Authors: Bado Lee, Youngmin Baek, Seung Shin, Jeonghun Baek, Chaeyoung Lee, Hwalsuk Lee
Description: Additional head on CRAFTS(Character Region Awareness For Text Detection) with model distillation. While training, pre-trained detector outputs character level heatmaps. The heatmap is then modified to be consistent with the gt(pixel-wise multiplication). The new gt is then used to train a duplicate model with an additional head of channels. Each channel corresponds to different languages. The different learning rates are applied for different languages to cope with data imbalance.
CLOVA-AI OCR team, NAVER/LINE Corp.