closed graphic entities. The drawing frame template files that pass the format verification are then preprocessed. This preprocessing includes layer cleaning (removing redundant entities from non-drawing frame layers), coordinate system normalization (aligning the drawing frame template's coordinate system to a standard coordinate system with the lower left corner as the origin), and scale normalization (scaling the drawing frame template's dimensions according to a preset reference scale to eliminate size differences between templates from different sources).

[0025] In the multidimensional feature extraction stage, three types of multidimensional features—geometric features, line features, and structural features—are extracted from the preprocessed frame template. Geometric features include the aspect ratio of the frame's outer rectangle, the area of ​​the outer rectangle, the perimeter of the outer rectangle, the rectangularity of the outer rectangle (the ratio of the area of ​​the outer rectangle to the area of ​​the circumscribed rectangle), and the distribution of corner coordinates. Line features include the line type of the frame's edges (solid line, dashed line, dotted line), the line width distribution, the number of line segments, the average length of the line segments, and the histogram of line segment directions. Structural features include the position coordinates of the title bar inside the frame, the area ratio of the title bar, the number of lines of text in the title bar, the number of tables embedded in the frame, the area ratio of the embedded tables in the frame, and corner marking features (such as whether there are cropping or folding marks at the corners).

[0026] In the standardized template feature vector conversion stage, the extracted geometric features, line features, and structural features are concatenated into an original feature vector according to a preset feature arrangement order. Each feature component in the original feature vector is then normalized according to its corresponding preset normalization parameter, mapping each feature component to the interval between 0 and 1, thus forming a standardized template feature vector. The standardized template feature vector, along with template identification information (template number, template name, import time), is stored in a template feature library. This template feature library uses a relational database for storage and management, supporting retrieval by template number and feature vector content.

[0027] The specific implementation of the vector extraction module is as follows: In the contour detection stage, the DWG drawing to be identified is parsed to extract the geometric information of all graphic entities, including lines, arcs, polylines, splines, etc. Based on the endpoint coordinates and connection relationships of the graphic entities, a contour tracking algorithm is used to identify closed shapes in the drawing. These closed shapes include closed polygons composed of line segments connected end to end and closed regions composed of closed polylines. The identified closed shapes are then subjected to area calculation and shape analysis. Closed shapes with an area exceeding a preset area threshold are designated as candidate frame regions, while closed shapes with too small an area (such as text boxes, table cells, and other non-frame regions) are excluded.

[0028] In the feature extraction and standardized candidate feature vector conversion stages, geometric features, line features, and structural features are extracted for each candidate bounding box region according to the same feature extraction rules as the feature library construction module. These features are then normalized using the same preset normalization parameters to form standardized candidate feature vectors. The standardized candidate feature vectors, along with the position coordinates of the candidate regions (the coordinates of the lower left and upper right corners of the candidate region in the drawing), are output to the target bounding box filtering module, providing input data for subsequent similarity comparison and target bounding box filtering.

[0029] The specific implementation of the target frame filtering module is as follows: In the target type feature vector determination stage, based on the recognition deviation data of each standardized template feature vector in the template feature library in the historical image recognition process, and according to the method described in S1 of Example 1, the target type feature vector that needs to be subjected to multidimensional feature similarity analysis is determined by calculating the number of recognition deviations and the probability of recognition deviations, and by hierarchically filtering according to the four-level deviation probability thresholds (preset deviation probability threshold, first deviation probability threshold, second deviation probability threshold, and third deviation probability threshold).

[0030] In the stage of determining the image recognition processing method, based on the updated results of feature analysis targets under different target type feature vectors, and according to the method described in S3 of Example 1, the image recognition processing method under each target type feature vector is determined by updating the existence judgment of the recognition weight value and judging one by one. That is, it is determined under what feature analysis target number threshold and similarity coefficient fluctuation range conditions to exclude features from the matching analysis feature set. The image recognition processing method includes method A (most stringent exclusion condition, feature analysis target number threshold is the largest), method B (most lenient exclusion condition, feature analysis target number threshold is the smallest) and method C (medium exclusion condition, feature analysis target number threshold is in the middle).

[0031] In the similarity comparison and target bounding box selection stages, the standardized candidate feature vectors output by the vector extraction module are compared with the target type feature vectors input into the similarity evaluation AI model. Specifically, the standardized candidate feature vectors are input into the first feature extraction sub-network of the similarity evaluation AI model, and the feature vectors of each target type are input into the second feature extraction sub-network. The model outputs the similarity coefficient between the candidate bounding box and the feature vectors of each target type. The target type feature vector with the highest similarity coefficient is used as the best matching template, and it is determined whether the highest similarity coefficient reaches or exceeds a preset similarity coefficient threshold. If the highest similarity coefficient reaches or exceeds the preset similarity coefficient threshold, the candidate bounding box is determined as the target bounding box, and the template identification information of the best matching template is associated and recorded with the target bounding box. If the highest similarity coefficient does not reach or exceed the preset similarity coefficient threshold, it is determined that there is no valid bounding box in the candidate region. In the similarity confirmation stage of matching analysis features, for matching analysis features whose matching weight values ​​meet the requirements, the feature values ​​of the candidate bounding box and the feature values ​​of the best matching template in the same feature dimension are input into the corresponding feature channel of the similarity evaluation AI model. The model outputs the similarity coefficient in the same feature dimension and determines whether the similarity coefficient is within the fluctuation range of the similarity coefficient of the matching analysis feature. Only when the similarity coefficient is within the fluctuation range and meets the preset similarity coefficient threshold requirement is the recognition result finally confirmed as accurate.

[0032] The specific implementation of the frame recognition module is as follows: In the dimension extraction stage, the drawing vector dimensions of the target drawing frame are extracted from the primitive entity data of the target drawing frame. The drawing vector dimensions include the length and width of the outer contour rectangle of the target drawing frame (expressed in drawing units). The actual physical dimensions of the template are obtained from the template identification information of the best matching template. The actual physical dimensions include the actual length and actual width of the outer contour rectangle of the template drawing frame (expressed in millimeters).

[0033] In the drawing scale calculation stage, the drawing scale is calculated using the formula: Drawing Scale = Actual Physical Length of Template / Vector Length of Target Drawing Frame. Based on the vector dimensions of the target drawing frame and the drawing scale, the actual physical dimensions of the target drawing frame are converted. The actual physical dimensions are calculated as follows: Actual Physical Length of Target Drawing Frame = Vector Length of Target Drawing Frame × Drawing Scale; Actual Physical Width of Target Drawing Frame = Vector Width of Target Drawing Frame × Drawing Scale.

