Herpetofauna Species Classification from Images with Deep Neural Network

Abstract

Camera-traps are noninvasive tools that can capture thousands of images of wildlife species per deployment. To conduct collaborative wildlife monitoring for conservation and to collect up to date information about wildlife species, integrated camera-sensor networking systems have been established at a large scale in Bastrop County, Texas. Species recognition from gathered images is a challenging assignment for computers due to a large amount of intra-class variability, viewpoint variation, lighting illumination, occlusion, background clutter, and deformation. Moreover, processing millions of captured images is daunting, expensive, and time-consuming as most of the images contain only background absent species of interest. This paper proposes a framework of automated wildlife species recognition by image classification using computer-vision techniques and machine learning algorithms. A Convolutional Neural Network (CNN) architecture has been suggested to classify any two species automatically. As an initial experiment, a binary CNN network has been trained and validated with a small public dataset of snakes, and toads/frogs to classify them within their group. The model evaluation achieved 76% accuracy on average for the test data that supports the prospects for the recommended model.