Deep Learning for Inexpensive Image Classification of Wildlife on the Raspberry Pi
Animal conservationists need unobtrusive methods of observing and studying wildlife in remote areas. Many commercial options for wildlife observation are expensive, obtrusive, or sub-optimal in remote environments. In this paper, we explore the viability of a Raspberry Pi-based camera system augmented with a deep learning image recognition model for detecting wildlife of interest. Unlike traditional sensor nodes that would have to transmit every captured image, localized image recognition enables only pictures of desired animals to be transferred to the user. For the purposes of this study, we use TensorFlow and Keras to create a convolutional neural network that runs on a Raspberry Pi 3B+. We trained the model on nearly 3,600 images gathered from publicly available image databases that are split into three classes. Our experiments suggest that our system can detect snow leopards with between 74 percent and 97 percent accuracy. We believe that our results show the viability of employing deep learning image recognition models on the Raspberry Pi to create an inexpensive system to observe wildlife.
Deep learning, Image recognition, Image databases, Snow, Computational modeling, Wildlife, Mobile communication
B. H. Curtin and S. J. Matthews, "Deep Learning for Inexpensive Image Classification of Wildlife on the Raspberry Pi," 2019 IEEE 10th Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON), New York, NY, USA, 2019, pp. 0082-0087, doi: 10.1109/UEMCON47517.2019.8993061.