LAUSR

Dairy cow face recognition using Neural Networks has several hurdles. For example, there are only a few instances of each individual. The positions and angles of the individuals in the image fluctuate considerably, the differences between individuals are not apparent, and the number of individuals that the network has not been trained on is enormous, etc. In this paper, an enhanced Siamese Neural Network is used to overcome these barriers. First, a combination of Dense Block and Capsule Network is employed as a feature extractor to keep the spatial information of features while expanding the feature extraction capabilities of the Convolutional Neural Network. Second, image pairings are processed through the Siamese Neural Network to obtain bivariate features. Finally, image recognition is achieved via the correlation analysis of bivariate features. We conduct comparison experiments with different networks on a small cow face dataset. The experimental results demonstrate that Siamese DB Capsule Network can learn abstract knowledge about distinct individuals and can be extended to unfamiliar cows for zero-shot learning.