Image Smoothing For Structure Extraction
   Linjia Chang, lchang10@illinois.edu Mentor: Jia-Bin Huang, jbhuang1@illinois.edu
Goal                                                                            Applications
                        ·Achieve Edge-aware image                               · Detail enhancement                   · Re-coloring
                        smoothing while being able to                           · Image composition                    · Stylization
                        distinguish texture/structure from                      · Object recognition                   · Video segmentation
                        general natural images                                  · Image denoise                        · Structure extraction

                                                                                Methods
                                                                                · Optimization with total variation regularization
                                                                                · - Robust loss function for texture removal
                                                                                · - Iterative reweighted L1 for sparsity[3]


Previous Related Work
· Gaussian Blur                         · L0 Gradient Minimization               · Domain Transformation[1]           · Structure Texture Extraction[2]

                 Pixel = weighted
                 average of
                 its neighbors


                                                                                                                        A major edge in a
                                                                                                                        local window
                                                                                                                        contributes more
                                    Enhances high-contrast edges by              Preserves the original distance:       similar-direction
                                    confining numbers of non-zero gradients      isometric transform                    gradients


Algorithm
                                                              · Idea: Image smoothing as a global optimization problem
                                     Huber Loss Function               Minimize S* = argmin ∑ λ||Sp – Ip|| + w||▽Sp||
                                                                                                 s
                                                                                Data Term                                           Regularization Term
                                                                                            Similar as previous
                                                                                            works but using Huber
                                                                                                                    Iteratively Reweighted L1
                                                    Solution Algorithm[4]
                                                                                            LF                      (Encourage Sparsity)
                                      1. Set dummy variables u and v                                  First solve the part without the
                                      S* = argmin ∑λ||Sp – Ip|| + w(|u|+|v|)+ β|(▽Spx-u)²+ (▽Spy-v)²| weight = λ||▽Sp||
                                              s
                                                                                                      And then introduce weight w
                                      2. Fix u, v and solve for S (convex)
                                      3. Fix S and solve for u, v (shrinkage)                                            w=1 / (|▽Sp| + ε)
   Test results using source code given by previous works




Things learnt from P.U.R.E.                                                       Future Work And Reference
Through the research this semester, I learnt:                                    Future works includes:
                                                                                 1.Using CVX to solve for the final algorithm
1.How to find/read/classify a paper in related fields.                           2.Testing algorithm effectiveness and efficiency

2. How to conduct a complete research from the                                   Reference:
beginning to the end.                                                            [1] Eduardo S. L. Gastal and Manuel M. Oliveira. "Domain

                                                                                 Transform for Edge-Aware Image and Video Processing".
3. The importance of doing experiments and testing                               SIGGRAPH 2011.
everything on my own.                                                            [2]Li Xu, et al. "Structure Extraction from Texture via Natural

                                                                                 Variation Measure”. SIGGRAPH Asia 2012
                                                                                 [3]Candes, E.J., et al. “Enhancing Sparsity by Reweighted ℓ1

Special thanks to: Mentor Jia-Bin Huang                                          Minimization”. Journal of Fourier Analysis and Applications,
                   P.U.R.E. Committee                                            2008
                                                                                 [4]Tom Goldstein, et al. “The Split Bregman Method for L1-

                                                                                 Regularized Problems”. SIAM Journal on Imaging
                                     Research Symposium                          Sciences, 2009

Image Smoothing for Structure Extraction

  • 1.
    Image Smoothing ForStructure Extraction Linjia Chang, lchang10@illinois.edu Mentor: Jia-Bin Huang, jbhuang1@illinois.edu Goal Applications ·Achieve Edge-aware image · Detail enhancement · Re-coloring smoothing while being able to · Image composition · Stylization distinguish texture/structure from · Object recognition · Video segmentation general natural images · Image denoise · Structure extraction Methods · Optimization with total variation regularization · - Robust loss function for texture removal · - Iterative reweighted L1 for sparsity[3] Previous Related Work · Gaussian Blur · L0 Gradient Minimization · Domain Transformation[1] · Structure Texture Extraction[2] Pixel = weighted average of its neighbors A major edge in a local window contributes more Enhances high-contrast edges by Preserves the original distance: similar-direction confining numbers of non-zero gradients isometric transform gradients Algorithm · Idea: Image smoothing as a global optimization problem Huber Loss Function Minimize S* = argmin ∑ λ||Sp – Ip|| + w||▽Sp|| s Data Term Regularization Term Similar as previous works but using Huber Iteratively Reweighted L1 Solution Algorithm[4] LF (Encourage Sparsity) 1. Set dummy variables u and v First solve the part without the S* = argmin ∑λ||Sp – Ip|| + w(|u|+|v|)+ β|(▽Spx-u)²+ (▽Spy-v)²| weight = λ||▽Sp|| s And then introduce weight w 2. Fix u, v and solve for S (convex) 3. Fix S and solve for u, v (shrinkage) w=1 / (|▽Sp| + ε) Test results using source code given by previous works Things learnt from P.U.R.E. Future Work And Reference Through the research this semester, I learnt: Future works includes: 1.Using CVX to solve for the final algorithm 1.How to find/read/classify a paper in related fields. 2.Testing algorithm effectiveness and efficiency 2. How to conduct a complete research from the Reference: beginning to the end. [1] Eduardo S. L. Gastal and Manuel M. Oliveira. "Domain Transform for Edge-Aware Image and Video Processing". 3. The importance of doing experiments and testing SIGGRAPH 2011. everything on my own. [2]Li Xu, et al. "Structure Extraction from Texture via Natural Variation Measure”. SIGGRAPH Asia 2012 [3]Candes, E.J., et al. “Enhancing Sparsity by Reweighted ℓ1 Special thanks to: Mentor Jia-Bin Huang Minimization”. Journal of Fourier Analysis and Applications, P.U.R.E. Committee 2008 [4]Tom Goldstein, et al. “The Split Bregman Method for L1- Regularized Problems”. SIAM Journal on Imaging Research Symposium Sciences, 2009