Image recognition-based leather product comparison identification method and system

Abstract

The application relates to the technical field of image recognition, and particularly discloses a leather product comparison and identification method and system based on image recognition. The application extracts a multi-dimensional standardized feature vector reflecting inherent physical and mechanical properties by collecting dynamic deformation videos of a to-be-inspected leather and a reference sample under controlled micro force, receives identification task description parameters, maps core discriminant features from a physical and mechanical feature knowledge base, dynamically instantiates a self-adaptive identification model through a meta-learning model, generates a feature weighting scheme and a dynamic decision threshold, calculates a mechanical feature matching degree and generates a visual preliminary report, adjusts the scheme in combination with user interactive correction instructions, updates the result and feeds back data optimization meta-models. The application upgrades the identification basis to essential mechanical properties, realizes task self-adaptive decision and man-machine collaborative optimization, can improve identification reliability and scene adaptability, makes the decision process transparent and interpretable, has a continuous optimization capability, and is suitable for multi-class requirements such as authenticity identification and traceability.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method and system for comparing and identifying leather products based on image recognition. Background Technology

[0002] With the widespread application of leather products in luxury goods, high-end apparel, and industrial components, the market demand for authentication, traceability, and damage assessment of leather products is increasing. Currently, image recognition-based comparison and authentication of leather products is the mainstream technical solution in the industry. Its typical implementation process is as follows: First, digital images of the leather product to be inspected are acquired under specific lighting and shooting conditions. Then, image processing algorithms such as local binary mode and convolutional neural networks are used to extract visual feature vectors such as texture, color, and pattern from the image. Next, the similarity between the feature vector of the sample to be inspected and the feature vector of the reference sample in the database is calculated. Finally, based on a preset fixed threshold, it is determined whether the two match, thus completing the authentication.

[0003] However, existing technologies have some shortcomings in practical applications, making it difficult to meet the needs of accurate identification in complex scenarios. They mainly rely on visual features such as texture and color, which are image appearance attributes rather than inherent physical and mechanical properties of leather. These features are easily affected by external factors such as light intensity, shooting angle, wear of the leather surface coating, and slight deformation, causing the same leather product to appear differently under different conditions, directly affecting the accuracy of the identification results. Furthermore, the decision-making strategies of existing technologies lack adaptability. Existing technologies use a globally uniform feature weight allocation method and a fixed similarity judgment threshold, which cannot adapt to diverse identification tasks. For example, authenticity identification needs to focus on the mechanical properties corresponding to the core process, while damage assessment needs to focus on deformation recovery ability. A single static strategy leads to a high misjudgment rate in complex scenarios.

[0004] Therefore, there is an urgent need for image recognition-based comparison and identification methods and systems for leather products to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a method for comparing and identifying leather products based on image recognition, comprising the following steps: Videos of the dynamic surface deformation process of the leather products to be inspected and the reference sample under controlled micro-force were collected. The video of the dynamic surface deformation process is processed to extract a multi-dimensional standardized feature vector that reflects the inherent physical and mechanical properties of the leather. The system receives user-defined identification task description parameters, which include the identification target type and the comparison strictness mode. Based on the identification target type, it determines at least one corresponding core discriminant feature from a predefined physical and mechanical feature knowledge base. Based on the identification task description parameters and the core discriminant features, an adaptive identification model is dynamically generated. The adaptive identification model is used to generate a feature weight vector according to the core discriminant features, to weight the multi-dimensional standardized feature vector, and to generate a decision threshold that matches the comparison strictness mode. Using the adaptive identification model, the mechanical feature matching degree between the product to be inspected and the reference sample is calculated, and a preliminary identification report containing the mechanical feature matching degree and the mechanical feature visualization comparison information related to the core discrimination feature is generated; The system receives interactive correction instructions from the user based on the preliminary identification report, adjusts the feature weight vector according to the interactive correction instructions and updates the identification results, and feeds the adjusted parameters back to the model for optimization.

[0006] Furthermore, the steps to obtain multi-dimensional standardized feature vectors include: Based on the video of the dynamic surface deformation process, the displacement field sequence and synchronous micro-force data of the monitoring points on the leather surface are obtained, and pressure-displacement time-varying relationship data are generated. Extract the time-domain mechanical response index set and the spatial domain mechanical distribution index set from the pressure-displacement time-varying relationship data; By fusing the time-domain mechanical response index set and the spatial-domain mechanical distribution index set, an original multi-dimensional feature vector is constructed. The original multi-dimensional feature vector is normalized to obtain the multi-dimensional standardized feature vector.

[0007] Furthermore, the time-domain mechanical response index set includes at least one of the following: maximum indentation depth, residual deformation depth, time required for a specified rebound ratio, and average rebound rate; and the spatial domain mechanical distribution index set includes at least one of the following: average strain value and strain distribution non-uniformity.

[0008] Furthermore, the steps for determining the core discriminative features include: The identification task description parameters are analyzed to obtain the identification target type and comparison strictness mode, and the use classification identifier of the leather products to be inspected is obtained. The product use sensitivity coefficient is obtained based on the aforementioned use classification identifier; Using the target type as an index, an initial candidate feature set is obtained from the physical and mechanical feature knowledge base; By combining the urgency parameter of the identification scenario with the sensitivity coefficient of the product's use, the features in the initial candidate feature set are dynamically weighted and sorted. The range of the number of candidate features is determined based on the comparison strictness mode, and the confidence level of the feature discrimination power of each candidate feature is calculated in combination with historical feedback data. Based on the dynamic weighted ranking result and the confidence level of the feature discrimination power, at least one core discriminative feature is selected from the initial candidate feature set.

[0009] Furthermore, the specific steps for dynamically generating the adaptive identification model include: The identification task description parameters and the core discriminant features are used as input conditions. Based on the input conditions, initialization parameters are generated, including the feature weight vector and a strictness bias, wherein the weight elements in the feature weight vector corresponding to the core discriminative features are initialized as dominant weight values. The decision threshold is calculated based on the strictness bias and historical statistical information related to the identification target type. The adaptive identification model is instantiated using the initialization parameters and the decision threshold.

