Research on Cultural Value Identification and Digital Inheritance Methods of Weinan Non-legacy Skills Based on Deep Learning in Higher Education Environment

Promoting the creative transformation and innovative development of the excellent traditional Chinese culture and improving the soft power of national culture is an important part of promoting the modernization of China’s cultural construction and cultural governance. And promoting Chinese-style modernization construction with vigorous development of high-quality higher education is a key initiative to realize the great rejuvenation of the Chinese nation. Local applied colleges and universities are an important part of higher education, and in the strategic layout of modernization, they shoulder the important mission of cultivating local high-quality and applied talents, and promoting local excellent traditional arts and culture.

Non-legacy arts and crafts are the most dynamic forms of expression in non-legacy culture, which are the traditional way of life and crystallization of the wisdom of local people, reflecting the diversity of regional culture [1-2]. Taking Weinan as an example, the long history of Weinan and the profound Qindong culture have given birth to a variety of folk arts, including Huayin Old Cavity, Huaxian Shadow, Bowl Cavity, Hancheng Yellow River Formation Drums, Agong Cavity, Huayin Mystery Hu, Heyang Wire Puppetry, Tongzhou Bangdang, Hancheng Rice-planting Songs, Heyang Jumping Opera and so on, all of them are precious and excellent local cultures, and they are the mainstay of the spiritual life of the Weinan people [3-7]. Not only are they the support and expression of people’s emotions, but they also play an important social function in marriage and funeral ceremonies and festivals [8-9]. However, in the present time of global economic integration, the survival and development of non-heritage music and opera arts in a multicultural context are facing severe tests [10-13]. As most of the non-heritage art has oral, living characteristics, mainly rely on oral teaching to inheritance, with the death of the old artists, took away a masterpiece, which led to a lack of non-heritage art inheritance successor, resource loss phenomenon is frequent, non-heritage art into the “endangered” predicament [14-16]. Therefore, vigorously rescue, protect and inherit the non-heritage art, improve its living space, cultivate non-genetic inheritance of talents, innovation of non-heritage art content form is an imminent historical task [17-19]. In the new journey of promoting the construction of Chinese-style modernized culture, local applied colleges and universities should give fuller play to regional resources and advantages in all aspects, combine digital technology, inherit local non-heritage cultural skills, continuously enhance historical initiative, and consciously integrate into the historical process of Chinese-style modernized culture construction [20-23].

This study focuses on the research of cultural value identification and digital inheritance of Weinan non-heritage skills based on deep learning in the university environment, and designs a deep learning model based on DL-CBAM. And the index mapping of non-heritage skills was constructed to visualize the digital resource themes of non-heritage skills, and a digital inheritance method from the perspective of digital resource protection of non-heritage skills was proposed. In addition, the database is crucial for recognizing cultural values and studying the effects of digital inheritance. Therefore, this paper utilizes the attribute map model to generate a dataset of Weinan non-heritage skills for subsequent validation experiments.

2

Research on identification and transmission of cultural values based on deep learning in higher education institutions

2.1

Deep Learning Algorithms for Recognizing Cultural Mechanisms of Non-Heritage Techniques

2.1.1

Convolutional Neural Networks

Convolutional neural network [24] belongs to local access multilayer neural network, the layer contains multiple neuron units, the network includes two parts, feature extraction and feature mapping, and four basic operations, convolution, activation, pooling, and full connectivity. 1)

Convolutional layer

Convolutional layer, as the core key part of the convolutional neural network, mainly through the weights in the convolutional layer the operation is a linear operation on two real variable functions. Its discrete definition is as follows: (1) $(f * g) (n) = \sum_{τ = - \infty}^{\infty} f (τ) g (n - τ)$

Its main role is to accomplish the extraction of the basic features of the data through a convolutional kernel with learning capability, which combines multiple basic features together to form the overall image characteristics. The output side of the convolutional layer is calculated as follows: (2) $x_{j}^{l} = f (\sum_{i, j} w_{i j}^{l} \otimes x_{i}^{l - 1} + b_{j}^{l})$ 2)

