method: CSN-ED2019-04-30

Authors: USTC-iFLYTEK

Description: Char Size Normalization Encode-Decode Recognition Model: A simple end-to-end encode-decode deep neural network, which is composed of a dense-net encoder and a GRU-based decoder with two dimensional attention module. The input image is normalized according to estimated char size. Data augmentation is also applied. Finally, we ensemble 12 models trained with different configures as our final model.