Multi-granularity Prediction for Scene Text Recognition

Scene text recognition (STR) has been an active research topic in computer vision for years. To tackle this challenging problem, numerous innovative methods have successively proposed and incorporating linguistic knowledge into STR models recently become a prominent trend. In work, we first draw inspiration from the recent progress Vision Transformer (ViT) to construct conceptually simple yet powerful model, which is built upon ViT outperforms previous state-of-the-art scene recognition, including both pure language-augmented methods. integrate knowledge, further propose Multi-Granularity Prediction strategy inject information language modality model implicit way, i.e. , subword representations (BPE WordPiece) widely-used NLP are introduced output space, addition conventional character level representation, while no independent (LM) adopted. The resultant algorithm (termed MGP-STR) able push performance envelop of even higher level. Specifically, it achieves average accuracy $$93.35\%$$ on standard benchmarks.