🎉 Our Team’s Paper Accepted to ICCV 2025! 🎉
Published:
We are thrilled to announce that our team’s paper, “MSA: Multi-view Character Representation for Open-set Chinese Text Recognition”, has been officially accepted to ICCV 2025 — one of the top-tier conferences in computer vision!
📌 Key Highlights:
Most existing methods treat open-set Chinese text recognition (CTR) as a single-task problem, primarily relying on prototype learning of glyphs or linguistic components to identify unseen characters. In contrast, humans recognize characters by integrating multiple cues — both linguistic and visual. Inspired by this, we propose a novel multi-task framework — MSA (Multi-view Structure-aware Architecture).
✅ Innovations: 🔹 SACE (Structure-Aware Component Encoding): Leverages a dynamic binary tree structure to emphasize core linguistic components, generating structure-aware and discriminative linguistic representations. 🔹 SAGE (Style-Adaptive Glyph Embedding): Employs glyph-centric contrastive learning to aggregate features across diverse font styles, enhancing model robustness to stylistic variations.
📊 Experimental Results: On the BCTR dataset, MSA achieves 1.3% and 6.0% accuracy improvements under closed-set and open-set settings, respectively — significantly outperforming current SOTA methods!
💻 Code will be released soon — stay tuned!