Wildlife Species Recognition Using Deep Learning

The recent deep learning techniques of the last decade are opening new applications and innovations for numerous diverse fields. In this research, we study the problem of wildlife species recognition based on camera trap snapshots and automated methods. In comparison to conventional image classification, this poses a harder problem due to noise and information redundancy. To this end, we apply state-of-the-art techniques from computer vision, and we consider the characteristics of this problem in regard to heterogeneous noisy data. We design a preliminary approach, in the form of a data pipeline, based on techniques such as image preprocessing, data augmentation, transfer learning, and convolutional neural network models. We introduce in this work a case study for the Integrated Management System of the Natural Environment (known as “SIGMedNat”) that collects data about Cantabria's wildlife. We analyze several factors and challenges for this case, as well as results from our preliminary approach for species recognition. This application can be useful to facilitate and improve the monitoring and tracking of wildlife animals for the purposes of observation and preservation.