The Application of Artificial Intelligence Technology in Intellectual Property Protection and Its Impact on the Cultural Industry

For patent data, the general data mining process [22] does not meet the requirements for data mining of patent data. It is necessary to analyze the patent data in detail, adjust the general process of data mining, and then provide a set of special processes applicable to patent data mining. The flow chart for patent intellectual property data mining is shown in Figure 1 below. The following special process of patent data mining is derived after analyzing the general process of data mining and the patent data set: (1) determining the target of patent data mining; (2) preparation and preprocessing of patent data; (3) selection and integration of machine learning models for patent data; (4) analysis and evaluation of patent data mining results; (5) discovery and application of patent data mining results.

Figure 1.

Patent data mining flow chart

The goal of this paper is to categorize, manage, and protect patent intellectual property. In the next better business, expand your business and occupy market share. Patent data is analyzed for feature extraction, and then the degree of influence of eigenvalue indicators on patent knowledge protection is determined.

Feature preprocessing is a very important step in data mining. Some of the feature data in our identified features are defective or do not meet the requirements, and feature preprocessing is to turn these non-compliant features into data that meets the requirements of data mining through a series of processing, and can be normal data mining data. 1)

Missing data in a dataset can reduce model fitting or may lead to model bias because the behavior and relationships of variables are not properly analyzed, which may lead to incorrect prediction or classification results. There are usually two ways to deal with missing values: deletion and filling.

The unit or size of the features differ greatly, or the variance of a feature is several orders of magnitude larger compared to other features, in which case it is easy to affect the final results, making some algorithms unable to learn other features. So we need to standardize. The following is the formula for normalization: 1z=x−μσ

where μ and σ are the mean and method of the patent data set, respectively. In the normalization process, the best results are achieved when the distribution of the original data set is normal. However, even if the data does not obey a normal distribution, this method can still be used, but the results will have a certain effect.

For patent data, the range, maximum and minimum values of the data are known, and there are no isolated points, which do not conform to the normal distribution, so this method is not very suitable for patent data.

Different features in the data set directly affect the value of the range and units. They may not be identical, and there may even be an excessive difference. In this case, the degree of convergence will fall very slowly. However, the normalization process can help accelerate the rate of convergence.

Normalization is mapping data to be between [0,1] by transforming the original data. Transforming the data to a decimal number between (0,1). After the data is normalized, the data that previously had too large a range gap is in the same order of magnitude, which removes the influence of dimensions and units. Improve the comparability of different data indicators. The formula for normalization is as follows: 2Xnorm=X−XminXmax−Xmin

Where max is the maximum value of the feature data and min is the minimum value of the feature data. Normalization has the disadvantage that it is not a problem when the data set is unchanged, but if the data set is changed and new data is added to the data set, it may result in a change in the maximum or minimum value, thus affecting the definition of the formula. This drawback is negligible for patent data because the minimum and maximum values of the data to be normalized are basically unchanged. 4)

Feature correlation is a way used to understand and analyze the relationship between multiple features, which can help to eliminate some correlations that are too large and some feature terms that do not meet the requirements. Using a matrix heat map to describe feature correlation can be a good way to show the data and correlation, and can also be a good way to analyze the degree of correlation between the data.

Feature fusion is the key to extracting features from the dataset. The original DP-MVCNN model only extracts the maximum component of each view feature, which does not retain the view feature information well. Especially, Patent-MNIST has only two view features. In order to better fuse the two view feature information, the Fusi(-) function is utilized to fuse the information pairs of the two channels of Max-Mean, which compensates for the detail information lost by max-pooling. The fusion method is shown in the following equation: 3Z= Fusion (max pooling ( vector 1, vecto 2) meanpooling ( vector 1, vector 2))

The above formula is a generalized form of fusion, where the fusion function Fusion(·) can be a column splicing cat(·) and summation sum(·) of the fusion method α , β can control the information of the two channels, which will be set to 1, and finally the fusion of the information generated by the shape descriptor of the design patent. As shown below: 4Z= Fusion cat (max pooling ( vector 1, vector 2) meanpooling ( vector 1, vector 2)) 5Z= Fusion sum (max pooling ( vector 1, vector 2) meanpooling ( vector 1, vector 2))

Classification module: finally, the generated shape descriptors are fed into the classifier to realize the classification of Patent-MNIST dataset.The appearance patent classification model uses Softmax function to classify the shape descriptors of appearance patents.The mathematical model of Softmax classifier is given in the following equation: 6yk = exp(xk)∑niexp(xi)

In this paper, we need to categorize different types of apparel patents using the appearance patent view structure, therefore, we need to construct the appearance patent view (DP View-Graph) first.

