Art work image recognition and copyright storage system fused with block chain

Abstract

The invention relates to the field of copyright evidence storage, in particular to an art work image recognition and copyright evidence storage system fused with a block chain, and the system comprises a feature vector extraction module; outputting a skill feature vector of which the dimension is matched with the visual feature vector; a feature focusing module; outputting the focused skill feature vector; a similarity identification module; a threshold value is preset to complete identification and judgment of the art works; a key parameter screening module; outputting a key parameter subset related to the skill core; a copyright evidence storage module; and fusing and generating a feature code with a fixed length through a self-defined Hash mapping rule, and chaining the feature code and completing copyright evidence storage. According to the method, key parameter subsets related to the skill core are screened through a recursive feature elimination algorithm, redundant information is eliminated, the data processing amount is reduced, meanwhile, a component analysis mechanism is introduced to map high-dimensional key parameters to a low-dimensional feature space, the feature processing efficiency is improved on the premise that the integrity of core information is kept, and the method is suitable for mass production. And both the identification precision and the operation performance are considered.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of copyright evidence storage, in particular to an art work image recognition and copyright evidence storage system fusing blockchain. BACKGROUND

[0002] In the field of art work copyright protection, traditional copyright evidence storage methods rely on paper registration, manual review or centralized database storage, which have problems such as complicated process, long time consumption, easy tampering or loss of stored information, and cannot meet the needs of rapid storage and efficient rights protection of art works in the digital age. At the same time, existing art work image recognition technology often only focuses on visual feature extraction, ignoring the core value of skill operation parameters in the creation process, such as creation method, material interaction and process details, resulting in insufficient accuracy in distinguishing between imitated works and original works - some imitated works can highly replicate the original works in visual presentation, but there are essential differences in skill core features, and traditional recognition methods are difficult to capture such implicit differences, which may easily lead to copyright disputes.

[0003] In addition, the feature parameter dimension of art works is complex, including visual pixels, structural dimension, skill operation and other types of high-dimensional data, and existing technologies lack effective feature focusing and key parameter screening mechanisms, which are easily disturbed by redundant information, resulting in low feature extraction efficiency and insufficient core information recognition; and there is a lack of standardized feature coding strategy in the copyright evidence storage process, and the stored information of different platforms is difficult to interconnect and verify, further reducing the credibility and practicality of copyright protection.

[0004] Therefore, how to realize the deep fusion recognition of visual features and skill features, and build an efficient and accurate copyright protection system combined with tamper-proof evidence storage technology, has become a technical problem to be solved in the field of art work copyright evidence storage. SUMMARY

[0005] The present application provides an art work image recognition and copyright evidence storage system fusing blockchain to solve the technical problems in the prior art.

[0006] The technical solution of the present application to solve the above technical problems is as follows: an art work image recognition and copyright evidence storage system fusing blockchain, the system comprising: a feature vector extraction module; for extracting visual features of art works by CNN, and outputting visual feature vectors with unified dimensions by a fully connected layer, and then extracting skill operation parameters obtained by a data acquisition device by CNN, and outputting skill feature vectors with matching dimensions with the visual feature vectors; The feature focusing module is used for receiving the skill feature vector and inputting the skill feature vector into a weight distribution layer to generate a feature correlation weight matrix, inputting the feature correlation weight matrix and the visual feature vector into a self-attention mechanism together, automatically positioning a skill core region by calculating a correlation degree between a visual region and the skill feature, and finally outputting a focused skill feature vector; The similarity recognition module is used for processing the focused skill feature vector and the visual feature vector through an alignment algorithm to generate a skill and visual fusion feature matrix, inputting the fusion feature matrix into an SVM, constructing a double-threshold classifier and outputting a similarity value, and then completing the recognition and determination of the art work according to a preset threshold value. The key parameter screening module is used for obtaining a deconstruction dimension parameter and a visual pixel parameter of the art work based on an image recognition technology after the recognition is passed, constructing an initial parameter data set, processing the initial parameter data set through a recursive feature elimination algorithm, and outputting a key parameter subset related to the skill core. The copyright evidence storage module is used for introducing a component analysis mechanism in the recursive feature elimination algorithm, mapping the key parameter subset to a low-dimensional feature space to generate a first necessary feature vector containing skill and visual core information, introducing a self-defined coding strategy, fusing a fixed-length feature code through a self-defined hash mapping rule, and completing copyright evidence storage by chaining the feature code.

[0007] In a preferred embodiment, the feature vector extraction module inputs the image of the art work into a convolutional neural network after uniformly scaling the image to a preset resolution, extracts visual features through a convolutional layer, samples and reduces dimensions through a pooling layer, finally maps the visual features to an initial feature vector through a fully connected layer, and outputs a standardized visual feature vector after normalization processing; after performing outlier rejection and standardization processing on the skill operation parameters, the skill operation parameters are input into the convolutional layer and the fully connected layer of the CNN, and the high-dimensional skill operation parameters are mapped to skill feature vectors with the same dimension as the visual feature vector. The feature focusing module takes the visual feature vector and the skill feature vector with the same dimension as input, introduces a skill feature guide weight layer in the self-attention mechanism, and outputs the focused skill feature vector.

