Presentation is loading. Please wait.

Presentation is loading. Please wait.

National Aerospace University of Ukraine IS&T/SPIE Electronic Imaging 2014 1 METRIC PERFORMANCE IN SIMILAR BLOCKS SEARCH AND THEIR USE IN COLLABORATIVE.

Similar presentations


Presentation on theme: "National Aerospace University of Ukraine IS&T/SPIE Electronic Imaging 2014 1 METRIC PERFORMANCE IN SIMILAR BLOCKS SEARCH AND THEIR USE IN COLLABORATIVE."— Presentation transcript:

1 National Aerospace University of Ukraine IS&T/SPIE Electronic Imaging 2014 1 METRIC PERFORMANCE IN SIMILAR BLOCKS SEARCH AND THEIR USE IN COLLABORATIVE 3D FILTERING OF GRAYSCALE IMAGES A.S. Rubel 1, V.V. Lukin 1, K.O. Egiazarian 2 1 Department of Transmitters, Receivers and Signal Processing, National Aerospace University, Kharkov, Ukraine 2 Department of Signal Processing, Tampere University of Technology, Finland

2 National Aerospace University of Ukraine Similarity search use in image processing 2 Collaborative and non-local filtering of remote sensing images Object recognition and tracking Motion estimation for video coding Computer vision UAV navigation Fractal compression

3 National Aerospace University of Ukraine Similarity metrics 3 Minkowski distance: Chebyshev distance: Manhattan: Euclidean distance: Mahalanobis distance: Hellinger distance: Bray-Curtis distance: Canberra : Cosine distance: Distance based on Pearson correlation:

4 National Aerospace University of Ukraine Test images 4 Weald San Diego Pentagon Airfield Bay For each test image we introduced groups of 33 identical blocks.

5 National Aerospace University of Ukraine Histograms of distance values with ϭ = 30 5 AWGN Spatially correlated noise

6 National Aerospace University of Ukraine Search performance estimate 6 ϭ = 5 ϭ = 30 Proposed rank estimate: Noise is one of the most destructive factors; Spatially correlated noise is more destructive than AWGN; Positions of detected block on sorted distances are important for further analysis.

7 National Aerospace University of Ukraine Search performance under AWGN 7 Spatial domainDCT spectrum domain Classical metrics (Euclidean distance and Manhattan) are not the best. Mahalanobis and Bray-Curtis distances have better performance in both domains, Canberra and Pearson correlation show high performance only in spatial domain.

8 National Aerospace University of Ukraine Table of performances under intensive AWGN ( ϭ = 30) 8 For AWGN case, Mahalanobis distance is the best metric among the considered. Bray-Curtis distance and Canberra have slightly worse performance. Search in spatial domain is preferable.

9 National Aerospace University of Ukraine Search performance under spatially correlated noise 9 Spatial domainDCT spectrum domain Mahalanobis distance and Pearson have better performance in spatial domains. Canberra and Bray-Curtis distance show high performance in the DCT spectrum domain. Classical metrics still are not the best.

10 National Aerospace University of Ukraine Table of performances under intensive spatially correlated noise ( ϭ = 30) 10 For spatial correlated noise case Bray-Curtis distance and Canberra are best metrics among considered. Search in DCT spectrum domain is preferable.

11 National Aerospace University of Ukraine BM3D filter 11 Classical Euclidean distance is used for additive Gaussian noise suppression by BM3D filter; BM3D uses DCT spectrum domain for search; Without similar blocks search BM3D turns into 2D DCT-filter; Search robustness becomes a substantial issue for this technique.

12 National Aerospace University of Ukraine Denoising performance of TID2013 images by IPSNR 12 AWGNSpatially correlated noise Improvement of PSNR, as performance criterion:

13 National Aerospace University of Ukraine Denoising performance of TID2013 images by IPSNR-HVSM 13 AWGNSpatially correlated noise Improvement of PSNR-HVSM, as performance criterion: V. Lukin, N. Ponomarenko, K. Egiazarian, “HVS-Metric-Based Performance Analysis Of Image Denoising Algorithms”, Proceedings of EUVIP, Paris, France, 2011, pp. 156-161.

14 National Aerospace University of Ukraine Spatially correlated noise ( ϭ = 15) 14 Noisy image BM3D with EuclideanBM3D with Bray-Curtis Noise-free image

15 National Aerospace University of Ukraine Conclusions 15 Classical metrics are not the best for both considered cases of the noise; Canberra, Mahalanobis and Bray-Curtis distances perform better than classical ones for AWGN in spatial domain; Canberra and Bray-Curtis distance are better for spatially correlated noise in the DCT spectrum domain; The use of Canberra and Bray-Curtis distance for the BM3D filter instead of default one provides better results for spatially correlated noise.

16 National Aerospace University of Ukraine Metric performance in similar blocks search and their use in collaborative 3D filtering of grayscale images 16 Thank you! Karen O. Egiazarian karen.egiazarian@tut.fi Vladimir V. Lukin vladimlukin@yahoo.com Aleksey S. Rubel edu.rubel@gmail.com


Download ppt "National Aerospace University of Ukraine IS&T/SPIE Electronic Imaging 2014 1 METRIC PERFORMANCE IN SIMILAR BLOCKS SEARCH AND THEIR USE IN COLLABORATIVE."

Similar presentations


Ads by Google