[0010] Furthermore, the specific steps for generating a preliminary identification report include: The feature weight vector is applied to calculate the weighted multi-dimensional standardized feature vectors of the product to be inspected and the reference sample respectively, so as to obtain the weighted feature vectors of the product to be inspected and the reference feature vectors. The similarity between the weighted feature vector to be detected and the reference feature vector is calculated and used as the mechanical feature matching degree. The mechanical feature matching degree is compared with the decision threshold to generate a preliminary text identification conclusion; Based on the core discriminative features, a visual comparison is generated that includes numerical comparison charts and feature distribution difference maps; The preliminary identification report is formed by integrating the mechanical feature matching degree, preliminary text identification conclusions, and visual comparison content.

[0011] Furthermore, the step of adjusting the feature weight vector and updating the identification result according to the interactive correction instruction includes: Display the preliminary identification report and receive the interactive correction instructions via interactive controls; In response to the weight adjustment command, the feature weight vector is updated in real time and the identification result is recalculated and updated. In response to the result overwrite instruction, the user-specified result is recorded as the final identification conclusion; The identification task description parameters, the final feature weight vector, and the final identification conclusion of this task are encapsulated to form a collaboratively optimized sample. The collaboratively optimized samples are used for incremental training and optimization of the model.

[0012] Furthermore, the present invention also discloses an image recognition-based leather product comparison and identification system, comprising: The acquisition module is used to acquire video of the dynamic surface deformation process of the leather product under inspection and the reference sample under controlled micro-force. The data processing module is used to process the video of the dynamic surface deformation process and extract a multi-dimensional standardized feature vector that reflects the inherent physical and mechanical properties of the leather. The receiving module is used to receive user-defined identification task description parameters, which include identification target type and comparison strictness mode, and determine at least one corresponding core discriminant feature from a predefined physical and mechanical feature knowledge base according to the identification target type. The generation module is used to dynamically generate an adaptive identification model based on the identification task description parameters and the core discriminant features. The adaptive identification model is used to generate a feature weight vector based on the core discriminant features to weight the multi-dimensional standardized feature vector and generate a decision threshold that matches the comparison strictness mode. The report acquisition module is used to calculate the mechanical feature matching degree between the product to be inspected and the reference sample using the adaptive identification model, and generate a preliminary identification report containing the mechanical feature matching degree and the mechanical feature visualization comparison information related to the core discrimination feature; The adjustment and optimization module is used to receive interactive correction instructions from the user based on the preliminary identification report, adjust the feature weight vector and update the identification result according to the interactive correction instructions, and feed the adjusted parameters back to the model for optimization.

[0013] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described image recognition-based leather product comparison and identification method.

[0014] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described image recognition-based leather product comparison and identification method.

[0015] The beneficial effects of this application are as follows: Firstly, this invention can elevate the identification criteria from unstable visual appearance features to stable and essential physical and mechanical properties. By collecting dynamic deformation videos under controlled micro-force, it extracts standardized feature vectors that reflect the feel and texture of leather, fundamentally eliminating the interference of external factors such as lighting, coating, and shooting angle. This significantly reduces the misjudgment rate caused by different appearances of the same object, and substantially improves the accuracy and reliability of the identification results.

[0016] Secondly, this invention addresses the shortcomings of rigid decision-making logic in existing technologies by introducing a meta-learning model. This model receives structured task description parameters and dynamically generates an adaptive identification model, including feature weight configuration adapted to the current task, similarity fusion algorithm, and dynamic decision threshold. This enables it to accurately serve various differentiated needs such as brand authenticity identification, process traceability, and damage assessment. There is no need to develop separate algorithms for different tasks, and the versatility and scenario adaptability are greatly improved. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a method flow proposed in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the system structure proposed in an embodiment of the present invention.

[0019] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] like Figure 1 As shown, this application provides a method for comparing and identifying leather products based on image recognition, including the following steps: S1, collect video of the dynamic surface deformation process of the leather product to be inspected and the reference sample under controlled micro-force; S2, analyze and process the video of the dynamic surface deformation process, and extract a multi-dimensional standardized feature vector that reflects the inherent physical and mechanical properties of leather. Each dimension of the multi-dimensional standardized feature vector is associated with a feature entry in a predefined physical and mechanical feature knowledge base. S3, receive the identification task description parameters input by the user. The identification task description parameters are used to define the identification target type and the comparison strictness mode of this comparison, and map at least one corresponding core discriminant feature from the physical and mechanical feature knowledge base according to the identification target type. S4, Based on the identification task description parameters and the core discriminant features, an adaptive identification model is dynamically instantiated through a meta-learning model. The adaptive identification model is used to generate a feature weight vector according to the core discriminant features. The elements in the feature weight vector that correspond to the core discriminant features in the multi-dimensional standardized feature vector are assigned dominant weight values ​​to form a weighting scheme for the standardized feature vector and generate a dynamic decision threshold that matches the comparison strictness mode. S5. Using the adaptive identification model, calculate the mechanical feature matching degree between the product to be inspected and the reference sample, and generate a preliminary identification report that includes the mechanical feature matching degree and a visual comparison of mechanical features related to the core discrimination features. S6, receive the user's interactive correction instruction based on the preliminary identification report, adjust the weighting scheme in real time according to the interactive correction instruction, update the identification result, and feed back the adjusted feature weight vector and the corresponding identification task description parameters to the meta-learning model for optimization.

[0022] As described in steps S1-S6 above, the physical and mechanical properties of leather are its inherent characteristics, which are not affected by external factors such as lighting, shooting angle, and surface coating. Different identification tasks have different focuses on features, and expert experience can effectively correct algorithm deviations. Existing identification methods based on image visual features rely on apparent attributes and often use fixed weights and thresholds, which cannot adapt to diverse tasks and the decision-making process is not transparent. This invention specifically constructs a complete process for adaptive modeling and human-machine collaborative optimization of mechanical feature extraction tasks, which solves the systemic defects of traditional methods.

[0023] Traditional methods, centered on image similarity calculation, are susceptible to environmental interference in extracting features such as texture and color. Fixed strategies cannot meet the diverse needs of tasks like authenticity verification and damage assessment, and simply outputting matching scores lacks interpretability, making it difficult to incorporate expert experience. This invention, through an integrated solution of acquiring essential features, dynamic model construction, and interactive correction, forms a closed loop from data acquisition to model optimization, comprehensively improving the reliability, adaptability, and credibility of the identification process.