[0034] In the map sheet specification determination stage, the actual physical dimensions of the target map frame are compared with a standard map sheet specification library. This library pre-stores the length and width data of various standard map sheet specifications (e.g., A0 is 841mm × 1189mm, A1 is 594mm × 841mm, A2 is 420mm × 594mm, A3 is 297mm × 420mm, and A4 is 210mm × 297mm). The actual physical length and width of the target map frame are then compared with the length and width of each specification in the standard map sheet specification library. The standard map sheet specification with the smallest deviation from the target map frame's actual physical dimensions is determined and identified as the target map sheet specification. If the deviation of the target map frame's actual physical dimensions from all standard map sheet specifications exceeds a preset deviation threshold, the map sheet specification is determined to be a non-standard specification, and the actual physical dimension data is recorded.

[0035] In the identification report output stage, the identification results of map sheet specifications, the calculation results of the actual physical size of the target map frame, the template identification information of the best matching template, the similarity coefficient, the drawing scale, and the position coordinate information of the candidate map frame area are integrated to generate an identification report and output it to the user interface. The identification report includes the position information of the target map frame, map sheet specifications, actual physical size, matching template information, similarity coefficient, and drawing scale.

[0036] Furthermore, the multidimensional features include geometric features, line features, and structural features.

[0037] Specifically, the target frame is a frame with a similarity coefficient greater than or equal to a similarity coefficient threshold. It should be noted that the method for determining the identification result of the map sheet size is as follows: The drawing scale is calculated using a formula. Based on the drawing vector size of the target drawing frame and the drawing scale, the actual physical size of the target drawing frame is converted and compared with the standard drawing size library to determine the drawing size specification.

[0038] Example 2 Secondly, such as Figure 2 As shown, this application provides a method for image recognition processing based on multi-source feature fusion and AI, applied to the aforementioned image recognition processing system based on multi-source feature fusion and AI, specifically including: S1 uses the identification data of the image tag to determine the identification deviation data of the image frame under different standardized template feature vectors. Using the identification deviation data of the image frame under different standardized template feature vectors, S1 determines the standardized template feature vector in the standardized template feature vector that is processed for similarity analysis based on multi-dimensional features and uses it as the target type feature vector. S2 determines the identification strategy for feature analysis targets in the target identification based on the target type feature vector and the degree of correlation between the identification deviation data of the frame under the target type feature vector and the identification deviation data of the model; S3 determines the image recognition processing method under the target type feature vector based on the update results of the feature analysis target under different target type feature vectors; S4 uses the image recognition processing method to determine the matching analysis features of the target under the target type feature vector, and uses the matching analysis features under different target type feature vectors and the fluctuation range of similarity under different matching analysis features to determine the updated target of the target type feature vector.

[0039] Specifically, the recognition deviation data of the frame under the standardized template feature vector is determined based on the number of recognition deviations of the frame under the standardized template feature vector.

[0040] It should be noted that, as Figure 3 As shown, the method for determining the target type feature vector is as follows: The core objective of this embodiment is to determine the target type feature vector requiring multi-dimensional feature similarity analysis by evaluating the recognition deviation data of standardized template feature vectors. Then, based on the correlation between the recognition deviation and model deviation of the target type feature vector, it determines the recognition strategy for the feature analysis target. Next, based on the update results of the feature analysis target, it determines the image recognition processing method. Finally, based on the matching analysis features and the similarity fluctuation range, it determines the update target for the target type feature vector. Its core logic uses image recognition deviation data as input. First, it filters the target type feature vectors through deviation probability. Then, it determines the recognition strategy for the feature analysis target through the recognition ratio and matching deviation ratio. Next, it determines the strictness of feature exclusion in the image recognition processing method by updating the recognition weight value. Finally, it determines the dynamic update target for the target type feature vector through the number of matching analysis features and the recognition matching weight value. The overall logic follows the process of "determining the target type feature vector → determining the feature analysis target recognition strategy → determining the image recognition processing method → ​​determining the target type feature vector update target."

[0041] Specifically, the recognition deviation data of the frame under the standardized template feature vector is determined based on the number of recognition deviations of the frame under the standardized template feature vector.

[0042] It should be noted that, as Figure 3 As shown, the method for determining the target type feature vector is as follows: In this embodiment, based on the recognition deviation data of the frame under different standardized template feature vectors, the overall recognition deviation risk of the AI ​​model is determined. Using the overall recognition deviation risk of the AI ​​model, the target type feature vector under the standardized template feature vector is determined, thereby reducing the difficulty of recognition processing and laying the foundation for improving the overall recognition reliability.

[0043] The standardized template feature vector refers to the standardized feature vector formed by extracting multi-dimensional features such as geometric features, line features, and structural features from the user-imported frame template; the recognition deviation data refers to the recorded data of recognition deviations that occur when using a certain standardized template feature vector for frame recognition processing; the recognition deviation risk refers to the degree of probability that the AI ​​model will exhibit recognition deviations under different standardized template feature vectors.

[0044] If a standardized template feature vector shows multiple identification errors in historical identification processes, then the identification error risk under this feature vector is high, and it needs to be included in the target type feature vector for special attention.

[0045] By assessing the overall identification bias risk based on the identified bias data and selecting target type feature vectors accordingly, identification resources are focused on feature vectors with higher bias risk, thereby reducing the difficulty of identification processing while improving the overall identification reliability.

[0046] S11 uses the recognition deviation data of the frame under different standardized template feature vectors to determine the number of recognition deviations of the frame under the standardized template feature vectors; The number of recognition deviations refers to the cumulative number of recognition deviations that occur when performing frame recognition processing under a certain standardized template feature vector.

[0047] Suppose that a standardized template feature vector has been used for multiple recognitions in historical recognition processes, and some of the recognition results do not match the actual bounding box. The number of times they do not match is the number of recognition deviations under that feature vector.

[0048] S12 Determine the recognition deviation probability of the frame under the standardized template feature vector based on the number of recognition deviations of the frame under the standardized template feature vector; It is understood that the recognition deviation probability of the frame under the standardized template feature vector is determined based on the ratio of the number of recognition deviations of the frame under the standardized template feature vector to the number of recognition processes.

[0049] The recognition deviation probability refers to the proportion of the number of recognition deviations to the total number of recognition processes under the standardized template feature vector, reflecting the frequency of recognition deviations under the feature vector.

[0050] Suppose that a standardized template feature vector has been processed multiple times in history, and some of these processing times resulted in recognition errors. The ratio of the number of errors to the total number of processing times is the recognition error probability for that feature vector.

[0051] The deviation probability is calculated by the ratio of the number of deviations to the number of recognition processes. The number of deviations is then converted into a normalized probability value, eliminating the influence of differences in the number of recognition processes under different feature vectors, thereby achieving comparability of deviation risks between different feature vectors.

[0052] S13 determines the target type feature vector in the standardized template feature vector based on the recognition deviation probability of the frame under different standardized template feature vectors.

[0053] Assuming that the deviation probabilities of each standardized template feature vector have been calculated, the target type feature vectors that need to be subjected to multidimensional feature similarity analysis are selected according to the distribution of the deviation probabilities and different deviation risk levels.

