Cow Body Condition Score Estimation with Convolutional Neural Networks

In this paper, we investigate cow body condition score estimation with convolutional neural networks and find that it is difficult for a convolutional neural network classifier to directly discriminate the categories that have very similar appearance in our dataset. In order to address this issue, we firstly recognize the region that contains discriminative features and input the region into a complete classifier. The results show that the method we proposed could improve the precision of cow body condition score estimation compared with applying single classifier directly, even though we discard a large part of input images. Our method could achieve 64.55% and 94.5% accuracy within 0.0 and 0.5 units of difference from true values respectively. We demonstrate that if a single network does not get expected results, cascade many networks that specialized to sub problems may work well.