Character Region Awareness Network for Scene Text Recognition

Recognizing text in natural scenes is still a very challenging task, due to arbitrary shapes, varying fonts, complex backgrounds and so on. Recently, some recognizers utilize Spatial Transform Network (STN) to rectify irregular text instances and achieve promising results. However, their robustness and accuracy are still limited, since rectification performance can be easily degraded by challenging samples. To tackle this issue, we propose a simple yet effective two-dimensional (2D) character attention module, which can enhance foreground text instances via character region awareness. By incorporating this with existing rectification pipeline, we build a novel scene text recognizer named Character Region Awareness Network (CRAN). Extensive experiments demonstrate that our CRAN outperforms previous methods nearly on all benchmarks of both regular and irregular text, particularly on SVT (+2.0%), SVTP (+1.5%) and CUTE80 (+2.1%).