MAGIC: Multi-granularity domain adaptation for text recognition

Domain gaps between synthetic text and real-world text restrict current text recognition methods. One solution is to align features through Unsupervised Domain Adaptation (UDA). Most existing UDA-based text recognition methods extract global and local features to alleviate domain differences, only focusing on character distribution gaps. However, notable distribution gaps in character combinations exert a pivotal influence diverse text recognition tasks. To this end, we propose a Multi-level And multi-Granularity domain adaptation with entropy loss guIded text reCognition model, named MAGIC. It integrates Global-level Domain Adaptation (GDA) to mitigate image-level domain drift and Local-level Multi-granularity Domain Adaptation (LMDA) local feature shifts. Particularly, we design a subword-level domain discriminator to align the subword features relating to each character combination. Moreover, multi-granularity entropy minimization is used to optimize the target domain data for better domain adaptation. Experimental results on several types of text datasets demonstrate the effectiveness of MAGIC.