Refining Scene Text Recognition via Multi-Domain Character Distance Perception

Scene Text Recognition (STR) faces persistent challenges due to diverse text appearances and complex backgrounds. While SVTRv2 demonstrates strong performance as a CTC-based model, and CDistNet achieves robustness through Multi-Domain Character Distance Perception (MDCDP), further enhancement remains possible. This paper proposes a novel architecture by integrating the MDCDP module into the SVTRv2 framework. MDCDP refines the recognition process by utilizing position embeddings to jointly query both visual and semantic features, enabling robust perception of character distance across domains. This integration allows SVTRv2 to leverage MDCDP’s strength in establishing accurate feature-character alignments under challenging conditions. Experimental results on standard English STR benchmarks (including IIIT5k, SVT, IC13, IC15, SVTP, and CUTE80) show our SVTRv2-MDCDP model achieves an average accuracy of $96.42 \%$. The results validate the effectiveness of synergizing SVTRv2’s efficient feature extraction with MDCDP’s multi-domain relational reasoning for robust text recognition.