[0054] By selecting target type feature vectors based on the probability of deviation, the identification resources are focused on feature vectors with a higher risk of deviation, thereby reducing the difficulty of identification and processing while improving the overall reliability of identification.

[0055] Specifically, the target type feature vector in the standardized template feature vector is determined based on the recognition deviation probability of the bounding box under different standardized template feature vectors, including: It should be noted that the order of the four deviation probability thresholds is as follows: third deviation probability threshold < second deviation probability threshold < first deviation probability threshold < preset deviation probability threshold.

[0056] Case 1: If there is no standardized template feature vector with an identification deviation probability greater than the preset deviation probability threshold, then the identification deviation risk is low. In this case, the target type feature vector in the standardized template feature vector is determined to be a standardized template feature vector with an identification deviation probability greater than the first deviation probability threshold. Case 1 refers to the target type feature vector selection scenario where the deviation probability of all standardized template feature vectors does not exceed the preset deviation probability threshold. In this case, a relatively lenient selection condition is adopted, and feature vectors with deviation probabilities exceeding the first deviation probability threshold are included in the target type.

[0057] Assuming that the deviation probabilities of all standardized template feature vectors are low and the overall deviation risk is low, feature vectors with deviation probabilities exceeding the first deviation probability threshold (but less than the preset deviation probability threshold) are selected as target type feature vectors.

[0058] When the overall deviation risk is low, a lenient screening strategy is adopted. When the overall deviation risk is low, only a few feature vectors with relatively high deviation probability need to be focused on, thereby avoiding the waste of processing resources caused by multi-dimensional feature analysis of all feature vectors.

[0059] Scenario 2: If there is a standardized template feature vector with an identification deviation probability greater than the preset deviation probability threshold, the standardized template feature vector with an identification deviation probability greater than the preset deviation probability threshold is used as the identification deviation risk feature vector. It is then determined whether the number of the identification deviation risk feature vectors meets the requirements. If yes, then the number of identification deviation risk feature vectors is large, and the target type feature vector in the standardized template feature vector is determined to be a standardized template feature vector with an identification deviation probability greater than the second deviation probability threshold. If not, then the target type feature vector in the standardized template feature vector is determined to be a standardized template feature vector with an identification deviation probability greater than the second deviation probability threshold and an identification deviation count greater than the preset deviation count threshold.

[0060] Case 2 refers to the target type feature vector screening scenario when there are feature vectors with a deviation probability exceeding a preset deviation probability threshold. In this case, the screening conditions need to be further refined based on the number of risk feature vectors. When the number of risk feature vectors is large, a more lenient screening condition (based solely on the deviation probability) is used; when the number is small, a more stringent screening condition with an additional deviation count condition is used.

[0061] If there are feature vectors whose deviation probability exceeds the preset deviation probability threshold and there are a large number of them, then feature vectors whose deviation probability exceeds the second deviation probability threshold (but is less than the first deviation probability threshold) will be selected as target type feature vectors. If the number of risk feature vectors is small, then a deviation frequency condition will be added to the deviation probability condition for selection.

[0062] When high-risk feature vectors exist, the screening criteria are determined according to the number of risk feature vectors. When there are many high-risk feature vectors, more lenient screening criteria are used to expand the coverage of target types. When there are few high-risk feature vectors, the deviation frequency condition is added to accurately screen the feature vectors that really need to be focused on, thereby achieving a precise match between the screening strategy and the deviation risk level.

[0063] This embodiment realizes a complete process from counting the number of identification deviations to calculating the deviation probability and then filtering the target type feature vector. Its core value is reflected in three aspects: First, by calculating the deviation probability by the ratio of the number of deviations to the number of identification processing, the influence of the difference in the number of identification processing under different feature vectors is eliminated; second, by using a four-level deviation probability threshold for hierarchical filtering, the strictness of the filtering is dynamically adjusted under different deviation risk levels; and third, by adding the deviation count condition, precise filtering is achieved when there are few high-risk feature vectors, thereby achieving a precise match between the target type feature vector filtering strategy and the deviation risk level.

[0064] A certain drawing recognition system has 10 standardized template feature vectors V1 to V10 pre-stored. The number of deviations and the number of recognition processes for each feature vector in historical recognition processes are as follows: V1: 3 deviations, 50 processing times, deviation probability 0.060; V2: 8 deviations, 50 processing times, deviation probability 0.160; V3: 2 deviations, 50 processing times, deviation probability 0.040; V4: 12 deviations, 50 processing times, deviation probability 0.240; V5: 1 deviation, 50 processing times, deviation probability 0.020; V6: 6 deviations, 50 processing times, deviation probability 0.120; V7: 15 deviations, 50 processing times, deviation probability 0.300; V8: 4 deviations, 50 processing times, deviation probability 0.080; V9: 9 deviations, 50 processing times, deviation probability 0.180; V10: 2 deviations, 50 processing times, deviation probability 0.040.

[0065] Step S11: The number of deviations for each standardized template feature vector is as described above.

[0066] Step S12: The deviation probability of each eigenvector has been calculated.

[0067] Step S13: Determine whether there are standardized template feature vectors with a deviation probability greater than the preset deviation probability threshold of 0.20: V4 deviation probability 0.24>0.20, V7 deviation probability 0.30>0.20, they exist, belonging to case 2.

[0068] Using V4 and V7 as risk feature vectors for identifying deviations, we determine whether the number of risk feature vectors meets the requirements (2 > the preset threshold of 1 for the number of risk feature vectors): 2 meets the requirements (too many).

[0069] The target type feature vectors were determined to be standardized template feature vectors with a deviation probability greater than the second deviation probability threshold of 0.05: V2 (0.16), V4 (0.24), V6 (0.12), V7 (0.30), and V9 (0.18), for a total of 5 target type feature vectors.

[0070] Specifically, such as Figure 4 As shown, the method for determining the identification strategy of the feature analysis target in the image recognition target is as follows: In this embodiment, based on the target type feature vector and the correlation between the recognition deviation data of the frame under the target type feature vector and the recognition deviation data of the model, the degree of improvement of the overall recognition deviation risk is determined by using the target type feature vector to perform frame recognition processing in a multi-source manner. Based on the degree of improvement of the overall recognition deviation risk, the recognition strategy of the feature analysis target in the sign recognition target is determined, thereby laying the foundation for optimizing the frame recognition processing strategy by realizing the similarity distribution of different feature analysis targets under different dimensions of features.

[0071] The feature analysis target refers to the image recognition object to be identified, and whether it is a valid image frame recognition result needs to be determined by the similarity between it and the feature vector of the target type; the recognition strategy refers to the decision rules for determining the screening conditions of the feature analysis target; the multi-source approach refers to the method of combining features from multiple dimensions for similarity analysis.

