Research on Pattern Recognition in the Integration of Traditional Cultural Elements with Modern Design Concepts

Semiotics is the study of symbols, symbols exist in any thing in our life, in the words of the master of semiotics, Zhao Yihang, “symbols are carrying perception”, semiotics is the study of the doctrine of meaningful activities.

To study semiotics, we should study the dichotomy proposed by Saussure, which is the universal semiotics of energy and reference. Able to refer to the existence of objective things or objects, refers to the things or objects behind the meaning of the reference, for example, red as a color is the objective existence of this is the able to refer to, and the color red represented by the passionate and intense emotion is his extension of the hidden meaning, which is the reference. Together, they constitute the theory of semiotics.

The dual-axis relationship was proposed by Saussure after he studied the dichotomy of semiotics, and he believed that all symbolic representational activities are carried out between the dual axes of combination and aggregation, with combination referring to the way symbols are composed, and aggregation referring to a kind of “associative relationship”, which is the combination of various possibilities. When the dual-axis relationship was introduced into China, it was translated as horizontal combination and vertical aggregation with horizontal and vertical coordinates, which is not rigorous enough in the opinion of Zhao Yiheng, so I will follow the translation of Mr. Zhao’s combination and aggregation axes here. As long as we make good use of the interaction between symbols and the combination of symbols, we can construct many different meanings. As far as design is concerned, there are different elements such as color, appearance, material and structure on the Combination Axis, and the corresponding Convergence Axis is the style, genre, emotion, etc., which may be derived from the arrangement and combination of these elements on the Combination Axis, and changes in the elements on the Combination Axis will lead to changes on the Convergence Axis. Grasping the relationship between the elements on the combination and aggregation axes and making free combinations can lead to new symbolic meanings.

First of all, it is necessary to clarify the intrinsic meanings of value orientation, social concepts and aesthetic emotions contained in the connotation of traditional cultural elements in order to better understand the significance of traditional cultural elements in modern art design [26]. On this basis, the inspiration and influence of traditional cultural elements on modern art design will be analyzed in depth, and how traditional cultural elements can convey rich cultural connotations in contemporary art works will be explored, as well as further triggering attention to and understanding of traditional culture. In addition, the diverse ways in which traditional cultural elements are presented in modern art and design works, as well as the cultural inheritance and innovation connotations carried by these ways, will also be explored.

In modern society, art design is no longer limited to traditional aesthetic standards, but pays more attention to designers’ personal aesthetic views and creativity in order to create works full of individuality and innovation. With the development of globalization, traditional cultural elements are no longer limited to specific regions and nationalities, and designers, while drawing on traditional cultural elements, also integrate design concepts and aesthetic views of other countries and regions to form new artistic styles and forms of expression [27]. Modern art design pays more attention to the actual value of the works, focusing on the functionality and practicality of the works, so that the art design works can be better adapted to the needs and lifestyle of modern society. With the continuous progress of science and technology, modern art design integrates digital technology, virtual reality, and other modern scientific and technological means, which opens up new expression and form possibilities for art design.

In the relationship between traditional cultural elements and modern art design, traditional cultural elements can provide inspiration and cultural heritage for modern art design. Artists can create new contemporary connotations and aesthetic values for their works by excavating and reinterpreting traditional cultural elements, which can provide more possibilities for modern life.This kind of innovation not only reflects the artist’s respect and understanding of traditional culture, but also enriches the expression and meaning of modern art design.The relationship between traditional cultural elements and modern art design is complementary, and the two complement each other by integrating and promoting the inheritance of traditional culture and the innovative development of modern art design.This relationship not only provides a richer cultural context for artistic creation, but also offers more possibilities for combining cultural traditions and contemporary art.

