Card quality evaluation method

Abstract

The invention discloses a card quality evaluation method, and relates to the technical field of image processing, and the method comprises the steps: obtaining an image of a to-be-evaluated card, and carrying out the preprocessing of the image; extracting a comprehensive feature vector from the preprocessed image, wherein the comprehensive feature vector comprises a color distribution parameter, a texture feature parameter and an edge definition parameter; inputting the comprehensive feature vector into a double-flow neural network model, and calculating through the double-flow neural network model to obtain a first card quality score; calibrating the first card quality score based on a preset expert score database to obtain a second card quality score; and according to the second card quality score, dividing the to-be-evaluated card to a predefined quality grade, and generating a quality evaluation report. The invention provides an innovative card quality evaluation method, and comprehensive, accurate and adaptive evaluation of the card quality is realized by comprehensively applying image processing, machine learning and expert knowledge.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a card quality assessment method. Background Art

[0002] With the booming card collection and trading market, the demand for accurate assessment of card quality is growing. However, existing card quality assessment methods mainly rely on manual judgment and have many defects. First, manual evaluation is often greatly affected by subjective factors, resulting in a lack of consistency and repeatability in the evaluation results. Second, traditional manual evaluation methods are time-consuming and labor-intensive, and it is difficult to meet the needs of large-scale card evaluation. In addition, manual evaluation is difficult to capture and quantify some subtle quality features, such as minor wear and slight fading, and the impact of these factors on the overall quality of the card is often ignored or under-evaluated. Summary of the invention

[0003] In view of the problems that the existing card quality assessment methods have strong subjectivity, low efficiency, incomplete assessment, etc., the present invention is proposed.

[0004] Therefore, the problem to be solved by the present invention is how to realize an objective, efficient and comprehensive card quality assessment method to overcome the limitations of manual assessment, improve the accuracy and consistency of assessment, and meet the needs of large-scale card assessment.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] In a first aspect, an embodiment of the present invention provides a card quality assessment method, which includes obtaining an image of a card to be evaluated and preprocessing the image; extracting a comprehensive feature vector from the preprocessed image, the comprehensive feature vector including a color distribution parameter, a texture feature parameter, and an edge clarity parameter; inputting the comprehensive feature vector into a two-stream neural network model, and obtaining a first card quality score through calculation by the two-stream neural network model; calibrating the first card quality score based on a preset expert scoring database to obtain a second card quality score; and classifying the card to be evaluated into predefined quality grades according to the second card quality score, and generating a quality assessment report.

[0007] As a preferred solution of the card quality assessment method described in the present invention, the method comprises the following steps: extracting a comprehensive feature vector from a preprocessed image: performing color space conversion on the preprocessed image, converting the RGB color space into the HSV color space; calculating the histogram data of hue, saturation and brightness in the HSV color space respectively, and combining the histogram data to form a color distribution parameter; constructing a multi-scale Gabor filter, filtering the image in the HSV color space, and obtaining a plurality of filter response graphs; performing statistical analysis on the filter response graph, extracting statistical features, and combining the extracted statistical features to form texture feature parameters; performing edge detection on the preprocessed image using the Canny edge detection algorithm, and obtaining an edge detection result; calculating the gradient amplitude and gradient direction of the image based on the edge detection result, and obtaining an edge clarity histogram, and using the edge clarity histogram as an edge clarity parameter; normalizing the obtained color distribution parameters, texture feature parameters and edge clarity parameters, and splicing them in a preset order to form a comprehensive feature vector characterizing the quality of the card.

[0008] As a preferred embodiment of the card quality evaluation method of the present invention, the first card quality score is obtained by calculating the two-stream neural network model, which includes the following steps: constructing a card sample data set; designing a two-stream neural network model; training the two-stream neural network model using the card sample data set; inputting the comprehensive feature vector of the card to be evaluated into the trained two-stream neural network model; calculating the global quality score and the local quality score respectively through the global feature stream and the local feature stream of the two-stream neural network model; and obtaining the first card quality score based on the weighted average of the global quality score and the local quality score.