How to utilize the view nodes to construct the DPView-Graph is the first problem to be solved in this paper. In this paper, the view vi generated from different viewpoints is regarded as the i st node of the graph, then the adjacency matrix S ∈ ▯ ^n*n of the view-graph is defined by the following function: 7Sij=ϕ(gij;θs) where g_ij =[v_i,v_j,v_i–v_j,‖v_i–v_j‖₂]∈▯¹⁰ denotes the relationship between two views and [] denotes two elements for vectorized stitching. ϕ is a nonlinear function and θ_s is a parameter. In this paper, ϕ is set as a 3-layer MLP and the output is a scalar S_ij.

In this paper, we utilize the k – Nearest Neighbor (KNN) method [23], which determines a fixed number of neighboring nodes for each node using the distance between nodes and retains only the edges between neighboring nodes, so the graph structure designed in this paper is a directed graph. The adjacency matrix is: 8Aij=Sij⋅ψ(vj∈N(vi)) where ψ(·) denotes whether v_j is a binary function within the KNN of v_i. A preserves a sparse adjacency matrix of k neighbors for S. The pre-trained ResNet-18 network [24–25] is used to extract features for the views, and each view is characterized as a 1024-dimensional feature vector, and there are four feature vectors, vector1, vector2, vector3 and vector4, for each appearance patent, and the Design Patent DPView-Graph is constructed on the basis of these feature vectors, and the initial Design Patent View-Graph is (G⁰).

For the constructed DPView-Graph, the number of nodes is changing in each View-GCN module, so that the View-Graph with N nodes in layer l is G^l , and the feature matrix is Finl . The KNN algorithm is utilized to determine the weight relationship between the view nodes. Then Graph Convolution (GCN) is defined as: 9Fl=χ(AlFinlWl) where Finl∈□Nl×d $$F_{in}^l \in { \square ^{{N_l} \times d}}$$ is the identity matrix, A^l. ∈▯^N_l×N_l is the adjacency matrix of G^l, W^l ∈▯^d×d is a learnable weight matrix, and χ is a parameterized nonlinear transformation function, LeakyReLU. The extrapolation process from Eq. reveals that the output F^l is still N × d - dimensional.

The first level feature graph is obtained by the View-GCN graph convolution module for feature learning on all views. For l-level View-GCN, which is embedding multi-view features on a G^lgraph, it is necessary to coarsen and update the l -level feature graph G^l coarsening to a G^l+1 feature graph with a smaller number of view nodes. Here, in this paper, the basic maximum pooling is first used as a baseline comparison, and the node features updated by the graph convolution operation are pooled according to the following formula: 10Fgloball=Maxpool(Gl),l=0,1,…,L−1 11F=[ Fglobal 0,Fglobal 1,…,Fglobal L−1 ]

The final global feature descriptor F is stitched together from the maximally pooled feature graph Fgloball at each level.

The construction of the appearance patent image to the View-Graph - the feature matrix and the adjacency matrix - is introduced first: 12Feature =( vector 1,…, vector 4)=CNN( view 1,…, view 4) 13A=Adj( vector 1,…, vector 4) 14G0= Patent_View_Graph (( Feature ,A))

Equation (12) denotes that images view₁ , view₂ , view₃ , view₄ are subjected to feature extraction by CNN to generate feature matrix Feature, where Feature ∈ R^4×d denotes a composition of four d-dimensional feature vectors. Eq. (13) represents the adjacency matrix A generated from the feature vectors, where the weights of A_ij can be calculated by Eq. (8). The feature matrix and adjacency matrix are computed to obtain the mathematical model G₀ based on the view-graph graph topology [26] of the appearance patent image.