[0008] In a preferred embodiment, the feature focusing module calculates the cosine similarity value of the skill feature vector and a core feature template library, maps the weight value range to the [0, 1] interval to generate a feature correlation weight matrix, reshapes the visual feature vector into a feature map format, and inputs the visual feature vector and the feature correlation weight matrix into the self-attention mechanism with the skill feature guide weight layer. The feature focusing module collects third-party digitized art resources, and obtains sample skill operation parameters and sample visual feature parameters of each art craft through binary dimension deconstruction, eliminates non-key dimensions, performs pixel normalization and light compensation on the sample visual feature parameters, classifies the sample skill operation parameters and sample feature parameters according to art craft types, constructs type sample subsets, lays a data foundation for feature extraction, each type sample subset screens sample key features through a recursive feature elimination algorithm, and reduces the dimensions of the screened sample key features, calculates the mean and standard deviation of all sample low-dimensional feature vectors in each type sample subset, generates sample low-dimensional feature vectors, and constructs core feature templates in a sample library. After the feature focusing module inputs the skill feature vector, the cosine similarity value between the skill feature vector and each sample low-dimensional feature vector in the core feature template is calculated one by one. Since the skill feature vector and each sample low-dimensional feature vector in the core feature template have been normalized, the cosine similarity value calculation is equivalent to the dot product of two vectors. The higher the cosine similarity value, the more the input skill feature vector matches the corresponding sample low-dimensional feature vector. All the obtained cosine similarity values are mapped to the [0, 1] interval as weight values, and are constructed into a feature correlation weight matrix. The feature correlation weight matrix and the feature map are input into the self-attention mechanism. The self-attention mechanism generates a query matrix composed of the feature map, a key matrix composed of the feature correlation weight matrix, and a value matrix through linear transformation. The attention score is obtained by calculating the dot product of the query matrix and the key matrix. The attention map is generated by the Softmax function, and the dimensions of the attention map are aligned with the value matrix to ensure that each feature position has a unique corresponding attention score. The attention map is used to update the value matrix by weighting, including multiplying the feature weight by the strengthening coefficient when the attention score exceeds the attention threshold, and multiplying the feature weight by the attenuation coefficient when the attention score is less than or equal to the attention threshold. Finally, the updated feature map is mapped to the focused skill feature vector through global average pooling.

[0009] In a preferred embodiment, the similarity recognition module ensures that the visual feature vector and the obtained skill feature vector pass the consistency test, and that both are homodimensional standardized vectors. If there is a dimensional deviation, zero padding or feature pruning is used to complete the calibration. Element-by-element multiplication is performed on the visual feature vector and the skill feature vector to obtain an interaction feature vector reflecting the degree of feature correlation. The original visual feature vector and skill feature vector are then spliced with the interaction feature vector according to the channel dimension to form a fusion feature vector. The fusion feature vector is reshaped into a two-dimensional skill and visual fusion feature matrix by a matrix reconstruction algorithm. In the skill and visual fusion feature matrix, the row dimension corresponds to the feature category, and the column dimension corresponds to the feature value, providing a dimensional guide for subsequent similarity calculation. Each row of the skill and visual fusion feature matrix is labeled with categories including visual basic features corresponding to the visual feature vector, skill core features corresponding to the focused skill feature vector, and feature interaction features corresponding to the interaction feature vector, providing a dimensional guide for subsequent similarity calculation. According to the labeled category labels, the skill and visual fusion feature matrix is split into a visual feature submatrix and a skill feature submatrix, corresponding to the labeled visual basic feature row vectors and the labeled visual basic feature row vectors, respectively. The visual feature submatrix and the skill feature submatrix are input into the pre-trained double-threshold classifier. It should be noted that the double-SVM classifier corresponds to independent SVM models, both using RBF kernel functions with a penalty coefficient of 1.0. The kernel function parameter is determined by cross-validation optimization. The SVM model calculates the similarity values of the visual feature submatrix and the skill feature submatrix with the core feature template through the ratio of vector dot product and modulus product, and calculates the similarity values through the cosine similarity formula, and outputs the visual feature similarity value and the skill feature similarity value, respectively. The output visual feature similarity value and skill feature similarity value are plotted as ROC curves. The minimum threshold value of the visual feature and the minimum threshold value of the skill feature are found based on the visual feature similarity value ROC curve and the skill feature similarity value ROC curve, and are used as the visual feature initial value and the skill feature initial value, respectively. After cross-validation of the visual feature initial value and the skill feature initial value, the accuracy is ensured to be more than 92%, including any of the following recognition settings: If both the visual feature similarity value and the skill feature similarity value exceed the visual feature initial value and the skill feature initial value, it is determined to be a non-imitation work; If the visual feature similarity value exceeds the visual feature initial value, and the skill feature similarity value is less than or equal to the skill feature initial value, it is determined to be an imitation work; If the visual feature similarity value is less than or equal to the visual feature initial value, and the skill feature similarity value is less than or equal to the skill feature initial value, it is determined to be a non-imitation work; If the visual feature similarity value and the skill feature similarity value are both less than or equal to the visual feature initial value and the skill feature initial value, it is determined that the work is a copycat work.

[0010] In a preferred embodiment, the key parameter screening module will obtain the structural dimension parameters and visual pixel parameters based on image recognition technology by identifying the determined artwork, the structural dimension parameters such as stitch type, thread angle, knife mark depth, stroke turning angle, etc., and construct an initial parameter data set containing parameter name, collection timestamp, and data identifier, fill in the missing values by the neighborhood interpolation method, smooth the noise in the visual pixel parameters by Gaussian filtering, and finally standardize the initial parameter data set; The standardized initial parameter data set is subjected to hierarchical random division by recursive feature elimination algorithm, and the random seed is fixed during the division process to ensure that the division result is reproducible and avoid the influence of data distribution deviation on the screening effect. After division, the obtained training set is input into the logistic regression base classifier, and the maximum skill feature discrimination is taken as the objective function, where the discrimination is quantified by the F value calculated by variance analysis. The larger the F value, the stronger the feature discrimination ability. After training, the absolute value of the first feature coefficient in the logistic regression base classifier is extracted and sorted in descending order according to the absolute value. At the same time, the removal threshold is determined by maximizing the validation set classification accuracy, and all first feature coefficients with absolute values less than the removal threshold are marked as redundant parameters, removed from the current initial parameter data set, and a first temporary parameter subset is formed. The temporary parameter subset is re-input into the logistic regression base classifier, and after training, the absolute value of the second feature coefficient in the logistic regression base classifier is extracted and sorted in descending order according to the absolute value. The removal threshold is removed again to remove redundant parameters and form a second temporary parameter subset. The operation is continuously performed until the accuracy of the validation set of adjacent two iterations fluctuates less than 1 percent, and the final obtained key parameter subset related to the skill core is output.

[0011] In a preferred embodiment, the key parameter screening module normalizes the absolute value of the feature coefficient obtained by the logistic regression base classifier in the final round of iteration as the parameter weight value of the key parameter.

