about me

个人简介

计算机视觉 机器学习

Education

Master of Science, College of Biomedical Engineering&Instrument Science,      Supervisor Prof. Hong Zhou
Zhejiang University(ZJU), China                                                                                           September 2018 - Today

     Currently focus on computer version: scene text detection and recognition. More specifically, design efficient and accurate network with CNN and RNN to extract text information from the real scene text pictures. For example, in traffic signs, commodity packages, and advertising posters.

    Weijia Wu, Jici Xing, Hong Zhou, et al. Texts as Lines: Text Detection with Weak Supervision. In AAAI, 2020. (reviewing)
    Xiaolei Sheng, Weijia Wu, Hong Zhou. Combined Features from Radiomics and CNN as Novel Potential Predictors for Predicting GeneMutant in GIST. IEEE TRANSACTIONS ON MEDICAL IMAGING(CiteScore:8.58). (reviewing)
    Weijia Wu, Jici Xing, Hong Zhou, et al. TextCohesion: Detecting Text for Arbitrary Shapes. In CVPR, 2020. (drafting)

Bachelor of Science, Electronic Information Engineering
Wenzhou University, China,                                                                                           September 2014 − June 2018

Professional Experience

“Yibao Cup” Algorithm design competition (2/240)                                                                                Feb 2019 May 2019

  • Task Description: Designing a segmentation algorithm to segment the overlapped text in the invoice, IOU was used to evaluate the performance of the network.
  • Technical Difficulty: It is difficult to segment tiny and complicated overlapped texts in the real scene. Segmentation label of the real scene is hard to obtain.
  • Key Technology: We proposed a novel recognition network integrated with segmentation task. It is difficult to segment overlapped text visually, and the segmentation task is to serve the recognition task. Therefore, we integrated the segmentation network and recognition network. Two-channel feature maps represent the different overlapped text information are extracted in the segmentation network and dealt by recognition network.
  • Highlight: Integrate the segmentation and recognition network, introduce RNN sequence information to enhance the segmentation network semantically, and simulate the human “logical reasoning ability”. The entire network does not need the segmentation mask but just requires text labels.

Texts as Lines: Text Detection with Weak Supervision. (submitted to AAAI2020)

  • Task Description: First proposed a scene text detector based on weakly supervised learning that significantly simplifies the annotation process without losing much precision. Unlike the other detectors using full masks, a line across the text region as the coarse mask is used in our method and could save up 80% time-consuming.
  • Technical Difficulty: The weak label loses the text edge information, and nearly all of the background information, which is rather problematic for supervised learning.
  • Key contribution: A text detector based on weakly supervised learning was proposed, which made full use of the model pre-trained on synthetic datasets. More specifically, the network pre-trained on synthetic data with full masks was used to enhance the coarse masks in the real image. And the enhanced masks were feedback to train our network.
    TextCohesion: Detecting Text for Arbitrary Shapes.
  • Task Description: We proposed a novel text detection network named TextCohesion. TextCohesion splits a text instance into five key components: the Text Skeleton, and four Directional Pixel Regions. These components are grouped together to detect text instance with arbitrary shape.
  • Technical Difficulty: Texts in the real world close to each other are difficult to segment. Arbitrary shapes texts are extremely challenging to capture accurately.
  • Key contribution: We propose a novel text detector named TextCohesion with three helpers: Text Skeleton, Directional Pixel Region and Candidate Filter, which outperformed state-of-the-art methods on two text benchmarks.

    Other Honors

    International Competition
       The first prize in the embedded design and development of the 8th “Blue Bridge Cup”                                 March 2017
    Provincial Competition
       The second prize in the first “Nanjiang Cup” college robot competition in Zhejiang Province                          August 2016
       The third prize in the 2016 Zhejiang Province 6th College Student Electronic Design Competition (TI Cup) September 2016
       The third prize in the 2017 Zhejiang Province 6th College Student Electronic Design Competition (TI Cup) September 2017
    School Competitions
       The second prize in the 8th Electronic Design Competition of Wenzhou University                                         May 2015
       The first prize in the 9th Electronic Design Competition of Wenzhou University                                              May 2016
       The first prize in the 10th Electronic Design Competition of Wenzhou University                                              May 2017
       The “Technology Innovation Star” of Wenzhou University September 2017

Email:wwj123@zju.edu.cn
Github:https://github.com/weijiawu
Kaggle:https://www.kaggle.com/beilin


|