The first graph convolution layer View_ GCN₁: 15F1=View−GCN1((A0Fin0W0)) 16G1= Pooling 1(F1)

Eq. (15) is the propagation model of View-GCN1, see Eq. (9) for detailed parameter settings. The output is F¹. The graph features after learning by graph convolution are denser than Feature. The pooling layer is as in Eq. (16), and the output generated by pooling is G¹.

The second graph convolution layer View_GCN₂: 17F2= View −GCN2((A1Fin1W1)) 18G2= Pooling 2(F2)

Eq. (17) is the propagation model of View-GCN2, see Eq. (9) for detailed parameter settings. The output is F² , and the graph features are denser than G¹ after learning by graph convolution. The pooling layer is shown in Eq. (18), and the output generated by pooling is G². It should be noted that Eqs. (16) and (18) are maximal pooling methods, and various pooling mechanisms are adopted in the experiments.

View-Graph generates two layers of features G¹ and G² after DP-MVGCN propagation model and the classification process is shown in the following equation: 19Y=Classifier(Fusion(G1,G2))

The above equation fuses the two layers of features G¹ and G² by the Fusion(·) function, and the Fusion (·) function can be either the cat(·) function or the sum (·) function, and the Classifier(·) classification function is three-layered in the DP-MVGCN model.

The Max-Mean based fusion pooling strategy is embedded in the graph convolutional model and pooling factors (α,β) are added to each layer of the pooling strategy: 20Fpool l= Fusion(α∗Maxpooling(Gl),β∗Meanpooling(Gl)) where Fpooll is the output of Max-Mean pooling. (α,β) is the pooling factor of layer l , while α, β are the fusion parameters of Max-pooling and Mean-pooling. For example, when the fusion method Fusion(·) is the splicing cat(·) function and the summation sum function, the Max-Mean pooling process is formulated as follows: 21Fsum =α∗Maxpooling(Gl)+β∗Meanpooling (Gl) 22Fcat =cat(α∗Maxpooling(Gl),β∗Meanpooling(Gl))

Parameters α and β control the ratio of Max-Mean pooling.

The DP-MVGCN classification model based on Max-Mean fusion pooling strategy is shown below: 23Y= Classifier( Fusion ((α∗Maxpooling(G1),β∗ Meanpooling (G1))(α∗Maxpooling(G2),β∗ Meanpooling (G2))))

In the proposed design patent apparel classification model in this paper, the level of apparel feature extraction performance is a key parameter that affects the performance of the design patent apparel classification model. It is difficult to capture the rich feature information in clothing when the feature extraction capability is insufficient, which affects the generation and expression of the clothing classification model. Meanwhile, the number of datasets has a significant impact on the design patent model. In order to realize and verify the feasibility of the apparel classification model proposed in this paper on design patent apparel classification, the following experiments are conducted.

This experiment is carried out on the garment design patent image classification dataset.The DP-MVGCN model extracts four feature vectors for vector1, vector2, vector3 and vector4. The training results of the feature extraction rate of the dataset are shown in Fig. 2. From the experimental results, it can be seen that when the training number of DP-MVGCN convolutional neural clothing image classification model for design patents reaches 70 times, the feature extraction rate of all four datasets reaches the maximum for the first time, and the extraction effect is greater than 95%. The main reason for such results is that the structure of the DP-MVGCN-trained convolutional neural clothing model is more reasonably designed, which therefore allows the model to learn and encode features better.

Figure 2.

Data collection characteristics extraction rate training results

This paper investigates the publication status of patents on intellectual property protection of apparel appearance published between 2010 and 2020, and finds that the patents on intellectual property protection of apparel appearance are classified into four datasets with four feature vectors, namely “vector1, vector2, vector3 and vector4”. The results of the survey on the number of published patent types between 2010 and 2020 are shown in Figure 3 below. The experimental results show that by 2020, the vector1 dataset has the most number of patents published, followed by vector3 and vector2 datasets, and vector4 dataset has the least number of patents. Their patent numbers are 595,700, 575,200, 560,100, and 514,900, respectively, but the change in the number of patents published in the same year for the four patent datasets is relatively small. In addition, analyzing the number of patents published in different years for the four datasets reveals that the number of patents published increases with the increase of the year, and the huge number of patents poses a challenge for their classification and protection. Therefore, this paper will classify the large patent data group by extracting features in order to provide more reasonable protection for intellectual property patents.