[0012] In a preferred embodiment, after obtaining the key parameter subset and the parameter weight value, the copyright evidence module reconstructs the original feature matrix according to the key parameter-parameter weight value dimension, constructs a column vector by arranging the parameter weight value in descending order of the feature coefficient, and performs row-by-row weighting on the original feature matrix and the weight column vector through matrix multiplication operation, that is, the corresponding row sum and the corresponding column element of the original feature matrix are multiplied by the corresponding column element of the weight vector to generate a weighted feature matrix. The matrix strengthens the feature expression of high correlation degree technical parameters and weakens the interference of low weight parameters through weight distribution; Calculate the mean of each row in the weighted feature matrix, subtract the mean of each column from the weighted feature matrix to obtain a centralized feature matrix, calculate the covariance between the key parameters based on the centralized feature matrix, construct a covariance matrix to reflect the linear correlation degree between the parameters, call the linear algebra library to decompose the covariance matrix to obtain the decomposition eigenvalue and the corresponding decomposition eigenvector, calculate the variance contribution rate of each decomposition eigenvalue to the total sum of all decomposition eigenvalues, accumulate the variance contribution rate in the order of all decomposition eigenvalues, stop accumulation when the general threshold is reached, which is 90%, and the recognition accuracy after dimensionality reduction can be adjusted through the validation set. Record the number of accumulated decomposition eigenvalues, extract the column vector corresponding to the number of decomposition eigenvalues from the original feature matrix, and construct a necessary feature vector matrix by sequentially extracting and collecting. Perform matrix multiplication on the centralized feature matrix and the necessary feature vector matrix to obtain a necessary feature matrix. After normalizing the necessary feature matrix, the first necessary feature vector has a length of 1. Output the first necessary feature vector corresponding to each key parameter to complete the introduction of the component analysis mechanism.

[0013] In a preferred embodiment, the copyright evidence module obtains the necessary feature matrix through the component analysis mechanism introduced in the recursive feature elimination algorithm. The copyright evidence module performs secondary fusion on the first necessary feature vector and the parameter weight value, multiplies each dimension value of the first necessary feature vector by the weight value of the corresponding parameter, and performs global scaling processing on the operation result to output the second necessary feature vector after weighting and scaling. The vector carries necessary core information and parameter weight association characteristics at the same time.

[0014] In a preferred embodiment, after obtaining the second necessary feature vector, the copyright evidence module introduces a self-defined encoding strategy, including: performing precision fixing processing on each dimension value of the second necessary feature vector to retain at least 8 significant digits after the decimal point, concatenating the processed values in dimension order, and using English comma as a separator between dimensions to form an original feature string. The second necessary feature vector is converted into a raw feature string format and input into a hash algorithm to generate an initial hash value.

[0015] In a preferred embodiment, the copyright evidence module randomly generates a creation identifier for the artwork, and embeds the creation identifier into the initial hash value in the hexadecimal string format, including: the creation identifier is randomly covered in the corresponding embedding bit of the initial hash value through bit operation, the embedding bit is first executed with 0 emptying, and then the identification bit is written through or operation, the hexadecimal hash value after embedding the identifier is output, the check code is converted into a hexadecimal string through a cyclic redundancy check algorithm, and the hexadecimal hash value after embedding the identifier is spliced to form a check string, the spliced check string is verified through the cyclic redundancy check algorithm in reverse, and the feature string with a fixed length of 72 bits is output after verification, the feature string is used as a feature code, the feature code is uploaded to the blockchain and saved, and the copyright evidence is completed.

[0016] The beneficial effects of the present application are: the recursive feature elimination algorithm is used to screen the key parameter subset related to the core of the skill, the redundant information is removed, the data processing amount is reduced, the component analysis mechanism is introduced to map the high-dimensional key parameters to the low-dimensional feature space, the feature processing efficiency is improved on the premise of retaining the integrity of the core information, and the recognition accuracy and operation performance are considered; The visual features and skill operation parameters are fused, the two types of feature vectors with matched dimensions are extracted through the CNN, and the skill core area is located by means of the self-attention mechanism, which makes up for the limitation of the traditional technology which only depends on the visual features, the double threshold classifier is combined to make a double determination on the visual similarity and the skill similarity, the implicit difference of the imitated work in the skill core features can be accurately captured, the misjudgment probability is significantly reduced, and a reliable basis is provided for the copyright ownership determination. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The flowchart for implementing the present application is shown. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0019] In the description of the present application, the terms "first", "second" are only for descriptive purposes and cannot be understood to indicate or imply relative importance or implicitly indicate the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise explicitly specified.

[0020] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or description". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can recognize that the present application can be implemented without using these specific details. In other examples, well-known structures and processes will not be described in detail to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope in accordance with the principles and characteristics disclosed in the present application.

[0021] As shown in the accompanying Figure 1 The present embodiment provides a fusion of art work image recognition and copyright evidence storage system based on blockchain, the system comprises: A feature vector extraction module is used to extract visual features of art works by CNN, and a visual feature vector with uniform output dimension is output by a fully connected layer. The skill operation parameters obtained by the data acquisition device are extracted by CNN, and a skill feature vector with matching output dimension is output. A feature focusing module is used to receive the skill feature vector and input it to a weight distribution layer to generate a feature correlation weight matrix. The feature correlation weight matrix and the visual feature vector are jointly input into a self-attention mechanism to automatically locate the skill core area by calculating the correlation degree of the visual area and the skill feature. Finally, the focused skill feature vector is output. A similarity recognition module is used to process the focused skill feature vector and the visual feature vector by an alignment algorithm to generate a skill and visual fusion feature matrix. The fusion feature matrix is input into an SVM to construct a double-threshold classifier and output a similarity value. Then, the identification of the art work is determined according to the preset threshold. A key parameter screening module is used to obtain the deconstruction dimension parameters and visual pixel parameters of the art work based on image recognition technology after identification, construct an initial parameter data set, process the initial parameter data set by a recursive feature elimination algorithm, and output a key parameter subset related to the skill core. The copyright evidence module is used for introducing a component analysis mechanism in a recursive feature elimination algorithm, mapping a key parameter subset to a low-dimensional feature space to generate a first necessary feature vector containing skill and visual core information, introducing a self-defined coding strategy, fusing a fixed-length feature code through a self-defined hash mapping rule, and chaining the feature code to complete copyright evidence.

