講演抄録/キーワード |
講演名 |
2021-03-04 13:25
A Study of Product Identification System Using Optical Character Recognition ○Shixi Chen・Nobuo Funabiki・Masaki Sakagami(Okayama Univ.)・Takashi Toshida・Kohei Suga(Astrolab) LOIS2020-48 |
抄録 |
(和) |
Recently, the optical character recognition (OCR) technology has been remarkably progressed due to the advancements of deep learning techniques. Besides, smartphones equipped with cameras have broadly spread among people around the world. As a result, the product identification from the product label photo using OCR becomes possible as the quick way to identify the product. However, the accuracy of OCR is still not 100%. Some characters are incorrectly recognized or missing in the recognition result, which must be considered for use. In this study, we propose a product identification system applying OCR of the label photo taken by a smart phone. The fuzzy search is adopted to improve the accuracy by finding the best-matching record in the database for the possibly incorrect key by OCR. Since this search takes inadmissibly long time when the database has a lot of records, we also propose the speedup method by limiting the matching records. For evaluations, we apply the proposal to 389 label photos. The results show that the CPU time is 15.39sec by the naïve search, and 0.99sec by the speedup one that limits the number of records to be searched into 0.24% of the naïve one, where the record hit rate is slightly reduced from 94.3% to 94.1%. |
(英) |
Recently, the optical character recognition (OCR) technology has been remarkably progressed due to the advancements of deep learning techniques. Besides, smartphones equipped with cameras have broadly spread among people around the world. As a result, the product identification from the product label photo using OCR becomes possible as the quick way to identify the product. However, the accuracy of OCR is still not 100%. Some characters are incorrectly recognized or missing in the recognition result, which must be considered for use. In this study, we propose a product identification system applying OCR of the label photo taken by a smart phone. The fuzzy search is adopted to improve the accuracy by finding the best-matching record in the database for the possibly incorrect key by OCR. Since this search takes inadmissibly long time when the database has a lot of records, we also propose the speedup method by limiting the matching records. For evaluations, we apply the proposal to 389 label photos. The results show that the CPU time is 15.39sec by the naïve search, and 0.99sec by the speedup one that limits the number of records to be searched into 0.24% of the naïve one, where the record hit rate is slightly reduced from 94.3% to 94.1%. |
キーワード |
(和) |
product identification / OCR / fuzzy search / regular expression / partial word matching / / / |
(英) |
product identification / OCR / fuzzy search / regular expression / partial word matching / / / |
文献情報 |
信学技報, vol. 120, no. 417, LOIS2020-48, pp. 6-11, 2021年3月. |
資料番号 |
LOIS2020-48 |
発行日 |
2021-02-25 (LOIS) |
ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
LOIS2020-48 |
|