ANALYSING MARINE ANIMALS CHARACTERISTICS USING CONVOLUTIONAL NEURAL NETWORKS

Abstract

The thesis explores the efficacy of convolutional neural network(CNNs) to categorize lobster images for improving lobster grading and traceability. Traceability ensures that lobsters are traceable to a sustainable source. Lobsters kept in unsuitable conditions such as extremely low temperatures or densely packed crates have low chances of survival leading to a lower grade ultimately affecting prices. The CNNs were able to achieve high accuracies for assessment of lobster traits. Attention mechanisms that learn to extract discriminating features were explored to improve the performance of CNNs. The attention augmented CNNs had similar accuracies compared to the vanilla CNN but were less sensitive to choice of architecture and learning rate. The attention CNNs could map landmarks on lobster images(for sizing) with an acceptable error of about 2cm. Additionally, siamese networks, that were explored for a black box approach towards uniquely identifying lobsters, were able to achieve a top-3 accuracy of about 84%.