Automated counting of large vertebrate species using AutoML technology

¹ St. Petersburg Federal Research Center of the Russian Academy of Science, 14 line V.O., 39, St. Petersburg, Russia
² United Directorate of Taimyr Nature Reserves, Talnakhskaya 22, Norilsk, Russia
³ Institute of Ecology and Evolution A. N. Severtsov of the Russian Academy of Science, Leninsky prospect, 33, Moscow, Russia

^* Corresponding author: arguzd@yandex.ru

Abstract

The purpose of the presented work is to develop an automation system for synthesizing models of automatic recognition of different animal species in photo and video images. The paper presents a system for recognizing and counting two large vertebrate species - reindeer (Rangifer tarandus) and white-cheeked goose (Branta bernicla) on aerial images. The AutoGenNet recognition system is based on a convolutional neural network (CNN) of Mask R-CNN architecture using the concept of automatic machine learning (AutoML). The created system is able to automate a number of stages of model creation for recognizing objects in images. In particular, the presented system utilizes transfer learning. This approach significantly reduces the amount of training data required. The CNN model is synthesized automatically based on the images marked up by AutoGenNet system. To learn the Mask R-CNN model and to test the recognition accuracy, we used the images of reindeer herds obtained during aerial surveys in Taimyr and the images of brant goose flocks taken in different regions of the Arctic zone of the Russian Federation. On average, the trained software correctly recognised 82% of reindeer on the test array. Correctly recognizable brant geese accounted for 65% across the entire data set tested. Considering that this model of different animal species recognition was created automatically, with minimal involvement of machine learning specialists, this result indicates the successful application of the AutoML approach.