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

Zhengmi Tang, Tomo Miyazaki, Zhijie Wang, Yongsong Huang, Jonathan Pradana Mailoa, Shinichiro Omachi

Abstract: Highlights•We propose VQ-STE, a text eraser that reduces inpainting residuals and artifacts.•We introduce MixSyn, a synthetic dataset with accurate masks and diverse samples.•VQ-STE achieves the state-of-the-art performance.