[0024] This invention acquires dynamic deformation videos of leather products under controlled micro-force, extracts standardized feature vectors reflecting the inherent physical and mechanical properties of leather, and dynamically instantiates an adaptive identification model by combining user-defined identification task description parameters. Through human-computer interaction correction and incremental model optimization, it achieves highly reliable, interpretable, and adaptable leather product comparison and identification for different scenarios.

[0025] The core logic and principle of this application's technical solution is to address the shortcomings of existing leather identification methods based on visual appearance features, such as unstable features, rigid decision-making, and uninterpretable processes. It focuses on obtaining the inherent physical and mechanical properties of leather. The main body employs a closed-loop process of data acquisition, feature extraction, task adaptation, model instantiation, report generation, and interactive optimization to achieve highly reliable, adaptive, and interpretable comparative identification. The logical chain begins with a data acquisition terminal equipped with a macro lens and a controllable micro-force excitation device, simultaneously acquiring video of the dynamic surface deformation process of the leather product under inspection and a reference sample under the same controlled micro-force. This ensures that the acquired data reflects the essential mechanical response of the leather and avoids optical interference. Next, high-precision motion analysis and strain field calculation are performed on the dynamic video, extracting multi-dimensional mechanical indicators in the time and spatial domains and standardizing them to form a multi-dimensional standardized feature vector associated with a predefined physical and mechanical feature knowledge base. This quantifies the abstract feel and texture of leather into comparable essential features. Subsequently, it receives structured task description parameters from the user, including the identification target type and comparison strictness mode, and then... The target type is accurately mapped from the knowledge base to the core discriminative features, enabling the identification process to focus on key indicators. Then, the task parameters and core discriminative features are input into a pre-trained meta-learning model, dynamically instantiating an adaptive identification model. This model generates a weighted scheme that highlights the weights of the core discriminative features and a dynamic decision threshold for the strictness of matching comparison, achieving precise adaptation of the identification strategy to the current task and breaking the limitations of traditional fixed strategies. Subsequently, the model is used to calculate the mechanical feature matching degree between the product to be inspected and the reference sample, generating a preliminary identification report that includes quantitative matching degree, preliminary conclusions, and a visual comparison of core features, making the identification process transparent and interpretable. Finally, user correction instructions are received through an interactive interface, adjusting the weighted scheme and updating the identification results in real time. At the same time, the task parameters, the adjusted weight vector, and the final conclusion are encapsulated as co-optimized samples and fed back to the meta-learning model for incremental training, enabling the system to continuously evolve during use. It incorporates leather identification experience to correct algorithmic biases and improves the adaptation accuracy of subsequent tasks through knowledge accumulation, ultimately constructing a complete identification logic supported by essential features, adaptive decision-making for tasks, and human-machine collaborative optimization.

[0026] In one embodiment, the step of extracting a multi-dimensional standardized feature vector that reflects the inherent physical and mechanical properties of leather includes: S21, based on the video of the dynamic surface deformation process, obtain the displacement field sequence and synchronous micro-force data of the monitoring points on the leather surface, and generate pressure-displacement time-varying relationship data based on the displacement field sequence and micro-force data; S22, extract the time-domain mechanical response index set and the spatial domain mechanical distribution index set from the pressure-displacement time-varying relationship data. The time-domain mechanical response index set includes the maximum indentation depth, residual deformation depth, time required for a specified rebound ratio, and average rebound rate. The spatial domain mechanical distribution index set includes the average strain value and strain distribution non-uniformity. S23, the time-domain mechanical response index set and the spatial-domain mechanical distribution index set are fused to construct the original multi-dimensional feature vector; S24, normalize the original multi-dimensional feature vector to obtain the multi-dimensional standardized feature vector.

[0027] As described in steps S21-S24 above, this application performs layered processing on the video of dynamic surface deformation process under controlled micro-force, extracts temporal and spatial mechanical indicators and integrates and standardizes them, constructs a multi-dimensional standardized feature vector associated with a physical and mechanical feature knowledge base, and achieves accurate quantification of the inherent physical and mechanical properties of leather.

[0028] The physical and mechanical properties of leather are the essential attributes that determine its texture and feel. They are not affected by external environmental factors such as lighting, shooting angle, and surface coating. However, the diversity of identification tasks requires comprehensive and comparable features. The visual appearance features extracted by existing technologies are greatly affected by the environment and lack multi-dimensional quantification and standardization of the mechanical properties of leather. As a result, the features cannot accurately reflect the essence, and the features between different samples are not directly comparable.

[0029] Traditional methods rely on single-dimensional visual features such as texture and color. These features are easily affected by interference and have a weak correlation with the essential properties of leather. Furthermore, they lack feature standardization, resulting in incomparable feature data from different samples with varying dimensions, and fail to establish a connection with a feature knowledge base, making it difficult to adapt to the core feature selection for subsequent tasks. This step, through multi-dimensional mechanical index extraction and standardization, transforms dynamic video data into comparable and correlated essential feature vectors. This ensures feature stability and provides a foundation for adaptive matching in subsequent tasks.

[0030] The video of the dynamic surface deformation process was recorded by a data acquisition terminal equipped with a macro lens and a controllable micro-force excitation device. During acquisition, the leather under test and the reference sample were subjected to the same controlled micro-force, the magnitude of which was precisely controlled by a miniature pressure feedback probe. The video was processed using a high-precision motion analysis algorithm. Monitoring points uniformly distributed on the leather surface were selected, and the position coordinates of each monitoring point at different times were tracked using inter-frame matching technology to generate a displacement field sequence. Simultaneously, micro-force data recorded by the miniature pressure feedback probe was acquired. The displacement field sequence and micro-force data were aligned according to the timestamp to generate pressure-displacement time-varying relationship data. This process realizes the transformation from visual video to mechanical response data, accurately capturing the dynamic deformation law of the leather after being subjected to force. Compared with traditional visual feature extraction, it ensures the correlation between features and the essential properties of the leather from the source.

