results.txt

小波变换算法

the lossless ts with "coder BP ZT 2" runs at about 500 cycles per byte

--------

	WT v1.0 :

Lena @ 0.250 , l=3 : shapiro : 33.2 , spiht: 33.5 , xiong dct : 32.2

	wavelet, bitplane, uniform quant = 24, l=4
		lena.512.raw  :  262144 ->    8104 =  0.247 bpb = 32.347 to 1
		error: av= 3.80,max= 38,mse= 27.493,rmse= 5.24,psnr= 33.77
		( -> .25  bpp and 33.9 @ l=7)
		( -> .252 bpp and 33.3 @ l=3, q=27)

	dct, bitplane, uniform quant = 35
		lena.512.raw  :  262144 ->    8233 =  0.251 bpb = 31.840 to 1
		error: av= 4.56,max= 74,mse= 41.003,rmse= 6.40,psnr= 32.04
			(my 'ezdct' @ quant = 29 : psnr = 31.88)

	got option: levels = 8
	got option: quantizer = 23.000000
	got option: coderN = 5 = order 2
	got option: stop rate = 0.250
	LL band       :       4 ->       6 = 12.000 bpb =  0.666 to 1
	H bands       :  262140 ->    8191 =  0.249 bpb = 32.003 to 1
	lena.512.raw  :  262144 ->    8197 =  0.250 bpb = 31.980 to 1
	error: av= 3.77,max= 59,mse= 27.296,rmse= 5.22,psnr= 33.80

	Q:\wt>e512 lena -qs.25
	chose Q = 23.700000 , best err = 6660558
	LL band       :    1024 ->     794 =  6.203 bpb =  1.289 to 1
	H bands       :  261120 ->    7398 =  0.226 bpb = 35.296 to 1
	lena.512.raw  :  262144 ->    8192 =  0.250 bpb = 32.000 to 1
	error: av= 3.77,max= 41,mse= 27.009,rmse= 5.20,psnr= 33.85
	forward:503.810,encode:0.150,decode:0.200,inverse:3.350

	got option: stop rate = 0.250
	got option: coderN = 6 = BitPlane Binary
	got option: quantizer = 23.000000
	LL band           :    1024 ->    1366 = 10.671 bpb =  0.749 to 1
	H bands           :  261120 ->    6830 =  0.209 bpb = 38.231 to 1
	lena.512.raw      :  262144 ->    8196 =  0.250 bpb = 31.984 to 1
	error: av= 3.73,max= 47,mse= 26.871,rmse= 5.18,psnr= 33.87
	forward:2.28478,encode:0.7662,decode:0.7332,inverse:3.30078

	(transposing:)
	Q:\wt>wt c:\ccc\images\lena.512.raw 512 512 -qu21 -c6 -l8 -s.25
	got option: quantizer = 21.000000
	got option: coderN = 6 = BitPlane Binary
	got option: levels = 8
	got option: stop rate = 0.250
	LL band           :       4 ->      11 = 22.000 bpb =  0.363 to 1
	H bands           :  262140 ->    8185 =  0.249 bpb = 32.026 to 1
	lena.512.raw      :  262144 ->    8196 =  0.250 bpb = 31.984 to 1
	error: av= 3.62,max= 47,mse= 24.899,rmse= 4.99,psnr= 34.20
	forward:2.29830,encode:0.7782,decode:0.7496,inverse:3.30192


---------

Lena @ 0.50 :

	tuned jpeg : 35.5 , EZW : 36.3 , SPIHT : 36.9 EZDCT : 36.0

	WaveCoder v1.5, (c) 1998 by cBloom
	got option: stop rate = 0.500
	got option: coderN = 6 = BitPlane Binary
	got option: quantizer = 11.000000
	LL band           :    1024 ->    1366 = 10.671 bpb =  0.749 to 1
	H bands           :  261120 ->   15018 =  0.460 bpb = 17.387 to 1
	lena.512.raw      :  262144 ->   16384 =  0.500 bpb = 16.000 to 1
	error: av= 2.72,max= 32,mse= 12.883,rmse= 3.59,psnr= 37.06
	forward:2.28844,encode:1.14957,decode:1.14385,inverse:2.29755

