Deep Image Quality Assessment Driven Single Image Deblurring

Ang Li,Jichun Li,Qing Lin,Chenxi Ma,Bo Yan

DOI: https://doi.org/10.1109/icme46284.2020.9102899

2020-01-01

Abstract:Motion deblurring is a challenging task in computer vision, which aims to recover the sharp image from the blurry one. Recently, deep learning based methods have made a significant improvement in the metric of PSNR due to the optimazation of the Mean Squared Error (MSE) loss function between deblurring results and sharp images. However, PSNR prefers smooth images and fails to evaluate the sharpness of deblurred images. To solve this problem, we firstly propose the deep learning based multi-scale non-reference quality assessment network (Deep DEBLUR-IQA) for assessing deblurred results. Moreover, we propose an efficient deblurring network, which is over 50 times faster than SOTA multi-scale networks. Finally, the combination of our Deep DEBLUR-IQA network and novel single image deblurring network can significantly increase subjective quality while maintaining satisfactory PSNR. Experiments show that the proposed method outperforms state-of-the-art methods, both qualitatively and quantitatively.