Authors: Feng Cheng, Lixin Gu, Qingjie Liu, Feng Han, Jingtao Han. Institution: 4Paradigm

Description: Detection model and recognition model are trained separately.
Detection method is based on mask-rcnn. Train-set: ICDAR2019 MLT task1 train-set.
Recognition method is based on convolutional transformer with ResNet50 backbone. T A simple voting is done to identify the script from the recognized transcription. Train-set: ICDAR2019 MLT task2 train-set & task4 Synthetic dataset & ICDAR2017 MLT task2 train-set.