[0072] If the target type feature vector accounts for a high proportion of the standardized template feature vector, and the number of biases under the target type feature vector accounts for a low proportion of the total number of biases in the model, it indicates that the introduction of the target type feature vector has limited effect on improving the bias risk, and a more stringent screening strategy should be adopted. Conversely, a more lenient screening strategy should be adopted.

[0073] The improvement of bias risk by multi-source identification is evaluated by the proportion of target type feature vectors and the degree of correlation with bias. The selection strategy of feature analysis targets is dynamically adjusted according to the degree of improvement, thereby achieving a balance between identification reliability and processing efficiency.

[0074] S21 determines the proportion of the target type feature vector in the standardized template feature vector based on the target type feature vector, and uses it as the recognition proportion; The recognition ratio refers to the proportion of the number of target type feature vectors to the total number of all standardized template feature vectors, reflecting the coverage of feature vectors that need to be analyzed in multiple dimensions.

[0075] Assuming there are multiple standardized template feature vectors, some of which are identified as target type feature vectors, the ratio of the number of target type feature vectors to the total number is the recognition ratio.

[0076] The calculation of the recognition ratio quantifies the coverage of the target type feature vector, providing a basis for determining the classification of subsequent recognition strategies.

[0077] Specifically, if the recognition ratio is less than a preset recognition ratio threshold, the recognition strategy for the feature analysis target in the tag recognition target is that as long as the similarity coefficient between the tag recognition target and the target type feature vector is greater than a preset similarity coefficient threshold, it is regarded as the feature analysis target. The preset recognition ratio threshold refers to the critical value for judging whether the coverage of the target type feature vector is sufficient. When the recognition ratio is low, the coverage of the target type feature vector is insufficient, and a lenient screening strategy is adopted, where as long as the similarity coefficient reaches the threshold, it is used as a feature analysis target.

[0078] If the recognition rate is lower than the preset threshold, it indicates that the target type feature vector has a small coverage range. In this case, a lenient strategy is adopted, and the target of the feature analysis is considered as long as the similarity coefficient between the target of the image recognition and any target type feature vector reaches or exceeds the preset similarity coefficient threshold.

[0079] When the recognition rate is low, a lenient screening strategy is adopted. When the target type feature vector coverage is insufficient, the screening threshold of the feature analysis target is lowered to ensure sufficient analysis samples, thereby avoiding the distortion of analysis results due to insufficient coverage.

[0080] It is also understood that if the recognition ratio is not less than the preset recognition ratio threshold, the process proceeds to step S22. S22 determines the proportion of the number of recognition errors under all target type feature vectors to the total number of recognition errors in the model based on the degree of correlation between the recognition error data of the bounding box under the target type feature vector and the recognition error data of the model, and uses it as the matching error ratio. The matching deviation ratio refers to the proportion of the sum of the number of recognition deviations under all target type feature vectors to the total number of recognition deviations of the AI ​​model, reflecting the degree of correlation between the target type feature vector and the model deviation.

[0081] If the sum of the number of deviations under the feature vectors of each target type is large and accounts for a high proportion of the total number of deviations in the model, then the proportion of matching deviations is high, indicating that the feature vectors of the target type are strongly correlated with the model deviations.

[0082] The correlation between the target type feature vector and the model bias is assessed by matching the bias ratio. When the correlation is strong, a stricter screening strategy should be adopted to improve the risk of bias. When the correlation is weak, a more lenient strategy can be adopted.

[0083] Specifically, it is determined whether the matching deviation ratio is less than a preset deviation ratio threshold. If so, the identification strategy of the feature analysis target in the image recognition target is that as long as the similarity coefficient between the image recognition target and the target type feature vector is greater than the preset similarity coefficient threshold, it is regarded as the feature analysis target. If not, proceed to step S23. The preset deviation ratio threshold refers to a critical value for determining whether the matching deviation ratio is low. When the matching deviation ratio is low, the correlation between the target type feature vector and the model deviation is weak, and a lenient screening strategy is adopted.

[0084] If the matching deviation ratio is lower than the preset deviation ratio threshold, it indicates that the target type feature vector is weakly correlated with the model deviation. In this case, a lenient strategy is adopted, and the target of feature analysis is used as long as the similarity coefficient reaches the threshold.

[0085] When the matching deviation ratio is low, a lenient screening strategy is adopted. When the target type feature vector is weakly correlated with the model deviation, multidimensional feature analysis has limited effect on improving the deviation risk, and only the similarity coefficient needs to be used as the screening condition.

[0086] S23 determines the identification strategy for the feature analysis target in the image recognition target based on the recognition ratio and the matching deviation ratio.

[0087] Specifically, based on the average of the recognition ratio and the matching deviation ratio, a recognition matching weight value is determined. It is then determined whether the recognition matching weight value is greater than a preset weight threshold. If so, the recognition strategy for the feature analysis target in the sign recognition target is that if the similarity coefficient between the sign recognition target and the target type feature vector is greater than a preset similarity coefficient threshold, and there is a sign recognition deviation, then it is regarded as a feature analysis target. If not, the recognition strategy for the feature analysis target in the sign recognition target is that if the similarity coefficient between the sign recognition target and the target type feature vector is greater than a preset similarity coefficient threshold, and there is a sign recognition deviation, or the number of feature analysis targets within the most recent preset time period is less than a preset target number threshold, then it is regarded as a feature analysis target.

[0088] The calculation method for the identification matching weight value is as follows: Identification matching weight value = (Identification ratio + Matching deviation ratio) / 2.

[0089] The preset weight threshold refers to the critical value for judging the overall level of the identification and matching weight values. When the weight value is high, both the identification ratio and the matching deviation ratio are high, and a screening strategy with additional deviation conditions is adopted; when the weight value is low, a more lenient screening strategy with additional deviation conditions or recent target quantity conditions is adopted.

[0090] If the calculated matching weight value is greater than the preset weight threshold, it indicates that the target type feature vector has a wide coverage and a strong correlation with the model deviation. In this case, strategy B is adopted, which requires both the similarity coefficient condition and the existence of deviation condition to be used as a feature analysis target. If the weight value is not greater than the preset weight threshold, strategy C is adopted, which requires both the similarity coefficient condition and the existence of deviation or insufficient number of recent targets to be used as a feature analysis target.

[0091] The screening strategy is determined by classifying the matching weight values. When both the identification ratio and the matching deviation ratio are high, stricter screening conditions (similarity coefficient + deviation) are adopted, and when both are low, more lenient screening conditions (similarity coefficient + deviation or insufficient number of targets) are adopted, thereby achieving a precise match between the screening strategy and the degree of improvement of multi-source identification.

[0092] This system implements a complete process from calculating the recognition ratio to evaluating the matching deviation ratio and determining the target recognition strategy through feature analysis. Its core value lies in three aspects: First, it judges the coverage of target type feature vectors by using the recognition ratio, and adopts a lenient strategy to ensure sufficient analysis samples when the coverage is insufficient. Second, it evaluates the correlation between target type feature vectors and model bias by using the matching deviation ratio, and adopts a lenient strategy to avoid unnecessary processing resource consumption when the correlation is weak. Third, it determines the screening strategy by classifying the recognition matching weight values, and dynamically adjusts the screening strictness when the coverage and correlation are different, thereby achieving a precise match between the screening strategy and the degree of improvement in multi-source recognition.

