Automatic Drug Pills Detection based on Convolution Neural Network

Abstract

In this study, an automatic drug pills detection system is proposed for visual system. Two stages, detection and classification, are included in the automatic drug pills detection system. In detection stage, the drug-pill localization is provided for pills location, architecture of deep convolution neural network has been applied to extract feature and construct feature pyramid with stronger semantics. The regression and the classification models are improved to output the position of the pills; The second stage is the drug pills classification, which uses the drug pills position output by the pills localization stage, the deep convolutional neural network is use to classify the pill types. The proposed database contains 131 categories of drug pill. There are total 1,680 images with 3144 annotations for localization and over 470,000 images for classification. The experiment result contains a verification dataset, includes 400 pills images with 2825 annotations. Finally, the experiment result is shown that, the top-1 accuracy rate is 79.4%. Top-3 and Top-5 accuracy are 88.3% and 91.8%. The proposed system has achieved well experiment result.