3.1 Action Recognition Classification Algorithm Combining Weight Offset and Dimension Reduction Retrieval

Aiming at the problem that the standard CNN cannot understand dynamic information in the action recognition classification task, a weight-offset module is designed. This module assigns a higher weight parameter to the convolutional layer and uses this particular convolutional layer as other convolutional layers. The weight offset was carried out based on information difference because the neural network is different from the dynamic information of the original data after learning. The network ability was optimized to understand the time-domain dynamic information and improve the classification accuracy of the model ^[16]. Fig. 1 shows the framework of the weight offset module.

In Fig. 1, the convolutional layer first sorts the learning potential of each dynamic input data, then calculates similarity and assigns more learning potential to dynamic data. The gradient of the weight parameters of each convolutional layer is improved, and the direction of optimization is changed. The update of weight parameters was modified by adding time domain constraints to the loss function.

This procedure was repeated by optimizing and updating the dynamic data with the lowest weight into the convolutional layer. In contrast, the dynamic data with the highest weight is used as the output. Fig. 2 presents the evolution of the data dynamic information throughout the training procedure.

In Fig. 2, the squares represent the dynamic data of the convolutional neural network input layer and several intermediate layer data. A closer color of the squares indicated a higher similarity of dynamic data between layers. The convolutional layer closest to the dynamic data and input layer was the dominant layer. The module achieved the effect of optimizing training by reducing the difference in time-domain dynamic information between the convolutional layer and the input layer as a whole and improving the network model ability to understand time-domain information. The dynamic information set of the convolutional layer data was ${\delta}$. The dynamic information set of the original data $T^{c}$ of the input layer was $T$. The numerical difference was $d$. The calculation method is expressed as formula (1).

$\delta $ is the medicalization of the dynamic information set. The dynamic information set of each data layer was calculated by the difference in frame numbers to obtain more comprehensive dynamic information and prevent valid information from being excluded, as shown in Eqs. (2) and (3).

Through the dynamic information extraction of Eqs. (2) and (3), the overfitting of the network can be avoided owing to the high local attention of the weight-offset module during the optimization process. The dynamic information of the feature data of the convolutional layer can be calculated according to Eqs. (4) and (5).

$u^{c}$ is the feature data. $n$is the total number of frames in the $i$$^{\mathrm{th}}$ data group. $t$ is any frame in the data group. The calculation process of the original data action information set of the input layer is the same as the above method, so it is omitted. For the different number of data channels in the convolutional layer, this study samples the feature data randomly with a large number of channels in proportion to ensure that the distribution of each data group remains consistent. When updating the weight parameters, the weight offset module needs to quantify the distance between $T^{c}$ and $T$, and takes the quantized result as the time domain constraint of the neural network loss function. The combination of time domain constraints and loss functions can jointly affect the gradient of the network. The distance between the sum $T^{c}$and the sum $T$ is minimized while ensuring the accuracy of the optimized model. The time domain difference between the data of the weight offset module and the original data of the input layer can be reduced. Eqs. (6) and (7) express the loss function and the update operations of weight parameters after adding time domain constraints, respectively.

$L_{p}\left(l,\hat{l}\right)$ is the loss function composed of the real value of the input sample and the output value of the model. $\lambda $is the constraint weight. $D\left(T,T^{c}\right)$ is the time domain constraint. $W$is the weight parameter of the convolutional layer. Because the original and feature data are not in the same data space, the common difference calculation method could not reproduce this cross-spatiality, resulting in inaccurate calculation results. Transfer learning was adopted for multi-domain adaptation, and the Maximum Mean Discrepancy (MMD) algorithm was used to quantify the distance of cross-spatial dynamic information accurately. The MMD algorithm maps the two datasets using a continuous function set in a certain sample space and judges whether the stones belong to the same space according to the difference in the mapping results. The selection of $F$ is the key to determining the accuracy of the calculation result. The definition of the MMD algorithm is shown in Eqs. (8) and (9).

$p$ and $q$ are the Borel probability distributions. $X$ and $Y$ are the sum of the data set $f$ obtained by the independent and identical distribution of the Borel probability distribution, respectively. $m$ is a function $n$ of the function set $F$. The weight-offset module uses the MMD algorithm of the regenerated kernel Hilbert space to map the two sets of action information sets in different spaces into the Hilbert space with the kernel function. The distance between the two data groups is equivalent to the quantified differences between two groups in action information across spatial-temporal domains. The completeness and regularity of the Hilbert space can ensure the stability of the calculation results. The deviation was not too large in the case of a large amount of sample data. The mapping process of the regenerated kernel Hilbert space is expressed as Eqs. (10) and (11).

$\varphi $ is the data that is mapped with a Gaussian kernel. $H$ is the Hilbert space. $x'$is the transpose of the input data. $\sigma $ is the sample standard deviation. The dynamic data of the original data of the input layer and the dynamic data of the weight-offset layer are used as the input cross-spatial data. The maximum mean difference between the two sets of time-domain dynamic information is used as the constraint, as shown in Eq. (12).

$T$ is the dynamic information set of the original data of the input layer. $T^{c}$ is the dynamic information set of the feature data of the weight-offset layer. The key frame of the $M$ image set is$I$, ${I_{{_{i}}}}_{i=1}^{M}$ is set as the input of the model. The number of image blocks is expressed as (13).

$n$ is the size of the image. $k$is the size of the convolution kernel. Each image, according to the column vector, is sorted. $X_{i}$ is the dimensional feature vector of any image in the image set $n$. The feature matrix for dimensionality reduction processing is extracted to form a new low-latitude feature library. The feature vector $F_{i}$ is extracted, and the similarity is calculated. The feature matrix of each row is converted as a vector. The similarity is judged by comparing the cosine angle between the vectors, which is expressed as formula (14).