[0009] As a preferred solution of the card quality assessment method described in the present invention, the design of a dual-stream neural network model includes the following steps: constructing a global feature stream network structure; designing a local feature stream network structure; adding adaptive average pooling layers after the convolution layers of the global feature stream and the local feature stream, respectively, to extract multi-scale feature information; designing a feature fusion module to splice and fuse the outputs of the global feature stream and the local feature stream to obtain a fused feature representation; based on the fused feature representation, designing two parallel fully connected layer networks, respectively used to calculate the global quality score and the local quality score.

[0010] As a preferred solution of the card quality assessment method described in the present invention, the method comprises the following steps: training the two-stream neural network model using a card sample data set comprises the following steps: selecting an anchor sample, a positive sample and multiple negative samples from the card sample data set; inputting the anchor sample, the positive sample and the negative sample into the two-stream neural network model respectively to obtain corresponding feature representations; calculating the cosine similarity between the anchor sample and the positive sample, and the cosine similarity between the anchor sample and the negative sample; constructing a contrast loss function in the form of InfoNCE loss; combining the contrast loss function and the mean square error loss of the quality score to construct an overall loss function; minimizing the overall loss function using the Adam optimizer to update the parameters of the two-stream neural network model.

[0011] As a preferred embodiment of the card quality evaluation method of the present invention, calibrating the first card quality score includes the following steps: constructing an expert rating database, which contains a large number of card samples of known quality levels and their corresponding expert ratings; using the comprehensive feature vector of the card to be evaluated, retrieving K samples most similar to the card to be evaluated in the expert rating database; performing weighted averaging on the expert ratings of the retrieved K similar samples to obtain a reference rating; calculating the deviation value between the first card quality score and the reference rating; designing an adaptive calibration function to calibrate the first card quality score based on the deviation value; outputting the calibrated first card quality score as the second card quality score; updating the expert rating database, and adding the current evaluation result and related information to the database.

[0012] As a preferred solution of the card quality evaluation method described in the present invention, the method comprises the following steps: establishing a deviation-weight mapping relationship, wherein the deviation-weight mapping relationship corresponds the deviation value to a corresponding weight coefficient; determining the corresponding weight coefficient for the current card to be evaluated according to the deviation-weight mapping relationship; constructing an adaptive calibration function, wherein the adaptive calibration function adopts a nonlinear transformation structure and takes the first card quality score as an input variable; setting an adjustable parameter in the adaptive calibration function, wherein the adjustable parameter is associated with the weight coefficient and is used to control the calibration amplitude; substituting the first card quality score into the adaptive calibration function to generate a calibration intermediate result; performing boundary constraints and normalization operations on the calibration intermediate result to ensure that the generated score is within a predefined valid range.

[0013] In a second aspect, an embodiment of the present invention provides a card quality assessment system, which includes a preprocessing module, which is used to obtain an image of a card to be evaluated and preprocess the image; a feature vector extraction module, which is used to extract a comprehensive feature vector from the preprocessed image, and the comprehensive feature vector includes a color distribution parameter, a texture feature parameter, and an edge clarity parameter; a model calculation module, which is used to input the comprehensive feature vector into a two-stream neural network model, and obtain a first card quality score through calculation of the two-stream neural network model; a calibration module, which is used to calibrate the first card quality score based on a preset expert scoring database to obtain a second card quality score; a quality assessment module, which is used to classify the card to be evaluated into a predefined quality grade according to the second card quality score, and generate a quality assessment report.

[0014] The beneficial effects of the present invention are as follows: the present invention provides an innovative card quality assessment method, which realizes a comprehensive, accurate and adaptive assessment of card quality by comprehensively using image processing, machine learning and expert knowledge. The method firstly uses multi-dimensional feature extraction technology to comprehensively capture key quality indicators such as color distribution, texture features and edge clarity of cards; then, it uses an innovatively designed two-stream neural network model, combined with global and local feature analysis, to improve the recognition ability of the overall quality and detailed features of the cards; further, by introducing a comparative learning strategy, the model's ability to identify quality differences of similar cards is enhanced; finally, an adaptive calibration mechanism based on an expert scoring database is adopted, which effectively integrates machine learning results and human expert experience, greatly improving the accuracy and reliability of the evaluation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 This is a framework flow chart of the card quality assessment method.

[0017] Figure 2 Flowchart of the feature extraction process of the card quality assessment method.

