Discriminative Blur Detection Features Jianping Shi , Li Xu, Jiaya Jia CVPR, 2014
Image Blurriness • Commonly occurred photo degradation • Visual effect by photographers • Important to analyze 10/20/2014 2
Image Blur Detection • Problem Definition – Finding blur pixels for a given input image • Potential application – Image segmentation, – Object detection, – Image quality assessment, – … 10/20/2014 3
Previous Works • De-convolution based approach = * = * 10/20/2014 4
Previous Works • Deconvolution based approach • Explicit blur detection – Gradient based: Levin 2007, Liu 2008 – Frequency based: Liu 2008, Chakrabarti 2010 – Matting based: Dai 2008, Dai 2009 10/20/2014 5
Our Blur Features • Image Gradient Distribution • Spectra in Frequency Domain • Local Filters 10/20/2014 6
Our Blur Features • Image Gradient Distribution Properties: 1. Peakedness 2. Heavy-tailedness 10/20/2014 7
Our Blur Features • Image Gradient Distribution – Previous L 0.8 norm 10/20/2014 8
Our Blur Features • Image Gradient Distribution – Peakedness Measure • Definition: Kurtosis • Proposition 1: Given the local blur model and kurtosis measure, it is guaranteed to have 𝐿 𝐶 𝑦 ≤ 𝐿(𝐽 𝑦 ) and 𝐿 𝐶 𝑧 ≤ 𝐿(𝐽 𝑧 ) . 10/20/2014 9
Our Blur Features • Image Gradient Distribution – Peakedness Measure • Kurtosis feature 10/20/2014 10
Our Blur Features • Image Gradient Distribution – Peakedness Measure • Kurtosis feature 10/20/2014 11
Our Blur Features • Image Gradient Distribution – Heavy-Tailedness Measure • Fit a Gaussian mixture model with two components to gradient magnitude 𝛼𝐶 𝛼𝐶 ∼ 𝜌 1 𝐻(𝛼𝐶|𝜈 1 , 𝜏 1 ) +𝜌 2 𝐻(𝛼𝐶|𝜈 2 , 𝜏 2 ) 10/20/2014 12
Our Blur Features • Image Gradient Distribution – Heavy-Tailedness Measure • Fit a Gaussian mixture model with two components to gradient magnitude 𝛼𝐶 • The heavy-tailedness feature is the larger variance 10/20/2014 13
Our Blur Features • Spectra in Frequency Domain – Average power spectrum 𝐾(𝜕) • It intuitively represents the strength of change • Blur attenuates high frequency components. • The power spectrum fall of faster for blur region 10/20/2014 14
Our Blur Features • Spectra in Frequency Domain – Average power spectrum 𝐾(𝜕) – Proposition 2: Given a natural image patch 𝑦 and its Gaussian or box blurred version 𝑧 by PSF 𝑙 , the fall-off speed of the average power spectrum on 𝑧 is several orders faster than that of 𝑦 . It is expressed as 10/20/2014 15
Our Blur Features • Spectra in Frequency Domain – Spectrum feature: – Proposition 3: Given a natural image patch 𝑦 , which is blurred by a PSF to form patch 𝑧 , the cumulated average power spectrum for the blurred patch is smaller than that for the sharp patch, i.e., 10/20/2014 16
Our Blur Features • Spectra in Frequency Domain 10/20/2014 17
Our Blur Features • Local Filters – Data driven approach based on our labeled dataset – Linear discriminative analysis – Learned feature 10/20/2014 18
Our Blur Features • Local Filters 10/20/2014 19
Visualizing Features in 3D 10/20/2014 20
Feature Covariance 10/20/2014 21
Multi-Scale Perception • Scale ambiguous 10/20/2014 22
Multi-Scale Perception • Fuse information in different scales • The input for each layer is the posterior of a naive Bayesian classifier for the set of features. 10/20/2014 23
Multi-Scale Perception • Fuse information in different scales 10/20/2014 24
Blur Detection Dataset 10/20/2014 25
Experiments • Visual comparisons • Quantitative comparisons • Applications enabled by blur detection 10/20/2014 26
Visual Comparison 10/20/2014 27
Quantitative Comparison 10/20/2014 28
Applications Based on Blur Detection • Blur Segmentation and Deblurring 10/20/2014 29
Blur Magnification 10/20/2014 30
Failure Case 10/20/2014 31
Conclusion • We have proposed several effective local blur features • We have integrated the local blur features into a multi-scale inference framework • Extensive experiments verified our method 10/20/2014 32
Recommend
More recommend
