Refining Scene Text Recognition via Efficient Local Attention Integration

Scene Text Recognition (STR) remains a significant challenge due to diverse text appearances and complex backgrounds. While advanced models like SVTRv2 demonstrate strong performance, optimizing encoder-stage feature representation offers further improvement. This paper enhances the SVTRv2 baseline by integrating an Efficient Local Attention (ELA) module into its encoder. The ELA module is designed to refine feature extraction by efficiently capturing critical local features with minimal computational overhead. Experimental results on standard English STR benchmarks (including IIIT5k, SVT, IC13, IC15, SVTP, and CUTE80) show our ELA-enhanced model achieves an average accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 6. 5 8 \%}$</tex>. This surpasses the SVTRv2 baseline and compares favorably with other state-of-theart methods. Ablation studies further confirm the ELA module's positive contribution.