[0018] Figure 3 Flowchart of the design and training of a two-stream neural network model for the card quality assessment method. DETAILED DESCRIPTION

[0019] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0020] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0022] Example 1, reference Figure 1 to Figure 3 , which is the first embodiment of the present invention, provides a card quality assessment method, comprising:

[0023] S1: Obtain an image of the card to be evaluated and preprocess the image.

[0024] The preprocessing includes but is not limited to the following operations: noise reduction, contrast adjustment and size standardization.

[0025] S2: Extract the comprehensive feature vector from the preprocessed image.

[0026] In this embodiment, the comprehensive feature vector includes color distribution parameters, texture feature parameters and edge definition parameters. The three parameters together cover the main aspects of card quality, including visual quality, printing accuracy and physical integrity. The extraction process is as follows:

[0027] S2.1: Perform color space conversion on the preprocessed image, converting the RGB color space into the HSV color space.

[0028] It should be noted that the HSV color space is more suitable for describing human perception of color. The HSV color space includes three channels: hue, saturation, and brightness.

[0029] S2.2: In the HSV color space, the histogram data of hue, saturation and brightness are calculated respectively, and the histogram data are combined to form color distribution parameters.

[0030] Among them, the hue histogram represents the distribution of different hue values, the saturation histogram represents the distribution of different saturation values, and the lightness histogram represents the distribution of different lightness values. These combined histogram data together constitute the color distribution parameter, which is used to characterize the overall color characteristics of the card.

[0031] Preferably, by comprehensively analyzing the color characteristics of the cards, quality problems such as color deviation, fading or uneven printing can be effectively identified, providing objective and comprehensive data support for the color quality assessment of the cards.

[0032] S2.3: Construct a set of multi-scale Gabor filters to filter the image in the HSV color space and obtain multiple filter response maps.

[0033] Furthermore, the multi-scale Gabor filter includes filters of different directions and frequencies.

[0034] S2.4: Perform statistical analysis on the filter response graph, extract statistical features, and combine the extracted statistical features to form texture feature parameters.

[0035] It should be noted that statistical features include energy, variance and entropy. Texture feature parameters are used to describe the texture complexity and detail information of the card.

[0036] Preferably, by performing statistical analysis on the filter response graph, the texture characteristics of the card can be quantitatively described, and printing quality problems such as blur, noise or missing patterns can be effectively detected, thereby providing an objective assessment of the surface quality of the card.

[0037] S2.5: Apply the Canny edge detection algorithm to perform edge detection on the preprocessed image to obtain an edge detection result.

[0038] Preferably, the Canny edge detection algorithm adopts an adaptive threshold selection method to dynamically adjust the threshold according to local characteristics of the image.

[0039] S2.6: Calculate the gradient amplitude and gradient direction of the image based on the edge detection result, obtain an edge clarity histogram, and use the edge clarity histogram as an edge clarity parameter.

[0040] It should be noted that the edge clarity parameter is used to quantify the clarity of the card edge.

[0041] S2.7: After normalizing the obtained color distribution parameters, texture feature parameters and edge clarity parameters, they are concatenated in a preset order to form a comprehensive feature vector representing the quality of the card.

[0042] Preferably, the present invention constructs a multi-dimensional comprehensive feature vector by comprehensively extracting the color distribution, texture features and edge clarity parameters of the card image. The HSV color space conversion is used to improve the accuracy of color analysis; the multi-scale Gabor filter is used to enhance the expression ability of texture features; and the Canny edge detection with adaptive threshold is introduced to improve the accuracy of edge analysis. This multi-dimensional feature extraction method not only comprehensively covers the key aspects of card quality, but also improves the accuracy and efficiency of evaluation through objective quantitative indicators.

[0043] S3: Input the comprehensive feature vector into the two-stream neural network model, and calculate the first card quality score through the two-stream neural network model.

[0044] Specifically, the following steps are included:

[0045] S3.1: Construct a sample card dataset.

[0046] Specifically, the card sample data set includes multiple card samples with known quality scores, and each sample includes a comprehensive feature vector of the card and a corresponding quality score.

[0047] S3.2: Design a two-stream neural network model.

[0048] The two-stream neural network model includes a global feature stream and a local feature stream. The global feature stream is used to process the overall comprehensive features, and the local feature stream focuses on the key areas of the card through the attention mechanism. The specific design process is as follows:

[0049] S3.2.1: Construct a global feature flow network structure.