Sigmoid activation function

The Sigmoid function [25] has an output characteristic mapping in the range (0, 1) and can be represented as a probability or used for input normalization. The mathematical expression for the Sigmoid function and its derivatives is given below: (3) $σ (z) = \frac{1}{1 + e^{- z}}$ (4) $σ' (z) = (1 - σ (z)) \cdot σ (z)$

3)

Pooling layer

The pooling layer is used to adjust the output of this layer by using different forms of nonlinear pooling functions. For a color pixel image, where the pixel values are set to matrix A, the pooling operation is performed on matrix A, which first needs to be chunked. If the chunks do not overlap and the size of each chunk is λ × τ, the formula for the ijth chunk is as follows: (5) $G_{λ, τ}^{A} (i, j) = {(a_{s t})}_{λ \times τ}$

where (i − 1) · λ + 1 ≤ s ≤ i · λ, (j − 1) · τ + 1 ≤ t ≤ j · τ.

Maximum downsampling is the use of a convolution kernel to take the maximum value in a selected image region as the output value after sampling that selection, and the maximum downsampling for $G_{i, i}^{A} (i, j)$ is defined as follows: (6) $\begin{array}{rcl} \max d o w n (G_{λ, τ}^{A} (i, j)) & = & \max {a_{s t}, (i - 1) \cdot λ + 1 \leq s \\ \leq i \cdot λ, (j - 1) \cdot τ + 1 \leq t \leq j \cdot τ} \end{array}$

Average downsampling is the process of first selecting a selected region and then calculating the average value of the selected image region as the output value after sampling that selection, the average downsampling for $G_{λ, τ}^{A} (i, j)$ is defined as follows: (7) $a v g d o w n (G_{λ, τ}^{A} (i, j)) = \frac{1}{λ \times τ} \sum_{s = (i - 1) \times λ + 1}^{i \times λ} \sum_{t = (j - 1) \times τ + 1}^{i \times τ} a_{s t}$ 4)

Fully Connected Layer

The fully connected layer, each layer in the fully connected layer is a multiple neuron tiling arrangement structure. It is mainly used to integrate the useful information with features in the convolutional layer or pooling layer, and then classify and summarize the information integrated from the neural network. Such a structure is generally built at the end in the CNN model, which is used to make a weighted sum of the feature information extracted from the previous design, i.e., on this basis, the learned distributional characteristics are mapped using the distributional form of one-dimensional vectors, and the obtained distributional characteristics are mapped into the said template space for classification, and the probability value of each category is output.

2.1.2

ResNet Recognition Network

The basic idea of Residual Network [26] (ResNet) is “constant mapping connection”, i.e., the output of each layer is the same, which has the advantage of reducing data redundancy, easier feature learning, and easier optimization of the residual mapping than the original mapping.

In the residual block, it is assumed that x is the input to the model and H(x) is the desired output. Because of the presence of Identity Mapping, H(x) = F(x) + x, F(x) is the real learning part of the model, where F(x) = H(x) − x.

The residual block in some cases it enhances the feature extraction in the network model with no side effects.

2.1.3

CBAM Attention Mechanism

The CBAM attention mechanism [27] pays attention to both spatial and channel dimensions.The CBAM attention mechanism model is shown in Fig. 1, with the channel attention mechanism on the left and the spatial attention mechanism on the right.

If the feature input at the beginning is F ∈ R^H×W×C and the number of network channels is C, at the very beginning the input features are passed through the maximum pooling layer and the global average pooling layer in the channel dimension to get the output features $F_{a v g}^{c}$ and $F_{\max}^{c}$ , based on which the features $F_{a v g}^{c}$ and $F_{\max}^{c}$ are downscaled, ReLU activated, and upscaled to get two new feature graphs. These two graphs are activated with Sigmoid function after summing operation, the steps of which are shown in equation (8): (8) $\begin{array}{rcl} M_{c} (F) & = & σ (M L P (A v g P o o l (F)) + M L P (M a x P o o l (F))) \\ = & σ (W_{1} (W_{0} (F_{a v g}^{c})) + W_{1} (W_{0} (F_{\max^{c}}))) \end{array}$

Where, $W_{0} \in R^{\frac{c}{r} \times c}$ , represents the process of dimensionality reduction. $W_{1} \in R^{c \times \frac{c}{r}}$ , represents the process of dimensionality raising, and r represents the multiplicity of dimensionality lowering or raising. σ represents the Sigmoid activation operation. After obtaining the weight information M_c(F), then multiply it by the input feature F to get the output feature F1 of the channel dimension.