Figure 3.

The number of patents published between 2010 and 2020

This experiment performs model training and evaluation on the dataset of design patent image descriptions. This experiment evaluates three models, DP-MVGCN, CNN+LSTM, and VGG16+LSTM, in terms of classification training effect. The training loss results of the models on the training dataset are shown in Figure 4. The results show that the training loss of DP-MVGCN image description model on 500 images is less than that of VGG16+LSTM and CNN+LSTM models. And the loss rate of DP-MVGCN model is extremely low as the number of images increases. From the generated results it is clear that the description results generated by DP-MVGCN model express more information in more detail than the results generated by CNN+LSTM and VGG16+LSTM models.

Figure 4.

The results of the training loss in the training data set

The classification features counted and observed in this paper are mainly: the description of the subject category of the design patent, the description of the subject color, the description of the material and the description of the design features of the corresponding category of the design patent. As can be seen from Table 1, the score situation on each evaluation index and the image description evaluation score of the image description model based on the DP-MVGCN coding clothing are higher than the other two models in general. 89.65, 83.26, 88.73) and VGG16+LSTM (89.95, 92.86, 85.55, 86.82) models. This shows the superiority of the DP-MVGCN model proposed in this paper. Therefore, it can be seen that, all other conditions being equal, the classification features have an impact on the performance of the design patent image description model, and the stronger the feature extraction ability of the classification features, the better it is for the performance of the design patent image description model. In addition, by comparing the results of the evaluation scores in Table 1, it can be seen that the setting of the width of the cluster search also has an impact on the performance of the design patent image description model. That is, under the same conditions, setting the appropriate search width can make the model obtain better evaluation scores, i.e., the generated description sentences are of better quality.

Secondly, this paper uses the Locarno-MNIST patent apparel dataset based on Locarno classification standard with a total of 500 pieces (jacket, short sleeve, pants, skirt) to classify 500 apparel jacket patent intellectual property datasets using the DP-MVGCN model. The classification results of the clothing appearance patent dataset are shown in Figure 5. It can be found that the classification effect of the four types of patents is obvious and all of them are more concentrated. The DP-MVGCN model has a good classification effect.

Figure 5

Classification of clothing appearance patent data set

This paper verifies the feasibility of the image description model in the classification of design patent images by counting the recognition accuracy rate of the description text generated by the image description model on the classification features on the design patent images. The statistical accuracy of classification features can be seen in Figure 6 below. Overall, compared to the other two models, the DP-MVGCN model proposed in this paper has the highest recognition accuracy of description features for classifying categories and colors of design patents (98.98% and 96.84%), and the model has a relatively lower overall accuracy for material and design features (93.01% and 94.99%), which has a certain amount of room for improvement. However, from the overall statistical results, it can be seen that the accuracy of the model for each categorical statistical feature is higher than the accuracy of random guessing under the corresponding categorical feature. Therefore, the description generated by the image description model for the design patent image has the ability to classify the image. In addition, it can also be seen from the figure below that the models with higher evaluation scores also have relatively better statistical recognition accuracies for categorical feature recognition.

Figure 6.

Classification characteristics statistics accuracy

The era of big data is the era of knowledge explosion and rapid development of information technology, corresponding to which intellectual property rights are presented in all aspects of the social economy with a very impressive volume of data. At the same time, along with the popularization of artificial intelligence, 5G, cloud computing, Internet of Things and other technologies, more data subjects and data types can be identified and recorded, which in turn give rise to a rich variety of intellectual property works. Apparel intellectual property data is doubling at an increasingly rapid rate, with the cumulative data volume jumping from terabytes to petabytes, EBs, and even ZBs. In the face of such a huge amount of apparel intellectual property data, how to protect it, to what extent, and what effect the protection will achieve need to be carefully evaluated, which requires public administrators to use the “big data” thinking, from the “big data” perspective to This requires public administrators to use “big data” thinking to collect and screen intellectual property data, and to categorize it for management and protection.

