A Dynamic Kernel Prior Model for Unsupervised Blind Image Super-Resolution

Zhixiong Yang,Jingyuan Xia,Shengxi Li,Xinghua Huang,Shuanghui Zhang,Zhen Liu,Yaowen Fu,Yongxiang Liu
2024-04-26
Abstract:Deep learning-based methods have achieved significant successes on solving the blind super-resolution (BSR) problem. However, most of them request supervised pre-training on labelled datasets. This paper proposes an unsupervised kernel estimation model, named dynamic kernel prior (DKP), to realize an unsupervised and pre-training-free learning-based algorithm for solving the BSR problem. DKP can adaptively learn dynamic kernel priors to realize real-time kernel estimation, and thereby enables superior HR image restoration performances. This is achieved by a Markov chain Monte Carlo sampling process on random kernel distributions. The learned kernel prior is then assigned to optimize a blur kernel estimation network, which entails a network-based Langevin dynamic optimization strategy. These two techniques ensure the accuracy of the kernel estimation. DKP can be easily used to replace the kernel estimation models in the existing methods, such as Double-DIP and FKP-DIP, or be added to the off-the-shelf image restoration model, such as diffusion model. In this paper, we incorporate our DKP model with DIP and diffusion model, referring to DIP-DKP and Diff-DKP, for validations. Extensive simulations on Gaussian and motion kernel scenarios demonstrate that the proposed DKP model can significantly improve the kernel estimation with comparable runtime and memory usage, leading to state-of-the-art BSR results. The code is available at
Image and Video Processing
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the Blind Super - Resolution (BSR) problem. Specifically, BSR aims to reconstruct high - resolution (HR) images from low - resolution (LR) images, and these LR images may be affected by unknown blur kernels. This problem is highly non - convex and ill - posed, and thus very challenging. ### Main Problems 1. **Requirement for Supervised Learning**: Most existing BSR methods rely on paired LR - HR datasets for supervised pre - training, which is often difficult to obtain sufficient labeled data in practical applications, especially in fields such as high - speed targets (e.g., satellites, airplanes) and medical images (e.g., a beating heart). 2. **Flexibility and Generalization Ability**: When dealing with different types of blur kernels, existing methods usually need to be retrained or fine - tuned for specific scenarios, resulting in limited flexibility and generalization ability. ### Solutions To solve the above problems, this paper proposes an unsupervised Dynamic Kernel Prior (DKP) model, which can achieve the BSR task without relying on paired LR - HR datasets. The core idea of the DKP model is to adaptively learn the dynamic kernel prior through the Random Kernel Sampling (RKS) module and the Prior Kernel Estimation (PKE) module, thereby achieving real - time blur kernel estimation. ### Model Features - **Unsupervised Learning**: The DKP model does not require pre - training and can be directly applied to existing image inpainting models, such as Deep Image Prior (DIP) and Diffusion Model, thus avoiding the need for expensive labeled data. - **Efficiency**: Through the Markov chain Monte Carlo (MCMC) sampling strategy and Network - based Langevin Dynamics (NLD), the DKP model can efficiently generate reasonable kernel priors and optimize blur kernel estimation. - **Flexibility**: The DKP model can be easily integrated into existing image inpainting models, is suitable for different types of blur kernels, and has good generalization ability. ### Experimental Verification To verify the effectiveness of the DKP model, the author conducted comparative experiments with existing BSR methods in two scenarios, including Gaussian blur kernels and Motion blur kernels. The experimental results show that without pre - training, the DKP model can significantly improve the accuracy of blur kernel estimation and reach the state - of - the - art level in image reconstruction quality. ### Summary The DKP model proposed in this paper solves the dependence on supervised learning in the BSR task, improves the flexibility and generalization ability of the model, and provides a brand - new solution for the BSR problem.