Context-Based Contrastive Learning for Scene Text Recognition

Pursuing accurate and robust recognizers has been a long-lasting goal for scene text recognition (STR) researchers. Recently, attention-based methods have demonstrated their effectiveness achieved impressive results on public benchmarks. The attention mechanism enables models to recognize with severe visual distortions by leveraging contextual information. However, recent studies revealed that the implicit over-reliance of context leads catastrophic out-of-vocabulary performance. On contrary superior accuracy seen text, are prone misrecognize unseen even good image quality. We propose novel framework, Context-based contrastive learning (ConCLR), alleviate this issue. Our proposed method first generates characters different contexts via simple concatenation operations then optimizes loss embeddings. By pulling together clusters identical within various pushing apart in embedding space, ConCLR suppresses side-effect overfitting specific learns more representation. Experiments show significantly improves generalization achieves state-of-the-art performance benchmarks recognizers.