GenDeg: Diffusion-Based Degradation Synthesis for Generalizable All-in-One Image Restoration

2024-11-26

arXiv.org

AI Summary

To address the poor generalization of All-in-One Image Restoration (AIOR) models under out-of-distribution degradations and the scarcity of real paired degradation data, this paper proposes GenDeg, a controllable degradation synthesis framework built upon latent diffusion models. GenDeg introduces a novel dual-conditioning mechanism that jointly controls degradation type and intensity, supporting six degradation categories—haze, rain, snow, motion blur, low-light, and raindrops—with continuous intensity adjustment. We further construct GenDS, the first large-scale synthetic-real hybrid dataset comprising over 750k samples. AIOR models trained on GenDS achieve state-of-the-art generalization across multiple benchmarks: average PSNR improves by 1.82 dB across unseen degradation types, and robustness on entirely unseen degradations increases by over 40%, demonstrating the effective transferability of controllably synthesized degradations to real-world restoration tasks.

Technology Category

• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications

Abstract

Deep learning-based models for All-In-One Image Restoration (AIOR) have achieved significant advancements in recent years. However, their practical applicability is limited by poor generalization to samples outside the training distribution. This limitation arises primarily from insufficient diversity in degradation variations and scenes within existing datasets, resulting in inadequate representations of real-world scenarios. Additionally, capturing large-scale real-world paired data for degradations such as haze, low-light, and raindrops is often cumbersome and sometimes infeasible. In this paper, we leverage the generative capabilities of latent diffusion models to synthesize high-quality degraded images from their clean counterparts. Specifically, we introduce GenDeg, a degradation and intensity-aware conditional diffusion model capable of producing diverse degradation patterns on clean images. Using GenDeg, we synthesize over 550k samples across six degradation types: haze, rain, snow, motion blur, low-light, and raindrops. These generated samples are integrated with existing datasets to form the GenDS dataset, comprising over 750k samples. Our experiments reveal that image restoration models trained on the GenDS dataset exhibit significant improvements in out-of-distribution performance compared to those trained solely on existing datasets. Furthermore, we provide comprehensive analyses on implications of diffusion model-based synthetic degradations for AIOR.

Problem

• Limited generalization of AIOR models to unseen data
• Insufficient diversity in existing degradation datasets
• Challenges in capturing real-world paired degradation data

Innovation

• Leverages latent diffusion models for degradation synthesis
• Introduces GenDeg, a degradation-aware conditional diffusion model
• Synthesizes 550k samples across six degradation types