ANALYSIS OF BODY CHANGES IN HIGH-LEVEL SPORTS DANCE MOVEMENTS BASED ON DEEP LEARNING BIOLOGICAL IMAGE VISUALIZATION TECHNOLOGY

This study proposes a method for dynamic feature analysis and visualization of high-level sports dance movements based on an improved Spatio-Temporal Graph Convolutional Network (ST-GCN). In response to the biomechanical characteristics of sports dance movements, where the trunk serves as the core of force generation and the limbs act as the extension of expression, the proposed model designs a dual-branch feature extraction architecture with parallel “core stability stream” and “peripheral high-frequency detail stream”. It also introduces an adaptive attention residual module based on the direction of dynamic energy flow. Furthermore, according to the intrinsic energy attributes of movements, the model dynamically divides a complete dance movement sequence into three phases: preparation, execution, and conclusion. A three-headed decoupled output structure is adopted for independent modeling and parallel analysis, achieving accurate characterization and quantitative evaluation of the fine-grained phase structure of complex dance movements. Experimental results on the AIST Dance Video Database and NTU RGB[Formula: see text]D Dataset demonstrate that the average joint localization errors of the proposed method reach 12.3[Formula: see text]mm and 14.7[Formula: see text]mm, respectively, while the action recognition accuracies hit 94.2% and 91.5%, respectively. The generated phased movement visualization results obtain a score of approximately 4.7 (on a 5-point scale) in expert evaluations. Even without visual input, the action recognition accuracy remains above 88.5%. Compared with the baseline methods, including the basic ST-GCN, Convolutional Neural Network-Long Short-Term Memory Network (CNN-LSTM), and nontemporal Graph Convolutional Network (GCN), the proposed method exhibits significant improvements in recognition accuracy, visualization effect, and generalization ability. In particular, when dealing with the long-term temporal dependencies and complex phase transitions unique to dance movements, it demonstrates superior recognition performance over current mainstream state-of-the-art models. This study provides effective technical support for the refined analysis, teaching, and evaluation of high-level dance movements.