VQ-STE: Scene text erasing with mask refinement and vector-quantized texture dictionary

Abstract

Scene text erasing (STE), which aims to remove text from natural images and restore a plausible background, has been extensively researched in recent years. Most existing STE methods employ segment-then-erase pipelines, either explicitly or implicitly. However, these methods still face challenges, such as inaccurate text segmentation, difficulty in large text removal, and the shortage of training data. To address the first two issues, we present a scene-text-erasing network, VQ-STE, and to mitigate the third issue, we introduce a high-quality synthetic dataset, MixSyn. VQ-STE comprises a lightweight text Mask Refinement Network (MRN) and a Texture Dictionary-based Inpainting Network (TDIN). The MRN refines the bounding box-level text region mask, producing a high-recall stroke-level text mask by incorporating data augmentation and multiple loss functions. The TDIN erases large text regions by replacing distorted features with those from a pre-learned Texture Dictionary. Moreover, our generated MixSyn dataset offers greater diversity in background, text appearance, layout, and annotation compared to existing synthetic datasets. VQ-STE performs effectively in one or two-step settings, i.e., with or without additional text bounding box information. Experimental results demonstrate that VQ-STE outperforms existing one-step and two-step methods in quantitative and qualitative evaluations on the SCUT-EnsText dataset.