Abstract
Indian Classical Dance (ICD) is a cultural art form much older than history, where the body movements convey emotions, spiritual meanings, or narrative meanings. Automated classification of ICD styles from visual data presents great challenges because of subtle inter-class variations, intra-class variations, and the complexity of human pose representation. Deep learning models have been applied to this kind of work, but the performance is usually limited due to suboptimal parameter tuning. Therefore, this work introduces a new approach to optimize Deep Belief Network (DBN) for accurate classification of ICD styles. The method proposes to use the newly improved Refined Chameleon Swarm Algorithm (RCSA). By incorporating the non-linear adaptive weight mechanism and the Bernoulli chaotic map, RCSA effectively adjusts its search strategy by balancing its tendency to explore with that of exploit, thereby overcoming CSA's tendency for rapid convergence and improving its global search capability in complex optimization landscapes. The RCSA fine-tunes the architecture and parameters of the DBN, thereby yielding a hybrid model of DBN/RCSA. The model is extensively validated on Indian Dance form Classification (ICD) and Bharatnatyam Dance Poses (BDP) datasets using 5-fold cross-validation against existing state-of-the-art techniques such as DCNN, PointNet, CNN-HPSGW, Transfer Learning, and MobileNet. Empirical results reveal that the DBN/RCSA model gave superior performance in this research with 95% accuracy, 94% in precision, sensitivity, and specificity, and 94% in the F1 score. Further ablation studies have established that RCSA is decisive in enhancing DBN's performance, corroborated by the results of confusion matrix analysis, which shows reasonably distinguishing ability across classes. This establishes DBN/RCSA as a new model for ICD classification with such accuracy and reliability that it may be widely challenged for application domains like cultural preservation, dance pedagogy, and digital heritage studies.