[0031] Based on the generated pressure-displacement time-varying relationship data, two sets of mechanical indicators are extracted. In the time-domain mechanical response indicator set, the maximum indentation depth is the maximum displacement of the monitoring point when the pressure reaches a set peak value, reflecting the leather's hardness; the residual deformation depth is the unrecovered displacement of the monitoring point after pressure release, reflecting the leather's plastic deformation capacity; the time required for a specified rebound ratio is the time required for the monitoring point's displacement to recover to a set proportion after pressure release, such as the time required to recover to 90% of the maximum indentation depth, reflecting the leather's rebound speed; and the average rebound rate is the average rate of displacement change at the monitoring point during the pressure release phase, further quantifying the rebound performance. In the spatial domain mechanical distribution indicator set, the average strain value is the arithmetic mean of the strain at all monitoring points, reflecting the overall strain level of the leather; and the strain distribution non-uniformity is the ratio of the standard deviation to the average strain value at each monitoring point, reflecting the spatial consistency of the leather's mechanical properties. These two types of indicators comprehensively characterize the physical and mechanical properties of the leather from both temporal and spatial dimensions, avoiding the limitations of single-dimensional indicators and making the feature vector more representative.

[0032] The extracted temporal mechanical response index set and spatial mechanical distribution index set are combined in a preset order to construct an original multi-dimensional feature vector. For example, the maximum indentation depth, residual deformation depth, time required for a specified rebound ratio, average rebound rate, average strain value, and strain distribution non-uniformity are arranged sequentially to form a six-dimensional original feature vector. This combination process integrates mechanical information from both temporal and spatial dimensions, comprehensively reflecting the inherent properties of leather and ensuring the comprehensiveness of the feature vector.

[0033] The original multi-dimensional feature vectors are processed using a Min-Max normalization algorithm, mapping each dimension's index value to the range of 0 to 1. The normalization formula is: the standardized feature value equals the original feature value minus the minimum value of that feature dimension, then divided by the difference between the maximum and minimum values ​​of that feature dimension. A physical and mechanical feature knowledge base stores the standard ranges and feature entries for various mechanical features of different types of leather. Each dimension of the normalized multi-dimensional standardized feature vector is associated with its corresponding feature entry in the knowledge base. This processing eliminates the dimensional differences between different indicators, enabling direct comparison of features across dimensions. Furthermore, by associating with the knowledge base, the practicality and adaptability of the feature vectors are improved.

[0034] In one embodiment, the step of mapping at least one corresponding core discriminant feature from the physical and mechanical feature knowledge base according to the identification target type includes: S31, receive the structured identification task description parameters through the interactive interface, and simultaneously obtain the use classification identifier of the leather product to be inspected; S32, parse the identification task description parameters to obtain the identification target type and the comparison strictness mode for this comparison, and at the same time query the preset use-feature association library according to the use classification identifier to extract the product use sensitivity coefficient; S33, Using the target type as the index, query the physical and mechanical feature knowledge base to obtain an initial candidate feature set; S34, introduce the identification scenario urgency parameter generated based on the current identification scenario information, and combine it with the product use sensitivity coefficient to dynamically weight and sort the priorities of each feature in the initial candidate feature set; S35. Based on the comparison strictness mode, determine the range of the number of candidate features, and combine the feedback data of feature discrimination effectiveness in historical identification tasks to calculate the feature discrimination effectiveness confidence of each candidate feature. S36. Based on the dynamic weighted sorting result and the feature discrimination power confidence, select the feature that meets the quantity range and has the highest comprehensive score from the initial candidate feature set, and map it to the corresponding at least one core discrimination feature.

[0035] As described in steps S31-S36 above, by receiving the structured identification task description parameters, and combining the purpose of the leather product to be inspected, the urgency of the identification scenario, and historical data, the core discriminative features that match the current task are accurately selected from the physical and mechanical feature knowledge base, providing targeted feature basis for the subsequent adaptive identification model, and ensuring that the identification process is positioned on key indicators.

[0036] Different identification tasks have fundamentally different focuses on leather characteristics. The intended use of the product determines the sensitivity of the features, the urgency of the identification scenario affects the priority of feature selection, and the accuracy of the core discriminative features directly determines the reliability of the identification results. Existing technologies use a fixed feature set for comparison, without distinguishing between task type and product use, treating all features equally. This leads to core features being masked by irrelevant features, increasing computational load and reducing identification accuracy. By using a multi-dimensional parameter fusion and dynamic screening mechanism, the problem of the lack of specificity in traditional feature selection can be addressed.

[0037] Traditional methods neglect the differences in identification tasks and the usage characteristics of artifacts, using a globally uniform feature set for comparison. This fails to focus on core indicators, leading to irrelevant features interfering with the identification results. Furthermore, they do not incorporate historical data to evaluate the effectiveness of features, resulting in insufficient discriminative power among the selected features. The steps described above, through an integrated process of structured task parsing, multi-factor weighted ranking, and confidence level screening, achieve accurate matching of core discriminative features, improving the targeting and efficiency of identification.

[0038] The interactive interface is a structured input interface provided by the system, supporting users to input standardized identification task description parameters. Simultaneously, the system obtains the intended use classification identifier of the leather product under inspection through user input or automatic identification based on the product's appearance features. For example, if a user inputs "authentication of brand A handbag handle," the system automatically identifies the intended use classification identifier as "handbag handle." This ensures the clarity of task information and product intended use, making feature selection more aligned with actual application scenarios and avoiding the lack of specificity caused by generic feature sets.

[0039] The system analyzes the parameters describing the identification task, extracts the target type and comparison strictness mode, and queries a pre-defined usage-feature association library based on the usage classification identifier. This association library is constructed using experimental data from a large number of leather products with different uses, storing the correspondence between various uses and the sensitivity of mechanical features. Based on the query results, usage sensitivity coefficients are extracted for each product; for example, the sensitivity coefficient for the rebound rate of a handbag handle is set to 0.8. This process transforms task requirements and product characteristics into quantitative indicators, providing data support for feature prioritization and ensuring that feature selection better aligns with the actual usage needs of the products.

[0040] The system queries a physical and mechanical feature knowledge base indexed by the target type. This knowledge base contains all predefined physical and mechanical feature entries for leather, quantification standards, and their correlation with the identification task. An initial candidate feature set is quickly obtained based on the index. For example, the initial candidate feature set for authenticity identification includes rebound rate, residual deformation depth, and strain distribution non-uniformity. This approach quickly narrows down the feature range, avoids the computational waste caused by global feature traversal, and ensures the relevance of the initial candidate features to the identification target.

