RefDeblur: Blind Motion Deblurring with Self-Generated Reference Image

Abstract

The challenge of blind motion deblurring is often tackled via two distinct paradigms: kernel-based and kernel-free methods. Each deblurring method provides inherent strengths. Kernel-based methods facilitate generating texture-detailed sharp images by closely aligning with the blurring process. In contrast, kernel-free methods are more effective in handling complex blur patterns. Building upon these complementary benefits, we propose a hybrid framework that decomposes a non-uniform deblurring task into two simpler tasks: a uniform kernel estimation, managed by our kernel-based method, and error prediction, handled by our kernel-free method. Our kernel-based method serves to generate a reference image with realistic texture details while our kernel-free model refines the reference image by correcting residual errors with preserving texture details. To efficiently build our kernel-based model, we consider the logarithmic fourier space that facilitates estimating a blur kernel easier by simplifying the relationship between blur and sharp samples. Furthermore, the regime under using a texture-detailed reference image allows for reducing the size of our kernel-free model without compromising performance. As a result, the proposed method achieves remarkable performance on several datasets such as RealBlur, RSBlur and GoPro, and comparable performance to state-of-the-art methods with a 75% reduction in computational costs.