A Light Deep Learning Based Method for Bank Serial Number Recognition

• Main Authors: Ardian Umam, 禹安銳
• Other Authors: Chuang, Jen-Hui
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/fkvqyu

碩士 === 國立交通大學 === 電機資訊國際學程 === 106 === Optical character recognition (OCR) is one of computer vision fields that attracts research interests due to its broad applications, such as barcode reading, document reading, licence plate recognition, and text-in-the-wild recognition. A mature technology on OCR is highly desired so that many tasks, such as the ones mentioned above, can be performed automatically by relying on machine instead of human who tends to work with limited time and possibilities of human errors. We are interested in text recognition, especially in bank serial number (SN) of Renminbi (RMB) bill, a paper currency used in China, which is often considered as one of the top five currencies used in the world. In general, the recognition of bank SN consists of two main stages: (i) region proposer to locate region of interest (ROI) which contains the bank SN itself, and (ii) character/text recognition inside the ROI. To the best of our knowledge, there are still few studies covering a full stage process, from ROI localization to character recognition process. This thesis work tries to propose a full stage of bank SN recognition system taking the benefit of recent deep learning advancement in vision recognition. We introduce Block-wise Prediction Networks (BPN) to treat the localization of bank SN as block-wise binary classification, which can be considered as a coarse version of dense/pixel-wise prediction used in semantic segmentation. The benefits include short execution time, which is equal to 85.22 ms in CPU, and the use of global features instead of local features to improve the segmentation. Our system then separates the localized bank SN ROI into individual characters, and feed only the character foregrounds to a softmax CNN classifier. Experimental results show that the proposed method can achieve 99.92% and 99.24% accuracy for character and SN of Renminbi (RMB), respectively, tested with 2,368 two sides images of 1,184 RMB bills.