LSLCD Algorithm Performance on Different Sampling Structures

The table lists the CPNSR value obtained on each of the 24 Kodak images. For each image, the training set used is the other 23 images of the Kodak set. The filters used in this comparison have a 11x11 size. For each image we indicate the best result in bold.

In general, the images sampled with a Bayer structure give a better objective demosaicking performance than the ones sampled with the diagonal stripe structure.

Looking at the images from a subjective point of view:

• At zoom-in, the fact that the image has been sampled on a diagonal stripe structure becomes obvious. The transition from one color to another causes color leaking and the -45 degree lines have a rainbow aspect when demultiplexed.
  • colour leaking at boundaries
  • introduction of what looks like a single signal having a 2 pixel wavelength
  • -45 degree white diagonal lines have a rainbow aspect - these types of lines are harder to estimate.
  • coloration on plane patterns; more obvious where the pattern is denser (see walls in image 1 and 3)
• These problems are noticed when demosaicking from a Bayer structure too but they are much less obvious:
  • pattern 'interference' / color leaking (small leaf detail, image 7; image 14)
  • aliasing next to vertical and horizontal lines (obvious in the window detail picture 1and also on the roof top in image 24)

  Clolor aliasing with stripe gives a rainbow effect at demosaicking. Color aliasing with Bayer is mostly due to the chroma 2 component making certain image portions look blue-orange.

No.	Thumb	Bayer	Stripe
1		38.67	38.81
2		40.81	38.93
3		43.37	39.46
4		41.34	39.59
5		38.30	35.31
6		40.42	38.96
7		42.86	40.43
8		35.66	36.48
9		42.85	40.99
10		42.85	41.10
11		40.33	38.68
12		43.46	42.02
13		34.89	32.91
14		36.75	34.63
15		40.03	38.88
16		44.09	42.56
17		42.00	40.34
18		37.80	36.05
19		40.97	39.95
20		41.25	39.67
21		39.28	38.82
22		38.67	37.36
23		43.15	40.89
24		35.60	34.17