[0050] Among them, the global feature flow adopts a multi-layer convolutional neural network structure to process the color distribution parameters, texture feature parameters and edge clarity parameters in the comprehensive feature vector.

[0051] S3.2.2: Design the local feature flow network structure.

[0052] Among them, the local feature flow adopts a residual network architecture and sets a channel attention mechanism in each residual block to extract the features of the key areas of the card.

[0053] S3.2.3: After the convolutional layers of the global feature stream and the local feature stream, adaptive average pooling layers are added to extract multi-scale feature information.

[0054] S3.2.4: Design a feature fusion module to concatenate and fuse the outputs of the global feature stream and the local feature stream to obtain a fused feature representation.

[0055] The feature fusion module first aligns the dimensions of the outputs of the global feature stream and the local feature stream, and then performs weighted fusion through an adaptive weight mechanism to obtain a preliminary fusion feature. Subsequently, a nonlinear transformation is applied to the preliminary fusion feature to enhance the expressiveness of the feature, and finally a fusion feature representation is obtained.

[0056] S3.2.5: Based on the fused feature representation, two parallel fully connected layer networks are designed to calculate the global quality score and the local quality score respectively.

[0057] Preferably, the dual-stream neural network model design proposed in the present invention realizes a comprehensive evaluation of the quality of cards by integrating the global feature stream and the local feature stream. The design introduces the channel attention mechanism and the adaptive average pooling layer, which effectively improves the ability to capture key areas and multi-scale features. The application of the feature fusion module and the parallel scoring structure further enhances the model's ability to comprehensively analyze the overall quality and local details of the cards. This innovative structure not only improves the accuracy and comprehensiveness of the evaluation, but also has good adaptability and can meet the quality evaluation requirements of different types of cards.

[0058] S3.3: Use the card sample dataset to train the two-stream neural network model.

[0059] Specifically, during the training process, a contrastive learning strategy is used to enhance the two-stream neural network model's ability to identify quality differences of similar cards. The training process is as follows: select an anchor sample, a positive sample, and multiple negative samples from the card sample data set; input the anchor sample, positive sample, and negative sample into the two-stream neural network model respectively to obtain the corresponding feature representation; calculate the cosine similarity between the anchor sample and the positive sample, as well as the cosine similarity between the anchor sample and the negative sample; construct a contrastive loss function in the form of InfoNCE loss; combine the contrastive loss function and the mean square error loss of the quality score to construct an overall loss function; use the Adam optimizer to minimize the overall loss function and update the parameters of the two-stream neural network model.

[0060] The sample selection process is as follows: calculate the quality score differences between all sample pairs in the card sample data set to obtain the quality score difference distribution; determine the positive sample selection threshold α1 and the negative sample selection threshold α2 based on the quality score difference distribution; in each training iteration, randomly select a sample from the card sample data set as an anchor sample, and select multiple positive samples and negative samples according to the positive sample selection threshold α1 and the negative sample selection threshold α2; form this anchor sample and its corresponding positive and negative samples into a training batch; during the training process, repeat the sample selection process every preset number of iterations, and update the positive sample selection threshold α1 and the negative sample selection threshold α2.

[0061] In addition, the positive sample selection threshold α1 is the 25th percentile of the quality score difference distribution, and the negative sample selection threshold α2 is the 75th percentile of the quality score difference distribution. Samples whose quality score difference with the anchor sample does not exceed α1 are selected as positive samples, and samples whose quality score difference with the anchor sample is greater than α2 are selected as negative samples.

[0062] S3.4: Input the comprehensive feature vector of the card to be evaluated into the trained two-stream neural network model.

[0063] S3.5: The global quality score and the local quality score are calculated respectively through the global feature stream and the local feature stream of the two-stream neural network model.

[0064] Among them, the global quality score and the local quality score are calculated by two parallel fully connected layer networks designed in step S3.2.6.

[0065] S3.6: Obtain a first card quality score based on a weighted average of the global quality score and the local quality score.

[0066] Among them, the weighted average adopts a fixed weight method.

[0067] S4: Based on a preset expert rating database, calibrate the first card quality rating to obtain a second card quality rating.

[0068] In a preferred embodiment, the calibration process includes the following steps:

[0069] S4.1: Construct an expert rating database, which contains a large number of card samples of known quality levels and their corresponding expert ratings.