After being processed by the spatial attention module, F1 undergoes two pooling operations and is merged into $M_{c} (F 1)$ . Then, a convolution operation is performed and the weighting coefficients are computed using the Sigmoid function, which is used to generate the output of the spatial attention module. Finally the weighting coefficients 2 are multiplied with the features F1 provided to this module to obtain a new feature map F2. The specific process is expressed in equation (9): (9) $\begin{array}{rcl} M_{c} (F 1) & = & σ (f^{7 \times 7} ([A v g P o o l (F 1); M a x P o o l (F 1)])) \\ = & σ (f^{7 \times 7} ([F_{a v g}^{s}; F_{\max^{s}}])) \end{array}$

Here, f^7×7 denotes a two-dimensional convolution operation with a convolution kernel size of 7 × 7, and σ denotes a Sigmoid activation function.

2.1.4

DL-CBAM-based model for recognizing the cultural value of performing arts and crafts

Aiming at the problem of the accuracy of image recognition of NRT, this paper proposes an image recognition network based on Deep Convolutional Block Attention Module (DL-CBAM). Adopting ResNet50 as the backbone network for the recognition of NRHs, introducing the high-dimensional attention mechanism module, combining with the characteristics and demands of the task of recognizing NRHs, focusing on the extraction of high-dimensional features, and fusing the residual module with CBAM in the deep convolution layer of the residual network, the overall structural flow of the proposed NRH recognition network based on the cultural value of NRHs based on the DL-CBAM module is shown in Fig. 2.

2.2

Research on Inheritance Methods Based on Digital Resource Conservation of NRH Techniques

The data resources of non-heritage skills include the digital resources generated in the whole process from the generation of non-heritage skills to the final formation of skills products, which not only involves the deep digital resources such as skills non-heritage bearers and skills non-heritage artifacts, but also involves the superficial digital resources such as the environment of non-heritage bearing and inheritance carriers of skills. Therefore, it is of great significance for cultural heritage to explore the theme of NRH digital resources.

2.2.1

Overall idea

Firstly, we study the theme mining scheme of digital resources of non-heritage skills from the theoretical level, synthesize the knowledge graph theory and the characteristics of non-heritage skills, explore the theme discovery method that integrates the two characteristics, and thus analyze the theme characteristics of digital resources of non-heritage skills. On this basis, the theme discovery scheme of non-heritage skills is explored from the empirical level, firstly, the non-heritage digital resources of traditional skills are mined from multi-dimensions, and then the themes of non-heritage digital resources are extracted and fused, and the analyzed themes of the non-heritage digital resources are visualized, and finally, a complete theme atlas containing the non-heritage digital resources of non-heritage skills will be obtained.

In addition, the theoretical-level subject discovery method of NRM digital resources guides the empirical-level subject discovery work. On the contrary, the correctness of the subject features of the non-heritage digital resources at the theoretical level is verified through the comparison of the non-heritage digital resources subjects at the empirical level, the theoretical model of non-heritage digital resources subject discovery based on knowledge graph is revised according to the problems encountered in the process of empirical validation, and then the subject mapping is regenerated based on the revised model until the empirical results and the theoretical model have been concluded in the analytical process.

2.2.2

Transmission methods

In this study, the digital resources of non-heritage skills are structured and expressed using the knowledge mapping method, which is used for theme mining. 1)

Grammatical structure analysis method. The digital resources of non-heritage skills include both the traditional skills data repositories that have been regularized and processed as well as the related digital resources disseminated on the Internet, whose structure is a mixture of ordered and disordered, with varying quality. The semantic structure analysis method is used to structure the digital resources, and by formulating appropriate rules for data resources, different types of non-heritage skills data resources such as text, pictures and videos are adjusted dimensionally, and statements are structured and classified according to their types, finally forming a structured non-heritage skills digital resource base.

