Comparative research of bounding box based and image segmentation based neural network models for individual cattle identification

Scientific and production association industrial capital Ltd, Tashkentskaya str., b. 17, Moscow, Russian Federation

^* Corresponding author: ya.shubin-ivan@yandex.ru

Abstract

Software and hardware based tools for collecting information about animals are an important part of most large agricultural enterprises. Various information collection methods and technologies help farmers to receive data about the health, behavior and location of animals therefore becoming an important tool for monitoring the area of animal captivity or even open spaces. This work is devoted to a comparative study of the effectiveness of methods for identifying cattle based on neural network technologies and machine learning. In this paper, the effectiveness of the bounding box-based method and the image segmentation-based method were investigated. Images of cattle, which can be obtained using computer vision methods, are used as input images for identifiers. The authors are aimed to determine a representative set of features which can be used for reliable and accurate cow identification using neural network based methods, estimate efficiency of those methods using precision, recall and F1- score quality metrics and suggest certain improvements that may improve the quality of identification for neural network methods, if possible. Neural network methods, the effectiveness of which was verified by quality metrics, were also tested on a set of images that did not participate in the formation of the training sample set.