[0041] The urgency parameter for the identification scenario is generated based on scenario information such as the time requirement and importance of the result. For example, the urgency parameter for urgent forensic identification is set to 0.9, and the urgency parameter for routine process comparison is set to 0.6. This parameter is combined with the sensitivity coefficient of the product's intended use, and a weighted summation algorithm is used to dynamically weight and sort the priorities of each feature in the initial candidate feature set. The priority score calculation formula is as follows: ; Among them, the Indicates priority score, the The application sensitivity coefficient is represented by the following. Indicates the use-sensitive weight, the The parameter representing the urgency of the scenario, the The priority score indicates the urgency of the scenario, with higher priority scores indicating higher priority features. By comprehensively considering both scenario and application factors, the feature ranking is made more aligned with actual identification needs, improving the rationality of core feature selection.

[0042] The number of candidate features is determined based on the comparison strictness mode: 2-3 features in high-precision mode, 3-5 features in normal mode, and 4-6 features in lenient mode. Simultaneously, historical authentication task databases are retrieved to extract the discriminative efficacy feedback data of each candidate feature in similar past tasks. The feature discriminative efficacy confidence score is calculated, where the confidence score is the number of correct discriminations divided by the total number of uses. For example, the confidence score for rebound rate in past authenticity authentication tasks is 0.92. Number constraints ensure feature refinement, and confidence score evaluation guarantees the effective discriminative ability of the features, avoiding invalid or redundant features from affecting the authentication results.

[0043] Based on the dynamic weighted ranking results and the confidence level of feature discriminative power, a comprehensive scoring algorithm is used to calculate the comprehensive score of each candidate feature. The formula for calculating the comprehensive score is as follows: ; Among them, the The overall score is represented by Q, which represents the priority score. The weighting coefficients representing the priority scores, the Indicates the confidence level, the This represents the confidence weight coefficient, which selects features from the initial candidate feature set that meet the quantity range and have the highest comprehensive score, and identifies them as core discriminant features. For example, in authenticating the handle of a brand A handbag, the rebound rate and residual deformation depth are selected as core discriminant features. Through multi-dimensional comprehensive evaluation, it is ensured that the core discriminant features not only meet the task requirements and product characteristics, but also possess high discriminant power, providing an accurate basis for the subsequent weight configuration of the adaptive authentication model and improving the reliability of the authentication results.

[0044] In one embodiment, the step of constructing a weighting scheme for the standardized feature vector and generating a dynamic decision threshold that matches the alignment strictness pattern includes: S41, the identification task description parameters and the core discriminative features are used as conditional inputs and input into the meta-learning model; S42, the conditional input is processed by the meta-learning model to generate and output initialization parameters, the initialization parameters including the feature weight vector and a strictness bias, wherein the weight element in the feature weight vector corresponding to the core discriminative feature is initialized as the dominant weight value; S43, the dynamic decision threshold is calculated based on the strictness bias and historical statistical information related to the identification target type; S44. Using the initialization parameters and the dynamic decision threshold, the adaptive identification model is instantiated.

[0045] As described in steps S41-S44 above, by inputting the identification task description parameters and core discriminative features into the pre-trained meta-learning model, a feature weighting scheme and dynamic decision threshold adapted to the current task are dynamically generated. This instantiates the adaptive identification model, enabling the identification strategy to accurately match task requirements and improving the relevance and reliability of the identification results. Different identification tasks have different requirements for feature importance, and the strictness of the comparison directly determines the leniency of the judgment standard. Fixed feature weights and decision thresholds cannot adapt to diverse tasks, leading to a weakening of the role of core features or a mismatch between the judgment standard and the task. The meta-learning-driven dynamic configuration mechanism specifically addresses the rigidity problem of traditional identification strategies.

[0046] Traditional methods employ globally uniform feature weights and fixed decision thresholds, neglecting task-specific differences. Core discriminative features fail to receive focused attention, and judgment criteria cannot be flexibly adjusted, resulting in low identification accuracy and poor adaptability. By using a meta-learning model for deep analysis of tasks and features, and dynamically generating specialized weighting schemes and decision thresholds, the identification strategy achieves task-adaptive behavior, ensuring the identification process focuses on key aspects and adheres to appropriate standards.

[0047] The meta-learning model consists of a task parsing layer, a weight generation layer, and a threshold calculation layer. It is pre-trained using sample data from different types of identification tasks. This sample data includes task descriptions, core features, optimal weights, and threshold information. The task description parameters and core discriminative features are used as input conditions. The task parsing layer performs structured processing on the input information, extracting key information such as task type, strictness requirements, and core feature attributes, providing a foundation for subsequent parameter generation. This process enables the model to accurately understand the current task requirements, ensuring that the generated parameters are highly aligned with the task, thus improving the strategy's relevance compared to traditional fixed-parameter models.

[0048] The weight generation layer of the meta-learning model generates feature weight vectors based on the correspondence between core discriminative features and multi-dimensional standardized feature vectors. It assigns dominant weight values ​​to the dimensions corresponding to the core discriminative features and auxiliary weight values ​​to non-core features, with a total weight sum of 1. For example, when the core feature is rebound rate, its corresponding weight is set to 0.6, and the total weight of other non-core features is 0.4. Simultaneously, the strictness bias is generated according to the comparison strictness mode: 0.15 in high-precision mode, 0.1 in normal mode, and 0.05 in relaxed mode. This highlights the role of core features, allowing the identification process to focus on key indicators. The strictness bias provides a basis for subsequent threshold calculations, ensuring that the judgment criteria match the task's strictness.

[0049] Historical statistical information is derived from past data of similar identification tasks stored in the system, including historical matching degree distribution and false positive rate. Based on the strictness bias and historical statistical information, a weighted calculation method is used to obtain the dynamic decision threshold, which is the sum of the baseline threshold for similar historical tasks and the strictness bias. For example, the baseline threshold for similar historical authenticity identification is 0.7, while the dynamic decision threshold in high-precision mode is 0.85. This dynamic threshold adapts to the strictness requirements of the current task and, compared to a fixed threshold, can more accurately distinguish matching results, reduce the false positive rate, and improve the accuracy of identification.

