The Application of Deep Learning in Dance Movement Design

This paper proposed deep learning for designing dance movements by estimating dance poses. In this study, we proposed Fusion-based Global Dance Pose Patterns with ResNet-152 an approach that resolves class imbalance existing in dance pose datasets with the help of high-resolution global feature fusion during pose estimation. A fusion layer has been added for discerning patterns on both the local and global levels, resulting in a considerable improvement in classifier performance thereby yielding a referred reliability for discriminative power to dance pose classification applications. Refinement in feature extraction and using deep new models such as ResNet-152 for pose recognition have helped to capitalize on model overfitting and worse generalization problems. This approach indeed goes a long way in making dance pose classification more accurate and efficient with possible real-time applications, in addition to a better understanding of the dance movements themselves. Experimental results indicate promising performances, with high accuracies (0.9870), precisions (0.9851), sensitivities (0.9873), F-measures (0.9861), and Kappa (0.9841) constituting proof of the model competency. The workflow comprises a stage for data collection, pre-processing is then applied using Gaussian filtering and histogram equalization to improve image features, the class imbalance is countered using SMOTE, feature extraction on HR-Net, and global and local feature fusion leads to robust pose estimations. ResNet-152 acts as a classifier with an SGD optimizer for better model parameter optimization. This system highly accurately predicts dance poses and efficiently approaches pose estimation in various dance application fields.