[0022] The above system is aimed at image recognition and copyright evidence of art works, and includes the following specific steps: S1, the art works are subjected to visual feature extraction by CNN, and the visual feature vectors of uniform output dimension are output by the fully connected layer, and the skill feature vectors of matching output dimension are output by CNN feature extraction of skill operation parameters obtained by the data acquisition device.

[0023] After the image of the art work is uniformly scaled to a preset resolution, it is input to the convolutional neural network, the visual features are extracted by the convolutional layer, the dimension is reduced by the pooling layer, and finally the visual features are mapped to the initial feature vector by the fully connected layer, and the standardized visual feature vector is output after normalization processing; After zero point calibration of the sensor by a standard calibration tool such as a force value calibration weight and an angle calibration instrument, and elimination of baseline drift, the sensor is preset with a sampling frequency, and starts working by activating a start signal, the sensor collects original data in real time, and associates the force, trajectory, angle and other parameters with the synchronously collected visual image or video frame through a time stamp, generates skill operation parameters associated with visual presentation, and provides a synchronous data source for subsequent feature extraction, the adaptive sensor includes an inertial measurement unit (IMU) for collecting operation angle and trajectory, a material sensing probe for collecting paint, silk thread and material interaction parameters, and a high-definition visual sensor for assisting in capturing details of the skill, after the skill operation parameters are subjected to outlier rejection and standardization processing, they are input to the convolutional layer and the fully connected layer of the CNN, and the high-dimensional skill operation parameters are mapped to the skill feature vectors consistent with the dimension of the visual feature vectors; The feature focusing module takes the visual feature vectors and the skill feature vectors of consistent dimension as input, introduces a skill feature guide weight layer in the self-attention mechanism, and outputs the focused skill feature vectors.

[0024] S2, the skill feature vectors are received and input to the weight distribution layer to generate a feature correlation weight matrix, the feature correlation weight matrix and the visual feature vectors are jointly input to the self-attention mechanism, the correlation degree between the visual area and the skill feature is calculated to automatically locate the skill core area, and finally the focused skill feature vectors are output.

[0025] By calculating the cosine similarity between the skill feature vector and the core feature template library, and mapping the weight values ​​to the [0,1] interval, a feature association weight matrix is ​​generated. This matrix reshapes the visual feature vector into a feature map format, and together with the feature association weight matrix, it is input into a self-attention mechanism that incorporates a skill feature-guided weight layer. This includes: The feature focusing module collects third-party digital art resources and performs binary-dimensional deconstruction to obtain the sample technique operation parameters and sample visual feature parameters for each art process. Non-critical dimensions are eliminated, and pixel normalization and light compensation processing are performed on the sample visual feature parameters. The sample technique operation parameters and sample feature parameters are classified according to the art process type to construct a categorized sample subset, which lays the data foundation for feature extraction. Each categorized sample subset is filtered for key features through a recursive feature elimination algorithm, and the dimensionality of the filtered key features is reduced. The mean and standard deviation of the low-dimensional feature vectors of all samples in each categorized sample subset are calculated to generate low-dimensional feature vectors of the samples, which are constructed as core feature templates and stored in the sample library. After the feature focusing module inputs the skill feature vector, it calculates the cosine similarity value between the skill feature vector and each sample low-dimensional feature vector in the core feature template. Since the skill feature vector and each sample low-dimensional feature vector in the core feature template have been normalized, the calculation of the cosine similarity value is equivalent to the dot product of the two vectors. The higher the cosine similarity value, the more it represents that the input skill feature vector matches the corresponding sample low-dimensional feature vector. All the obtained cosine similarity values ​​are mapped to the [0,1] interval as weight values ​​to construct a feature association weight matrix. The feature association weight matrix and feature map are input into the self-attention mechanism. The self-attention mechanism generates a query matrix composed of feature maps, a key matrix composed of feature association weight matrices, and a value matrix through linear transformation. The attention score is obtained by calculating the dot product of the query matrix and the key matrix. The attention score is used to generate an attention map through the Softmax function, and the dimension of the attention map is aligned with the value matrix to ensure that each feature position has a unique corresponding attention score. The value matrix is ​​then updated with weights using the attention map, including multiplying the weights of features with attention scores exceeding the attention threshold by an enhancement coefficient, and multiplying the weights of features with attention scores less than or equal to the attention threshold by a decay coefficient. Finally, global average pooling is used to map the updated feature map into a focused feature vector.

[0026] It should be noted that the feature weights in the weighted update refer to the weights of the skill feature vector and the visual feature vector in the feature association weight matrix. In some specific implementations, the specific operation is as follows: The attention map and value matrix output from the self-attention mechanism are aligned in dimension through a dimension broadcast mechanism to ensure that each feature position has a unique attention score. A binary mask matrix is constructed by iterating through all elements in the attention map. If the attention score is greater than 0.7, the corresponding mask position is assigned a value of 1, marking it as a core skill area. If the attention score is less than or equal to 0.7, the corresponding mask position is assigned a value of 0, marking it as a low-weight area. The region with a value of 1 in the mask matrix is associated with the corresponding position in the value matrix. The feature in this region is subjected to a weight gain operation, i.e., the original feature value is kept unchanged or multiplied by a gain coefficient. The specific coefficient is dynamically adjusted according to the concentration of the core feature template to ensure that the core region feature strength is not weakened and the original information is retained. For the region with a value of 0 in the mask matrix, the feature value at the corresponding position in the value matrix is multiplied by a fixed attenuation coefficient to reduce the interference of non-core regions on the overall feature extraction. In this application, the attention threshold is 0.7. In the visual feature extraction scenario of the self-attention mechanism, 0.6-0.8 is the mainstream core region screening threshold interval. 0.7 is the middle value, which can be compatible with the concentration difference of different art skills. It avoids the threshold being too high, causing core region selection to be missed, or too low, causing noise to be mixed in. In the pre-training of the core feature template using the attention mechanism, an accurate attention threshold can be further obtained.

[0027] S3, the focused skill feature vector and the visual feature vector are processed through an alignment algorithm to generate a skill and visual fusion feature matrix. The fusion feature matrix is input into an SVM to construct a double-threshold classifier and output a similarity value. The identification of the art work is completed according to a pre-set threshold.

