Image EnhancementDenoising & Contrast Enhancement Mingzhu Long - PowerPoint PPT Presentation

Image Enhancement—Denoising & Contrast Enhancement Mingzhu Long 2016.02.05

MENU n Image Denoising Ø Algorithm Classification Ø Examples n Image Contrast Enhancement Ø Algorithm Classification Ø Examples n Image Quality Assessment

MENU n Image Denoising Ø Algorithm Classification Ø Examples n Image Contrast Enhancement Ø Algorithm Classification Ø Examples n Image Quality Assessment

Image Denoising Ø Single Image Denoising Ø Multi-image Denoising Ø Spatial Domain Ø motion compensation Ø Local Ø Non-motion Ø Nonlocal compensation Ø Transform Domain Ø Data adaptive transform Ø Non-data adaptive transform from spatial domain to transform domain, from local to nonlocal methods, from pointwise to multipoint methods, from single image to multi-image.

Examples--NLM The NL-means not only compares the grey level in a single point but also the geometrical configuration in a whole Neighborhood. This fact allows a more robust comparison than neighborhood filters. A nonlocal algorithm for image denoising. CVPR 2005.

Examples--NLM Noisy image(standard deviation 20) ， Gauss filtering, anisotropic filtering, Total variation, Neighborhood filtering and NLmeans algorithm A nonlocal algorithm for image denoising. CVPR 2005.

Examples--BM3D Step 1)Basic estimate. Step 2)Final estimate: a) Block-wise estimates. a) Block-wise estimates. Grouping. Grouping. Collaborative hard-thresholding. Collaborative Wiener filtering. b) Aggregation. b) Aggregation. 时，效果明显下降。计算量，存储。 时，效果明显下降。计算 量，存储。 Image Denoising by Sparse 3-D Transform-Domain Collaborative Filtering. TIP 2007.

Examples--BM3D Image Denoising by Sparse 3-D Transform-Domain Collaborative Filtering. TIP 2007.

Examples--Multi-image How to deciding whether a patch A1 is similar to a patch B1 in the reference image I1? We find their corresponding patches A2 and B2 ,respectively, in the second image I2. If A1 is similar to B1, A2should also be similar to B2. Multiple view image denoising. CVPR 2009.

Examples--Multi-image Noisy patch BM3D Our method Ground truth Multiple view image denoising. CVPR 2009.

MENU n Image Denoising Ø Algorithm Classification Ø Examples n Image Contrast Enhancement Ø Algorithm Classification Ø Examples n Image Quality Assessment

Image Contrast Enhancement Intensity Histogram Transformation Equalization Global Logarithmic histogram transformation equalization Local Gamma ‹ histogram transformation equalization Contrast stretching 8/16

Results-Log 9/16

Results-Gamma Gamma correction depends on a suitable selection of the index. It seems that one index can not take into account both the details in the dark area and lighter area. 10/16

Results-HE 11/16

Results-HELoc 12/16

Results It seems that CE&NLM is clearer than NLM&CE The histogram makes some areas too bright enhancement 6/16

Solutions n Combination of different methods[1] or index[2]! Ø Brightness estimation Ø Threshold estimation n Transformation domain contrast enhancement？ [1] Gabriel Zahi. Automatic Detection of Low Light Imagesina Video Sequence Shot under Different Light Conditions. 2015. [2] Jinfang Shi. A Novel Image Enhancement Method Using Local Gamma Correction with Three-level 13/16 Thresholding. 2011

MENU n Image Denoising Ø Algorithm Classification Ø Examples n Image Contrast Enhancement Ø Algorithm Classification Ø Examples n Image Quality Assessment

Image Quality Assessment n Subjective Evaluation n Objective Evaluation Ø Full-reference Evaluation Method Ø PSNR, MSE, SSIM, Entropy ， etc. Ø Semi-reference Evaluation Method Ø Non-reference Evaluation Method Ø Evaluation methods for some certain distortions Ø General evaluation method

图像质量评价 n Mean Square Error n Normalized Mean Square Error n Mean Absolute Error

图像质量评价 n Normalized Mean Absolute Error n Peak Signal-to-Noise Ratio a n Structural Similarity

THANKS ！ Q&A Mingzhu Long 2016.02.05