CoT-SD: Chain-of-Thought Semantic Denoising


Yanlin Jiang (Beijing University of Technology), Yuchen Liu (Beijing University of Technology), Mingren Liu (Alibaba Cloud)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)

Abstract

Recent advances in zero-shot image denoising have largely been driven by self-supervised learning and generative models. However, these methods often struggle to preserve fine-grained structural details in real-world noisy images. Inspired by the human ability for slow thinking in problem-solving, which involves breaking down tasks and addressing them step by step, this paper proposes Chain-of-Thought Semantic Denoising (CoT-SD), a novel zero-shot denoising framework that leverages multi-step self-reflective reasoning and CLIP-based semantic alignment to progressively enhance image quality. Unlike conventional single-step denoising approaches, CoT-SD formulates denoising as an iterative reasoning process in which both the encoder and the decoder progressively decompose the problem, refine their understanding, and iteratively generate cleaner images. The CLIP-guided decoder ensures that denoised outputs align with high-level image semantics, enabling deeper image understanding for reflection, thereby improving the quality of the denoised images. Experimental results demonstrate that our CoT-SD outperforms other state-of-the-art (SOTA) dataset-free denoisers. The code will be made publicly available.

Citation

@inproceedings{Jiang_2025_BMVC,
author    = {Yanlin Jiang and Yuchen Liu and Mingren Liu},
title     = {CoT-SD: Chain-of-Thought Semantic Denoising},
booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
publisher = {BMVA},
year      = {2025},
url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_588/paper.pdf}
}


Copyright © 2025 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection