Automatic Multi-Pill Detection and Recognition System includes two Convolution Normal Network (CNN) models, and a pill dataset:
1.Multi-pill localization: The localization model is built based on the Feature Pyramid Network and Convolution Normal Network (CNN). The pill detector estimates coordinates of one to more pills; and the pill images are cropped from the input image, according to the estimated result. The accuracy of pill localization achieves 95％ and the execution time is 0.07s.
2.Pill recognition: A CNN model is used to classify the pill images, which are captured by the multi-pill localization task. The accuracy achieves 90％ and the execution time is 0.02s for classification task of each pill. Furthermore, A Standard Operation Procedure (SOP) is designed for the collection of pill images. The drug detection system can be customized, according to the categories of drug provided.
3.Pill dataset: the pill dataset is constructed from images collected in a local medical center. Training, validation and testing datasets are included in the database, which contains a total of 2,429,753 images under 612 drugs categories.