Pig Face Recognition Based on Multi Supervising Forms

With the development of intelligent breeding, non-contact pig face recognition based on deep learning has been widely studied, enabling accurate and intensive management. Existing face recognition network training in limited categories cannot meet the requirements for breeding short-growth period edible livestock. Moreover, using a single supervising function is hard to distinguish pig faces that are more similar to humans. Therefore, this paper improves Inner origin loss, which can adaptively adjust repulsive circles, allowing for fast and efficient pig face feature extraction. In addition, we propose the TACO loss which can learn facial features to separate pig faces. The experiments conducted on our dataset indicate that in the 1200 pairs of pig face verification pairs composed of 15 pigs, the accuracy of open-set is 96.08%. This study provides a technical reference for pig face recognition research.