[0093] Continuing with the calculation results of S1, 5 out of the 10 standardized template feature vectors are target type feature vectors (V2, V4, V6, V7, V9). The total number of deviations in the AI ​​model during historical recognition processing is 62, including 8 deviations for V2, 12 for V4, 6 for V6, 15 for V7, and 9 for V9. The sum of the number of deviations under the target type feature vectors is 8 + 12 + 6 + 15 + 9 = 50.

[0094] Step S21: Recognition ratio = 5 / 10 = 0.50.

[0095] Determine if the recognition ratio of 0.50 is less than the preset recognition ratio threshold of 0.40: If 0.50 is not less than 0.40, proceed to S22.

[0096] Step S22: Matching deviation ratio = 50 / 62 = 0.806.

[0097] Determine if the matching deviation ratio of 0.806 is less than the preset deviation ratio threshold of 0.70: If 0.806 is not less than 0.70, proceed to S23.

[0098] Step S23: Identify matching weight value = (0.50 + 0.806) / 2 = 1.306 / 2 = 0.653.

[0099] Determine whether the identification matching weight value of 0.653 is greater than the preset weight threshold of 0.60: If 0.653 is greater than 0.60, the identification strategy for the feature analysis target is determined to be Strategy B, that is, as long as the similarity coefficient between the image identification target and the target type feature vector is greater than the preset similarity coefficient threshold of 0.85, and there is an identification deviation of the image, then it is regarded as the feature analysis target.

[0100] It is understandable that the method for determining the image recognition processing method under the target type feature vector is as follows: In this embodiment, based on the number of updated targets under different target type vectors, the reliability of identification of the similarity distribution interval under different features under different target type vectors is determined. Based on the reliability of identification of the similarity distribution interval under different target type vectors and different features, the image recognition processing method under target type feature vectors is determined. Thus, while ensuring the overall recognition reliability, timely updating processing of features that do not meet the similarity requirements is also achieved, reducing the difficulty of similarity analysis and processing.

[0101] It is understandable that the method for determining the image recognition processing method under the target type feature vector is as follows: In this embodiment, based on the number of updated targets under different target type feature vectors, the reliability of identification of the similarity distribution interval under different features under different target type feature vectors is determined. Based on the reliability of identification of the similarity distribution interval under different target type feature vectors, the image recognition processing method under target type feature vectors is determined. Thus, while ensuring the overall recognition reliability, timely updating processing of features that do not meet the similarity requirements is also achieved, reducing the difficulty of similarity analysis and processing.

[0102] The image recognition processing method refers to the decision rules for determining which features belong to the matching analysis features and which do not when performing image frame recognition under the target type feature vector; the matching analysis features refer to the feature dimensions that need to be focused on for similarity comparison during the image recognition process; the number of updated feature analysis targets refers to the number of feature analysis targets that are updated or adjusted under the target type feature vector.

[0103] If a large number of feature analysis targets are updated under a certain target type feature vector, it indicates that the feature distribution under that feature vector is relatively active, and more stringent feature exclusion conditions need to be adopted to reduce the processing difficulty of similarity analysis.

[0104] The activity level of the feature distribution is assessed by analyzing the number of target updates, and the image recognition processing method is determined accordingly. The strictness of feature exclusion is dynamically adjusted based on the stability of the feature distribution, thereby reducing the processing difficulty of similarity analysis while ensuring the reliability of recognition.

[0105] S31 determines the number of updated features of the target type under the feature analysis target based on the update results of the feature analysis target under different target type feature vectors; The update count of the feature analysis target refers to the cumulative number of times the feature analysis target set has been updated and adjusted under a certain target type feature vector.

[0106] If the feature analysis target under a certain target type feature vector has been adjusted multiple times recently, the large number of updates indicates that the feature distribution under that feature vector is relatively active.

[0107] By analyzing the number of target updates using statistical features, the activity level of feature distribution is quantified into a comparable count, providing basic data for determining the weight values ​​for subsequent updates.

[0108] S32 determines the update recognition weight value under the target type feature vector based on the updated number of the target under the feature analysis of the target type feature vector; The update recognition weight value under the target type feature vector is determined according to the preset recognition weight value corresponding to the number range of the update quantity of the target under the target type feature vector. The more updates, the higher the update recognition weight value.

[0109] If the number of updated target features under a certain target type feature vector is in a high range, the corresponding preset recognition weight value is high, indicating that the feature distribution under this feature vector is more active, and more stringent feature exclusion conditions need to be adopted.

[0110] By updating the quantity mapping to update the identification weight value, discrete update counts are transformed into comparable weight values, providing a quantitative basis for determining the classification of subsequent image tag identification processing methods.

[0111] S33 utilizes the updated recognition weight values ​​under different target type feature vectors to determine the image recognition processing method under the target type feature vector.

[0112] It is understandable that the image recognition processing method under the target type feature vector is determined by updating the recognition weight values ​​under different target type feature vectors, specifically including: S331, using the updated identification weight values ​​under different target type feature vectors, determine whether there are target type feature vectors whose updated identification weight values ​​are greater than a preset weight threshold. If yes, proceed to step S332. If no, determine that the image recognition processing method under the target type feature vector is as follows: if the number of feature analysis targets under the target type feature vector is greater than the feature analysis target number threshold, and the fluctuation range of the similarity coefficient under the feature does not meet the requirements, then determine that the feature does not belong to the matching analysis feature. The preset weight threshold refers to the critical value for determining whether the updated identification weight value is too high. When there are no target type feature vectors with weight values ​​exceeding the threshold, the feature distribution of all feature vectors is relatively stable, and the strictest feature exclusion condition is adopted.

[0113] Assuming that the updated recognition weight values ​​of all target type feature vectors do not exceed the preset weight threshold, then method A is adopted. The feature is excluded only when the number of target objects for feature analysis exceeds the maximum threshold for the number of target objects for feature analysis and the fluctuation range of the similarity coefficient does not meet the requirements.

[0114] When there are no high update weight values, the strictest feature exclusion strategy is adopted. When the feature distribution is relatively stable, more sufficient feature analysis targets and more significant fluctuation anomalies are required to exclude features, thereby avoiding feature mis-exclusion due to insufficient data.

[0115] S332 determines whether the updated recognition weight values ​​of different target type feature vectors are all greater than the preset weight threshold. If so, the image recognition processing method under the target type feature vector is determined to be that the feature does not belong to the matching analysis feature if the number of feature analysis targets under the target type feature vector is greater than the second feature analysis target number threshold and the fluctuation range of the similarity coefficient under the feature does not meet the requirements. If not, the image recognition processing method under the target type feature vector is determined to be that the feature does not belong to the matching analysis feature if the number of feature analysis targets under the target type feature vector is greater than the third feature analysis target number threshold and the fluctuation range of the similarity coefficient under the feature does not meet the requirements.