[0050] The generated feature weight vector, dynamic decision threshold, and corresponding similarity fusion algorithm are integrated to instantiate the adaptive identification model. This model incorporates task-specific comparison logic, calculating the matching degree based on the weighted feature vector and determining the result according to the dynamic threshold. The instantiation process equips the adaptive identification model with the core functionalities required for the current task, ensuring that the identification process strictly adheres to the task-adaptive strategy. Compared to general-purpose models, this reduces irrelevant computations and improves identification efficiency and accuracy.

[0051] In one embodiment, the step of generating a preliminary identification report that includes the mechanical feature matching degree and a visual comparison of mechanical features related to the core discriminant features includes: S51, Apply the weighting scheme to perform weighted calculations on the multi-dimensional standardized feature vector of the product to be inspected and the multi-dimensional standardized feature vector of the reference sample, respectively, to obtain the weighted feature vector to be inspected and the weighted reference feature vector; S52, calculate the similarity between the weighted feature vector to be detected and the weighted reference feature vector, and use it as the mechanical feature matching degree; S53, compare the mechanical feature matching degree with the dynamic decision threshold to generate a preliminary text identification conclusion; S54, Based on the core discrimination features, generate a visual comparison of the mechanical features, including numerical comparison charts and feature distribution difference maps; S55, integrate the mechanical feature matching degree, the preliminary text identification conclusion, and the mechanical feature visualization comparison content to form the preliminary identification report.

[0052] As described in steps S51-S55 above, the mechanical feature matching degree between the product to be inspected and the reference sample is calculated by applying the weighted scheme of the adaptive identification model. A preliminary text identification conclusion is generated by combining the dynamic decision threshold. Simultaneously, a visual comparison of the core discriminative features is generated. The results are integrated to form a preliminary identification report that combines quantitative data, clear conclusions and intuitive evidence, providing a complete and interpretable basis for subsequent human-machine collaborative correction.

[0053] The credibility of identification results depends not only on quantitative matching data, but also on intuitive display of feature differences. Users need to clearly understand the core reasons for matching or non-matching. However, traditional technologies only output a single matching score, lacking explanation of the results and visualization of feature comparisons, creating a black box for decision-making. Users find it difficult to trust the results and cannot make targeted corrections. By integrating quantitative calculation, threshold determination, visualization generation, and report integration into a unified process, the problem of the inexplicability of traditional identification results can be specifically solved.

[0054] Traditional methods only provide a final match result or a simple score, without showing differences in key features. Users cannot trace the reasons for the results or judge their reasonableness. Furthermore, they lack structured report presentation, resulting in fragmented and unintuitive information. By employing a collaborative process of weighted matching degree calculation, dynamic threshold determination, and visualization of core features, we achieve quantification, clarity, and interpretability of identification results, improving their credibility and practicality.

[0055] The weighting scheme is a feature weight configuration generated by the adaptive identification model, where the dimensions corresponding to the core discriminative features are assigned dominant weight values. This scheme is applied to perform weighted calculations on the multi-dimensional standardized feature vectors of the specimen to be inspected and the multi-dimensional standardized feature vectors of the reference sample. The calculation method involves multiplying the value of each dimension of the feature vector by its corresponding weight value to obtain new vector elements, which are then combined to form the weighted feature vector to be inspected and the weighted reference feature vector. For example, if the core discriminative features are rebound rate and residual deformation depth, with weights of 0.6 and 0.3 respectively, and the sum of the weights of other features is 0.1, then multiplying the value of each feature dimension by its corresponding weight yields the weighted vector. This ensures that the core features dominate the matching degree calculation, making the calculation results more closely aligned with the core needs of the current identification task, avoiding interference from irrelevant features, and improving the targeting and accuracy of the matching degree calculation.

[0056] The cosine similarity algorithm is used to calculate the similarity between the weighted detection feature vector and the weighted reference feature vector. This algorithm measures the similarity of the vectors by calculating the cosine of the angle between the two vectors. The formula is that the similarity equals the dot product of the two vectors divided by the product of their magnitudes. For example, if the dot product of the weighted detection feature vector and the reference feature vector is 0.78, the magnitude of the detection vector is 0.95, and the magnitude of the reference vector is 0.92, the calculated similarity is 0.88, which is the mechanical feature matching degree. Transforming the differences in feature vectors into quantified matching data makes the identification results more accurate and convincing compared to traditional fuzzy feature comparison.

[0057] The dynamic decision threshold is a judgment standard adapted to the current task, generated based on the identification task description parameters and core discriminative features. For example, the dynamic decision threshold is 0.85 in high-precision mode and 0.75 in normal mode. The calculated mechanical feature matching degree is compared with this dynamic decision threshold. If the matching degree is greater than or equal to the dynamic decision threshold, a preliminary text identification conclusion of matching is generated; if the matching degree is less than the dynamic decision threshold, a preliminary text identification conclusion of non-matching is generated. For example, if the mechanical feature matching degree is 0.88 and the dynamic decision threshold is 0.85, a preliminary text identification conclusion of matching is generated; if the matching degree is 0.82 and the dynamic decision threshold is 0.85, a preliminary text identification conclusion of non-matching is generated. This ensures that the identification conclusion matches the comparison strictness mode of the current task, avoiding conclusion bias caused by a fixed threshold, and making the conclusion more closely aligned with task requirements.

[0058] Core discriminant features are the key mechanical characteristics most relevant to the current identification task. Based on these features, visual comparison content is generated, including numerical comparison charts and feature distribution difference maps. The numerical comparison charts, in the form of bar charts or line graphs, intuitively display the specific numerical differences between the tested sample and the reference sample in each core discriminant feature; for example, displaying the rebound rate values ​​of both side-by-side. The feature distribution difference maps, in the form of heat maps, highlight the areas of difference in the spatial distribution corresponding to the core discriminant features; for example, a heat map comparison of strain distribution non-uniformity. This transformation of abstract feature data into intuitive visual information allows users to clearly perceive the differences in core features, solving the problem of uninterpretable results from traditional methods and providing intuitive evidence for users to judge the reasonableness of the results.

