Single-image reflection removal via self-supervised diffusion models

2024-12-26

Journal of Supercomputing

AI Summary

To address the lack of paired ground-truth data for single-image glass reflection removal, this paper proposes the first diffusion-based framework incorporating self-supervision. Built upon denoising diffusion probabilistic models (DDPMs), the method employs a dual-branch latent-space encoder to disentangle reflection and transmission components. It further introduces a contrastive self-supervised loss and physically inspired reflection priors to eliminate reliance on manual annotations and mitigate synthetic artifacts inherent in data synthesis. Evaluated on real-world benchmarks—including Real20 and UG2+—the approach achieves state-of-the-art performance, improving PSNR by over 2.1 dB. It also demonstrates superior visual fidelity and structural consistency compared to existing supervised and unsupervised methods, exhibiting strong generalization to diverse real-world scenarios.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Unsupervised & Self-Supervised Learning
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Security and Privacy: Data transparency and provenance
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Problem

• Reflection Removal
• Transparent Objects
• Training Data

Innovation

• Cyclical Consistency
• Denoising Diffusion Probabilistic Models (DDPM)
• Reflection Removal