[0116] The second threshold for the number of target features analyzed is less than the threshold for the number of target features analyzed, and the third threshold for the number of target features analyzed is less than the threshold for the number of target features analyzed but greater than the threshold for the number of target features analyzed. When the updated recognition weight value is high, the most lenient feature exclusion condition (minimum second threshold) is used; when the weight value is low, a moderately strict exclusion condition (middle third threshold) is used.

[0117] If the updated recognition weight values ​​of all target type feature vectors are greater than the preset weight threshold, indicating that the feature distribution is active, then method B is adopted, and features are excluded only if the number of target analysis objects exceeds the smaller second threshold and the fluctuation does not meet the requirements. If the weight value is not greater than the preset weight threshold, then method C is adopted, and features are excluded only if the number of target analysis objects exceeds the middle third threshold and the fluctuation does not meet the requirements.

[0118] By updating the identification weight values ​​and judging each feature individually and determining the feature exclusion conditions in a hierarchical manner, the significance lies in using more lenient exclusion conditions when the feature distribution is active to quickly eliminate unstable features, and using moderately strict exclusion conditions when the feature distribution is stable to retain more features for analysis, thereby achieving a precise match between the feature exclusion strategy and the activity level of the feature distribution.

[0119] It should be noted that the fluctuation range of the similarity coefficient under the aforementioned feature is the fluctuation range of the similarity coefficient between the target and the target type feature vector under the aforementioned feature for accurate feature analysis of the target identification result.

[0120] The fluctuation range does not meet the requirements when the difference between the endpoints of the fluctuation range is greater than the preset fluctuation difference threshold. In other words, when the similarity coefficient fluctuates greatly, it indicates that the similarity analysis of the feature is unstable.

[0121] The preset fluctuation difference threshold is determined by the following reference factors: the number of feature dimensions, the distribution of typical fluctuation ranges, and the level of recognition reliability requirements.

[0122] This embodiment realizes a complete process from the statistical analysis of the number of updated targets to the determination of the updated identification weight value and the determination of the image recognition processing method. Its core value is reflected in three aspects: First, it assesses the activity level of feature distribution by updating the number of targets, providing a quantitative basis for feature exclusion strategies; second, it dynamically adjusts the strictness of feature exclusion under different activity levels by using two-level analysis of the existence judgment and individual judgment of the updated identification weight value; and third, through the hierarchical design of the three-level threshold of the number of targets for feature analysis, it achieves accurate matching of feature exclusion strategies when the activity level of feature distribution is different, thereby reducing the processing difficulty of similarity analysis while ensuring the reliability of recognition.

[0123] Continuing with the calculation results of S2, the feature vectors for the five target types are V2, V4, V6, V7, and V9. The number of feature analysis target updates under each target type feature vector are as follows: V2 updated 3 times, V4 updated 8 times, V6 updated 2 times, V7 updated 10 times, and V9 updated 5 times.

[0124] The mapping relationship between the update quantity range and the preset recognition weight value is as follows: 0 to 3 times corresponds to 0.3, 4 to 7 times corresponds to 0.5, and more than 8 times corresponds to 0.8.

[0125] Step S31: The number of updates for each target type feature vector is as described above.

[0126] S32 steps: V2 updates 3 times → weight value 0.3; V4 updates 8 times → weight value 0.8; V6 updates 2 times → weight value 0.3; V7 updates 10 times → weight value 0.8; V9 updates 5 times → weight value 0.5.

[0127] Step S331: Determine if there is a target type feature vector with an updated identification weight value greater than the preset weight threshold of 0.5: V4 weight value 0.8>0.5, V7 weight value 0.8>0.5, if present, proceed to S332.

[0128] Step S332: Determine whether the updated recognition weight value of the feature vector of each target type is greater than the preset weight threshold of 0.5. If not, then the number of target analysis targets is greater than the third threshold of 12 and the fluctuation does not meet the requirement → exclude the feature.

[0129] Specifically, the method for determining the update target of the target type feature vector is as follows: In this embodiment, by utilizing the matching analysis feature data under different target type feature vectors and the fluctuation range of the similarity coefficient under the matching analysis features, the accuracy of the identification and analysis processing of the image tag under different target type feature vectors combined with features of different dimensions is determined. This accuracy in identifying and analyzing the image tag under different target type feature vectors combined with features of different dimensions determines the update target of the target type feature vector, thereby further realizing the update and identification processing of the target type feature vector and ensuring the reliability of the image tag identification and processing.

[0130] The update target refers to the standardized template feature vector that needs to be added or adjusted, and is included or removed from the target type feature vector set; the matching analysis feature data refers to the feature dimension data that is determined as matching analysis features in the process of image recognition.

[0131] If the number of matching analysis features under a certain target type feature vector is large and the similarity coefficient fluctuates within a small range, it indicates that the multidimensional feature recognition accuracy under this feature vector is high, and therefore no update is required. Otherwise, a new standardized template feature vector needs to be incorporated into the target type feature vector.

[0132] By analyzing feature data and fluctuation ranges through matching, the accuracy of multidimensional feature recognition is evaluated, and the update target is determined accordingly. This is significant because when the recognition accuracy is insufficient, the set of feature vectors for the target type is adjusted in a timely manner, thereby ensuring the continuous reliability of the image and tag recognition process.

[0133] S41, using the matching analysis feature data under different target type feature vectors, determine the number of matching analysis features under different target type feature vectors; The number of matching analysis features refers to the number of feature dimensions that are determined as matching analysis features under a certain target type feature vector.

[0134] If multiple feature dimensions are identified as matching analysis features under a certain target type feature vector, then a large number of matching analysis features indicates that the multidimensional feature recognition capability under this feature vector is strong.

[0135] By statistically analyzing the number of features, the multidimensional feature recognition capability of feature vectors for each target type is quantified, providing basic data for determining subsequent updated targets.

[0136] It is understandable that in the above steps, the average number of matching analysis features under different target type feature vectors is obtained, and it is determined whether the average number of matching analysis features under different target type feature vectors is less than the preset matching feature number threshold. If so, it is determined that it is difficult to effectively improve the recognition reliability of the frame under the target type feature vector by using matching analysis features. Therefore, the update target of the target type feature vector is determined to be the standardized template feature vector with a recognition deviation probability greater than the third deviation probability threshold. If not, proceed to step S42. The preset threshold for the number of matching features refers to a critical value used to determine whether the average number of matching analysis features is sufficient. When the average is lower than the threshold, it indicates that the matching analysis features are insufficient to effectively improve the reliability of recognition, and more standardized template feature vectors need to be included in the target type.