[0059] The mechanical feature matching degree, preliminary textual identification conclusions, and visual comparisons of mechanical features are integrated according to a pre-defined structure to form a preliminary identification report. The report first presents the preliminary textual identification conclusions and the mechanical feature matching degree, then displays the visual comparisons of the core discriminant features, ensuring clear information hierarchy and logical coherence. For example, the report may begin by clearly stating a mismatch, with a mechanical feature matching degree of 0.82, followed by a bar chart comparing rebound rate values ​​and a heat map showing strain distribution differences. By integrating multi-dimensional identification information, the report provides users with comprehensive and systematic preliminary identification results, facilitating quick access to core conclusions and allowing for in-depth understanding of the basis for the results through visualizations.

[0060] In one embodiment, the step of adjusting the weighting scheme in real time according to the interactive correction instruction, updating the identification result, and feeding back the adjusted feature weight vector and the corresponding identification task description parameters to the meta-learning model for optimization includes: S61, Display the preliminary identification report and activate the weight adjustment control and conclusion overlay control to receive the interactive correction instruction; S62, in response to the weight adjustment command input through the weight adjustment control, the feature weight vector is updated in real time and the identification result is recalculated and updated; S63, in response to the result overwrite instruction input through the conclusion overwrite control, the user-specified result is recorded as the final identification conclusion; S64, At the end of the identification session, the identification task description parameters, the final feature weight vector and the final identification conclusion of this task are encapsulated to form a collaboratively optimized sample; S65, the collaboratively optimized samples are added to the training dataset of the meta-learning model for subsequent incremental training and model optimization.

[0061] As described in steps S61-S65 above, by providing an interactive correction interface to receive user feedback, the weighting scheme is adjusted in real time and the identification results are updated. At the same time, the interactive data is encapsulated into collaborative optimization samples to feed back into the meta-learning model, thus constructing a closed loop of human-machine collaborative iteration and continuous model evolution, thereby improving the accuracy of the identification results and the long-term adaptability of the system.

[0062] Preliminary identification results may be biased due to initial algorithm configuration. Domain experience in leather identification can effectively correct the judgment of fuzzy intervals, and the model needs to be continuously optimized through actual application data. The current technology decision-making process is closed and cannot be interfered with, and experience in the field of leather identification cannot be involved. The model lacks a dynamic optimization mechanism. Through an integrated process of interactive adjustment and incremental learning, the problems of traditional methods being unable to intervene and lacking evolutionary ability can be specifically solved.

[0063] The preliminary identification report is displayed through a visual interface, including the mechanical feature matching degree, preliminary text identification conclusions, and a visual comparison of core discriminant features. Simultaneously, weight adjustment controls and conclusion overlay controls are activated. The weight adjustment controls are sliding interactive components, allowing users to directly drag and adjust the weight values ​​of each feature. The conclusion overlay controls are result input components, allowing users to directly specify the final identification conclusion. For example, displaying a preliminary report on the identification of the handle of a brand A handbag, activating the corresponding slider controls and conclusion input box provides users with an intuitive interactive entry point. This allows for targeted adjustments based on professional experience, breaking the limitations of closed decision-making in traditional algorithms.

[0064] When a user inputs a weight adjustment command via the weight adjustment control, the system responds in real time and updates the feature weight vector. For example, if a user believes that the system's weight of 0.5 for the linearity of the rebound curve is too high and drags the slider to lower it to 0.2, the system immediately recalculates the mechanical feature matching degree, updating the original 85% confidence level to 92%, and simultaneously updating the preliminary text identification conclusion and visual comparison content. This enables real-time fusion of experience and algorithms in the field of leather identification, quickly correcting deviations in the initial weight configuration, and effectively improving the accuracy of identification results, especially when the results are in an ambiguous range.

[0065] After the authentication session concludes, the key data from this task is automatically encapsulated, including the task description parameters, the final adjusted feature weight vector, and the final authentication conclusion, forming a collaborative optimization sample. For example, the task parameters for "authentication of the handle of a brand A handbag," the adjusted weight vector (rebound rate 0.6, residual deformation depth 0.3, rebound curve linearity 0.2), and the final conclusion of "mismatch" are encapsulated. This transforms the interactive experience of a single authentication session into structured data, providing high-quality samples for model optimization and ensuring the integrity and usability of the data.

[0066] Co-optimized samples are added to the training dataset of the meta-learning model, which then employs an incremental learning algorithm for subsequent training. This algorithm iteratively updates model parameters using mini-batch samples, eliminating the need to retrain the entire model. For example, newly generated co-optimized samples can be merged with existing training data, and stochastic gradient descent can be used to update the parameters of the model's weight generation and threshold calculation layers. This allows the model to learn the implicit rule that bounce time is more important than curve shape in authenticity verification. This enables continuous model evolution; as application scenarios increase and interactive data accumulates, the model's task adaptability and initial configuration accuracy continuously improve, reducing the correction costs for subsequent verifications.

[0067] like Figure 2 As shown, the present invention also discloses an image recognition-based leather product comparison and identification system, comprising: Acquisition module 1 is used to acquire video of the dynamic surface deformation process of the leather product under inspection and the reference sample under controlled micro-force. Data processing module 2 is used to process the video of the dynamic surface deformation process and extract a multi-dimensional standardized feature vector that reflects the inherent physical and mechanical properties of leather. The receiving module 3 is used to receive user-defined identification task description parameters, which include identification target type and comparison strictness mode, and determine at least one corresponding core discriminant feature from a predefined physical and mechanical feature knowledge base according to the identification target type. The generation module 4 is used to dynamically generate an adaptive identification model based on the identification task description parameters and the core discriminant features. The adaptive identification model is used to generate a feature weight vector according to the core discriminant features to weight the multi-dimensional standardized feature vector and generate a decision threshold that matches the comparison strictness mode. The report acquisition module 5 is used to calculate the mechanical feature matching degree between the product to be inspected and the reference sample using the adaptive identification model, and generate a preliminary identification report containing the mechanical feature matching degree and the mechanical feature visualization comparison information related to the core discrimination feature; The adjustment and optimization module 6 is used to receive interactive correction instructions from the user based on the preliminary identification report, adjust the feature weight vector and update the identification results according to the interactive correction instructions, and feed the adjusted parameters back to the model for optimization.

[0068] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described image recognition-based leather product comparison and identification method.

[0069] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described image recognition-based leather product comparison and identification method.

[0070] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0071] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0072] The above description is merely a preferred embodiment of the present invention and does not limit the scope of this application. Any equivalent results or equivalent process transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of protection of this application.