[0070] Among them, the expert rating database covers cards of different types and qualities to ensure the comprehensiveness of the calibration process.

[0071] S4.2: Using the comprehensive feature vector of the card to be evaluated, retrieve the K samples most similar to the card to be evaluated in the expert rating database.

[0072] Wherein, K is a preset positive integer, and the similarity calculation adopts the cosine similarity method.

[0073] S4.3: Take a weighted average of the expert scores of the K similar samples retrieved to obtain a reference score.

[0074] It should be noted that the weight allocation is based on the comprehensive similarity between the sample and the card to be evaluated. The higher the similarity, the greater the weight.

[0075] S4.4: Calculate the deviation value between the first card quality score and the reference score.

[0076] The deviation value reflects the degree of difference between the initial score and the expert score.

[0077] S4.5: Design an adaptive calibration function to calibrate the first card quality score based on the deviation value.

[0078] Specifically, a deviation-weight mapping relationship is established, and the deviation-weight mapping relationship corresponds the deviation value to the corresponding weight coefficient; according to the deviation-weight mapping relationship, the corresponding weight coefficient is determined for the current card to be evaluated; an adaptive calibration function is constructed, and the adaptive calibration function adopts a nonlinear transformation structure and takes the first card quality score as an input variable; an adjustable parameter is set in the adaptive calibration function, and the adjustable parameter is associated with the weight coefficient to control the calibration amplitude; the first card quality score is substituted into the adaptive calibration function to generate a calibration intermediate result; and boundary constraints and normalization operations are performed on the calibration intermediate result to ensure that the generated score is within a predefined valid range.

[0079] Preferably, by introducing the bias-weight mapping relationship and the nonlinear adaptive calibration function, dynamic calibration of cards of different types and qualities is achieved, which improves the accuracy and adaptability of the scoring. At the same time, the boundary constraints and normalization operations ensure the consistency and comparability of the output scores.

[0080] S4.6: Output the calibrated first card quality score as the second card quality score.

[0081] It should be understood that the second card quality score combines the output of the two-stream neural network model and expert experience and has higher reliability.

[0082] S4.7: Update the expert scoring database and add the current assessment results and related information to the database.

[0083] For example, the feature vector of the card to be evaluated, the quality score of the first card, the quality score of the second card, and other information can be stored in the database for continuous optimization of subsequent evaluations. During the update process, the following data cleaning and screening steps are performed: remove duplicate samples; remove outliers (such as samples with scores that significantly deviate from the normal range).

[0084] Furthermore, the updating process of the expert rating database can be performed periodically to balance system performance and computing resource consumption.

[0085] S5: According to the second card quality score, the cards to be evaluated are classified into predefined quality levels, and a quality evaluation report is generated.

[0086] Furthermore, the present embodiment also provides a card quality assessment system, including a preprocessing module, used to obtain an image of a card to be evaluated and preprocess the image; a feature vector extraction module, used to extract a comprehensive feature vector from the preprocessed image, the comprehensive feature vector including color distribution parameters, texture feature parameters and edge clarity parameters; a model calculation module, used to input the comprehensive feature vector into a two-stream neural network model, and obtain a first card quality score through calculation of the two-stream neural network model; a calibration module, used to calibrate the first card quality score based on a preset expert scoring database to obtain a second card quality score; a quality assessment module, used to classify the card to be evaluated into a predefined quality grade according to the second card quality score, and generate a quality assessment report.

[0087] In summary, the present invention provides an innovative card quality assessment method, which realizes a comprehensive, accurate and adaptive assessment of card quality by comprehensively using image processing, machine learning and expert knowledge. The method firstly uses multidimensional feature extraction technology to comprehensively capture key quality indicators such as color distribution, texture features and edge clarity of cards; then, it uses an innovatively designed two-stream neural network model, combined with global and local feature analysis, to improve the recognition ability of the overall quality and detail features of the cards; further, by introducing a comparative learning strategy, the model's ability to identify quality differences between similar cards is enhanced; finally, an adaptive calibration mechanism based on an expert scoring database is adopted, which effectively integrates machine learning results and human expert experience, greatly improving the accuracy and reliability of the assessment results. This multi-level and multi-angle assessment method can not only objectively quantify the quality of cards, but also has good adaptability and scalability, and can meet the quality assessment needs of different types of cards.

