Authors: Guokun Wang(王国坤), Jingyi Shen(沈静逸), Yue Wu(吴岳), Chang Zhou(周昌), Jianqiang Huang(黄建强)

Affiliation: Alibaba

Description: The Training method is based on transformer, which used in both encoder and decoder, multiple loss is combined for better accuracy. Our training data consists of serveral public datasets including CTW, LSVT, RCTW, ReCTS, Baidu Scene Text Recognition contest data. We Train the model on the whole dataset at first, and finetuned on the ReCTS for several epochs.

Authors: Jie.Li(李杰),BaoLin.Zhang(张保林),KeJie.Liu(刘克捷),Yuan.Hu(胡源)

Affiliation: UNISINSIGHT

Description: First, the training method is based on crnn framework, which takes SE-ResNet with multi-scale features as the backbone, and uses the BiLSTM and attention mechanism to integrate multiple model results. At the same time, the training data used include LSVT, ReCTS, RCTW, ArT and other public free data.

Authors: Qiang Zeng(曾强)，Zhaolin You(游照林)，Yuanyuan Chen(陈媛媛)，Jianping Xiong(熊剑平)

Affiliation: ZHEJIANG DAHUA TECHNOLOGY CO.,LTD

Description: Our training data included ReCTS, LSVT, RCTW, ART and some high-quality artificial synthetic data. We used the CRNN framework for text recognition, and different structures of multi-scale feature extraction backbone such as SA-ResNet, SE-ResNet were used.We used an efficient shuffle attention method which combine spatial attention with channel attention.Meanwhile we used multi-model fusion to predict the final result.