Accurate system for automatic pill recognition using imprint information. Issue 12 (1st December 2015)

• Record Type: Journal Article
• Authors: Yu, Jiye; Chen, Zhiyuan; Kamata, Sei‐ichiro; Yang, Jie
• Is Part Of: IET image processing. Volume 9:Issue 12(2015)
• Page Start: 1039
• Page End: 1047
• Publication Date: 2015-12-01
• Subjects: feature extraction -- image recognition -- drugs -- image segmentation -- medical computing -- image sampling -- transforms; imprint information -- contemporary medicine -- adverse pill events -- high‐accuracy automatic pill recognition system -- pill imprint -- imprint extraction -- description parts -- modified stroke width transform -- coherent strokes -- image segmentation -- Loopy belief propagation -- incoherent coarse stroke problem -- two‐step sampling distance sets -- imprint partition -- noise points -- query pill images; Image processing -- Periodicals; 621.36705
• DOI: 10.1049/iet-ipr.2014.1007
• Languages: English
• ISSNs: 1751-9659

Abstract

Abstract : With rapidly advancing of contemporary medicine, it is necessary to help people identify various kinds of pills to prevent the adverse pill events. In this study, a high‐accuracy automatic pill recognition system is proposed for accurate and automatic pill recognition. As pill imprint is main distinction between different pills, this system proposes algorithms on both imprint extraction and description parts to make use of imprint information. First, proposed modified stroke width transform is adopted to extract the imprint by detecting coherent strokes of imprint on the pill. Moreover, image segmentation by Loopy belief propagation is also added on printed imprint pills to solve the incoherent and coarse stroke problem. Second, a new descriptor named two‐step sampling distance sets is proposed for accurate imprint description and successfully cut down the noise on extracted imprint. This strategy is based on the imprint partition – partitions the imprint on the basis of separated strokes, fragments and noise points. Recognition experiments are applied on extensive databases and result shows 90.46% rank‐1 matching accuracy and 97.16% on top five ranks when classifying 12 500 query pill images into 2500 categories.