2)

Semantic relationship analysis. Semantic relationship analysis firstly needs to be done through dependent syntactic analysis and semantic role labeling, and then RDF triad modeling is completed in order to facilitate the construction of the semantic network of the digital resources of non-heritage skills.

3)

Semantic value analysis method. The topic discovery model of digital resources of non-legacy skills is constructed using extracted semantic relations to promote the realization of the semantic value of digital resources. The knowledge graph of NRT digital resources is constructed through dependent syntactic analysis, semantic role annotation and RDF triad extraction, and the theme discovery of NRT digital resources based on semantic theme graph is accomplished under the guidance of the formulated theme discovery rules. The collected multivariate heterogeneous non-heritage skills digital resources are analyzed and processed, from which the multi-level semantic information of different themes of digital resources is extracted, and the themes of digital resources are extracted on the basis of semantic relations.

2.2.3

Program implementation

The implementation scheme of non-heritage skills digital resource discovery based on knowledge graph is carried out under the guidance of the theoretical framework of theme discovery by comprehensively applying data collection and analysis tools and knowledge graph analysis and visualization tools, and the implementation path of the theme mining specific scheme is shown in Figure 3.

According to the constructed theme mining program implementation path of non-heritage skills digital resources, firstly, it is necessary to collect existing multi-source non-heritage resources with the help of web crawler tools to get non-heritage digital resources. Then the digital resources are filtered and cleaned with the help of software and manual power, so as to get the initial digital resource dataset. Secondly, the text data in the initial digital resource dataset is processed through pre-processing such as de-duplication, and the pre-processed text set is integrated to form a non-heritage corpus. Once again, the dependency set of NRL digital resources is obtained through semantic dependency analysis. After that, the entity-entity relationships in the dependency set are derived by the DL-CBAM naming recognition algorithm. And the semantic relations of non-legacy digital resources consisting of RDF triples of digital resources and their weights are derived through topic extraction. Finally, the thematic mapping of NRM is formed by the Gephi visualization and presentation tool.

2.2.4

Ontology Construction of Knowledge Graph for Non-Heritage Techniques

The creation of knowledge graph mainly includes the following links, firstly, obtaining original data, taking diverse data as data source, from which we realize the refining of data such as entities and attributes and relationships, and deposit them into the schema layer and data layer of our knowledge base. Secondly, knowledge fusion involves integrating and disambiguating new knowledge or data to construct the knowledge graph.

In the construction of knowledge graph nodes, we assume that we exist N node, and the average degree of each node is M. We construct the association between intangible cultural heritage based on the following process: 1. Construct a regular ring dot matrix containing N nodes. 2. This regular ring dot matrix satisfies the following characteristics, each node is connected to M nodes, where there are M/2 on one side if and only if $0 < | i - j | \mod (N - 1 - \frac{M}{2}) \leq \frac{M}{2}$ $0 < \left| {i - j} \right|\,\bmod \,\left( {N - 1 - {M \over 2}} \right) \le {M \over 2}$ , and there exists an edge $(n_{i}, n_{j})$ . The clustering coefficient of this graph satisfies the following equation: (10) $C^{'} (β) = \frac{3 * n u m b e r o f t r i a n g l e s}{n u m e b r o f c o n n e c t e d t r i p l e s}$ $$C'\left( \beta \right) = {{3*number\,of\,triangles} \over {numebr\,of\,connected\,triples}}$$

2.3

Realization of graphical database storage for Weinan’s non-heritage skills

2.3.1

Database design

In this paper, the attribute graph model is used as the data structure of Weinan NRT. When using the graph data model to characterize the Weinan NRT data, the entities, attributes, and relationships in the knowledge model are represented by the structure of the graph data model, and the fields in the Weinan NRT data model graph are all of the Neo4j default storage type, i.e., character type, and the Weinan NRT graph data model is shown in Figure 4.

The Weinan NRT data model is represented by the attribute graph model, which defines a total of 7 types of nodes, with the attributes of the nodes inside the rounded rectangle, representing the attributes defined on the nodes in the graph model, and the names of the nodes inside the white box in the upper right corner of the rectangle, and the 7 types of nodes associated with each other through the edges, forming the Weinan NRT data model.

