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

2026-06-26

AI Summary

This work addresses the significant performance degradation of scene text detection under domain shift and realistic imaging degradations by proposing TextDS, a novel framework based on a dual-encoder architecture that leverages vision foundation models without requiring large-scale scene text pretraining. TextDS introduces two key innovations: Stepwise Low-Rank Adaptation (SWLoRA), which enables efficient parameter fine-tuning with dynamic early exiting, and Common Subspace Fusion (CSF), which enhances cross-domain feature consistency. With only 4.9 million trainable parameters, TextDS achieves competitive detection performance across multiple cross-domain and degraded scenarios, substantially improving both robustness and computational efficiency.

Abstract

In real-world deployments, scene text detectors inevitably face distribution shifts beyond the training distribution. Prior work often depends on large-scale scene-text pretraining, yet evaluation under cross-domain changes and real-world imaging degradations remains limited. We propose TextDS, an efficient framework for scene text detection under distribution shifts. First, we propose a data-efficient dual-encoder design with visual foundation models, eliminating the reliance on large-scale scene-text pretraining. Second, we introduce Step-wise LoRA adaptation (SWLoRA), which performs progressive low-rank refinement with a dynamic early-exit mechanism for effective feature adaptation. Third, we propose Common Subspace Fusion (CSF) to align and fuse the two branches in a shared subspace while retaining complementary, shift-robust information. Finally, we construct adverse-condition scene text detection datasets to address the gap in evaluating under imaging degradation. Experiments show that TextDS achieves competitive performance in scene text detection, demonstrating robustness across domains and adverse imaging conditions with only 4.9M trainable parameters.

Problem

• scene text detection
• distribution shifts
• cross-domain
• imaging degradations
• robustness

Innovation

• distribution shift
• parameter-efficient adaptation
• visual foundation model
• low-rank adaptation
• feature alignment