Chinese herbal recognition by Spatial-/Channel-wise attention

Chinese herbal recognition has played an essential role in traditional Chinese medicine. Chinese herbal recognition is essentially an image classification task. However, unlike general image classification tasks, due to the particularity of Chinese medicine, traditional Chinese herbal medicine recognition pays more attention to the details of identification objects. However, the details that rely on the delicate paper size cannot be completely distinguished. In many cases, there is a large variety of different types of Chinese medicine. Paying too much attention to the details will cause the model to fall into partial dependence. So many times users need to use global features. In order to solve this challenge, we propose an effective way to model the mechanism by exploiting Spatial-/Channel-wise attention. Besides, we also leverage a dynamic adaptation mechanism that helps the model to balance global and detailed information. We verified the effectiveness of the proposed method via a series of experiments.