A Comparative Study for Single Image Blind Deblurring Wei-Sheng Lai UC Merced Jia-Bin Huang UIUC Zhe Hu UC Merced Narendra Ahuja UIUC Ming-Hsuan Yang UC Merced Hi everyone. My name is Wei-Sheng Lai. Today I am going to talk about a comparative study for single image blind deblurring.

Single Image Blind Deblurring Algorithms: Datasets: Real images: Fergus et al. 2006 Shan et al. 2008 Cho & Lee 2009 Krishnan et al. 2011 Whyte et al. 2011 Hirsch et al. 2011 Xu et al. 2013 Zhong et al. 2013 Sun et al. 2013 Michaeli et al. 2014 Pan et al. 2014 Perrone et al. 2014 Levin et al. 2009 However, real-world blurred images are much more complicate. For example, the scene depth variation, non-linear camera response functions, saturation and compression artifacts are common factors in real images but not considered in previous datasets. Kohler et al. 2012 Sun et al. 2013 Depth variation Camera response functions Saturation Compression artifacts

Our Goal Performance evaluation on real-world blurred images a dataset of real images large scale comparative study In this work, we provide the first large scale performance evaluation of blind motion deblurring algorithms on real-world blurred images. We first collect a dataset of 100 real blurred images from different sources.// such as Google, Flickr, and photos captured by ourselves. Then, we conduct a user study and comparative analysis to understand the relative strengths of state-of-the-art algorithms.

User-Study Evaluate on 14 methods, 100 images 14 2 =910 comparisons per image collect about 100k paired comparisons from 2000 subjects In our user study, we show two images side by side, and ask users to select which one looks better. We evaluate totally 14 methods on 100 images. We collect user votes from around 2000 human subjects (on Amazon Mechanical Turk).

From Paired Comparisons to Full Ranking Fit votes to the Bradley-Terry Model (B-T Model) 𝑀 𝑖𝑗 =#times that users choose method i over method j 𝑆 𝑖 = the B-T score of method i Cumulative Frequency 𝑀 12 101 𝑀 13 80 ⋯ 𝑀 45 25 B-T Model To obtain the full ranking of each algorithm, we need convert the pairwise comparisons to real-value scores. We use the Bradley-Terry model to convert the user votes to the B-T scores, and plot the cumulative frequency of the B-T scores for each method. //In this figure, the curve on the right has better performance. 𝑆 1 3.14 𝑆 2 1.59 ⋯ 𝑆 5 2.65

Comparing Real and Synthetic Datasets We also conduct the baseline user-study on synthetic datasets.// for both uniform and non-uniform blurred images. We show the scatter plots of the BT scores between real and synthetic datasets to understand their performance difference. //Third, we show the scatter plot the BT scores on real and synthetic datasets. //In these figures, the methods on the lower-right triangle have better performance on the synthetic dataset. But their performance is not as competitive in the real dataset. //As a result, we say that the performance of those methods on the lower-right are over-estimated on the synthetic dataset.

Comparing Image Quality Metrics We also compute the correlation between user subject scores and several image quality metrics, including full-reference and no-reference metrics. Full-reference metrics No-reference metrics

Observations 𝛻𝑥 1 𝛻𝑥 2 , 𝑥 0 > Image priors: sparse priors are more robust than edge prediction methods Image formations: Datasets: performance on synthetic datasets does not reflect the performance on real images Quality metrics: IFC/VIF > PSNR/SSIM; none of no-reference metrics are applicable 𝛻𝑥 1 𝛻𝑥 2 , 𝑥 0 > Finally, if you need a quality metric to evaluate the performance on synthetic images, IFC and VIF are better than PSNR and SSIM. For real blurred images, none of existing no-reference metrics are suitable to evaluate the quality of deblurred results.

Conclusions First large scale comparative study on real-world images quantitatively evaluate the progress of the field identify potential research directions Code, datasets and results are available: bit.ly/deblur_study Poster #22 In conclusion, we provide the first comparative study of blind motion deblurring on real-world blurred images. We evaluate the progress of this field, and identify potential directions for future research. Our code and datasets are available on our project website. If you are interested in more details, please come to our poster at section 22.