method: PATECH_CHENGDU_OCR_V22019-08-05

Authors: JunKun Zhou, Ming Guan, YuBin Xiao, ZhengNan Luo, MingTao Wang, JiuLin Li

Description: After confirmed with the committee, the total field doesn't need to recognize charactors such as "RM","$" in version V1. This recognization would be supported in version V2, and improves the accuracy with about 5%.Four indepent procedures were adopted for the four fields in this solution, EAST/PSENET integration for detection, ResNet/Attention BiLSTM based for recognization, and multiple-model integration is used. The field extraction is based on the NLP classfication and we autocorrect the final results by a lexicon which is made by the train data set.