[0137] If the average number of matching analysis features under the feature vectors of each target type is low, indicating that the overall multidimensional feature recognition capability is insufficient, then the standardized template feature vector with a deviation probability exceeding the third deviation probability threshold (less than the second deviation probability threshold) will be used as the update target.

[0138] When the number of matching features is insufficient, an expansion and update strategy is adopted. When the multidimensional feature recognition capability is insufficient, more feature vectors with higher risk of deviation are included in the target type to expand the analysis scope, thereby improving the overall recognition reliability.

[0139] S42, based on the fluctuation range of the similarity of the target type feature vector under different matching analysis features, determine the fluctuation range of the similarity coefficient under the matching analysis features, and use the fluctuation range to determine the identification matching weight value of the matching analysis features; The identification matching weight value of the matching analysis feature is determined based on the preset matching weight value corresponding to the deviation between the endpoints of the fluctuation range, wherein the larger the deviation, the smaller the identification matching weight value.

[0140] If the endpoint deviation of the similarity coefficient fluctuation range under a certain matching analysis feature is small, the corresponding identification matching weight value is high, indicating that the similarity analysis of this feature is relatively stable and reliable. Conversely, if the deviation is large, the weight value is low.

[0141] By mapping the deviation of the fluctuation range endpoints to identify matching weight values, the degree of fluctuation of the similarity coefficient is quantified into comparable weight values, providing a quantitative basis for determining the subsequent update target.

[0142] S43, determine the update target of the target type feature vector based on the number of matching analysis features under different target type feature vectors and the identification matching weight values ​​of different matching analysis features.

[0143] Furthermore, based on the number of matching analysis features under different target type feature vectors and the identification matching weight values ​​of different matching analysis features, the update target of the target type feature vector is determined, specifically including: S431, determine the identification matching coefficient under the target type feature vector based on the number of matching analysis features under the target type feature vector and the identification matching weight value of different matching analysis features, and determine whether there is a target type feature vector with an identification matching coefficient greater than a preset matching coefficient threshold. If yes, proceed to step S432. If no, determine that the update target of the target type feature vector is a standardized template feature vector with an identification deviation probability greater than a third deviation probability threshold. The matching coefficient is calculated as follows: Matching coefficient = (Number of all matching analysis features / Number of all features) / Average of the maximum matching weight value. The more matching analysis features and the larger the maximum matching weight value, the higher the matching coefficient, indicating a stronger multi-dimensional feature recognition capability under the target type feature vector.

[0144] The preset matching coefficient threshold refers to the critical value for judging whether the matching coefficient is high. When there are no target type feature vectors with coefficients exceeding the threshold, it indicates that the multi-dimensional feature recognition capability of all target type feature vectors is insufficient, and more feature vectors need to be included in the updated target.

[0145] Assuming that the matching coefficients of all target type feature vectors do not exceed the preset matching coefficient threshold, the standardized template feature vectors whose deviation probability exceeds the third deviation probability threshold and whose deviation count exceeds the preset deviation count threshold are used as update targets. The deviation count condition is added to accurately filter the feature vectors that really need to be focused on.

[0146] When a high matching coefficient is not available, an update strategy with additional deviation count conditions is adopted. When the overall multidimensional feature recognition capability is insufficient, the feature vectors that need to be included in the target type are accurately screened by the dual conditions of deviation probability and deviation count, thereby avoiding mis-inclusion caused by screening only by deviation probability.

[0147] S432 determines the basic identification matching weight value based on the identification matching coefficients of different target type feature vectors and the proportion of target type feature vectors in the standardized template feature vectors. It then determines whether the basic identification matching weight value is greater than the preset matching weight value. If so, it determines that the update target of the target type feature vector is that no update processing of the target type feature vector is required. If not, it determines that the update target of the target type feature vector is the standardized template feature vector whose identification deviation probability is greater than the third deviation probability threshold and whose identification deviation count is greater than the preset deviation count threshold.

[0148] The basic identification matching weight value is calculated as follows: Basic identification matching weight value = Average identification matching coefficient + Proportion of target type feature vector. When both the average identification matching coefficient and the proportion are high, it indicates that the overall multi-dimensional feature recognition capability is strong, and no update is required.

[0149] The preset matching weight value refers to the critical value for determining whether the basic recognition matching weight value is sufficient. When the weight value exceeds the preset value, it indicates that the overall recognition capability is sufficient and no update is required. Otherwise, feature vectors with a deviation probability exceeding the third deviation probability threshold and a recognition deviation count greater than the preset deviation count threshold should be included in the update target.

[0150] If the calculated basic recognition matching weight value is greater than the preset matching weight value, it indicates that the overall multi-dimensional feature recognition capability is sufficient, and no update processing is required. If the weight value is not greater than the preset matching weight value, the standardized template feature vector with a deviation probability exceeding the third deviation probability threshold and a recognition deviation number greater than the preset deviation number threshold will be used as the update target.

[0151] The significance of further refining the update strategy by using basic identification matching weight values ​​is that when there is a high identification matching coefficient but the overall identification ability is insufficient, it is possible to determine whether an update is needed based on a comprehensive evaluation of the average and proportion of the identification matching coefficients, thereby ensuring the reliability of identification while avoiding unnecessary update processing.

[0152] It is understandable that if the identification matching weight value of the matching analysis feature of the target type feature vector meets the requirements, then when performing the identification analysis of the image, in addition to considering the similarity of the template feature vector, it is also necessary to consider whether the similarity of the matching analysis feature that meets the requirements is within the fluctuation range. Only when the fluctuation range is within the requirements and the similarity meets the requirements can the identification result be determined to be accurate.

[0153] If a certain matching analysis feature has a high identification matching weight value, then in addition to comparing the similarity of the template feature vector, it is also necessary to compare whether the similarity of the matching analysis feature is within the fluctuation range when identifying the image. Only when both meet the requirements can the identification result be determined to be accurate.

[0154] Incorporating matching analysis features into the recognition result confirmation criteria, and improving the accuracy of the recognition results through dual verification of multi-dimensional features, thereby avoiding misjudgments where the similarity of a single feature meets the standard but there are actual deviations.

[0155] This embodiment realizes a complete process from statistical analysis of the number of matching features to determination of matching weight values ​​and then to target type feature vector update target determination. Its core value is reflected in three aspects: First, it evaluates the overall multi-dimensional feature recognition capability by using the average number of matching analysis features, and adopts an expansion update strategy when the capability is insufficient; second, it dynamically adjusts the strictness of the update strategy when the recognition capability is different by judging the existence of the matching coefficient and the multi-layer analysis of the basic identification matching weight values; third, it achieves accurate screening of update targets by adding deviation frequency conditions and proportions for comprehensive evaluation, thereby ensuring the reliability of recognition while avoiding unnecessary update processing.