$x_{i}$ and $y_{i}$ are two vector elements. $n$ is the number of vector elements. A closer value of the calculation result indicates that the two vectors are closer. Finally, the value range of the calculation result is normalized, and the normalized interval is expressed as Eq. (15).

$S$ is sorted by size and output of the search results. According to the similar metric of the feature vector and feature library of the image to be retrieved, the feature index is performed, the corresponding image is found, and the top-ranked image is displayed according to the decreasing law.

3.2 An Online-offline Blended Teaching Model Integrating Deep Learning for Chinese Painting Education

Blended teaching refers to integrating multiple teaching modes and teaching methods, which are usually assisted by digital information technology to improve teachers’ teaching effects and students’ learning efficiency ^[17]. To a certain extent, Chinese painting education is more about cultivating students’ emotional identification with China’s traditional culture rather than painting skills and aesthetic cognition ^[18]. Fig. 3 shows the online-offline blended teaching mode of Chinese painting in colleges and universities.

The blended teaching model is divided into three parts (Fig. 3). The first part analyzes teaching needs, the second part is online learning, and the third part is an online classroom. According to the Chinese painting learning environment and resources in colleges and universities, students' psychological characteristics, learning needs, and content must be compared and analyzed. Students' emotional needs in learning Chinese painting are the key part. Through online learning methods, including micro-video learning and collaborative learning online discussion, students can have emotional expression and accumulation. Teachers can leave questions and role-playing activities after online courses so that students can summarize and reflect on themselves and maintain emotional continuity. Finally, the results exchange, debate, and discussion are conducted in offline classes. The primary purpose is to facilitate the generation and expression of new emotions and evaluate the learning content. This part runs through the whole learning process instead of being placed only in the third part. Learning emotional Chinese painting courses requires students to master the basic knowledge of Chinese painting. It cultivates students’ understanding of the traditional culture contained in Chinese painting and enhances their emotional identification with Chinese traditional culture ^[19]. This model mainly uses qualitative research and quantitative research methods to conduct a fine-grained analysis of students’ reflective experiences by evaluating the students’ reflection after class. Fig. 4 shows the flow of the blended learning mode integrating deep learning.

The teaching process is not a one-way process but a cyclic process, as shown in Fig. 4. Through reflection and summary, teachers can analyze students’ emotions reflected by students’ movements and expressions during class through the CNN model and adjust teaching methods and content in time. In this process, the function of the weight offset module in CNN is to pre-set teaching objectives, in which CNN focuses on monitoring the actions and expressions related to the target and assigns higher weights to related images and videos, which is more accurate. Finally, whether students meet the established learning goals in the classroom is determined.

4.1 Practical Application Analysis of Online-offline Blended Learning Mode Integrating Deep Learning

An indirect assessment is more important because the sentiment toward traditional culture occurs at a higher level of sentiment classification (Huang D et al. 2020) ^[20]. Learning indicators, student behavior, personal reflection experience, and theoretical achievements are collected through observation. Content analysis method is used for in-depth analysis. In a university, students in Chinese painting elective courses who volunteered to participate in the experiment are selected as research subjects. The experimental group adopts an emotional online-offline blended learning model integrating deep learning, while the control group adopts the traditional classroom-teaching model. Among them, there are 41 people in the experimental group, including 29 males and 12 females. There were 47 people in the control group, including 21 males and 26 females. The indicators examined included theoretical knowledge, emotional development, and learning attitude. Among them, theoretical knowledge included classroom questions, homework, half-term exams, and final exams. Emotional development involves cultural development, moral development, and aesthetic development. The learning attitudes include attendance, class concentration, and online or offline self-learning content completion. Fig. 7 shows the scores of the two groups of members.

The average score differences between the control and experimental groups on theoretical knowledge and learning attitude are insignificant (Fig. 7), with ${-}$0.19 and 0.46 respectively. On the other hand, the differences in standard deviation are ${-}$1.10 and -2.78, respectively. It shows that the control group members deviate more from the mean value. Table 1 lists the test of the three indicators in the two groups.

The theoretical knowledge score was $F=4.03$ $\left(P=0.047<0.05\right)$ (Table 1), reaching the significance level of 0.05. Under the opposite hypothesis, $H1\colon \sigma _{\times 1}^{2}\neq \sigma _{\times 2}^{2}$, the statistical value $t$ was 3.21. The value of the probability of significance is$p=0.001<0.01$. The difference was quite significant. Regarding theoretical knowledge, the students in the control group have significant individual differences in their mastery of theoretical knowledge. The trend in the learning attitude index is the same as that of the theoretical knowledge index, i.e., the average score was similar. The sum of the standard deviation $F$ is larger, indicating that the individual differences in the learning attitude of the students in the control group are also obvious. In the sentiment index, the mean of the control group was 1.12; the standard deviation was 1.33; the mean absolute error was 0.201. The mean of the experimental group was 2.76; the standard deviation was 2.38; the mean absolute error was 0.383. The average score of the experimental group was higher than that of the control group, but the individual differences were also greater than that of the control group. The subordinate indicators were evaluated for further explanation, as shown in Table 2.

As shown in Table 2, the average number of words in the test group was 1312.72, and that of the control group was 1016.43, a 29.15% increase. The appreciation index of the test group was 5.73, and that of the control group was 5.11, a 12.13% increase. The moral development index of the test group was 6.29, and that of the control group was 5.62, an 11.92% increase. The social development index of the test group was 4.14, and the social development index of the control group was 1.97, a 113.19% increase. The inspirational development index of the test group was 4.28, and that of the control group was 3.67, an increase of 16.62%. Overall, the two groups have little difference in appreciation and moral development, but there are large differences in social and inspirational development.