Shape-Adaptive Transforms Filtering

EXPERIMENT RESULTS

(“Shape-Adapted” DCT)

The results in this page have been obtained by replacing the SA-DCT with the more complex “Shape-Adapted” DCT (obtained by orthogonalization). Such results are only marginally better than those obtained with the SA-DCT, and visually the two estimates are almost indistinguishable.

Denoising

Anisotropic LPA-ICI + Gilge's “Shape-Adapted” DCT denoising
Results (PSNR, dB), noise is white additive Gaussian with variance σ².
(Click on the numbers to see the corresponding images)

σ	Lena 512x512	Boats 512x512	House 256x256	Peppers 256x256	Cameraman 256x256	Barbara 512x512	Hill 512x512	Lake 512x512
5	38.62	37.17	39.45	38.14	38.26	37.57	37.04	36.62
10	35.71	33.69	36.09	34.64	34.12	33.50	33.44	32.86
15	34.00	31.85	34.26	32.62	31.84	31.30	31.60	31.06
20	32.75	30.54	33.01	31.19	30.32	29.84	30.40	29.82
25	31.76	29.52	32.01	30.07	29.22	28.75	29.51	28.87
30	30.95	28.68	31.15	29.16	28.33	27.87	28.81	28.10
35	30.25	27.98	30.42	28.41	27.59	27.10	28.24	27.44
50	28.69	26.36	28.78	26.75	25.94	24.94	26.93	25.94
75	26.78	24.70	26.64	24.71	23.92	23.30	25.50	24.26
100	25.44	23.58	25.08	23.30	22.53	22.38	24.49	23.06

Deblurring

Pointwise “Shape-Adapted” DCT regularized deconvolution
Results (ISNR, dB) for four different deblurring experiments: (Click on the numbers to see the corresponding images)

Experiment #	1	2	3	4
*Anisotropic LPA-ICI* + Gilge's “Shape-Adapted” DCT**	8.58	8.29	6.34	4.55

Experiment 1 - Cameraman 256x256, blur: 9x9 uniform "box-car" blur, Gaussian white noise: BSNR 40dB (σ²=0.308)
Experiment 2 - Cameraman 256x256, blur: 15x15 1/(x²+y²), x,y=-7,...,7, kernel, Gaussian white noise: σ²=2
Experiment 3 - Cameraman 256x256, blur: 15x15 1/(x²+y²), x,y=-7,...,7, kernel, Gaussian white noise: σ²=8
Experiment 4 - Lena 512x512, blur: 5x5 separable [1, 4, 6, 4, 1]/16 filter, Gaussian white noise: σ²=49