Deep learning image recognition of cow behavior and an open data set acquired near an automatic milking robot

Abstract

Production animals enjoying good health and well-being are more productive and have a higher output quality. Several technical solutions have been used to monitor the animals’ welfare: those based on computer vision pro-vide cost-efficient and scalable options. In this work, we performed a continuous two-month image acquisition of cows in front of an automatic milking station and divided the data into ten different classes related to the most im-portant activities appearing in the images. The data consisted of almost 19 hours of video, equivalent to more than 1.7 million still images. Based on these imaged, we then implemented a convolutional neural network classifier to recognize the cows’ behavior. The network was tested using cross-validation methodology and achieved an 86% precision rate and 85% recall rate in the classification. The data and the Python program code used in this study are made available. An image data set that directly resembles the harsh conditions inside a barn and can be used for deep learning purposes has not been previously made available.