How to Win Chinese OCR Competition?

Speaker:        Dr. Hwalsuk LEE
                LINE/Naver OCR Team

Title:          "How to Win Chinese OCR Competition?"

Date:           Monday, 4 November 2019

Time:           4:00pm - 5:00pm

Venue:          Lecture Theater F (near lift no. 25/26), HKUST


Abstract:

In this talk, Dr. Hwalsuk Lee will share his recent achievement on winning
the Chinese OCR competition at ICDAR 2019 using novel text detector and
recognition algorithms.

First, he will introduce CRAFT, which can detect individual characters
even when character-level annotations are not given. This method provides
the character region score and the affinity score that, together, fully
cover various text shapes in a bottom-up manner. (CVPR 2019)

Then, he will analyze the contributions of existing scene text recognition
(STR) methods, which have been evaluated with inconsistent benchmarks,
leading to difficulties in determining whether and how a proposed module
improves the STR baseline model. In response to this issue, his team has
introduced a common framework as well as consistent datasets. The team has
provided a fair comparison, and has analyzed which modules bring the
highest accuracy, speed, and size gains. (ICCV 2019, oral)

Based on these two papers, and with the help of an advanced synthetic data
generation method, Clova-OCR team achieved the top rank in "Latin and
Chinese scene text recognition task" for ICDAR2019 Robust Reading
Challenge on Arbitrary-Shaped Text.


********************
Biography:

Hwalsuk Lee received a Ph.D. degree in Electrical Engineering from KAIST,
Korea, in 2006. His main research interest is in the area of computer
vision, especially text detection and recognition, including document
image analysis and generative modeling.

He is leading the Clova-OCR team since 2017. Recently, his team won 11
tasks in the ICDAR 2019 challenges. He publishes his work on top AI venues
such as CVPR, ICCV, and ICDAR. He is a steering committee member of
Tensorflow Korea, one of the world's most active AI communities.