[0028] After the visual feature vector and the obtained skill feature vector pass the consistency test, it is ensured that both are standardized vectors of the same dimension. If there is a dimensional deviation, it is calibrated through zero padding or feature pruning. The visual feature vector and the skill feature vector are multiplied element by element to obtain an interactive feature vector reflecting the feature correlation. The original visual feature vector and skill feature vector are concatenated with the interactive feature vector in the channel dimension to form a fusion feature vector. The fusion feature vector is reshaped into a two-dimensional skill and visual fusion feature matrix by a matrix reconstruction algorithm. In the skill and visual fusion feature matrix, the row dimension corresponds to the feature category, and the column dimension corresponds to the feature value, providing dimension guidance for subsequent similarity calculation. Each row of the skill and visual fusion feature matrix is labeled with categories such as visual basic features corresponding to the visual feature vector, skill core features corresponding to the focused skill feature vector, and interactive features corresponding to the interactive feature vector, providing dimension guidance for subsequent similarity calculation. According to the labeled category label, the skill and visual fusion feature matrix is split into a visual feature sub-matrix and a skill feature sub-matrix, which correspond to the labeled visual basic feature row vector and the labeled visual basic feature row vector, respectively. The visual feature sub-matrix and the skill feature sub-matrix are input into the pre-trained double-threshold classifier. It should be noted that the double-SVM classifier corresponds to an independent SVM model, both of which use an RBF kernel function with a penalty coefficient of 1.0. The kernel function parameter is determined by cross-validation optimization. The similarity values of the visual feature sub-matrix and the skill feature sub-matrix to the core feature template are calculated by the SVM model through the ratio of vector dot product to modulus product. The similarity values are calculated by the cosine similarity formula, and the visual feature similarity value and the skill feature similarity value are output respectively. The output visual feature similarity value and the skill feature similarity value are plotted as ROC curves, and the specific process is as follows: Generate all candidate thresholds in the [0, 1] interval with a fixed step size. Traverse each candidate threshold, calculate the number of real categories as non-imitation works that are correctly judged as non-imitation works under the candidate threshold, and mark it as true positive. The number of real categories as imitation works that are misjudged as non-imitation works is false positive. The number of real categories as imitation works that are correctly judged as imitation works is marked as true negative. The vector of real categories as non-imitation works that are misjudged as imitation works is marked as false negative. The true positive rate is calculated as the ratio of true positive to the sum of true positive and false negative. Similarly, the false positive rate is calculated. The ROC curve is plotted with the true positive rate as the horizontal axis and the false positive rate as the vertical axis. The area under the curve is calculated. If the area under the curve is greater than 0.95, it is considered that the feature index in the visual feature sub-matrix or the skill feature sub-matrix has effective distinguishing ability. The minimum threshold is found based on the visual feature similarity value ROC curve and the skill feature similarity value ROC curve, and the visual feature minimum threshold and the skill feature minimum threshold are found respectively as the visual feature initial value and the skill feature initial value. After cross-validation of the visual feature initial value and the skill feature initial value, the accuracy is ensured to be greater than 92%. Any of the following identification settings is included: If both the visual feature similarity value and the skill feature similarity value exceed the visual feature initial value and the skill feature initial value, it is determined as non-imitation work. If the visual feature similarity value exceeds the visual feature initial value and the skill feature similarity value is less than or equal to the skill feature initial value, it is determined as imitation work. If the visual feature similarity value is less than or equal to the visual feature initial value and the skill feature similarity value is less than or equal to the skill feature initial value, it is determined as non-imitation work. The visual feature similarity value and the skill feature similarity value are both less than or equal to the visual feature initial value and the skill feature initial value, and it is determined that the work is a copy.

[0029] S4, after identification, the deconstruction dimension parameter and the visual pixel parameter of the art work are obtained based on image recognition technology, an initial parameter data set is constructed, the initial parameter data set is processed by a recursive feature elimination algorithm, and a key parameter subset related to the skill core is output.

[0030] The art work identified by recognition is based on image recognition technology to obtain structure dimension parameters and visual pixel parameters, such as needle type, silk angle, knife mark depth, stroke turning angle, etc. An initial parameter data set containing parameter name, collection timestamp and data identification is constructed, missing values are filled in by neighborhood interpolation method, noise in visual pixel parameters is smoothed by Gaussian filter, and finally the initial parameter data set is standardized; The standardized initial parameter data set is divided into layers by recursive feature elimination algorithm, and the random seed is fixed during the division process to ensure that the division result is reproducible and avoid data distribution deviation affecting the screening effect. After division, the obtained training set is input into a logistic regression base classifier, and the maximum skill feature discrimination is taken as the objective function, wherein the discrimination is quantified by F value calculated by variance analysis, and the larger the F value, the stronger the feature discrimination ability. After training, the first feature coefficient in the logistic regression base classifier is extracted, the absolute value is taken, and the first feature coefficient is arranged in descending order according to the absolute value from large to small. At the same time, the removal threshold is determined by maximizing the classification accuracy of the validation set, and the first feature coefficient with absolute value less than the removal threshold is marked as redundant parameter, which is removed from the current initial parameter data set to form a first temporary parameter subset. The temporary parameter subset is re-input into the logistic regression base classifier, and after training, the second feature coefficient in the logistic regression base classifier is extracted, the absolute value is taken, and the second feature coefficient is arranged in descending order according to the absolute value from large to small. The removal threshold is removed again to remove redundant parameters and form a second temporary parameter subset. The operation is continuously operated until the accuracy of the validation set of adjacent two iterations fluctuates less than 1 percent, and the finally obtained key parameter subset related to the skill core is output.

[0031] The structural dimension parameter focuses on the operation essence characteristics of the non-heritage technology, and the visual pixel parameter reflects the explicit presentation effect of the work. The divided training set is input into the logistic regression base classifier, the maximum of the technology feature distinction degree is taken as the objective function, the F value calculated by the variance analysis is the ratio of the inter-group variance to the intra-group variance, the feature distinction ability is quantified, and the higher the F value is, the greater the contribution of the feature to the identification of the technology category is. After the training is completed, the feature coefficient absolute value of the classifier is extracted and arranged in descending order, the classification accuracy of the verification set is maximized to determine the redundancy parameter elimination threshold, the features with absolute values lower than the threshold are marked as redundant and eliminated, the first temporary parameter subset is formed through preliminary screening, and the loop operation is performed. Gradually strip the redundant features through the feedback loop of multiple training and verification.

