Reproduce “Text Change Detection in Multilingual Documents Using Image Comparison” on OpenResearch

About this paper

Doyoung Park, Naresh Reddy Yarram, Sunjin Kim, Minkyu Kim, Seongho Cho, Taehee Lee

Researchers at Samsung SDS developed a language-independent Text Change Detection (TCD) model that directly compares text image areas using a two-way semantic segmentation approach. The model achieves an average F1 score of 71.5% and IoU of 55.6% on a new multilingual segmentation benchmark dataset, outperforming state-of-the-art semantic segmentation and general change detection models.

Key insights

• Direct image comparison, when enhanced with multi-scale attention and correlation marginalization, can robustly detect text changes in multilingual documents without relying on language-specific OCR.
• Using a sparse 4D correlation map and marginalization significantly improves performance in localizing changes and reduces the need for extensive preprocessing like explicit text alignment.
• Two-way segmentation, generating maps for changes from source to target and target to source, provides valuable insights into the nature of modifications (e.g., insertions vs. deletions).

Summary by alphaXiv.