Deep Sheep: kinship assignment in livestock from facial images

Lech Szymanski, Michael Lee

Abstract: For the non-farmer folk all sheep might look the same, but they are in fact morphologically quite different; including when it comes to facial features. Image analysis has already demonstrated that computer-based facial recognition in livestock is very accurate. We investigate the viability of deep learning for assigning kinship in livestock for use in genetic evaluation- given two images of sheep faces, our proposed model predicts their genetic relationship. In this work we present two CNN models: one for face detection (reporting 80% accuracy) and one for kinship detection (reporting 68% balanced accuracy).