A Comparative Study for Single Image Blind Deblurring

By Wei-Sheng Lai, Jia-Bin Huang, Zhe Hu, Narendra Ahuja, Ming-Hsuan Yang. 2016

418 citations recorded by the source corpus.

Numerous single image blind deblurring algorithms have been proposed to restore latent sharp images under camera motion. However, these algorithms are mainly evaluated using either synthetic datasets or few selected real blurred images. It is thus unclear how these algorithms would perform on images acquired "in the wild" and how we could gauge the progress in the field. In this paper, we aim to bridge this gap. We present the first comprehensive perceptual study and analysis of single image blind deblurring using real-world blurred images. First, we collect a dataset of real blurred images and a dataset of synthetically blurred images. Using these datasets, we conduct a large-scale user study to quantify the performance of several representative state-of-the-art blind deblurring algorithms. Second, we systematically analyze subject preferences, including the level of agreement, significance tests of score differences, and rationales for preferring one method over another. Third, we study the correlation between human subjective scores and several full-reference and noreference image quality metrics. Our evaluation and analysis indicate the performance gap between synthetically blurred images and real blurred image and sheds light on future research in single image blind deblurring.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

The Perception-Distortion Tradeoff (2017, background): “blind deblurring with large blurs [20, 25]), the perceptual metric should be human-opinionscores or the loss of a discriminator trained to distinguish the algorithms’ outputs from natural images.”
Unsupervised Domain-Specific Deblurring via Disentangled Representations (2019, result): “[16], and the results are shown in Fig.”
DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better (2019, methodology): “The Lai dataset [22] has real-world blurry images of different qualities and resolutions collected in various types of scenes.”
Deep Image Harmonization (2017, methodology): “After obtaining all the pairwise results, we use the Bradley-Terry model (B-T model) [2, 12] to calculate the global ranking score for each method.”
Deep Image Deblurring: A Survey (2022, methodology): “[61] X X 108 300 Both Single image Not divided GoPro[86] X X 3,214 3,214 Non-uniform Single image 2,103/1,111 HIDE [116] X × 8,422 8,422 Non-uniform Single image 6,397/2,025 Blur-DVS [46] X X 2,178 2,918 Non-uniform Single image 1,782/396”
Benchmarking Single-Image Dehazing and Beyond (2017, background): “As the same for peer benchmarks [44], [45], the subjective”
Deblurring via Stochastic Refinement (2021, methodology): “Traditional deblur-ring methods formulate the problem as one of blind deconvolution [9,10,16,19,29,37,39,60,74,78].”
Real-World Blur Dataset for Learning and Benchmarking Deblurring Algorithms (2020, background): “[23] introduced a dataset of 100 real blurred images for benchmarking deblurring methods.”
Fast and Accurate Image Super-Resolution with Deep Laplacian Pyramid Networks (2017, methodology): “Finally, we collect the results from 71 participants. To obtain a global score for each method, we ﬁt the results of paired comparisons to the Bradley-Terry (BT) model [56]. We refer readers to [54], [55] for details of the BT model. We normalize the BT scores to zero means and show the results in T”
Dynamic Scene Deblurring With Parameter Selective Sharing and Nested Skip Connections (2019, methodology): “The first image is from the synthetic dataset [15].”
Neural Blind Deconvolution Using Deep Priors (2019, methodology): “Experiments are conducted on two widely used benchmarks [18,19] as well as real-world blurry images to evaluate our SelfDeblur.”
Human-Aware Motion Deblurring (2019, background): “Most representative datasets [23, 39] were constructed on the basis of a simplified camera-driven assumption that camera disturbance dominates the cause of blur [45].”
Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring (2016, background): “Lai et al. [20] generated synthetic dataset by convolving nonuniform blur kernels and imposing several common degradations.”