Towards toxic and narcotic medication detection with rotated object detectors

BACKGROUND: Recent years have witnessed the advancement of deep learning vision technologies and applications in the medical industry. Intelligent devices for specific medication management could alleviate workload of medical staff by providing assistance services to identify drug specifications and locations. METHODS: In this work, object detectors based on the you only look once (YOLO) algorithm are tailored for toxic and narcotic medication detection tasks in which there are always numerous of arbitrarily oriented small bottles. Specifically, we propose a flexible annotation process that defines a rotated bounding box with a degree ranging from 0° to 90° without worry about the long-short edges. Moreover, a mask-mapping-based non-maximum suppression method has been leveraged to accelerate the post-processing speed and achieve a feasible and efficient medication detector that identifies arbitrarily oriented bounding boxes. RESULTS: Extensive experiments have demonstrated that rotated YOLO detectors are highly suitable for identifying densely arranged drugs. Six thousand synthetic data and 523 hospital collected images have been taken for training of the network. The mean average precision of the proposed network reaches 0.811 with an inference time of less than 300 ms. CONCLUSIONS: This study provides an accurate and fast drug detection solution for the management of special medications. The proposed rotated YOLO detector outperforms its YOLO counterpart in terms of precision.