Bupleurum Seeds Recognition with Attention Mechanism

The research on intelligent classification of Chinese herbal seeds with deep learning are usually coarse-grained categories classification. However, these methods only focus on the global feature rather than the local texture details, which leading to a poor performance for the seeds of same genus. In order to solve the problem, in this paper, the Multiple Attentional Pyramid (APN) Networks is applied to complete the fine-grained classification task on the Chinese herbal seeds of the same genus. Specifically, an attention mechanism is applied to the process of information fusion in the pyramid network, where the pyramid mechanism enables the APN networks to focus on information at different scales and the attention mechanism enables the APN network to focus on the local texture details of Chinese herbal seeds. We applied the APN network to discriminate between different species of bupleurum seeds on our bupleurum seeds dataset including 4 kinds of bupleurum seeds. Experiments shows the APN network achieve the better performance (98% accuracy) compared to the traditional networks.