Unleashing the Sea Turtle Recognition Power: A Battle of Activation Functions in ResNet50

This paper examines the role of activation functions in deep learning models for sea turtle recognition using the ResNet50 model. We compare six activation functions, namely Sigmoid, ReLU, Leaky ReLU, PReLU, ELU, and Swish, to identify their impact on performance. The results show that Swish and PReLU are highly effective, closely followed by Sigmoid. Interestingly, PReLU outperforms others in terms of training duration. These findings emphasize the importance of selecting the right activation function for sea turtle recognition. Researchers and practitioners can use these insights to improve accuracy and reliability in similar tasks.