Reproduce “MDiff4STR: Mask Diffusion Model for Scene Text Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yongkun Du, Miaomiao Zhao, Songlin Fan, Zhineng Chen, Caiyan Jia, Yu-Gang Jiang
MDiff4STR recognizes scene text in difficult images by predicting masked character slots from image features and surrounding characters, then revisiting predictions over three denoising steps. It trains with six inference-aligned masking patterns and token-replacement noise to address unfamiliar masks and confident errors. On OST, its ablation reaches 84.25% accuracy, versus 81.92% without token-replacement noise; it cannot change sequence length.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.