DegAE: A New Pretraining Paradigm for Low-Level Vision
Yihao Liu, Jingwen He, Jinjin Gu, Xiangtao Kong, Yu Qiao, Chao Dong
Abstract
Self-supervised pretraining has achieved remarkable success in high-level vision, but its application in low-level vision remains ambiguous and not well-established. What is the primitive intention of pretraining? What is the core problem of pretraining in low-level vision? In this paper, we aim to answer these essential questions and establish a new pretraining scheme for low-level vision. Specifically, we examine previous pretraining methods in both high-level and low-level vision, and categorize current low-level vision tasks into two groups based on the difficulty of data acquisition: low-cost and high-cost tasks. Existing literature has mainly focused on pretraining for low-cost tasks, where the observed performance improvement is often limited. However, we argue that pretraining is more significant for highcost tasks, where data acquisition is more challenging. To learn a general low-level vision representation that can improve the performance of various tasks, we propose a new pretraining paradigm called degradation autoencoder (De-gAE). DegAE follows the philosophy of designing pretext task for self-supervised pretraining and is elaborately tailored to low-level vision. With DegAE pretraining, SwinIR achieves a 6.88dB performance gain on image dehaze task, while Uformer obtains 3.22dB and 0.54dB improvement on dehaze and derain tasks, respectively.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
8 top-tier papers have cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
- Editprint: General Digital Image Forensics via Editing Fingerprint with Self-Augmentation Training cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- Uformer: A General U-Shaped Transformer for Image Restoration cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- Uformer: A General U-Shaped Transformer for Image Restoration cites Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
- Time Without Time: Pseudo-Temporal Representation for Space-Time Super-Resolution cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- Restormer: Efficient Transformer for High-Resolution Image Restoration cites Uformer: A General U-Shaped Transformer for Image Restoration
- Restormer: Efficient Transformer for High-Resolution Image Restoration cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- Restormer: Efficient Transformer for High-Resolution Image Restoration cites Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
- FoundIR: Unleashing Million-Scale Training Data to Advance Foundation Models for Image Restoration cites Restormer: Efficient Transformer for High-Resolution Image Restoration
- Swin Transformer: Hierarchical Vision Transformer using Shifted Windows cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- UniRestorer: Universal Image Restoration via Adaptively Estimating Image Degradation at Proper Granularity cites Restormer: Efficient Transformer for High-Resolution Image Restoration
- UniRestorer: Universal Image Restoration via Adaptively Estimating Image Degradation at Proper Granularity cites FoundIR: Unleashing Million-Scale Training Data to Advance Foundation Models for Image Restoration
- UniRestorer: Universal Image Restoration via Adaptively Estimating Image Degradation at Proper Granularity cites Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
Cited by top-tier papers8
- FoundIR: Unleashing Million-Scale Training Data to Advance Foundation Models for Image RestorationHao Li, Xiang Chen, Jiangxin Dong, Jinhui Tang et al.ICCV 2025 · 15 citations
- UniRestorer: Universal Image Restoration via Adaptively Estimating Image Degradation at Proper GranularityJingbo Lin, Zhilu Zhang, Wenbo Li, Renjing Pei et al.ICLR 2026 · 8 citations
- Depth Prompting for Sensor-Agnostic Depth EstimationJin-Hwi Park, Chanhwi Jeong, Junoh Lee, Hae-Gon JeonCVPR 2024 · 4 citations
- Editprint: General Digital Image Forensics via Editing Fingerprint with Self-Augmentation TrainingHaiwei Wu, Kemou Li, Yuanman Li, Jiantao ZhouCVPR 2026
- Time Without Time: Pseudo-Temporal Representation for Space-Time Super-ResolutionHee Min Choi, Hyoa Kang, Suji Kim, Dokwan Oh et al.CVPR 2026
Builds on19
- Swin Transformer: Hierarchical Vision Transformer using Shifted WindowsZe Liu, Yutong Lin, Yue Cao, Han Hu et al.ICCV 2021 · 31,683 citations
- An Image is Worth 16x16 Words: Transformers for Image Recognition at ScaleAlexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn et al.ICLR 2021 · 21,477 citations
- Restormer: Efficient Transformer for High-Resolution Image RestorationSyed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat et al.CVPR 2022 · 3,348 citations
- Uformer: A General U-Shaped Transformer for Image RestorationZhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou et al.CVPR 2022 · 1,970 citations
- FFA-Net: Feature Fusion Attention Network for Single Image DehazingXu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie et al.AAAI 2020 · 1,828 citations
Related papers
- Image-free Pre-training for Low-Level VisionSiyang Wang, Jinghao Zhang, Jie Huang, Feng ZhaoACM MM 2024 · 1 citation
- Universal Image Restoration Pre-training via Degradation ClassificationJiakui Hu, Lujia Jin, Zhengjian Yao, Yanye LuICLR 2025
- Learning Semantic Degradation-Aware Guidance for Recognition-Driven Unsupervised Low-Light Image EnhancementNaishan Zheng, Jie Huang, Man Zhou, Zizheng Yang et al.AAAI 2023 · 19 citations
- Text-DIAE: A Self-Supervised Degradation Invariant Autoencoder for Text Recognition and Document EnhancementMohamed Ali Souibgui, Sanket Biswas, Andrés Mafla, Ali Furkan Biten et al.AAAI 2023 · 31 citations
- MixMAE: Mixed and Masked Autoencoder for Efficient Pretraining of Hierarchical Vision TransformersJihao Liu, Xin Huang, Jinliang Zheng, Yu Liu et al.CVPR 2023