The objective of an action generation network is to create adapter execution decisions that are relevant to the target task for every layer of the adaptive network [28]. The gating network is integrated into each 3×3 conv layer of the feature extractor f_ϕ(·) that has been pre-trained through the source domain to generate discrete probabilities that each task-relevant adapter is selected. Specifically, the discrete probability of each task-related adapter in layer l can be denoted as {zii}i=0K$\{z_{i}^{i}\}_{i=0}^{K}$. In gated networks, the input features are first subjected to global average pooling (GAP) to compress the spatial dimensions of the features. If hι−1∈ℝNS×C×W×H${{h}_{\iota -1}}\in {{\mathbb{R}}^{{{N}_{S}}\times C\times W\times H}}$ denotes the input features at layer l in the feature extractor. Then, the output u_l−1 of global average pooling can be expressed as: (1) ul−1=GAP(hl−1)\[{{u}_{l-1}}=GAP({{h}_{l-1}})\] where u_l−1 can also be denoted as ul−1=[ ul−11,ul−12,⋯,ul−1Ns ]${{u}_{l-1}}=\left[ u_{l-1}^{1},u_{l-1}^{2},\cdots ,u_{l-1}^{{{N}_{s}}} \right]$ and ul−1i$u_{l-1}^{i}$ denotes the output feature map of support sample i at layer l–1 of the model. Next, the prototype features vl−1=[ vl−11,vl−12,⋯,vl−1N ]${{v}_{l-1}}=\left[ v_{l-1}^{1},v_{l-1}^{2},\cdots ,v_{l-1}^{N} \right]$ of all the classes in the target task can be calculated, where v_l–1ⁿ denotes the class center of class n and N denotes the number of classes in the target task, specifically, the class center of class n can be obtained by calculating the average value of embedded features of all the samples belonging to this class, which can be expressed as follows: (2) vl−1n=1|Sn|∑i∈Snul−1i\[v_{l-1}^{n}=\frac{1}{|{{S}_{n}}|}\sum\limits_{i\in {{S}_{n}}}{u_{l-1}^{i}}\] where S_n denotes the set of support samples belonging to category n. v_l−1 is the set containing all the class features whose statistical information is representative of the entire target task. Next, through the two fully connected layers, the target task features are compressed into an K + 1 - dimensional vector that represents the unnormalized probability scores of the adapters associated with the execution of the K + 1 kinds of tasks, which can be expressed as follows: (3) z¯l=Wl2σ(Wl1vl−1)\[{{\bar{z}}_{l}}=W_{l}^{2}\sigma (W_{l}^{1}{{v}_{l-1}})\] where z¯l∈ℝK+1${{\bar{z}}_{l}}\in {{\mathbb{R}}^{K+1}}$, σ denote the ReLU activation functions. Wl1∈ℝd×C,Wl2∈ℝ(K+1)×d$W_{l}^{1}\in {{\mathbb{R}}^{d\times C}},W_{l}^{2}\in {{\mathbb{R}}^{(K+1)\times d}}$, d are the dimensions of the hidden layer and z¯i${{\bar{z}}_{i}}$ can be denoted as z¯l=[ z→l0,z¯l1,⋯,z¯lK ]${{\bar{z}}_{l}}=\left[ \vec{z}_{l}^{0},\bar{z}_{l}^{1},\cdots ,\bar{z}_{l}^{K} \right]$ . Finally, the gated network normalizes the output by the Softmax function and outputs the probability distribution z_l, which can be expressed as the probability of all task-related adapters being executed in layer l: (4) zl=softmax(z¯l)\[{{z}_{l}}=softmax({{\bar{z}}_{l}})\] where zl=[ zl0,zl1,⋯,zlK ]${{z}_{l}}=\left[ z_{l}^{0},z_{l}^{1},\cdots ,z_{l}^{K} \right]$. zlk$z_{l}^{k}$ denotes the possibility of selecting the task-related adapter Module-k at the lth convolutional layer. And z = [z₁,z₂,⋯,z_L], L denote the number of layers of 3×3 conv in the backbone network.

Training: The generation of task-relevant adapter execution decisions is optimized by combining reinforcement learning algorithms and task training mechanisms, and the prediction results of the adaptation network constructed based on the “actions” generated by the action generation network are the rewards for optimizing the action generation network. Inspired by the self-criticism sequence training method, the objective of reinforcement learning training can be expressed as follows: (5) L(θ)=−(r(zs)−r(zm))log(zs)\[L(\theta )=-(r({{z}^{s}})-r({{z}^{m}}))\log ({{z}^{s}})\] where θ is a parameter of the gate network and θ={ Wl1,Wl2 }\[\theta =\left\{ W_{l}^{1},W_{l}^{2} \right\}\],l ∈ {1,2,⋯,L}. zs=[ z1s,...,zLs ]${{z}^{s}}=\left[ z_{1}^{s},...,z_{L}^{s} \right]$ denotes the discrete probability corresponding to each layer of random sampling actions, and accordingly, r(z^s) denotes the prediction result of the adaptation network constructed under random sampling actions; furthermore, zm=[z1m,...,zLm]${{z}^{m}}=[z_{1}^{m},...,z_{L}^{m}]$, and: (6) zlm=argmax (zl)\[z_{l}^{m}=\arg \max \text{ }({{z}_{l}})\] where arg max(·) denotes the maximum discrete probability of returning the output of the gate network at layer l, then r(z^m) denotes the prediction result of the adapted network constructed under the action corresponding to the maximum discrete probability. Therefore, in training, if the return of the randomly selected action is higher than the return under the maximum probability corresponding action, the corresponding probability of the randomly selected action will be increased, otherwise it is reduced, and through the meta-training of a large number of tasks, it makes the model adaptive in generating task-related parametric policies for the target task, and then constructing the adaptive network model.

The network architecture of an adaptive network consists of a series of residual blocks (RBs), each of which contains an adaptive module (AM), a BN layer, and a ReLU function [29]. In the adaptive module, the 3×3 conv is parallel to the task-related adapter, which is determined by the corresponding layer gate network of the action generation network.

