Graduate Students XIAO Shanyu et al. Won ICDAR 2019 Best Student Paper Runner-Up Award
Recently, master candidate XIAO Shanyu and Ph.D. candidate YAN Ruijie from the Department of Electronic Engineering at Tsinghua University published a paper entitled "Deep Network with Pixel-level Rectification and Robust Training for Handwriting Recognition" at the 15th International Conference on Document Analysis and Recognition (ICDAR 2019), and won the Best Student Paper Runner-Up Award. The authors of the paper include XIAO Shanyu, PENG Liangrui, YAN Ruijie, and WANG Shengjin. XIAO Shanyu, YAN Ruijie and their advisor, associate professor PENG Liangrui, attended the conference held in Sydney, Australia, September 22-25. The first author XIAO Shanyu made an oral presentation at the conference, and answered the questions raised by the attendees.
ICDAR has been the largest biennial international conference in the Optical Character Recognition (OCR) research field organized by the International Association for Pattern Recognition since 1991. ICDAR 2019 has received 403 paper submissions and accepted 52 papers for oral presentations. The acceptance rate for oral presentation is about 13%. There are also 176 papers accepted for poster presentations. The total number of attendees was more than 500.
Handwriting Recognition is one of the artificial intelligence technologies to convert handwritten document images into full-text retrievable text files. It is one of the most challenging topics in optical character recognition research field because of the variations in handwritings and the lack of sufficient labeled samples. The method proposed in this paper helps deep neural networks to learn invariant feature representation by incorporating convolutional neural network based pixel level rectification, and improves the generalization ability of deep neural network models via several regularization techniques. The proposed method has outperformed other previously reported methods on three public handwriting datasets. The proposed method is also useful for other tasks including scene text recognition.
