Enhancing scene text super-resolution via dense residual reconstruction and asymmetric recurrent modeling

Scene text image super-resolution (STISR) aims to recover structurally intact characters that remain readable under severe degradation. When thin strokes are erased, boundaries blur, and adjacent characters merge, the reconstruction backbone must preserve the weak structural evidence that determines character identity. While recent prior-guided methods have improved text-aware guidance extraction, three issues in the reconstruction backbone remain less explored. First, simple residual connections fail to preserve weak stroke evidence across deep layers. Second, limited visual diversity makes stronger backbones prone to channel co-adaptation. Third, uniform recurrent modeling overlooks the asymmetric contextual needs of different reconstruction stages. We address these issues through three design choices: text residual-in-residual dense blocks (TRRDBs) for denser feature reuse, output-side channel dropout to prevent co-adaptation, and asymmetric recurrent modeling (ARM) that introduces BiLSTM only at the late stage where contextual integration is most needed. Experiments on TextZoom show consistent recognition improvements across multiple recognizers, with larger gains on hard samples (+ 2.9%, + 3.1%, + 3.2% on ASTER, MORAN, CRNN). Ablation studies show that the three components address distinct aspects of reconstruction, and cross-dataset evaluation indicates generalization beyond the training benchmark. These results suggest that reconstruction backbone design is an effective approach to improving recognition-oriented text image restoration.