[0032] After the feature coefficient absolute value obtained by the logistic regression base classifier in the final iteration round is normalized, the parameter weight value is taken as the key parameter.

[0033] In the recursive feature elimination algorithm, the component analysis mechanism is introduced, the key parameter subset is mapped to the low-dimensional feature space to generate the first necessary feature vector containing the technology and visual core information, and the self-defined coding strategy is introduced. A fixed-length feature code is generated by fusing the self-defined hash mapping rule, the feature code is chained, and the copyright evidence is completed.

[0034] After obtaining the key parameter subset and the parameter weight value, the original feature matrix is reconstructed according to the dimension of the key parameter-parameter weight value, the parameter weight value is arranged in descending order of the feature coefficient to construct a column vector, and the original feature matrix and the weight column vector are weighted row by row through matrix multiplication operation, that is, the corresponding row and column elements of the original feature matrix are multiplied by the corresponding column elements of the weight vector, to generate a weighted feature matrix. The matrix strengthens the feature expression of the high-correlation technology parameter and weakens the interference of the low-weight parameter through weight distribution; The mean of each row in the weighted feature matrix is calculated, each column of the weighted feature matrix is subtracted by the mean of the corresponding column to obtain a centralized feature matrix, the covariance between the key parameters is calculated based on the centralized feature matrix, and a covariance matrix is constructed to reflect the linear correlation degree between the parameters. The decomposition feature value and the corresponding decomposition feature vector are obtained by decomposing the covariance matrix by calling the linear algebra library, the variance contribution rate of each decomposition feature value to the total sum of all decomposition feature values is calculated, the variance contribution rate is accumulated in the order of all decomposition feature values, and when the general threshold is reached, the general threshold is 90%. The recognition accuracy after dimensionality reduction of the verification set can be adjusted, the accumulation is stopped, and the number of accumulated decomposition feature values is recorded. The column vector corresponding to the number of decomposition feature values in the original feature matrix is extracted in order and collected to construct a necessary feature vector matrix; The centering feature matrix is subjected to matrix multiplication operation with the necessary feature vector matrix to obtain a necessary feature matrix, and after normalization processing of the necessary feature matrix, the first necessary feature vector is made to have a length of 1, and the first necessary feature vector corresponding to each key parameter is output, and the introduction of the component analysis mechanism is completed.

[0035] When performing the row-by-row weighting operation through matrix multiplication, each row of sample data of the original feature matrix is multiplied with the weight column vector element by element, so that the feature values of high-weight parameters are amplified and the feature values of low-weight parameters are suppressed, thereby highlighting the expression strength of the core skill feature and reducing the interference of secondary parameters; subsequently, the mean offset influence caused by the dimension difference is eliminated by subtracting the corresponding column mean value, to obtain a centering feature matrix, and after the eigenvalues and eigenvectors are obtained by decomposing the covariance matrix by calling a linear algebra library, the information carrying capacity of each eigenvector is measured by calculating the variance contribution rate of the eigenvalue, and the corresponding eigenvector set is accumulated to form a necessary feature vector matrix, thereby realizing the stripping of high-correlation redundant parameters and the focusing of core features.

[0036] In the component analysis mechanism, the core lies in projecting the high-dimensional centering feature into a low-dimensional space spanned by the necessary feature vectors to obtain a necessary feature matrix containing only core information, and performing normalization processing on the necessary feature matrix to adjust the element values of the first necessary feature vector by the vector length calculation formula, i.e., the length is the square root of the sum of squares of elements, so that the length of the first necessary feature vector is strictly 1, and the first necessary feature vector corresponding to each key parameter is finally output, which not only retains the linear correlation information of the core skill feature, but also has a standardized numerical form.

[0037] The copyright evidence module performs secondary fusion of the first necessary feature vector and the parameter weight value, multiplies each dimension value of the first necessary feature vector by the weight value of the corresponding parameter, and performs global scaling processing on the operation result to adjust the mean value of the weighted vector to 0.5, thereby enhancing the recognition degree of subsequent hash coding, and outputs the second necessary feature vector after weighting and scaling, which simultaneously carries necessary core information and parameter weight correlation characteristics.

[0038] After taking the first necessary feature vector output by the component analysis mechanism and carrying the linear correlation information of the core skill of the art work and the parameter weight value output by the recursive feature elimination algorithm and reflecting the correlation strength of the feature and the skill as double inputs, a dimension-by-dimension weighting operation is performed, each dimension value of the first necessary feature vector is multiplied by the corresponding parameter weight value, so that the high-weight core skill feature is directionally amplified in the vector, and the interference of the low-weight feature is further weakened.

[0039] To improve the recognition and stability of subsequent hash coding, a global scaling process is performed on the weighted vector, and the vector mean is adjusted to 0.5 through the mean adjustment formula to avoid the risk of coding collision caused by the overall shift of the vector value. The final output of the second necessary feature vector realizes the binding of the skill core information and the feature weight, providing a feature carrier with both accuracy and discrimination for subsequent coding.

[0040] After obtaining the second necessary feature vector, a custom coding strategy is introduced, including: performing precision fixing processing on each dimension value of the second necessary feature vector, retaining at least 8 significant digits after the decimal point, concatenating the processed values in dimension order, and using English comma as the separator between dimensions to form the original feature string. Selecting a hash algorithm as the coding tool, the original feature string format of the second necessary feature vector is input into the hash algorithm to generate an initial hash value, and the initial hash value is converted to hexadecimal coding to obtain the initial hash value in hexadecimal string format.

[0041] Among them, precision fixing processing is performed to retain at least 8 significant digits after the decimal point, eliminating coding bias caused by floating-point precision differences in different devices or environments, and ensuring that the same feature vector is converted to a unified string. Then, the values are concatenated in dimension order separated by English commas to form a structured original feature string, providing a standardized input format for the hash algorithm.

[0042] Selecting a hash algorithm to encode the original feature string generates a binary initial hash value, which is then converted to a string format through hexadecimal coding — the principle of hexadecimal conversion is to map 4-bit binary numbers to 1-bit hexadecimal numbers, which not only compresses the data volume, but also facilitates subsequent bit operation operations.

