Visual Attention-Guided Approach to Monitoring of Medication Dispensing Using Multi-location Feature Saliency Patterns

This paper is dedicated to the development of a computer vision-based system for medication (pills and capsules) identification and counting in order to increase the productivity of medication dispensing and maintain its high safety. The algorithmic basis of the system is the attentive vision approach to robust and fast object detection in images. It consists in time-efficient image analysis by a multi-scale visual attention operator to detect feature-point areas located inside the pill and capsule regions. The attention operator combines a spatial saliency filter with a temporal change (novelty) detector in order to robustly detect salient and object-relevant feature points. The medication recognition algorithm involves a set of image descriptors at the feature-point areas called the multi-location feature-saliency pattern, which fully discriminates between different types of medication. The method detects pills and extracts area-based descriptors without any image pre-segmentation procedure due to the proposed multi-scale attention operator.