An Automated Animal Classification System: A Transfer Learning Approach

Abstract

Animal classification from images obtained by various techniques in forest become an important task to carry out focused distribution and abundance estimation. In the following paper a frame work based on Transfer Learning (TL) in a Convolutional Neural Network is proposed for the construction of an automated animal identification system. The framework is used to analyze & identify focal species in the images. A dataset of 6,203 camera trap images of 11 species including Wild pig, Barking deer, Chital, Elephant, Gaur, Hare, Jackal, Jungle cat, Porcupine, Sambhar, Sloth bear was obtained. Superior performance can be achieved by using Transfer learning in Deep Convolutions Neural Network (DCNN) for species classification. The accuracy achieved by the proposed model on the test dataset is 96% in 18 epochs by using batch-size of 32. This, in turn, can speed up research findings, construct more efficient and reliable animal monitoring systems, and consequently, save the time and effort of the Indian scientists. Therefore, having the potential to make significant impacts in the classification and analysis of camera trap images of the site under observation.

Keywords: Abundance, Distribution, Images, Convolutions Neural Network, Deep Learning, Transfer Learning, Animal Classification, VGG-16