Separable bilateral filtering for fast video preprocessing
Download as PPT, PDF2 likes4,038 views
This document summarizes research on separable bilateral filtering for fast video preprocessing. Bilateral filtering reduces noise while preserving edges but has high computational complexity. The researchers propose a separable implementation that approximates the original filter with linear complexity. They apply separable bilateral filtering to video noise reduction and show it achieves better compressed video quality than full-kernel filtering with the same computation. The separable approach makes real-time bilateral filtering possible for applications like video preprocessing.
Separable bilateral filtering for fast video preprocessing
- 1. ICME 2005 IEEE International Conference on Multimedia & Expo July 6-8, 2005, Amsterdam, The Netherlands Separable bilateral filtering for fast video preprocessing Tuan Q. Pham & Lucas J. van Vliet Delft University of Technology The Netherlands 1 Quantitative Imaging Group Faculty of Applied Sciences
- 2. Contents 1. Bilateral filtering: Edge-preserving filter High computational complexity 2. Separable implementation: Good approximation of the original filter Linear computational complexity 3. Application to video preprocessing: Noise reduction Better compressed video
- 3. Gaussian filtering revisited Gaussian filtering: weights depend on distance to the center pixel * = Noisy step edge Gaussian weights Gaussian filtered result Adaptive filtering: avoid filtering across edges . = Noisy step edge Edge-preserving weights Edge-presered filtered result
- 4. Gaussian vs bilateral filtering Gaussian filtering: weights depend on distance to the center pixel * = Noisy step edge Gaussian weights Gaussian filtered result Bilateral filtering: weights depend on both spatial closeness and photometric similarity . = Noisy step edge Edge-preserving weights Edge-presered filtered result
- 5. How bilateral filtering works? Every sample is a weighted average of its neighbors: 1 O ( s0 ) = ∑ w ( s, s0 ) . I ( s ) K s Local modes The weight w = w s . wt is product of two Gaussian weights: after bilateral − ( s − s0 ) 2 filtering Spatial proximity: w s = exp 2σ s2 − ( I ( s ) − I ( s0 ) ) 2 Tonal similarity: wt = exp 2σ t 2
- 6. Example: Gaussian filtering Noisy input: PSNR = 39.1 dB Gaussian filtered: PSNR = 67.9 dB
- 7. Example: Bilateral filtering Noisy input: PSNR = 39.1 dB Bilateral filtered: PSNR = 41.6 dB
- 8. Computational complexity Bilateral filtering kernel is space-variant → complexity is: O ( Nm d ) where N: number of pixels in the image m: size of filtering kernel (m ≈ 7 is good enough) d: image dimensionality Previous attempt for fast bilateral filtering: Piecewise-linear: approximate bilateral filtering with M Gaussian filtering - Durand & Dorsey (SIGGRAPH 2002) → complexity is: O ( N log( N ) . M ) M ≈ 17 for 8-bit images Our approach: separable bilateral filtering O ( Nmd )
- 9. Is bilateral filter separable? Gaussian filter is space-invariant and separable: g(x) g(y) ∑ w ( s − s ) . I ( s) s 0 O( s ) = s * = ∑ w (s − s ) 0 s 0 s y x w s ( s ) = gauss( x ) . gauss( y ) Kernel center Bilateral filter is NOT separable: I(s0) Space-variant kernel due to local I(s) intensity dependency However, even a highly non-linear filter like median filter is approximately separable (Narendra – PAMI 1981)
- 10. Separable bilateral filtering result Separable bilateral filtering is a good approximation of full kernel filtering: noisy Erika σnoise = 10 bilateral filtered in x-dimension followed by y-dimension filtering
- 11. Separable bilateral filtering result Separable bilateral filtering is a good approximation of full kernel filtering: noisy Erika σnoise = 10 bilateral filtered in x-dimension followed by y-dimension filtering Image size Brute-force Durrand 2002 Separable Aniso. diffusion 256x256 4.46 0.37 0.21 2.76 512x512 17.88 1.59 0.89 26.02 61x61x61 5.29 4.20 0.45 5.25 256x256x212 56 min Out of memory 50.3 Out of memory
- 12. How separable bilateral filtering works? Pros: extremely fast (fixed spatial weight + LUT for tonal weight) Cons: effective filtering kernel is a slightly distorted
- 13. Performance of separable bilateral filtering Very good approximation of Almost linear execution time the full-kernel filter per pixel
- 14. MPEG-1 Foreman with bilateral preprocessing 16 without preprocessing with 3x3x3 full-kernel with 9x9x5 separable 15 RMSE 14 13 0 400 800 1200 bit-rate (K bits/s) Better RMSE is achieved with separable bilateral filtering compared to full-kernel bilateral filtering with the same computation requirement
- 15. Less artifact with Bilateral preprocessing
- 16. Conclusions Separable implementation of bilateral filtering: Very good approximation of the original filter Much faster than the original or other approximations Applications in video preprocessing: Improved quality of compressed video Reduced processing time → real-time possibility
- 17. Literature C. Tomasi and R. Manduchi, Bilateral fitering for gray and color images, Proc. of ICCV, USA, 1998, 839-846. T.Q. Pham and L.J. van Vliet, Separable bilateral filtering for fast video preprocessing, Proc. of ICME’05. F. Durrand and J. Dorsey, Fast bilateral filtering for the display of high dynamic range images, Proc. of SIGGRAPH’02, 2002, 844-847. P. Perona and J. Malik, Scale-space filtering and edge detection using anisotropic diffusion, PAMI, vol. 12, no. 7, 1990, 629-639. R. v.d. Boomgaard and J. v.d. Weijer. On the equivalence of local- mode finding, robust estimation and mean-shift analysis as used in early vision tasks. In Proc. of ICPR, pages 927-930, 2002.