Construction of a Multimodal Learning-Based Emotion Computing Model for Dance Movements, Clothing, and Aesthetic Education

Abstract

Modeling aesthetic perception from dynamic visual information requires simultaneous understanding of human motion and the physical behavior of textile materials, making multimodal feature fusion a critical challenge in intelligent perception systems. Similar information integration strategies have also attracted increasing attention in electromagnetic sensing and heterogeneous signal interpretation, where multiple data sources must be jointly analyzed to characterize complex targets. This study proposes a multimodal emotion computing framework (DC-AEM) that combines dance kinematics and garment dynamics for computational aesthetic assessment. The framework employs a GCN-LSTM network to encode spatio-temporal skeletal motion and constructs a dedicated costume representation by integrating color distribution, texture descriptors, shape characteristics, and optical-flow-based dynamic features through a convolutional architecture. A cross-modal attention mechanism is introduced to adaptively associate movement patterns with context-dependent fabric behavior, generating a unified representation for predicting gracefulness, power, and fluidity. Experimental evaluation on a ballet performance dataset demonstrates that the proposed model significantly outperforms movement-only and simple fusion baselines, achieving superior correlation and lower prediction error across all aesthetic dimensions. The results confirm that dynamic textile characteristics provide essential complementary information for affective perception and establish an effective multimodal analysis strategy that may offer methodological reference for intelligent visual sensing, electromagnetic perception systems, and data-driven costume design.