method: Ssubm190429_nonchar_thres002019-04-29

Authors: ShaoAn Lu

Description: Our approach is a combination of a YOLOv3 for character detection and a ResNet50 for recognition. The ResNet50 model was trained on the 1500 most frequent characters and an additional non-character class. In this submission, we also applied a post-processing scheme that corrects erroneous predictions of "一(Chinese)" and "-(dash)". In addition, for those textlines that our model did not predict any character, we replaced those predictions with that from another ResNet50 which does not contain non-character class in its output.

Name: Shao-An Lu (呂紹安)
Organization: awesome techno