E-Swish Activations Function in ResNet Architectures for Enhanced Sea Turtle Individual Recognition

Siti NurFarahim Shaharudin, Wan Nural Jawahir Hj Wan Yussof, Muhammad Suzuri Hitam, Ezmahamrul Afreen Awalludin, Nur Baini Ismail, Mohd Erman Safawie Che Ibrahim
dx.doi.org/10.17576/apjitm-2025-1401-16

Abstract

Traditional sea turtle monitoring methods often rely on physical tagging, which can pose challenges such as tag loss and potential harm to the animals. Conversely, deep learning methods offer non-invasive, efficient, and accurate alternatives for sea turtle identification, crucial for effective conservation efforts. This study evaluates the effectiveness of E-Swish compared to the Swish activation function within various ResNet architectures?ResNet18, ResNet34, ResNet50, ResNet101, and ResNet152?for non-invasive sea turtle recognition via deep learning. Utilizing a specialized dataset from Pulau Redang, Malaysia, which includes images affected by environmental noise, we applied advanced convolutional neural network (CNN) configurations. Our methodology included preprocessing techniques to emphasize distinctive scute patterns. We used a comprehensive set of metrics for performance evaluation: accuracy, precision, recall, and F1-score. Notably, the ResNet34 model equipped with E-Swish achieved the highest test accuracy of 88.20%, outperforming its Swish counterpart, underscoring the superior performance and learning efficiency of E-Swish. These findings suggest E-Swish's potential to significantly enhance CNN applications in wildlife conservation, setting new benchmarks for adaptive activation functions in advancing technological tools for effective, ethical, and scalable biodiversity monitoring.

keyword

Sea Turtle Recognition, Convolutional Neural Networks, E-Swish Activation Function, ResNet Architectures, Wildlife Monitoring

Area

Pattern Recognition