[0156] This embodiment implements a signature recognition processing method based on multi-source feature fusion and AI through the complete process from S1 to S4. S1 determines the target type feature vectors requiring multi-dimensional feature similarity analysis by using a four-layer threshold grading screening based on the probability of recognition deviation, focusing recognition resources on feature vectors with higher deviation risk. S2 determines a differentiated recognition strategy for feature analysis targets through multi-layer judgments of recognition ratio and matching deviation ratio, achieving an adaptive balance between recognition reliability and processing efficiency. S3 determines the strictness of feature exclusion in the signature recognition processing method through two-level analysis of the existence judgment and individual judgment of updated recognition weight values, reducing the processing difficulty of similarity analysis while ensuring recognition reliability. S4 determines the dynamic update target of the target type feature vector through multi-layer judgments of the number of matching analysis features and recognition matching coefficient, achieving continuous optimization of the target type feature vector set. Overall, this method constructs a complete technical closed loop from deviation risk assessment to identification strategy optimization and dynamic update of feature vectors through the coordinated linkage of target type feature vector screening, feature analysis target identification strategy determination, hierarchical determination of image recognition processing methods, and dynamic determination of updated targets. Its core value is reflected in four aspects: First, through hierarchical screening with four layers of deviation probability thresholds, it achieves precise matching between the target type feature vector screening strategy and the deviation risk level; second, through multi-layered judgment of identification matching weight values, it achieves precise matching between the feature analysis target screening strategy and the degree of improvement in multi-source identification; third, through the hierarchical design of three-level feature analysis target quantity thresholds, it achieves precise matching between the feature exclusion strategy and the activity level of feature distribution; and fourth, through multi-layered analysis of identification matching coefficients and basic identification matching weight values, it achieves precise matching between the updated target screening strategy and multi-dimensional feature recognition capabilities.

[0157] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0158] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0159] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A signature recognition and processing system based on multi-source feature fusion and AI, characterized in that, Specifically, it includes: The feature library construction module is responsible for format validation and preprocessing of the imported frame templates, extracting multi-dimensional features of the frame templates, converting them into standardized template feature vectors, and storing them in the template feature library. The vector extraction module is responsible for locating candidate frame regions of the DWG drawing to be identified through contour detection and closed shape recognition, extracting features and converting them into standardized candidate feature vectors. The target frame filtering module is responsible for determining the standardized template feature vector based on multi-dimensional features for similarity analysis and recognition processing in the standardized template feature vector based on the recognition deviation data of the frame under different standardized template feature vectors, and using it as the target type feature vector. Based on the update results of feature analysis targets under different target type feature vectors, the module determines the image recognition processing method under the target type feature vector. Using AI algorithms and image recognition processing methods, the module compares the candidate feature vector with the template feature vector, calculates the similarity, and filters out the target frames. The drawing frame recognition module extracts the drawing vector dimensions and the actual physical dimensions of the template of the target drawing frame, integrates the recognition results of the drawing sheet specifications and the calculation results of the actual physical dimensions of the target drawing frame, and outputs a recognition report.

2. The image signature recognition processing system based on multi-source feature fusion and AI as described in claim 1, characterized in that, The multidimensional features include geometric features, line features, and structural features.

3. The image signature recognition processing system based on multi-source feature fusion and AI as described in claim 1, characterized in that, The target frame is a frame with a similarity coefficient greater than or equal to the similarity coefficient threshold.

4. The image signature recognition processing system based on multi-source feature fusion and AI as described in claim 1, characterized in that, The method for determining the identification result of the map sheet size is as follows: The drawing scale is calculated using a formula. Based on the drawing vector size of the target drawing frame and the drawing scale, the actual physical size of the target drawing frame is converted and compared with the standard drawing size library to determine the drawing size specification.

5. A method for image recognition processing based on multi-source feature fusion and AI, applied to the image recognition processing system based on multi-source feature fusion and AI as described in any one of claims 1-4, characterized in that, Specifically, it includes: Using the identification data of the image tag, the identification deviation data of the image frame under different standardized template feature vectors is determined. Using the identification deviation data of the image frame under different standardized template feature vectors, the standardized template feature vectors that are analyzed and identified based on multi-dimensional features are determined, and they are used as the target type feature vectors. Based on the target type feature vector and the degree of correlation between the recognition deviation data of the frame under the target type feature vector and the recognition deviation data of the model, the recognition strategy of feature analysis target in the tag recognition target is determined; Based on the update results of feature analysis targets under different target type feature vectors, the image recognition processing method under the target type feature vectors is determined; The image recognition processing method is used to determine the matching analysis features of the target under the target type feature vector. By utilizing the matching analysis features under different target type feature vectors and the fluctuation range of similarity under different matching analysis features, the updated target of the target type feature vector is determined.

6. The image signature recognition processing method based on multi-source feature fusion and AI as described in claim 5, characterized in that, The recognition deviation data of the frame under the standardized template feature vector is determined based on the number of recognition deviations of the frame under the standardized template feature vector.

7. The image signature recognition processing method based on multi-source feature fusion and AI as described in claim 5, characterized in that, The method for determining the target type feature vector is as follows: The number of recognition deviations of the frame under the standardized template feature vector is determined by the recognition deviation data of the frame under different standardized template feature vectors. The recognition deviation probability of the frame under the standardized template feature vector is determined based on the number of recognition deviations of the frame under the standardized template feature vector. The target type feature vector in the standardized template feature vector is determined by the recognition deviation probability of the frame under different standardized template feature vectors.

8. The image signature recognition processing method based on multi-source feature fusion and AI as described in claim 7, characterized in that, The recognition deviation probability of the frame under the standardized template feature vector is determined based on the ratio of the number of recognition deviations of the frame under the standardized template feature vector to the number of recognition processing operations.

9. The image signature recognition processing method based on multi-source feature fusion and AI as described in claim 5, characterized in that, The method for determining the identification strategy of the feature analysis target in the image recognition target is as follows: Based on the target type feature vector, determine the proportion of the target type feature vector in the standardized template feature vector, and use it as the recognition proportion; Based on the correlation between the recognition deviation data of the bounding box under the target type feature vector and the recognition deviation data of the model, the proportion of the recognition deviation number under all target type feature vectors in the total recognition deviation number of the model is determined, and this proportion is used as the matching deviation ratio. Based on the recognition ratio and the matching deviation ratio, the identification strategy for the feature analysis target in the image recognition target is determined.

10. The image signature recognition processing method based on multi-source feature fusion and AI as described in claim 5, characterized in that, The method for determining the update target of the target type feature vector is as follows: By using matching analysis feature data under feature vectors of different target types, the number of matching analysis features under feature vectors of different target types can be determined; Based on the fluctuation range of the similarity of the target type feature vector under different matching analysis features, the fluctuation range of the similarity coefficient under the matching analysis features is determined, and the identification matching weight value of the matching analysis feature is determined using the fluctuation range. The update target of the target type feature vector is determined by the number of matching analysis features under different target type feature vectors and the identification matching weight values ​​of different matching analysis features.

Images