2.3.2

Generation of Weinan NRL Techniques Dataset

When using the Neo4j graph database to store the data of Weinan non-legacy skills, the data should be organized according to the attribute graph data model before storing the data into the Neo4j graph database. According to the designed data model of Weinan NRL techniques, these data are organized into a more structured dataset.

The data in the dataset mainly came from the “Project Profile” field on the China Nonheritage Network and the Weinan Nonheritage Network, but some of the data on the techniques were not comprehensive enough, so in order to make the Weinan Nonheritage techniques dataset more comprehensive, it was partially supplemented by other encyclopedic websites. These organized data were then used to form the Weinan Intangible Cultural Heritage Skills Dataset. The Weinan Nongbei Techniques dataset contains 1798 pieces of data and is 661kB in size.

2.3.3

Database map data importation

After organizing Weinan’s NRM knowledge into a data structure that conforms to the graph data model, it is necessary to import the dataset into the Neo4j software. Download the installation package from the official website that is compatible with your computer environment. Register for a personalized Neo4j account. Fill in your name, email address, nationality, and other required information to complete the registration process and download the installation package. Follow the system instructions to complete the installation, select a good folder for future storage of map data, configure the environment variables, and complete all setup operations. After installing the Neo4j software, it was necessary to import the Weinan NRT dataset into the graph database to complete the storage of the Weinan NRT. Through the non-heritage skills knowledge meta, the connotation and extension of skills knowledge are accurately expressed through the medium of images.

3

Identification of cultural value of skills and analysis and evaluation of inheritance effect under deep learning

The hardware configuration given for the experimental environment of this paper is: Intel core(TM) i7-3770 CPU @3.4GHz processor with 16G of RAM. The software environment based on this paper is: under Windows 10 64-bit operating system in a development environment like Anaconda. Utilizing languages like Python with the help of several toolkits like Librosa, Numpy, and others.

3.1

Performance of DL-CBAM Deep Learning Model for Cultural Value Recognition

The results of the cultural value recognition experiments in this study were evaluated using classical metrics. The classical evaluation metrics include accuracy rate, recall rate and the reconciled mean of the two, F1, which is calculated as follows: (11) $\begin{array}{l} p r e c i s i o n = \frac{T P}{T P + F P} \\ r e c a l l = \frac{T P}{T P + F N} \\ F 1 = \frac{2 p r e c i s i o n \times r e c a l l}{p r e c i s i o n + r e c a l l} \end{array}$

TP denotes entities that are recognized and matched with labeled entities, FP denotes entities that are recognized but not matched with labeled entities, and FN denotes labeled entities that are not recognized.

The CNN model was chosen as the benchmark method to test the recognition performance of the Weinan NRT cultural mechanism with the DL-CBAM model in the graph database of Weinan NRT constructed above under different numbers of images (10, 30, 50, 80, 100, 150, 200) for comparison with the DL-CBAM model.

The results of the recognition performance of the two models for the non-heritage skills are shown in Figure 5. As can be seen from the figure, the methods used in this paper, the measured values on the evaluation indexes are higher than the benchmark methods. Although the recognition accuracy of this model decreases with the increase of the number of images, it still has a high recognition accuracy. When the number of images is 200, the recognition accuracy also remains above 90%, and the recall of the model as well as the F1 value also show high measured values, which are 93.97% and 92.66%, respectively. It is experimentally verified that the DL-CBAM model has a much better effect in recognizing the cultural value of NRL techniques, which provides strong support for the subsequent digital inheritance.

3.2

Identification of the value of cultural emotional expression of Weinan’s non-heritage skills and techniques

To achieve a better effect on cultural value recognition, this section utilizes the DL-CBAM model to study the recognizing and classifying emotional value of NRM skills based on NRM skill images. The selected NRT image segments contain joy, sadness, peace, longing, and irony, totaling five emotions.

