Abstract: Most fine-grained recognition only extracts global feature to classify and ignores visual differences of parts caused by attitude angle and pose. So this paper proposes a fine-grained recognition method by combining part feature with global feature. First, the paper does the target pose clustering to show the same visible part of object in the same pose, then extracts parts feature of object and combines global feature to classify in every pose. The proposed model is validated by the experimental results on the bird database CUB_200-2011, which has a significant effect on attitude and visual angle. The results show that the proposed method has better performance than the existing methods.

Key words: fine-grained； pose clustering, part feature, global feature