Chinese Herbal Recognition Databases Using Human-In-The-Loop Feedback

Traditional Chinese medicine identification plays an important role in the development of traditional Chinese medicine. Traditional Chinese medicine identification mostly relies on researchers' experience, so traditional Chinese medicine identification is still challenging. Using the computer identification of traditional Chinese medicine seems an effective method, but no dataset can train models. The lack of a dataset is the challenge of traditional Chinese medicine identification by use computers. This paper proposes a method for constructing a Chinese medicine dataset based on human-in-the-loop. This method uses a manual intervention labeling method to realize a labeling mode that saves labour resources. First, we use a web crawler to collect data from the Internet, then use a pre-model to remove some irrelevant data, next, we iterative data annotation based on the classification confidence, finally, we will obtain a dataset named CH42 that annotation by human-computer collaboration. Besides, we designed a backbone network for explicitly modeling interdependencies between channels. The CH42 contains 42 types of Chinese medicine data, a total of 6,112 pictures, the model automatically labeled about 64% of the data. We sampled 6 sets of data and found 6 mislabeled data from 1458 pictures. The model labeling accuracy rate is about 98.6%.