Image-based Animal Recognition based on Transfer Learning

Abstract

Automatic image-based recognition systems have been widely used to solve different computer vision tasks. In particular, animals' identification in farms is a research field of interest for the computer vision and the agriculture community. It is then necessary to develop robust and precise algorithms to support detection, recognition, and monitoring tasks to enhance farm management. Traditionally, deep learning approaches have been proposed to solve image-based detection tasks. Nonetheless, databases holding many instances are required to achieve competitive performances, not mentioning the hyperparameters tuning issues. In this paper, we propose a transfer learning approach for image-based animal recognition. We enhance a pre-trained Convolutional Neural Network model for animal classification from noisy and low-quality images. First, a dog vs. cat task is tested from the well-known CIFAR database. Further, a cow vs. no cow database is built to test our transfer learning approach. The achieved results show competitive classification performance using different types of architectures compared to state-of-the-art methodologies.