On farm automatic sheep breed classification using deep learning

Abstract Automatic identification of breeds of sheep can be valuable to the sheep industry. Sheep producers need to identify different breeds of sheep to estimate the commercial value of their flock. In many situations however, farmers find it challenging to identify the breeds of sheep without a great deal of experience. DNA testing is an alternative method for breed identification. However, it is not practical for real time assessment of large quantities of sheep in a production environment. Hence, autonomous methods that can efficiently and accurately replicate the identification ability of a sheep breed expert, while operating in a farm environment are beneficial to the industry. Our original contributions in this field include: setting up a prototype computer vision system in a sheep farm, building a database compromising 1642 sheep images of four breeds captured on a farm and labelled by an expert with its breed and training a sheep breed classifier using machine learning and computer vision to achieve an average accuracy of 95.8% with 1.7 standard deviation. This classifier could assist sheep farmers to accurately and efficiently differentiate between breeds and allow more accurate estimation of meat yield and cost management.