Confusion matrix is a visualization tool, and the experiments in this section use the confusion matrix to demonstrate the classification accuracy and classification error of NRT emotional expressions. In all the confusion matrices in this paper, the emotion labels on the horizontal axis represent the real emotions of NRT, and the emotion labels on the vertical axis represent the predicted classifications obtained from testing NRT, where the numerical values in the matrices represent the probability of judging the predicted classifications of the NRT image fragments into the corresponding vertical axis categories. And the larger the probability, the darker the color, and conversely, the smaller the probability, the lighter the color.

Fig. 6 shows the confusion matrix of using the DL-CBAM model to realize the classification of the recognition of the emotional value of the non-heritage arts and crafts. From the figure, it can be seen that the accuracy of this paper’s model for the classification of the emotional expression of NRM skills is high, and through the analysis of the accuracy of the recognition of the five emotions, it is found that among all the emotions, the classification of the NRM skills for the cheerful emotion is the best, with an accuracy of 89%, and the classification of the sad emotion is the worst, with an accuracy of 80% or more as well. It illustrates the superiority of this paper in utilizing the DL-CBAM model to extract the features of NRM techniques to achieve classification.

3.3

Evaluation of Theme Mining Based on NRM Knowledge Metadata

On the basis of the identification of cultural values of NRM skills, this section extracts the knowledge items of NRM skills-related knowledge meta-knowledge. The extracted NRM knowledge meta-knowledge is categorized into seven categories, such as characters, objects, concepts, events, actions, spaces, and documents, etc., and the same indicators as above are used to evaluate the performance of NRM knowledge source knowledge item extraction.

Figure 7 shows the results of the measured evaluation metrics for non-legacy knowledge meta-extraction. As shown in the figure, the performance of the extraction system is significantly improved by the deep convolutional block attention module-based knowledge meta-knowledge item extraction method adopted in this paper due to the incorporation of multiple constraints of non-legacy terminology, feature words and semantic relationships, etc. Compared with the traditional CNN model-based extraction method, the accuracy of the knowledge meta-knowledge extraction method proposed in this paper is improved by 7.71%~18.16%, and the recall rate is improved by 4.67%~10.35%, and the F1 value is improved by 3.71%~12.03%.

Using the DL-CBAM model for faceted topic analysis of the extracted knowledge items of knowledge elements, set a=40/K, β=0.02 according to the empirical values, and calculate its optimal topic structure for the knowledge elements according to the following formula and on the basis of this, carry out the topic-based importance calculation of the knowledge items: (12) $D - E I D F - L E N (a) = D \times L {og}_{2} (1 + \frac{P (a)}{P {(a)}^{'} + 0.001}) \times L o g_{2} L E N$

where D is the document frequency of term a, P(a) denotes the word frequency of candidate term a in the intangible culture domain corpus, P(a)′ denotes the word frequency of candidate term a in the generalized corpus, and LEN is the term word length, with 0.001 in the denominator in order to avoid a zero denominator.

The importance threshold is set to the peak importance of the knowledge meta-knowledge items × 0.5. By taking the manually extracted knowledge meta-knowledge items by experts as the benchmark standard, defining the filtering efficiency = the number of noisy knowledge items filtered out / the number of knowledge items that should be filtered, and the error rate = the number of correct knowledge items filtered out / the number of filtered knowledge items, the subject matter filtering performance of the model in this paper for nonfolklore is shown in Fig. 8.

From the data depicted in the figure, it can be known that the knowledge meta-knowledge item topic filtering algorithm of the model proposed in this paper embodies good filtering performance on different categories of knowledge meta-knowledge item sets of non-legacy, with an average filtering efficiency of 91.85% for redundant and noisy data, and an average error rate of only 3.02%. Experimentally, it is proved that the model proposed in this paper not only has good filtering performance of noisy data, but also can better recognize the knowledge content of different topic dimensions, which can be used as an effective method for topic mining of digital resources of non-legacy arts and crafts. The average optimal number of topics refers to the number of topics that adequately represent the distribution of topics in the collection of NRT documents. By determining the average optimal number of topics, the paper avoids the limitation of setting the number of topics artificially and obtains a better classification effect of the model.

