A Novel Deep Learning Approach for Automatic Indian Classical Dance Style Classification

In computer vision, the classification and recognition of complex human actions from unregulated web and offline video sequences is still a tremendous challenge. Leveraging Convolutional Neural Networks (CNNs), this work introduces a classification approach based on deep learning to classify Indian classical dance movements. Ten subjects performed 200 diverse dance poses and mudras within a range of Indian classical dances and background scenes, as observed in recordings made from controlled offline sources. In addition, YouTube was utilized to capture online video data that recorded ten different respondents’ performances. Dance postures are depicted by 60 frames per video clip. Eight sample groups—the last two samples were reserved for testing—were utilized to train the proposed model on different data subsets. To achieve optimal recognition accuracy, we systematically developed and evaluated a series of CNN architectures. Our best model surpassed other classifier models that had been reported earlier on the same dataset with a recognition accuracy of 93.86% through exhaustive hyperparameter optimization. These findings demonstrate the capability of CNN-based approaches to automate dance movement classification, pushing computer vision applications and cultural heritage preservation.