🖼️ Image Restoration¶

💬 ACL2025 · 3 paper notes

🔥 Top topics: Image Restoration ×3 · Adversarial Robustness ×2

A Self-Denoising Model for Robust Few-Shot Relation Extraction: This paper proposes a Self-Denoising Model (SDM) to address the issue of support set label noise in few-shot relation extraction. Through the co-training of a label correction module and a relation classification module, SDM automatically corrects noisy labels and achieves more robust relation prediction, significantly outperforming baselines even in noise-free scenarios.
DiffuseDef: Improved Robustness to Adversarial Attacks via Iterative Denoising: DiffuseDef inserts a diffusion denoiser layer between the encoder and the classifier. During training, it learns to predict noise in hidden states. During inference, it adds noise to the hidden representations, iteratively denoises them, and performs ensemble averaging. This plug-and-play approach significantly enhances the robustness of text classification models under both black-box and white-box adversarial attacks.
PreP-OCR: A Complete Pipeline for Document Image Restoration and Enhanced OCR Accuracy: Proposes the PreP-OCR two-stage pipeline: first restoring historical document images using a ResShift model trained on synthetically degraded data (employing multi-directional patch extraction and median fusion), and then applying ByT5 for semantic post-OCR error correction. It reduces CER by 63.9-70.3% across 13,831 pages of real-world historical documents.