University of Groningen Institute of Mathematics and Computing Science Universitá degli Studi di Roma Tre Dipartimento di Elettronica Applicata Well Posed non-Iterative Edge and Corner Preserving Smoothing For Artistic Imaging Giuseppe Pápari, Nicolai Petkov, Patrizio Campisi
Photographical image
Output of the proposed operator
Smoothing out texture while preserving edges Input imageGaussian smoothingProposed operator Co ntents Kuwahara Filter and Generalizations Limitations Proposed Operator Results and Comparison Discussion
Kuwahara Filter and Generalizations Four local averages: Four local standard deviations: Kuwahara output For each pixel, value of m i that corresponds to the minimum standard deviation Generic pixel of the input image
Kuwahara Filter and Generalizations Edge Only the most homogeneous region is taken into account. No smoothing across the edge (x,y) = 1 Central pixel on the white side of the edge (x,y) = 0 Central pixel on the black side of the edge
Kuwahara Filter and Generalizations Local averaging Smoothing Flipping due to Minimum Variance Criterion Edge Preserving
Kuwahara Filter and Generalizations An example Input image Kuwahara output Artifacts on texture
Kuwahara Filter and Generalizations Generalizations Number and shape of the sub-regions »Pentagons, hexagons, circles »Overlapping Weighted local averages (reducing the Gibbs phenomenon) »Gaussian-Kuwahara New class of filters (Value and criterion filter structure) »N local averages and local standard deviations (computed as convolutions) »Criterion: minimum standard deviation Connections with the PDEs theory and morphological analysis
Kuwahara Filter and Generalizations LimitationsLimitations Proposed Operator Results and Comparison Discussion
Limitations Artifacts (partially eliminable with weighted averages) Not mathematically well defined Equal standard deviations s i Devastating instability in presence of noise
Limitations Simple one-dimensional example Input signal I(t) I(t) = kt Local averages I t tTtTt+T Negative offset 1D Kuwahara filtering Two sub-windows w 1 and w 2 t w2w2 t*t* I(t)I(t) w1w1
Limitations Simple one-dimensional example Input signal I(t) I(t) = kt Local averages I t tTtTt+T Negative offset Positive offset 1D Kuwahara filtering Two sub-windows w 1 and w 2 t w2w2 t*t* I(t)I(t) w1w1
I(t) = kt Local standard deviations Equal standard deviations I t tTtTt+TLimitations Simple one-dimensional example Local averages m 1 (t), m 2 (t) Input signal I(t) Local std. dev. s 1 (t), s 2 (t) 1D Kuwahara filtering Two sub-windows w 1 and w 2 t w2w2 t*t* I(t)I(t) w1w1
Limitations Input imageKuwahara filtering Proposed approach Synthetic two-dimensional example
KuwaharaLimitations Natural image example Input image Gauss-Kuwahara Shadowed area Depleted edge Our approach
Limitations Ill-posedness of the minimum variance criterion. Devastating effects in presence of noisy shadowed areas. We propose Different weighting windows w i A different selection criterion instead of the minimum standard deviation
Kuwahara Filter and Generalizations Limitations Proposed OperatorProposed Operator Results and Comparison Discussion
Proposed Operator Gaussian mask divided in N sectors N weighting windows N local averages and local standard deviations computed as convolutions Weighting windows
Proposed Operator Selection criterion q Only the minimum s i survives Criterion and value Output: » Weighted average of m i » Weights equal to proportional to (s i ) q (q is a parameter) Normalization High variance small coefficient (s i ) q No undetermination in case of equal standard deviations!
Proposed Operator Particular cases Equal standard deviations: s 1 = s 2 = … = s N Gaussian smoothing One standard deviation is equal to zero: s k = 0 Several values of s i are equal to zero = Arithmetic mean of the corresponding values of m i.
Proposed Operator Edge Half of the sectors have s i = 0. The other ones are not considered An example Edgeless areas: All std. dev. similar Gaussian smoothing (no Gibbs phenomenon) Corner preservation Automatic selection of the prominent sectors
Proposed Operator Color images 3 sets of local averages and local standard deviations, one for each color component with Same combination rule Not equivalent to apply the operator to each color component separately
Proposed Operator Independence on the color space Input imageRGBYC r C b L*a*b*
Proposed Operator Why independence? Linear transform. independent Nonlinear transf. almost independent for homogeneous regions Local averages
Proposed Operator Why independence? Linear transform. independent Nonlinear transf. almost independent for homogeneous regions Local averages Low for homogeneous regions. The degree of homogeneity of a region does not depend on the color space. Local standard deviations
Kuwahara Filter and Generalizations Limitations Proposed Operator Results and ComparisonResults and Comparison Discussion
Results and comparison Existing algorithm for comparison Kuwahara filter and generalizations Bilateral filtering Morphological filters Median filters
Input image
Proposed approach
Gauss-Kuwahara filter
Input image (blurred)
Proposed approach (deblurred)
Bilateral filtering (not deblurred)
Input image
Proposed approach
Morphological closing (Struct. elem.: Disk of radius 5px)
Morphological area open-closing
Input image
Morphological area open-closing
Proposed approach
Input image
Proposed approach
Kuwahara Filter
Morphological area open-closing
Input image
Proposed approach
Bilateral Filtering
Input image
proposed aproach
5 5 median filter
Results and comparison Larger set of results and Matlab implementation available at Graphical interface
Kuwahara Filter and Generalizations Limitations Proposed Operator Results and Comparison DiscussionDiscussion
Discussion Edge/corner preserving smoothing Undetermination for equal standard deviation » Instability in presence of noise » Discontinuities in presence of shadowed areas Criterion and value filter structure » Local averaging Smoothing » Minimum variance criterion Edge preserving
Discussion Proposed approach » Different windows » Different criterion Mathematically well defined operator Adaptive choice of the most appropriate sub-regions. Our approachGauss-KuwaharaKuwahara
Discussion Limitations » Lines are thinned » Small objects are not preserved
References G. Papari, N. Petkov, P. Campisi Artistic Edge and Corned Preserving Smoothing To appear on IEEE Transactions on Image Processing, 2007 G. Papari, N. Petkov, P. Campisi Edge and Corned Preserving Smoothing for Artistic Imaging Proceedings SPIE 2007 Image Processing: Algorithms and Systems, San Jose, CA