[0043] Finally, a random creation identifier for the artwork is given, and the creation identifier is embedded into the obtained initial hash value in hexadecimal string format, including: embedding the creation identifier randomly in the corresponding embedding bits of the initial hash value through bit operation, first performing AND 0 empty on the embedding bits, then performing OR identifier bit writing, outputting the hexadecimal hash value after embedding the identifier, converting the check code to a hexadecimal string through the cyclic redundancy check algorithm, concatenating it with the hexadecimal hash value after embedding the identifier to form a check string, and verifying the concatenated check string through the cyclic redundancy check algorithm. If the verification is passed, a feature string with a fixed length of 72 bits is output, which is used as a feature code, uploaded to the blockchain, and saved, completing the copyright evidence.

[0044] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0045] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0046] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0047] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0048] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0049] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0050] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A system for image recognition and copyright registration of artworks integrating blockchain technology, characterized in that, The system includes: Feature vector extraction module; used to extract visual features from artworks using CNN, and output visual feature vectors with uniform dimensions by fully connected layers. Then, the technical operation parameters obtained by the data acquisition device are used for feature extraction by CNN, and output technical feature vectors with dimensions matching the visual feature vectors. Feature focusing module: It receives the skill feature vector and inputs it into the weight allocation layer to generate a feature association weight matrix. The feature association weight matrix and the visual feature vector are input into the self-attention mechanism. By calculating the correlation between the visual region and the skill feature, the core region of the skill is automatically located, and the focused skill feature vector is finally output. The similarity recognition module is used to process the focused technical feature vector and visual feature vector through an alignment algorithm to generate a technical and visual fusion feature matrix. The fusion feature matrix is ​​then input into an SVM to construct a dual-threshold classifier and output a similarity value. Finally, the recognition and judgment of the artwork is completed based on a preset threshold. The key parameter filtering module is used to obtain the deconstruction dimension parameters and visual pixel parameters of the artwork based on image recognition technology after the recognition is passed, construct an initial parameter dataset, process the initial parameter dataset through a recursive feature elimination algorithm, and output a subset of key parameters related to the core of the technique. The copyright registration module introduces a component analysis mechanism into the recursive feature elimination algorithm. It maps a subset of key parameters to a low-dimensional feature space to generate the first necessary feature vector containing the core information of the technique and vision. It also introduces a custom encoding strategy and generates a fixed-length feature code through a custom hash mapping rule. The feature code is then put on the blockchain to complete the copyright registration.

2. The art image recognition and copyright registration system integrating blockchain as described in claim 1, characterized in that, The feature vector extraction module scales the image of the artwork to a preset resolution and then inputs it into a convolutional neural network. Visual features are extracted through the convolutional layer, sampled and reduced in dimensionality through the pooling layer, and finally mapped to an initial feature vector by the fully connected layer. After normalization, a standardized visual feature vector is output. After outlier removal and standardization of the skill operation parameters, they are input into the convolutional and fully connected layers of the CNN to map the high-dimensional skill operation parameters into skill feature vectors with the same dimension as the visual feature vectors. The feature focusing module takes visual feature vectors and technical feature vectors of the same dimension as input, introduces a technical feature guiding weight layer into the self-attention mechanism, and outputs the focused technical feature vector.

3. The art image recognition and copyright registration system integrating blockchain as described in claim 1, characterized in that, The feature focusing module calculates the cosine similarity between the technique feature vector and the core feature template library, maps the weight values ​​to the [0,1] interval, generates a feature association weight matrix, reshapes the visual feature vector into a feature map format, and inputs it along with the feature association weight matrix into a self-attention mechanism that introduces a technique feature-guided weight layer, including: The feature focusing module collects third-party digital art resources and performs binary-dimensional deconstruction to obtain the sample technique operation parameters and sample visual feature parameters for each art process. Non-critical dimensions are eliminated, and pixel normalization and light compensation processing are performed on the sample visual feature parameters. The sample technique operation parameters and sample feature parameters are classified according to the art process type to construct a categorized sample subset. Each categorized sample subset is filtered for key features through a recursive feature elimination algorithm, and the dimensionality of the filtered key features is reduced. The mean and standard deviation of the low-dimensional feature vectors of all samples in each categorized sample subset are calculated to generate low-dimensional feature vectors of the samples, which are then constructed as core feature templates and stored in the sample library. After the feature focusing module inputs the skill feature vector, it calculates the cosine similarity value between the skill feature vector and each sample low-dimensional feature vector in the core feature template, and maps all the obtained cosine similarity values ​​to weight values ​​to construct a feature association weight matrix. The feature association weight matrix and feature map are input into the self-attention mechanism. The self-attention mechanism generates a query matrix composed of feature maps, a key matrix composed of feature association weight matrices, and a value matrix through linear transformation. The attention score is obtained by calculating the dot product of the query matrix and the key matrix. The attention score is used to generate an attention map through the Softmax function, and the dimension of the attention map is aligned with the value matrix. The value matrix is ​​then updated with weights using the attention map, including multiplying the weights of features with attention scores exceeding the attention threshold by an enhancement coefficient, and multiplying the weights of features with attention scores less than or equal to the attention threshold by a decay coefficient. Finally, global average pooling is used to map the updated feature map into a focused technique feature vector.