[0088] Example 2, reference Figure 1 to Figure 3 , which is the second embodiment of the present invention, and this embodiment provides a card quality evaluation method. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0089] To verify the effectiveness of the present invention, the research team collected 10,000 card images of different types and qualities on a well-known card collection platform as an experimental dataset. These cards cover multiple popular series, including Magic: The Gathering, Yu-Gi-Oh, and Pokémon. In the preprocessing stage, all images were uniformly adjusted to a resolution of 1024×1024 pixels and denoised using Gaussian filtering. Color space conversion was implemented using the OpenCV library, and the hue, saturation, and brightness were divided into 18, 3, and 3 bins, respectively, when calculating the HSV histogram. The multiscale Gabor filter was set with 4 scales and 6 directions, generating a total of 24 filters. The low and high thresholds of Canny edge detection were set to 50 and 150, respectively.

[0090] The two-stream neural network model was trained using 8,000 card images, with 1,000 images each for the validation set and test set. The global feature stream uses a 5-layer convolutional neural network, with each layer followed by BatchNormalization and ReLU activation functions. The local feature stream uses ResNet-18 as the infrastructure, and adds a SE (Squeeze-and-Excitation) attention mechanism to each residual block. During training, the batch size was set to 64, the initial value of the learning rate was 0.001, and the cosine annealing strategy was used for adjustment. In contrastive learning, 1 positive sample and 5 negative samples were selected for each anchor sample, and the temperature parameter was set to 0.07. The model training lasted 100 epochs, and finally achieved a classification accuracy of 95.2% on the test set.

[0091] The expert rating database contains 5,000 samples rated by senior card identification experts. During the calibration process, K = 5 most similar samples were selected for weighted averaging. The adaptive calibration function is implemented using cubic spline interpolation to ensure the continuity and smoothness of the rating. In order to comprehensively evaluate the performance of the method of the present invention, the research team compared it with two prior art methods. The results are shown in Table 1:

[0092] Table 1 Comparison table between the present invention and two prior art methods

[0093] Evaluation Methodology Accuracy Mean absolute error Processing time (seconds / sheet) Method of the present invention 95.2％ 0.18 0.85 Traditional CNN method 88.7％ 0.43 0.62 Expert manual evaluation 93.5％ 0.22 180

[0094] As can be seen from the table data, the method of the present invention is significantly better than other automated methods in terms of accuracy and mean absolute error, and even exceeds the performance of expert manual evaluation. Although the processing time is slightly longer than the traditional CNN method, this slight increase in time is acceptable considering its significantly improved accuracy. It is particularly noteworthy that compared with expert manual evaluation, this method not only maintains considerable accuracy, but also greatly improves efficiency, reducing the evaluation time for each card from 3 minutes to less than 1 second, which is of great significance for large-scale card evaluation tasks.

[0095] In general, the method of the present invention has shown obvious advantages in accuracy, efficiency and scalability, and provides a powerful automated solution for the field of card quality assessment.

[0096] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A card quality assessment method, characterized in that: include, Obtaining an image of the card to be evaluated, and preprocessing the image; Extracting a comprehensive feature vector from the preprocessed image, wherein the comprehensive feature vector includes a color distribution parameter, a texture feature parameter, and an edge definition parameter; Inputting the comprehensive feature vector into a two-stream neural network model, and calculating a first card quality score through the two-stream neural network model; Based on a preset expert rating database, calibrate the first card quality rating to obtain a second card quality rating; According to the second card quality score, the cards to be evaluated are classified into predefined quality levels, and a quality evaluation report is generated.

2. The card quality assessment method according to claim 1, wherein: The method of extracting a comprehensive feature vector from the preprocessed image comprises the following steps: Perform color space conversion on the preprocessed image, converting the RGB color space into the HSV color space; In the HSV color space, histogram data of hue, saturation and brightness are calculated respectively, and the histogram data are combined to form color distribution parameters; Constructing a multi-scale Gabor filter to filter the image in the HSV color space to obtain a plurality of filter response graphs; Performing statistical analysis on the filter response graph, extracting statistical features, and combining the extracted statistical features to form texture feature parameters; Apply the Canny edge detection algorithm to perform edge detection on the preprocessed image to obtain the edge detection result; Calculating the gradient amplitude and gradient direction of the image based on the edge detection result to obtain an edge clarity histogram, and using the edge clarity histogram as an edge clarity parameter; After the obtained color distribution parameters, texture feature parameters and edge clarity parameters are normalized, they are spliced ​​in a preset order to form a comprehensive feature vector that characterizes the quality of the card.

