Few-shot Classification with Feature Branches for Cow Face Recognition

Abstract

Chinese government has carried out necessary policies to reduce mass losses in cattle farming in recent years. One of the most important and effective measures is to encourage farmers to buy cattle insurance. Thus, cattle recognition becomes the uttermost primary requirement for advanced technology. Compared to other approaches of cattle recognition, artificial intelligence (AI) technology yields fewer negative effects and has lower costs. Yet, currently limited numbers of cattle pictures hinder the direct application of state-of-the-art AI techniques for cattle recognition, leading to unsatisfied results, such as overfitting. In this paper, we propose a novel AI cattle face recognition method, which uses few-shot classification to overcome the problem of limited numbers of cattle pictures. Our model extracts two elements from the limited number of cattle pictures respectively, i.e. the shared and the private features, independent of each other. This model is likely to learn from a small number of samples and to classify images more accurately. In addition, we incorporate self-supervised learning to augment the model’s learning capacity. The training process of our model uses few-shot learning method. Against our cattle face dataset, this model outperforms other traditional few-shot classification methods.