Real-World Blur Dataset and Multiscale Algorithmic for Enhanced Image Deblurring
Xun Li, Dongming Zhou · 2025
In this paper, we present a novel real-world blur dataset, named MTRBlur, comprising both out-of-focus and motion-blurred images captured from diverse static and dynamic scenes, respectively. This dataset addresses the limitations of existing synthetic and real blur datasets by offering a more comprehensive and realistic representation of real-world blurred images. To effectively restore these blurred images, we propose a multi-input and multi-scale encoder-decoder network, MMTNet, built upon the Transformer structure. Our approach demonstrates the necessity of utilizing real blur datasets for achieving authentic deblurring and the efficacy of our network in restoring such images. Extensive experiments across low-level and high-level computer vision tasks show that our dataset and method significantly enhance the quality of deblurring for actual blurred images. Our contributions include the introduction of the MTRBlur dataset, the proposal of the MMTNet network, and the evaluation of our low-level computer vision task using target detection methods in high-level tasks.