The colorful forms and types of intellectual property data are a result of the multilayered structure. The forms of data have expanded from traditional text, images, and digital forms to include audio, video, geographic location information, analog signals, and so on, resulting in more and more personalized data. In addition, apparel log files, social media, search engines, and sensor networks all generate diverse data, constantly expanding the source and scope of intellectual property data. Therefore, utilizing big data to classify and protect apparel intellectual property rights more effectively, and checking and balancing the Internet through big data, fundamentally solves the problem.

In addition, if the huge amount of intellectual property data, including data of poor quality or even invalid, redundant or even erroneous data, are protected without differentiation, such protection mechanism is not reasonable and effective, and also results in a waste of resources. Therefore, public administrators should utilize big data technology to quickly complete the screening of massive amounts of data and classify them into different categories, so as to effectively utilize and protect intellectual property data with the fastest speed and highest efficiency.

As clothing intellectual property involves many industries and fields, it is not only directly related to intellectual property organs, but also closely linked with scientific and technological organs, judicial organs, cultural organs, educational organs, etc., and lacks a unified coordinated management mechanism for intellectual property. Intellectual property rights focus on the protection of patents, trademarks, and copyrights, which are managed by the Intellectual Property Office, the Industry and Commerce Bureau, and the Press and Publication Bureau of the three government departments. Different government departments for the protection of intellectual property rights of different policies, different ways, different law enforcement personnel, is very likely to cause “no one to manage the management, should not be managed by the cross-management” of the chaos, is not conducive to the development of intellectual property rights of the garment industry. Traditional culture for the protection of intellectual property rights is not enough to understand, the people also generally lack of awareness of intellectual property rights protection, which leads to the protection of intellectual property rights has been insufficient for a long time, especially for the protection of intellectual property rights is the lack of due social awareness. Therefore, this paper takes clothing design as an example, after classifying a huge amount of intellectual property rights, this paper explores the protection effect of intellectual property rights and the protection strength before and after classification from three different paths: judicial, administrative and social.

In this paper, the weights of the indices of different protection paths were analyzed and measured using the hierarchical analysis method to indicate the relative importance of the indices in the index system. First of all, in order to compare the importance of the indexes at the same level, we designed the “AHP Expert Consultation Questionnaire on Garment Intellectual Property Protection Index”, which is used to assist the respondents in judging the relative importance of the indexes. In this paper, the comparative measures of “judicial protection, administrative protection and social protection” are categorized into five levels: most important, relatively important, equally important, relatively unimportant and least important, which correspond to scores of 5, 4, 3, 2 and 1 respectively. In this paper, questionnaires were issued to experts and scholars in the field of apparel intellectual property protection and government staff, and 110 valid questionnaires were issued and recovered from experts and 107 valid questionnaires from the government, which accounted for 97.27% of the valid questionnaires. The weight model of every index in the Garment Intellectual Property Protection Index is calculated. Table 2 displays the indexes and weights of the clothing intellectual property protection index.

Time series data analysis is a commonly used analytical method in the field of social sciences, and we have utilized the Apparel Appearance Intellectual Property Protection Situation Awareness System to conduct a time series analysis of intellectual property protection of apparel appearance patents with the aim of comparing the data changes in different years. Detecting and early warning of the changing trends in the development of intellectual property protection for clothing, verifying the degree of coincidence between the indicator system we designed and the information system and the historical process that has taken place, and thus verifying the validity and credibility of the indicator system and the information system. The data in this paper are obtained from official websites, government work reports, court work reports, industry surveys, apparel questionnaires and other forms of data, and the unclassified apparel appearance IPR protection index from 2010-2020 is analyzed in a simple simulation. During the period from 2020 to 2023, the evaluation will be carried out on the impact of classified protection on three protection paths following the classification of apparel appearance intellectual property rights. The trend of patent intellectual property protection index of different paths is shown in Figure 7. According to the calculation, the judicial protection index, administrative protection index and social protection index in 2010-2020 are between 26.98-31.96 points, 26.5732.06 points and 21.35-31.18 points respectively. Overall, the judicial protection, administrative protection, and social protection indices are climbing steadily from low to high, but the magnitude of the climb is relatively small.

