China leads optical character detection tech

China has grabbed all top five posts of a prestigious global ranking in optical character recognition (OCR) field, The Paper reported on Wednesday.

The ranking of International Conference on Document Analysis and Recognition (ICDAR), an important index in OCR field, was reset by Chinese tech company Cloudwalk's algorithm Pixel-Anchor, which took the top spot in the latest single-scale results on ICDAR 2015 released on Tuesday.

The other four positions out of top five on ICDAR 2015 ranking were grabbed by China's Nanjing University and Nanjing University of Science and Technology with PSENet algorithm, Face++ with Mask Text algorithm, SenseTime with FOTS algorithm and Alibaba with IncepText algorithm, respectively.

ICDAR2015 is a pure English text detection sub-data set, used to test algorithm of text detection in natural scene in robust reading competitions established by ICDAR in 2003, aiming to improve text detection and recognition level in natural scene.

ICDAR2017 MLT is another sub-data set including nine languages such as Latin, English, Chinese, Korean, Japanese and Arabic of a complete data set for the competition.

ICDAR2015 and ICDAR2017 are two authorized data sets in field of text detecting in natural scene and basically all articles will be detected there, said Li Yuan, an algorithm engineer with Cloudwalk.

Chinese tech company Cloudwalk's algorithm Pixel-Anchor ranked first on list of results on ICDAR 2017 MLT Chinese and the fourth on a comprehensive list of results on ICDAR 2017 MLT Global, with the top three winners still to submit published thesis.

Comparing with traditional optical character recognition, text detection and recognition in natural scene faces challenges of background interruption, characters blurring and degeneration and unexpected lights but the technology can be applied in more broader fields such as photo analysis, license plate recognition and image advertisement filtering.

Focusing on text recognition and application, the ICDAR is one of the professional conferences organized by International Association for Pattern Recognition.

The robust reading competitions, with open testing data set online and the results being submitted anytime, has attracted over 3,500 participating teams from 89 countries and regions, including Google, Microsoft, Amazon and Facebook.