Methodological approaches and algorithms for recognizing and counting animals in aerial photographs

Abstract

Introduction: The complexity of recognition and counting of objects in a photographic image is directly related to variability of related factors: physical difference of objects from the same class, presence of images similar to objects to be recognized, non-uniform background, change of shooting conditions and position of the objects when the photo was taken. Most challenging are the problems of identifying people in crowds, animals in natural environment, cars from surveillance cameras, objects of construction and infrastructure on aerial photo images, etc. These problems have their own specific factor space, but the methodological approaches to their solution are similar. Purpose: The development of methodologies and software implementations solving the problem of recognition and counting of objects with high variability, on the example of reindeer recognition in the natural environment. Methods: Two approaches are investigated: feature-based recognition based on binary pixel classification and reference-based recognition using convolutional neural networks. Results: Methodologies and programs have been developed for pixel-by-pixel recognition with subsequent binarization, image clustering and cluster counting and image recognition using the convolutional neural network of Mask R-CNN architecture. The network is first trained to recognize animals as a class from the array of MS COCO dataset images and then trained on the array of aerial photographs of reindeer herds. Analysis of the results shows that feature-based methods with pixel-by-pixel recognition give good results on relatively simple images (recognition error 10–15%). The presence of artifacts on the image that are close to the characteristics of the reindeer images leads to a significant increase in the error. The convolutional neural network showed higher accuracy, which on the test sample was 82%, with no false positives. Practical relevance: А software prototype has been created for the recognition system based on convolutional neural networks with a web interface, and the program itself has been put into limited operation.