Shape Recognition Using Unconstrained Pill Images Based on Deep Convolution Network

The correctness of pill identification and inspection based on visual appearance is a very important issue. The shape of the pill is the most popular visual feature. To locate an individual pill and extract the pill shape from the image, we adopt two object detection models, namely you only look once v3 (YOLOv3) and Mask Regional Convolutional Neuron Network (Mask R-CNN). In this research, we trained each algorithm on our pill image dataset. We analyzed the efficiency of these two models to determine the best pill shape recognition model that can be employed in real world usage conditions. Our results showed that Mask R-CNN successfully identifies shapes with a F1-Score up to 99.5S%, and YOLOv3 yields a F1-Score up to 97.50%. Both models correctly located an individual pill in an image with 98% accuracy.