Chinese materia medica resource images screening method study

Chinese materia medica resource survey provides an important basis for the development of traditional Chinese Medicine (TCM) industry. During the Chinese materia medica resource survey process, millions of materia medica plant images are collected. The collected image dataset includes some images that are unqualified for image analysis, i.e. they can’t be used to build medicinal plant classifier model. It is a burdensome work to identify the unqualified Chinese materia medica resource images manually. How to screen the unqualified images automatically is an important task of Chinese materia medica resource survey. Image recognition techniques developed quickly in recent years. Outlier detection is a kind of unsupervised method to find the unqualified images automatically. Lots of research work has been done on the topic. Extracted features and correlation metric play important roles on the outlier image detection result. For improving the image screening performance, a novel outlier detection method is proposed in this paper. Convolutional neural network (CNN) is used to extract the complicated features of Chinese materia medica resource images. Extended entropy is introduced into the calculation of information loss that is used to measure the distance between images. Based on the extracted image features and correlation metric, a novel outlier detection method based on clustering is proposed here. The efficiency of the screening method is illustrated with a practical example.