Enhancing Yoga Posture Recognition with Deep Learning: A Customized Activation Function Approach

Image recognition has been one of the most important fields of image processing and computer vision. Yoga pose estimation is one of the encountered challenges in this area. It is of utmost importance to perform yoga correctly, as any incorrect posture can be unproductive. This paper lays the foundation for building and discussing convolutional neural networks to accurately classify yoga poses into 10 different classes. This paper generates an output indicating whether the user has executed the prescribed pose accurately or not. While earlier research has explored the efficacy of activation functions, studies that center specifically on the classification of yoga poses are not commonly encountered. The main focus of this paper is to examine and compare different activation functions along with the development of a custom activation function. The existing activation functions were insufficient in capturing the underlying patterns within the data effectively, leading to the requirement of a personalized activation function. This customized activation function is developed by exponentiating the Rectified Linear Unit activation function to the power of 1.25. For this experiment, we obtained a collection of images from Google Images and pre-processed the dataset accordingly. The customized activation function was formulated after reasonable fine-tuning and refinement, which yielded the most optimal results among all other models, with accuracy of 80.46% and a root mean square error value of 0.17.