A real-time visual inspection system for automated prescription dispensing

We present an image processing system that is designed to confirm the correctness of automated prescription orders, without any special pill handling or barcodes, using only a PC equipped with a frame grabber and an overhead camera. Training consists of averaging several categories of features (color, shape, texture, glossiness, and markings) from multiple image samples of each target pill type. During online operation, this system extracts sets of features from the sample pill image and eliminates pill types from the candidate list whose features are outside certain tolerances from the expected target pill. If there is exectly one candidate left, and that candidate is the target pill type, then the sample is accepted as correct. If there is more than one candidate left, or if the target pill type has been pruned out of the candidate list, then the sample is sent to pharmacists for manual verification.