Related Experiment Video
Updated: Jan 16, 2026
06:54
Photorealistic Learned Landscapes for Augmented Reality
Published on: June 27, 2025
684
Image restoration driven by dual-scale prior
Weimin Yuan1, Cai Meng1, Xiangzhi Bai1
1Image Processing Center, Beihang University, Beijing, 100191, China.
Summary
This study introduces a novel dual-scale prior (DSP) model for image restoration (IR). DSPIR, an effective IR method, enhances image denoising and inpainting by combining non-learning and learning-based priors.
Area of Science:
- Computer Vision
- Image Processing
- Machine Learning
Background:
- Advanced imaging increases demand for high-quality images.
- Image degradation from noise and data loss hinders quality.
- Existing image restoration methods have limitations in flexibility and generalization.
Purpose of the Study:
- To introduce a novel dual-scale prior (DSP) model for image restoration.
- To integrate the strengths of non-learning and learning-based priors.
- To develop an effective image restoration method (DSPIR) using the DSP model.
Main Methods:
- Developed a dual-scale prior (DSP) model combining group-scale physical prior (NSS) and image-scale deep denoising prior.
- Incorporated DSP into the maximum a posteriori (MAP) principle to create DSPIR.
- Solved DSPIR using alternating minimization and alternating direction method of multipliers.
Main Results:
- The DSP model effectively preserves edges and removes noise.
- DSPIR demonstrates robustness across various degradation types.
- Extensive evaluations show DSPIR outperforms state-of-the-art methods in denoising and inpainting.
Conclusions:
- The dual-scale prior model offers a powerful approach to image restoration.
- DSPIR achieves superior performance in image denoising and inpainting.
- The proposed method shows significant improvements over existing techniques.
