Image Classification with Convolutional Neural Networks Using Gulf of Maine Humpback Whale Catalog

Abstract

While whale cataloging provides the opportunity to demonstrate the potential of bio preservation as sustainable development, it is essential to have automatic identification models. This paper presents a study and implementation of a convolutional neural network to identify and recognize humpback whale specimens by processing their tails patterns. This work collects datasets of composed images of whale tails, then trains a neural network by analyzing and pre-processing images with TensorFlow and Keras frameworks. This paper focuses on an identification problem, that is, since it is an identification challenge, each whale is a separate class and whales were photographed multiple times and one attempts to identify a whale class in the testing set. Other possible alternatives with lower cost are also introduced and are the subject of discussion in this paper. This paper reports about a network that is not necessarily the best one in terms of accuracy, but this work tries to minimize resources using an image downsampling and a small architecture, interesting for embedded system.