End-to-end identification of pharmaceutical blister packages based on one-side handheld images

摘要

Image-based identification of handheld pharmaceutical blister packages occurred during prescription dispensing process is challenging in that the packages are held in arbitrary positions with partial occlusion. Previous solutions rely on combining two complementary images of the handheld packages taken in opposite views, which are then processed in two stages: one for cropping the packages and the other for classification. In contrast, this paper presents solutions that rely on only one side images and require only one end-to-end deep learning network. In particular, Mask R-CNN and FOTS are utilized. Our experiments show that, given a pool of about 230 types of pharmaceutic blister packages commonly found in adult lozenges prescription stations, these two end-to-end solutions attain F1 scores of 99% and 96% for identification performance, comparable to that by previous two-staged architectures. Additional advantages of the new solutions include more compact architectures, faster runtime performance, with less training data and training process involved.