4. The art image recognition and copyright registration system integrating blockchain as described in claim 3, characterized in that, The similarity recognition module performs a consistency check on the visual feature vector and the obtained skill feature vector, then performs an element-wise multiplication operation on the visual feature vector and the skill feature vector to obtain an interaction feature vector that reflects the feature correlation. The original visual feature vector and skill feature vector are then concatenated with the interaction feature vector along the channel dimension to form a fused feature vector. The matrix reconstruction algorithm reshapes the fused feature vector into a two-dimensional skill and visual fusion feature matrix, with each row of the skill and visual fusion feature matrix labeled with a category label. Based on the labeled category tags, the skill and visual fusion feature matrix is ​​split into a visual feature sub-matrix and a skill feature sub-matrix. The visual feature sub-matrix and the skill feature sub-matrix are respectively input into a pre-trained dual-threshold classifier. The SVM model calculates the similarity value between the visual feature sub-matrix and the skill feature sub-matrix and the core feature template by the ratio of the vector dot product to the modulus product, and outputs the visual feature similarity value and the skill feature similarity value respectively. ROC curves are plotted for the output visual feature similarity values ​​and skill feature similarity values, respectively. Minimum thresholds for both visual and skill features are then found and used as initial values ​​for the visual and skill features, respectively. After cross-validation of the initial visual and skill feature values, any of the following recognition criteria are applied: If both the visual feature similarity value and the technical feature similarity value exceed the initial values ​​of the visual feature and the technical feature, the work is determined to be a non-imitative work. If a work has a visual feature similarity value that exceeds the initial visual feature value, and a technical feature similarity value that is less than or equal to the initial technical feature value, it is determined to be an imitation work. If a work has a visual feature similarity value less than or equal to the initial visual feature value, and a technique feature similarity value less than or equal to the initial technique feature value, it is determined to be a non-imitative work. If both the visual feature similarity value and the technical feature similarity value are less than or equal to the initial values ​​of the visual feature and the technical feature, the work is judged to be an imitation.

5. The art image recognition and copyright registration system integrating blockchain as described in claim 1, characterized in that, The key parameter filtering module will obtain structural dimension parameters and visual pixel parameters based on image recognition technology from the identified artworks, construct an initial parameter dataset with fields including parameter name, collection timestamp, and data identifier, fill in missing values ​​using neighborhood interpolation, smooth the noise in the visual pixel parameters using Gaussian filtering, and finally standardize the initial parameter dataset. The standardized initial parameter dataset is hierarchically randomly partitioned using a recursive feature elimination algorithm, with a fixed random seed during the partitioning process. After partitioning, the resulting training set is input into a logistic regression base classifier, with the objective function being to maximize the discriminative power of the skill features. After training, the absolute values ​​of the first feature coefficients in the logistic regression base classifier are extracted and sorted in descending order of absolute value. At the same time, the elimination threshold is determined by maximizing the classification accuracy of the validation set. All first feature coefficients with absolute values ​​lower than the elimination threshold are marked as redundant parameters and removed from the current initial parameter dataset to form the first temporary parameter subset. The temporary parameter subset is used as the new input to the logistic regression base classifier. After training, the absolute values ​​of the second feature coefficients in the logistic regression base classifier are extracted and sorted in descending order of absolute value. The redundant parameters are then removed again using the determined removal threshold to form the second temporary parameter subset. This process is repeated until the accuracy fluctuation of the validation set between two adjacent iterations is less than 1 percentage point. Finally, the key parameter subset related to the core of the technique is output.

6. The art image recognition and copyright registration system integrating blockchain as described in claim 5, characterized in that, The key parameter screening module normalizes the absolute values ​​of the feature coefficients obtained by the logistic regression base classifier in the final iteration round and uses them as the parameter weight values ​​of the key parameters.

7. The art image recognition and copyright registration system integrating blockchain as described in claim 6, characterized in that, After obtaining the key parameter subset and parameter weight values, the copyright certificate module reconstructs the original feature matrix of the key parameter subset according to the dimension of key parameter - parameter weight values, constructs the parameter weight values ​​into a column vector by arranging the feature coefficients in descending order, and generates a weighted feature matrix by performing matrix multiplication operations to weight the original feature matrix and the weight column vector row by row. Calculate the mean of each row in the weighted feature matrix, subtract the mean of the corresponding column from the weighted feature matrix to obtain the centered feature matrix, calculate the covariance between key parameters based on the centered feature matrix, construct the covariance matrix, call the linear algebra library to decompose the covariance matrix to obtain the decomposed eigenvalues ​​and corresponding decomposed eigenvectors, calculate the variance contribution rate of each decomposed eigenvalue to the sum of all decomposed eigenvalues, accumulate the variance contribution rates in the sorted order of all decomposed eigenvalues, stop accumulating after reaching a general threshold, record the number of accumulated decomposed eigenvalues, extract the column vectors corresponding to the number of decomposed eigenvalues ​​in the original feature matrix, and set them to construct the necessary eigenvector matrix; The centered feature matrix and the necessary feature vector matrix are multiplied to obtain the necessary feature matrix. After normalization of the necessary feature matrix, the first necessary feature vector is made to have a magnitude of 1. The first necessary feature vector corresponding to each key parameter is then output, thus completing the introduction of the component analysis mechanism.

8. A system for image recognition and copyright registration of artworks integrating blockchain technology as described in claim 6, characterized in that, The copyright certificate module obtains the necessary feature matrix through the component analysis mechanism introduced in the recursive feature elimination algorithm. The copyright certificate module performs a secondary fusion of the first necessary feature vector and the parameter weight value, multiplies each dimension value of the first necessary feature vector by the corresponding parameter weight value, and performs global scaling processing on the calculation result to obtain the second necessary feature vector.

9. A blockchain-integrated art image recognition and copyright registration system according to claim 8, characterized in that, After obtaining the second necessary feature vector, the copyright certificate module introduces a custom encoding strategy, including: performing precision fixing on each dimension value of the second necessary feature vector, concatenating the processed values ​​in dimensional order, and using English commas as separators between dimensions to form the original feature string; A hash algorithm is selected as the encoding tool. The original feature string format after the second necessary feature vector is converted is input into the hash algorithm to generate an initial hash value. The initial hash value is then converted into hexadecimal encoding to obtain the initial hash value in hexadecimal string format.

10. A blockchain-integrated art image recognition and copyright registration system according to claim 9, characterized in that, The copyright registration module assigns a random creation identifier to the artwork and embeds it into an initial hash value in hexadecimal string format. This includes: randomly overwriting the corresponding embedding bit in the initial hash value using bitwise operations; clearing the embedding bit with 0 and then writing the identifier bit; outputting the hexadecimal hash value after embedding the identifier; converting the checksum into a hexadecimal string using a cyclic redundancy check (CRUD) algorithm; concatenating the checksum with the hexadecimal hash value after embedding the identifier to form a checksum string; verifying the concatenated checksum string using a CRUD algorithm in reverse calculation; outputting a fixed-length feature string as a feature code; uploading the feature code to the blockchain and saving it to complete the copyright registration.

Images