Figure 7.

Different path patent intellectual property protection index trend

It is worth noting that from 2020 onwards, this paper measures the protection index of clothing appearance after categorization. It has been found that after classification, the protection indexes of all three protection routes show a sharp increase. Compared with before 2020, the protection indexes of judicial protection, administrative protection and social protection in 2023 increased by 1.27 times, 1.29 times and 1.26 times respectively compared with 2020. It can be seen that in terms of judicial protection, the number of laws and regulations issued by the National People's Congress, the State Council, and local governments on the protection of intellectual property rights of garments is increasing, the trial of intellectual property rights of garments by people's courts at all levels across the country is increasing, the rate of fulfillment of judicial judgments and the rate of implementation in place are improving, and the transparency and credibility of judicial protection are generally recognized by all sectors of the society. In terms of administrative protection, the national administrative authorities have continuously carried out special actions to protect apparel intellectual property rights, strengthened the support for the apparel intellectual property rights industry, cracked down on the infringement of apparel intellectual property rights, and promoted the “one-stop-shopping” reform to improve the level of administrative services. In terms of social protection, social resources are gradually tilted to the field of apparel intellectual property rights, apparel intellectual property rights protection of economic and social benefits are increasing, and the whole society's awareness of respect for and protection of intellectual property rights has been significantly enhanced. Obviously, the protection of patent knowledge can be significantly enhanced after classification.

The protection of intellectual property rights under artificial intelligence technology can enhance the degree of cultural industry aggregation, optimize the allocation of resources based on the factors of knowledge production, and then achieve the effect of the final development of the cultural industry.

Most studies show that intellectual property protection in the digital economy can enhance the degree of cultural industry agglomeration. First, the lack of intellectual property protection will lead to a loss of trust within the cultural industry cluster, which will eventually lead to the collapse of the cluster. On the contrary, it can be proved that the stronger the strength of intellectual property protection, the more aggregation of cultural industry clusters occurs. Secondly, when intellectual property protection is weak, it will lead to copying or simply imitating each other without innovation within the same cultural cluster, which will make the original intellectual property rights holders lose their enthusiasm for innovation and take the lead in removing the cultural cluster. At the same time, due to the 'lemon market' effect, original intellectual property rights are excluded from the market, and piracy becomes more popular among consumers, eventually leading to the gradual disappearance of the cultural cluster. On the contrary, it can also be proven that the higher the intensity of IPR protection, the more clustered cultural industries are. Clustering of cultural industries can facilitate the sharing of information and various elements, making communication between them smooth and fast, and is conducive to the development of industrial upgrading in the form of collective upgrading.

Intellectual property protection under artificial intelligence technology can promote the self-production of knowledge production factors in subcultural communities, so as to optimize the resource allocation of knowledge production factors and thus achieve the effect of the development of cultural industries. This kind of “self-production” is regarded as “industrial aggregation” in cyberspace. Under the digital economy, subculture clusters are formed on the Internet due to the involvement of artificial intelligence technology. Since this segmentation is categorized by the same point of view, each cluster shares a common ideological consensus, i.e., a subculture. “Self-production” is considered to be an important factor in the development of the culture industry because in the digital economy, subculture tribes are formed by gathering on the basis of communities, which are active and stable within the tribes. At the same time, the “self-production” of subcultural communities can optimize the allocation of knowledge production factor resources. As subcultural communities can promote more subjects to participate in re-innovation, it is also one of the ways to properly allocate knowledge production factors. In addition, subcultures can interact with each other and enhance the exchange and cooperation between them. When divided into different subcultural clusters, each field is not in contact with the other, so when they cross each other, it is easier to create sparks and obtain cross-border innovation.