Title:
END-TO-END RECOGNITION METHOD FOR SCENE TEXT IN ANY SHAPE
Document Type and Number:
WIPO Patent Application WO/2019/192397
Kind Code:
A1
Abstract:
An end-to-end recognition method for a scene text in any shape, comprising: extracting a text feature by means of a feature pyramid network for generating a candidate text box by a region extraction network; and then adjusting the position of the candidate text box by means of a fast region classification and regression branch to obtain more accurate text bounding box position information; inputting the text bounding box position information into a segmentation branch to obtain a predicted symbol sequence by means of a pixel voting algorithm; finally, processing the predicted symbol sequence by means of a weighted edit distance algorithm to locate the most matching word of the predicted sequence in a given dictionary to obtain the final text recognition result. The method can detect and recognize scene texts in any shape in a natural image, comprising a horizontal text, a multi-directional text, and a curved text, and can perform an end-to-end training completely. The detection and recognition method has a strong practical application value.
Inventors:
BAI XIANG (CN)
LV PENGYUAN (CN)
LIAO MINGHUI (CN)
YAO CONG (CN)
CHU JIAJIA (CN)
LV PENGYUAN (CN)
LIAO MINGHUI (CN)
YAO CONG (CN)
CHU JIAJIA (CN)
Application Number:
PCT/CN2019/080354
Publication Date:
October 10, 2019
Filing Date:
March 29, 2019
Export Citation:
Assignee:
UNIV HUAZHONG SCIENCE TECH (CN)
International Classes:
G06K9/34
Foreign References:
| CN108549893A | 2018-09-18 | |||
| CN104751153A | 2015-07-01 | |||
| CN106778757A | 2017-05-31 | |||
| US20150063688A1 | 2015-03-05 |
Other References:
YANG CHUN: "china doctoral dissertations full text database", ELECTRONIC JOURNALS, 15 February 2018 (2018-02-15)
Attorney, Agent or Firm:
LIUJIA CHINA IP LAW OFFICE (CN)
Download PDF: