Yoga Pose Classification Using CNN with PReLU Activation

This research uses the power of machine learning for yoga pose classification. A convolutional neural network (CNN) solution for Yoga Pose estimation is proposed in this research. This approach is firmly grounded on a diverse dataset sourced from Kaggle, comprising five quintessential yoga poses - Downdog, Goddess, Plank, Tree, and Warrior2. The distinguishing feature of this model lies in the evaluation of various activation functions namely, Sigmoid, Rectified Linear Unit (ReLU) and Parametric Rectified Linear Unit (PReLU) and found PReLU giving the expected results. This evaluation lends flexibility to the network and augments its learning ability. The final architecture of Yoga Pose Prodigy was designed, encompassing convolutional layers, PReLU activations, max-pooling layers, and densely connected layers. The pivotal layer leads to a five-unit output layer, employing the softmax activation for precise multi-class classification. In the quest for optimal performance, the proposed model undergoes rigorous training, validation, and testing phases. During this process, the proposed system attains a remarkable accuracy of 97.8%, marking a significant milestone in the realm of yoga pose recognition.