Machine Learning for Adaptive Bilateral Filtering I. Frosio 1, K. Egiazarian 1,2, and K. Pulli 1,3 1...

download Machine Learning for Adaptive Bilateral Filtering I. Frosio 1, K. Egiazarian 1,2, and K. Pulli 1,3 1 NVIDIA, USA 2 Tampere University of Technology, Finland.

If you can't read please download the document

Transcript of Machine Learning for Adaptive Bilateral Filtering I. Frosio 1, K. Egiazarian 1,2, and K. Pulli 1,3 1...

  • Slide 1
  • Machine Learning for Adaptive Bilateral Filtering I. Frosio 1, K. Egiazarian 1,2, and K. Pulli 1,3 1 NVIDIA, USA 2 Tampere University of Technology, Finland 3 Light, USA
  • Slide 2
  • Motivation (denoising) void denoise(float *img){ for (int y = 0; y < ys; y++){ for (int x = 0; x < xs; x++){ img(y*xs+x) = } }
  • Slide 3
  • Motivation: (Gaussian filter) t(x)d(x)
  • Slide 4
  • Bilateral filter Motivation (bilateral filter) C. Tomasi and R. Manduchi, Bilateral filtering for gray and color images, ICCV, 1998. t(x)d(x)
  • Slide 5
  • Motivation: (choice of the parameters) dd dd rr d(x)
  • Slide 6
  • Motivation (use intuition?) d scales with resolution r scaled with grey level dynamics possible automatic design of parameter values [BF, Tomasi and Manduchi, ICCV, 1998] image noise std n d = [1.5, 2.1], independently from n r = k n [BF, Zhang, Gunturk, TIP, 2008] local signal variance s 2 (x,y) d = [1.5, 2.1], independently from n r (x,y) = n 2 / s (x,y) [ABF, Qi et al., AMR, 2013] dd dd rr
  • Slide 7
  • Motivation (use machine learning!) dd rr
  • Slide 8
  • Framework (adaptive denoising) 3 features 6 unknowns d (x,y) r (x,y)
  • Slide 9
  • Framework (learning) Training images {t j } j=1..N Noise model (AWGN) Local image features, f x,y Image quality model (PSNR) Adaptive bilateral filter,
  • Slide 10
  • Entropy features FlatGradientTextureEdges 0.0 bit6.0 bits1.0 bit5.6 bits Shannons entropy i(x,y)
  • Slide 11
  • Entropy features FlatGradientTextureEdges eiei 0.0 bit6.0 bits1.0 bit5.6 bits egeg 0.0 bit 1.2 bit5.5 bits i(x,y) ||grad[i(x,y)]||
  • Slide 12
  • Entropy features eiei egeg
  • Slide 13
  • Framework (complete) Training images {t j } j=1..N Noise model (AWGN) Local image features f x,y Image quality model (PSNR) Adaptive bilateral filter, d (x,y) Logistic functions Local image features f x,y r (x,y) Noisy image EABF Filtered image
  • Slide 14
  • Results - PSNR BFBF [Zhang]ABF [Qi]EABF d (x,y) optimal1.8 our r (x,y) optimal 2n2n n 2 /(0.3 s ) our n = 5 36.1336.063636.27 n = 10 31.4531.431.4431.1 n = 20 27.1127.0927.3626.4 n = 30 25.072525.3224.68 n = 40 23.9423.6924.1123.76 n = 5 36.2936.0435.9236.4 n = 10 32.5332.1732.0332.81 n = 20 28.9628.4828.7529.51 n = 30 26.9626.3126.9827.62 n = 40 25.6324.7225.6826.17 n = 5 36.536.0735.9336.54 n = 10 32.6432.2332.1532.78 n = 20 29.3228.8129.1729.65 n = 30 27.7126.8427.6828.01 n = 40 26.6625.3326.5326.82 BFBF [Zhang]ABF [Qi]EABF d (x,y) optimal1.8 our r (x,y) optimal 2n2n n 2 /(0.3 s ) our n = 5 37.6937.537.237.81 n = 10 3433.7633.6634.37 n = 20 30.3129.6430.1731.11 n = 30 28.227.1228.229.24 n = 40 26.8625.4626.8327.78 n = 5 38.1737.8637.6438.45 n = 10 34.6434.0934.1835.17 n = 20 31.230.0231.0532.08 n = 30 29.3727.7829.2430.21 n = 40 28.225.9727.8328.77 n = 5 37.8137.7437.3737.86 n = 10 34.7534.3134.2234.98 n = 20 31.2730.431.2231.8 n = 30 29.1427.8529.329.75 n = 40 27.7925.9527.7128.3
  • Slide 15
  • Results - PSNR BF BF [Zhang] ABF [Qi] EABF d (x,y) optimal1.8 our r (x,y) optimal 2n2n n 2 /(0.3 s ) our n = 5 37.136.8836.6837.22 n = 10 33.3332.9932.9533.53 n = 20 29.6929.0729.6230.09 n = 30 27.7426.8227.7928.25 n = 40 26.5125.1926.4426.93 average n = 5 40 30.8830.1930.731.21 +1.01dB +0.51dB
  • Slide 16
  • Results Image quality Ground truth Noisy 20.11 dB BF [Zhang] 30.02 dB ABF [Qi] 31.05 dB EABF 32.08 dB
  • Slide 17
  • Machine learning vs. intuition: d (x,y), r (x,y) n = 20 BF [Zhang et al.]ABF [Qi et al.]EABF d (x,y) 1.8 r (x,y) 2s n [0.6, 2.6] [71, 85] [20, 110]
  • Slide 18
  • Machine learning vs. intuition: d = d ( n ) Flat areaEdge area
  • Slide 19
  • Machine learning vs. intuition: r = r ( n ) Flat areaEdge area
  • Slide 20
  • Conclusion Learning optimal parameter modulation strategies through Machine Learning is feasible. Modulation strategies are complicated But effective. PSNR
  • Slide 21
  • Conclusion Training images {t j } j=1..N Noise model (AWGN) Local image features, f x,y Image quality model (PSNR) Adaptive bilateral filter,
  • Slide 22
  • Conclusion Your training images Noise model (AWGN) Local image features, f x,y Image quality model (PSNR) Adaptive bilateral filter,
  • Slide 23
  • Conclusion Your training images Your noise model Local image features, f x,y Image quality model (PSNR) Adaptive bilateral filter,
  • Slide 24
  • Conclusion Your training images Your noise model Your local features, f x,y Image quality model (PSNR) Adaptive bilateral filter,
  • Slide 25
  • Conclusion Your training images Your noise model Your local features, f x,y Your image quality model, Q Adaptive bilateral filter,
  • Slide 26
  • Conclusion Your training images Your noise model Your local features, f x,y Your image quality model, Q A different adaptive filter,
  • Slide 27
  • Conclusion A general FRAMEWORK based on MACHINE LEARNING for the development of ADAPTIVE FILTERS