	Wavelet
	got option: quantizer = 11.100000
	using default coder : BitPlane
	lena.512.raw  :  262144 ->   16385 =  0.500 bpb = 15.999 to 1
	error: av= 2.80,max= 22,mse= 13.670,rmse= 3.70,psnr= 36.81
 
	DCT
	using default coder : BitPlane
	quantize : 14.74,
	lena.512.raw  :  262144 ->   16381 =  0.499 bpb = 16.002 to 1
	error: av= 3.04,max= 43,mse= 16.479,rmse= 4.06,psnr= 36.00

	e512 lena -qs.5 -l4
	LL band       :    1024 ->     927 =  7.242 bpb =  1.104 to 1
	H bands       :  261120 ->   15459 =  0.473 bpb = 16.891 to 1
	lena.512.raw  :  262144 ->   16386 =  0.500 bpb = 15.998 to 1
	error: av= 2.77,max= 29,mse= 13.424,rmse= 3.66,psnr= 36.89
	forward:16:040,encode:0:240,decode:0:270,inverse:3:220

---------------------------

Lossless coding :	of Lena 512

got option: use S+P transform	(l == 4)
BitPlane      :  262144 ->  141539 =  4.319 bpb =  1.852 to 1
order1 l==3   :  262144 ->  139918 =  4.269 bpb =  1.873 to 1
order1        :  262144 ->  139372 =  4.253 bpb =  1.880 to 1
order1 l==5   :  262144 ->  139287 =  4.250 bpb =  1.882 to 1
S&P's paper(l=3):262144 ->  136643 =  4.17  bpb

v1.5 :
BP Bin Fast   :  262144 ->  137323 =  4.190 bpb =  1.908 to 1
BP BIn        :  262144 ->  136774 =  4.174 bpb =  1.916 to 1


got option: quantizer = 4.000000
got option: do lossless, 2 = SigMap-Big
using default coder : BitPlane
lossy         :  262144 ->   51373 =  1.567 bpb =  5.102 to 1
deltas        :  262144 ->   99255 =  3.029 bpb =  2.641 to 1
lossless      :  262144 ->  150628 =  4.596 bpb =  1.740 to 1

	very disappointing.

checa 256

	delta	:		order-1		order-0
	quant=4	:		3.719 		3.728 bpb
	quant=40:		4.716		4.724 bpb

these don't seem right.  Casual inspection shows the
deltas for quant small (4) are totally random, while
for quant large (40) the deltas have shape patterns
that mimic the wavelets.  We should see a more
striking variation between order-1 & 0 here

The problem seems to be that for high quant, the deltas
carry shape information, but the context (wavelet) data
doesn't, while for low quant, the wavelets carry shapes,
and the deltas don't.

--------------

	-ts -l2 , v1.5 , LL band (64x64)

Checa	wt 		:    4096 ->    2466 =  4.816 bpb =  1.660 to 1
Checa	\DPCM   :    4096 ->    2349 =  4.587 bpb =  1.743 to 1

Lena	wt      :    4096 ->    2981 =  5.822 bpb
Lena	\DPCM	:    4096 ->    2992 =  5.843 bpb

	-ts -l1 , v1.5 , LL band (128x128)

Checa   wt      :   16384 ->    9050 =  4.418 bpb
Checa   \dpcm   :   16384 ->    8017 =  3.914 bpb

Lena    wt      :   16384 ->   10790 =  5.268 bpb
Lena    \dpcm   :   16384 ->   10000 =  4.882 bpb


so DPCM (aka CALIC-badly-cloned) beats our ass on 128-wide LL bands,
but that's quite irrelevant.  Our cheezy ass bitch does dandily on the
important tiny LL's.