Related Experiment Video
Updated: Jan 10, 2026
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
999
Unified-Removal: A Semi-Supervised Framework for Simultaneously Addressing Multiple Degradations in Real-World Images
1State Key Laboratory of Networking and Switching Technology, BUPT, No. 10 Xitucheng Road, Haidian District, Beijing 100876, China.
Journal of Imaging
|November 26, 2025
Summary
Uni-Removal is a novel framework for image restoration domain adaptation. It uses dual-phase learning to bridge the gap between synthetic and real-world data, improving performance on tasks like dehazing and deraining.
Area of Science:
- Computer Vision
- Machine Learning
- Image Processing
Background:
- Domain adaptation is crucial for image restoration, as models trained on synthetic data perform poorly on real-world images.
- Existing methods struggle with the domain discrepancy between training and testing environments.
Purpose of the Study:
- To introduce Uni-Removal, a two-stage framework for unified image restoration domain adaptation.
- To bridge the domain gap using dual-phase representation learning.
Main Methods:
- Stage 1: Multi-teacher knowledge distillation with Instance-Grained Contrastive Learning (IGCL) for representation consistency.
- Stage 2: Cluster-Grained Contrastive Learning (CGCL) for output distribution calibration and alignment with real-world image characteristics without paired supervision.
Main Results:
- Uni-Removal outperforms state-of-the-art methods in real-world dehazing, deraining, and deblurring.
- Achieved competitive denoising performance on the SIDD benchmark.
- Significantly improved downstream object detection by 4.36 mAP.
Conclusions:
- Uni-Removal effectively addresses domain adaptation challenges in unified image restoration.
- The framework demonstrates strong generalization and practical utility for computer vision systems.