3.4

Effectiveness of digital inheritance of skill culture based on deep learning

In this section, the study of inheritance effects is divided into three levels: 1)

The cognitive level, which is based on external information acting on people’s perceptions and memory systems to study people’s cognition of Weinan’s non-heritage arts and culture.

2)

The attitudinal level, which mainly refers to the changes in emotions caused by people’s perceptions, analyzes respondents’ attitudes towards the dissemination of Weinan’s non-heritage arts and crafts to find out the effects of their digital inheritance.

3)

The behavioral level, which is mainly expressed through people’s words and actions, analyzes whether the respondents are willing to participate in the digitalization process of disseminating and passing on Weinan’s non-heritage arts and crafts.

The study randomly selected 100 students from the campus and used a 5-level scale (ratings of 1 to 5, representing incremental levels of satisfaction) to investigate the effect of the transmission of Weinan non-heritage arts and crafts culture in the university.

Figure 9 shows the results of the students’ survey at the cognitive, attitudinal, and behavioral levels. According to the data in the figure, the mean value of the students’ understanding of the Weinan non-heritage arts and culture is about 4.29, indicating that the familiarity is on the familiar side, but has not yet reached the state of complete understanding. The data analysis of the students’ satisfaction with the dissemination status of Weinan’s non-heritage arts and culture shows that the mean value is about 4.70, indicating a relatively high level of satisfaction, i.e., the model of this paper has played a positive and effective role in the excavation of Weinan’s non-heritage digital resource themes, and has performed well for the inheritance of the non-heritage arts and culture. The behavioral level of analysis centered on three aspects of the research discussion: the extent to which students recommended Weinan’s non-heritage skills and culture, the extent to which they participated in the inheritance, and the extent to which they were influenced by the effects of the dissemination. The average score for this item was approximately 4.00, with students increasing their interest in the Weinan NRL culture through the model and willing to recommend it to each other to promote cultural inheritance.

4

Conclusion

In this paper, the DL-CBAM deep learning model is constructed as the main algorithm for cultural value recognition and digital inheritance of Weinan’s non-heritage skills in the university environment. Based on the knowledge map of non-heritage arts and crafts, the cultural digital inheritance program is designed to specifically analyze the model’s cultural value recognition performance and the digital inheritance effect. This paper mainly proves the research argument through the following conclusions: 1)

The accuracy, recall and F1 value of the DL-CBAM model for cultural value recognition of non-heritage skills are all higher than 90%.

2)

The model in this paper has the highest precision rate of 89% for the value of “cheerful” emotions of NRM culture.

3)

The model is more effective in extracting non-heritage knowledge elements, which is significantly better than the benchmark model.

4)

Through the model of this paper, students’ familiarity with Weinan’s non-legacy arts and culture is 4.29, which improves students’ satisfaction and interest in Weinan’s non-legacy arts and culture, and promotes cultural digital inheritance.

The current application of deep learning technology in the field of non-heritage is still in the preliminary stage, and many applications are only used as a recording tool for the digital dissemination of non-heritage culture. In the future, deep learning technology can be used to realize the three-dimensional and comprehensive storage of knowledge related to non-heritage skills and culture that will be lost, and to realize the inheritance of non-heritage skills and culture.

Lingua:: Inglese

Frequenza di pubblicazione:: 1 volte all'anno
Argomenti della rivista:: Scienze biologiche, Scienze della vita, altro, Matematica, Matematica applicata, Matematica generale, Fisica, Fisica, altro

Feed RSS della rivista

Research on Cultural Value Identification and Digital Inheritance Methods of Weinan Non-legacy Skills Based on Deep Learning in Higher Education Environment

Ruiyao Ren

Ruiyan Ren

Pubblicato online: 24 mar 2025

Ricevuto: 13 nov 2024

Accettato: 27 feb 2025

DOI: https://doi.org/10.2478/amns-2025-0708

Parole chiaveConvolutional Network, ResNet Recognition, CBAM Attention Mechanism, Cultural Value Recognition, Digital Inheritance

© 2025 Ruiyao Ren et al., published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Parole chiave
Convolutional Network, ResNet Recognition, CBAM Attention Mechanism, Cultural Value Recognition, Digital Inheritance