Claims

1. A method for the identification of leather articles by comparison based on image recognition, characterized in that, Includes the following steps: Videos of the dynamic surface deformation process of the leather products to be inspected and the reference sample under controlled micro-force were collected. The video of the dynamic surface deformation process is processed to extract a multi-dimensional standardized feature vector that reflects the inherent physical and mechanical properties of the leather. The system receives user-defined identification task description parameters, which include the identification target type and the comparison strictness mode. Based on the identification target type, it determines at least one corresponding core discriminant feature from a predefined physical and mechanical feature knowledge base. Based on the identification task description parameters and the core discriminant features, an adaptive identification model is dynamically generated. The adaptive identification model is used to generate a feature weight vector according to the core discriminant features, to weight the multi-dimensional standardized feature vector, and to generate a decision threshold that matches the comparison strictness mode. Using the adaptive identification model, the mechanical feature matching degree between the product to be inspected and the reference sample is calculated, and a preliminary identification report containing the mechanical feature matching degree and the mechanical feature visualization comparison information related to the core discrimination feature is generated; The system receives interactive correction instructions from the user based on the preliminary identification report, adjusts the feature weight vector according to the interactive correction instructions and updates the identification results, and feeds the adjusted parameters back to the model for optimization.

2. The image recognition-based comparison identification method of leather products according to claim 1, characterized in that, The steps to obtain multi-dimensional standardized feature vectors include: Based on the video of the dynamic surface deformation process, the displacement field sequence and synchronous micro-force data of the monitoring points on the leather surface are obtained, and pressure-displacement time-varying relationship data are generated. Extract the time-domain mechanical response index set and the spatial domain mechanical distribution index set from the pressure-displacement time-varying relationship data; By fusing the time-domain mechanical response index set and the spatial-domain mechanical distribution index set, an original multi-dimensional feature vector is constructed. The original multi-dimensional feature vector is normalized to obtain the multi-dimensional standardized feature vector.

3. The image recognition-based comparison identification method for leather products according to claim 2, characterized in that, The time-domain mechanical response index set includes at least one of the following: maximum indentation depth, residual deformation depth, time required for a specified rebound ratio, and average rebound rate. The spatial domain mechanical distribution index set includes at least one of the following: average strain value and strain distribution non-uniformity.

4. The image recognition-based comparison identification method of leather products according to claim 1, characterized in that, The steps to determine the core discriminative features include: The identification task description parameters are analyzed to obtain the identification target type and comparison strictness mode, and the use classification identifier of the leather products to be inspected is obtained. The product use sensitivity coefficient is obtained based on the aforementioned use classification identifier; Using the target type as an index, an initial candidate feature set is obtained from the physical and mechanical feature knowledge base; By combining the urgency parameter of the identification scenario with the sensitivity coefficient of the product's use, the features in the initial candidate feature set are dynamically weighted and sorted. The range of the number of candidate features is determined based on the comparison strictness mode, and the confidence level of the feature discrimination power of each candidate feature is calculated in combination with historical feedback data. Based on the dynamic weighted ranking result and the confidence level of the feature discrimination power, at least one core discriminative feature is selected from the initial candidate feature set.

5. The image recognition based comparison identification method of leather products according to claim 1, characterized in that, The specific steps for dynamically generating the adaptive identification model include: The identification task description parameters and the core discriminant features are used as input conditions. Based on the input conditions, initialization parameters are generated, including the feature weight vector and a strictness bias, wherein the weight elements in the feature weight vector corresponding to the core discriminative features are initialized as dominant weight values. The decision threshold is calculated based on the strictness bias and historical statistical information related to the identification target type. The adaptive identification model is instantiated using the initialization parameters and the decision threshold.

6. The image recognition based comparison identification method of leather products according to claim 1, characterized in that, The specific steps for generating a preliminary appraisal report include: The feature weight vector is applied to calculate the weighted multi-dimensional standardized feature vectors of the product to be inspected and the reference sample respectively, so as to obtain the weighted feature vectors of the product to be inspected and the reference feature vectors. The similarity between the weighted feature vector to be detected and the reference feature vector is calculated and used as the mechanical feature matching degree. The mechanical feature matching degree is compared with the decision threshold to generate a preliminary text identification conclusion; Based on the core discriminative features, a visual comparison is generated that includes numerical comparison charts and feature distribution difference maps; The preliminary identification report is formed by integrating the mechanical feature matching degree, preliminary text identification conclusions, and visual comparison content.

7. The image recognition based comparison identification method of leather products according to claim 1, characterized in that, The step of adjusting the feature weight vector and updating the identification result according to the interactive correction instruction includes: Display the preliminary identification report and receive the interactive correction instructions via interactive controls; In response to the weight adjustment command, the feature weight vector is updated in real time and the identification result is recalculated and updated. In response to the result overwrite instruction, the user-specified result is recorded as the final identification conclusion; The identification task description parameters, the final feature weight vector, and the final identification conclusion of this task are encapsulated to form a collaboratively optimized sample. The collaboratively optimized samples are used for incremental training and optimization of the model.

8. An image recognition based leather product comparison identification system characterized in that, include: The acquisition module is used to acquire video of the dynamic surface deformation process of the leather product under inspection and the reference sample under controlled micro-force. The data processing module is used to process the video of the dynamic surface deformation process and extract a multi-dimensional standardized feature vector that reflects the inherent physical and mechanical properties of the leather. The receiving module is used to receive user-defined identification task description parameters, which include identification target type and comparison strictness mode, and determine at least one corresponding core discriminant feature from a predefined physical and mechanical feature knowledge base according to the identification target type. The generation module is used to dynamically generate an adaptive identification model based on the identification task description parameters and the core discriminant features. The adaptive identification model is used to generate a feature weight vector based on the core discriminant features to weight the multi-dimensional standardized feature vector and generate a decision threshold that matches the comparison strictness mode. The report acquisition module is used to calculate the mechanical feature matching degree between the product to be inspected and the reference sample using the adaptive identification model, and generate a preliminary identification report containing the mechanical feature matching degree and the mechanical feature visualization comparison information related to the core discrimination feature; The adjustment and optimization module is used to receive interactive correction instructions from the user based on the preliminary identification report, adjust the feature weight vector and update the identification result according to the interactive correction instructions, and feed the adjusted parameters back to the model for optimization. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images