Applications of classical and deep learning techniques for polar bear detection and recognition from aero photography

Abstract

The problem of detecting polar bears on the image taken from the plane is essential for ecologists who are tracking the disappearing population of the arctic inhabitants. The main challenge for this problem is to detect the white bear on the white ice. This paper covers the approaches which have shown valuable results for contrast objects captured from the plane, like cars, ships, and many others, instead of the polar bears that look blurry on the ice. However, the introduced approach based on both statistical and machine learning methods made it possible to build a tool that increases the semi-automatic bear detection rate a dozen times. The source data consists of 7360 × 4912 px aerial images, each image covering about the 21600 sq.m. of ice. On average, only one bear appears on every 1000 photos. The best-fit parameters for the solution gave a result of about 100% by recall metric and 51% by precision metric. The main strength of this solution is that it allows for finding almost all bears with a moderate amount of false-positive detections.