TextDS: Parameter-Efficient Representation Alignment for Scene Text Detection under Distribution Shifts

In real-world deployments, scene text detectors inevitablyface distribution shifts beyond the training distribution. Prior work of-ten depends on large-scale scene-text pretraining, yet evaluation undercross-domain changes and real-world imaging degradations remains lim-ited. We propose TextDS, an efficient framework for scene text detec-tion under distribution shifts. First, we propose a data-efficient dual-encoder design with visual foundation models, eliminating the reliance onlarge-scale scene-text pretraining. Second, we introduce Step-wise LoRAadaptation (SWLoRA), which performs progressive low-rank refinementwith a dynamic early-exit mechanism for effective feature adaptation.Third, we propose Common Subspace Fusion (CSF) to align and fuse thetwo branches in a shared subspace while retaining complementary, shift-robust information. Finally, we construct adverse-condition scene text de-tection datasets to address the gap in evaluating under imaging degrada-tion. Experiments show that TextDS achieves competitive performancein scene text detection, demonstrating robustness across domains andadverse imaging conditions with only 4.9M trainable parameters. Thecode is publicly available at https://github.com/ZChenDang/TextDS