OTSNet: A unified observation-thinking-spelling network for robust scene text recognition

Scene Text Recognition (STR) remains challenging because many existing methods decouple visual and linguistic optimization, often relying on large-scale pre-training to bridge the visual-semantic gap. Inspired by human reading behavior, we propose OTSNet, a unified "Observation-Thinking-Spelling" framework. OTSNet contains three coordinated stages: (1) Observation: a Dual Attention Macaron Encoder (DAME) applies differential attention as an attention sharpening mechanism to emphasize glyph-relevant regions while suppressing background noise; (2) Thinking: a Position-Aware Module (PAM) and a Semantic Quantizer (SQ) transform visual features into discrete glyph-level semantic units; and (3) Spelling: a Multi-Modal Collaborative Verifier (MMCV) performs autoregressive decoding by jointly exploiting visual features, quantized semantic units, and character embeddings. Unlike decoupled pipelines, OTSNet aligns visual and semantic representations without gradient blocking and supports end-to-end training from scratch. Without pre-training, OTSNet achieves 83.5% on Union14M-L and 79.1% on OST, attaining the highest aggregate accuracy on both robustness-oriented benchmark suites among the methods compared. These results demonstrate strong robustness under complex and irregular scene text conditions.