3. The card quality assessment method according to claim 1, wherein: The step of calculating the first card quality score by the two-stream neural network model comprises the following steps: Construct a card sample dataset; Design a two-stream neural network model; Using the card sample data set to train the two-stream neural network model; Input the comprehensive feature vector of the card to be evaluated into the trained two-stream neural network model; Calculating a global quality score and a local quality score respectively through the global feature stream and the local feature stream of the two-stream neural network model; The first card quality score is obtained based on the weighted average of the global quality score and the local quality score.

4. The card quality assessment method according to claim 3, characterized in that: The design of the two-stream neural network model includes the following steps: Construct a global feature flow network structure; Design local feature flow network structure; After the convolution layers of the global feature stream and the local feature stream, an adaptive average pooling layer is added respectively to extract multi-scale feature information; Design a feature fusion module to concatenate and fuse the outputs of the global feature stream and the local feature stream to obtain a fused feature representation; Based on the fused feature representation, two parallel fully connected layer networks are designed to calculate the global quality score and the local quality score respectively.

5. The card quality assessment method according to claim 3, wherein: The training of the two-stream neural network model using the card sample data set comprises the following steps: Selecting an anchor sample, a positive sample, and multiple negative samples from the card sample dataset; Inputting the anchor sample, the positive sample and the negative sample into the two-stream neural network model respectively to obtain corresponding feature representations; Calculating the cosine similarity between the anchor sample and the positive sample, and the cosine similarity between the anchor sample and the negative sample; Construct a contrast loss function using the InfoNCE loss form; Combine the contrast loss function and the mean square error loss of the quality score to construct the overall loss function; The Adam optimizer is used to minimize the overall loss function and update the parameters of the two-stream neural network model.

6. The card quality assessment method according to claim 1, wherein: The calibrating the first card quality score comprises the following steps: Build an expert rating database, which contains a large number of card samples with known quality levels and their corresponding expert ratings; Using the comprehensive feature vector of the card to be evaluated, searching the expert rating database for K samples that are most similar to the card to be evaluated; Take a weighted average of the expert scores of the K similar samples retrieved to get a reference score; Calculate the deviation between the quality score of the first card and the reference score; Designing an adaptive calibration function to calibrate the first card quality score based on the deviation value; Outputting the calibrated first card quality score as the second card quality score; Update the expert scoring database and add the current evaluation results and related information to the database.

7. The card quality assessment method according to claim 6, characterized in that: The calibrating the first card quality score based on the deviation value comprises the following steps: Establishing a deviation-weight mapping relationship, wherein the deviation-weight mapping relationship corresponds the deviation value to a corresponding weight coefficient; According to the deviation-weight mapping relationship, a corresponding weight coefficient is determined for the current card to be evaluated; Constructing an adaptive calibration function, wherein the adaptive calibration function adopts a nonlinear transformation structure and takes the first card quality score as an input variable; Setting an adjustable parameter in the adaptive calibration function, the adjustable parameter being associated with the weight coefficient and used to control the calibration amplitude; Substituting the first card quality score into the adaptive calibration function to generate an intermediate calibration result; Boundary constraints and normalization operations are performed on the calibration intermediate results to ensure that the generated scores are within a predefined valid interval.

8. A card quality assessment system, based on the card quality assessment method according to any one of claims 1 to 7, characterized in that: Also includes, A preprocessing module, used to obtain an image of a card to be evaluated and preprocess the image; A feature vector extraction module is used to extract a comprehensive feature vector from the preprocessed image, wherein the comprehensive feature vector includes a color distribution parameter, a texture feature parameter and an edge definition parameter; A model calculation module, used for inputting the comprehensive feature vector into a two-stream neural network model, and calculating a first card quality score through the two-stream neural network model; A calibration module, configured to calibrate the first card quality score based on a preset expert rating database to obtain a second card quality score; A quality assessment module is used to classify the cards to be assessed into predefined quality levels according to the second card quality scores, and generate a quality assessment report.

Images