Aditya Katre

Classifying Marine Mammals Using Convolutional Neural Networks

Started Apr. 1, 2024

Abstract or project description

Deep learning continues to advance image recognition capabilities rapidly, providing cutting-edge support for wildlife conservation initiatives. In this study, we develop a Convolutional Neural Network (CNN) framework to distinguish individual whales and dolphins extracted from the Happywhale dataset, demonstrating high precision in marine mammal identification. Our model achieves robust feature extraction and enhanced class separability by leveraging an EfficientNetB5 backbone with an ArcFace loss function. Multiple data augmentation techniques, including random cropping, grayscale conversion, and color manipulation, are used to improve the model’s adaptability across various imaging conditions. We incorporate a k-nearest neighbors (KNN) algorithm at the inference stage to refine predictions, especially when assigning labels to new individuals. By combining these strategies, we were able to boost classification accuracy, reaching a Mean Average Precision at 5 of 0.88. The results show how effective deep learning can be for fine-grained image identification tasks in marine mammal conservation. Beyond accuracy, the model offers real potential to simplify research workflows and support long-term conservation efforts for marine life.