In the adaptation network, f_ϕl(·) denotes the 3×3 conv in the lth adaptive module, which is pre-trained in the source domain. hl−1∈ℝNS×C×W×H${{h}_{l-1}}\in {{\mathbb{R}}^{{{N}_{S}}\times C\times W\times H}}$ denotes the input feature of the lth adaptive module, if the action generation network generates the decision selection task related adapter Module-0, it means that the features of this layer do not need to perform the transformation of the feature distribution, and the output of the adaptive module is h_l = f_ϕl(h_l−1). When the kth task-related adapter is selected, the features learned by the task-related adapter will be fused with the features extracted by 3×3 conv to form the final features, which can be expressed as: (7) hl=fϕl(hl−1)+1k∑i=1kfψli(hl−1)\[{{h}_{l}}={{f}_{{{\phi }_{l}}}}({{h}_{l-1}})+\frac{1}{k}\sum\limits_{i=1}^{k}{{{f}_{\psi _{l}^{i}}}}({{h}_{l-1}})\] where fψli(⋅)${{f}_{\psi _{l}^{i}}}(\cdot )$ denotes the ith task-specific layer in the task-related adapter, 1×1 conv in the experiment.

TRAINING: Based on the two classes of actions generated by the action generation network with discrete probabilities of z^s and z^m, two adaptive networks Madapt1$M_{adapt}^{1}$ and Madapt2$M_{adapt}^{2}$ can be constructed , In training, firstly, the two adaptive networks are trained using the categorical cross-entropy loss of the support samples, which can be expressed as: (8) Ladapt(ψ,π)=1NS∑(x,y)∈Sl(fϕ,ψ,π(x),y)\[{{L}_{adapt}}(\psi ,\pi )=\frac{1}{{{N}_{S}}}\sum\limits_{(x,y)\in S}{l}({{f}_{\phi ,\psi ,\pi }}(x),y)\] where f_ϕ,ψ,π(·) denotes the classification probability vector of the output of the adaptation network with a dimension equal to the number of classes in the support set, and π denotes the parameters of the linear classifier. Then, the reward r of the action generation network can be obtained by predicting the classification accuracy of the query sample Q, which can be expressed as: (9) r(zs)=Nadapt1NQ,r(zm)=Nadapt2NQ\[r({{z}^{s}})=\frac{N_{adapt}^{1}}{{{N}_{Q}}},r({{z}^{m}})=\frac{N_{adapt}^{2}}{{{N}_{Q}}}\] where Nadapt1$N_{adapt}^{1}$ and Nadapt2$N_{adapt}^{2}$ denote the number of query samples correctly predicted by the Madapt1$M_{adapt}^{1}$ and Madapt2$M_{adapt}^{2}$ networks, respectively.

Chinese traditional culture is rich in connotation, in order to give full play to its value in modern art design, designers should change their design concepts, explore the essence of traditional culture in depth, and strive to show the charm of traditional culture, and in this way, expand the spread of Chinese culture. On the one hand, Chinese modern art designers should eliminate the constraints of foreign culture, accept traditional cultural elements with an open mind, and emphasize the role of traditional cultural elements in art design. Using the model of this paper to extract and identify and categorize various types of traditional cultural elements, actively understand traditional cultural elements, and on the basis of correctly grasping their connotations, take the essence and remove the dregs, and refine the elements that are valuable to modern art design, so as to promote the effective dissemination of traditional cultural elements while adding a cultural flavor to art.

Traditional cultural elements bring creative inspiration to modern art design. Therefore, it is necessary to develop traditional cultural elements reasonably and playfully portray the charm of the unique oriental culture in the new era. If traditional cultural elements want to realize the harmonious development with modern art design, designers first need to break through the barriers between tradition and modernity, interpret traditional culture with modern thinking, and enhance the public’s acceptance of traditional culture. On the one hand, the designer’s use of traditional cultural elements should be based on respect, which does not mean the wholesale acceptance of traditional cultural elements, but on the basis of a full understanding of the needs of the times and the characteristics of the design object, using the model in this paper, the traditional cultural elements should be optimized and combined or adjusted appropriately, so as to make them compatible with modern aesthetics. At the same time, at the beginning of artistic design, the designer should have a clear outline of the design work, so that the work and the cultural elements have a high degree of consistency in conveying the mood.

The aesthetic experience brought by art design can give people a sense of spiritual pleasure. Incorporating traditional cultural elements in design not only meets people’s needs for cultural heritage of art design, but also makes art design more innovative. Traditional cultural elements have been infused over the years and are still flourishing, leading the trend of the times with their profound cultural meaning. Therefore, maximizing the benefits of traditional cultural elements has become the key to art design innovation. On the one hand, the connotation of traditional cultural elements is reasonably applied to modern art design, which can effectively enhance the level of design, guide people to deeply perceive the traditional culture, and appreciate the charm of the fusion of tradition and the times.