[argyllcms] Re: ImageTarget version 0.1.1 released

Alastair M. Robinson wrote:

What sort of degree of blurring / downsampling / denoising did you use?


Klaus Karcher wrote:
Before reporting on the results, I have to discuss some strange side effects I noticed in conjunction with different source profiles (I will do that in a separate mail).
I just finished my experiments and want to give a brief summary of the results:
I used the 16bit/chanel data of my oil on canvas scan, transformed it to MinkaRGB (L* TRC). For comparison, I applied some filters also in Gamma-1-ProPhoto-RGB, but transformed the results back to MinkaRGB to avoid side effects with cctiff. 

I tested 5 methods to restrict the source gamut (all in Photoshop CS3):

- Gaussian Blur, r = 0.5 ... 10
- Bicubic downsampling to 90 ... 10%
- Median filter, r = 1 ... 10
- "Dust and Scratches" r = 6, threshold = 25
- "Reduce Noise", "Strength" = 10, all other options set to zero
- "Dust and Scratches" had only slight effects on gamut mapping (slightly increased overall saturation, no objectionable clipping). 
Similar results can be obtained with downscaling to approx. 60%.
- "Reduce Noise" had nearly no (at least no positive) influence on gamut mapping. The VRML file suggests that "Reduce Noise" deforms the gamut rather than reducing it. viewgam -i confirm this impression: 
Intersecting volume = 194686.1 cubic units
'FrauCLMr.gam' volume = 223546.7 cubic units, intersect = 87.09%
'FrauCLMr_RN10.gam' volume = 234535.9 cubic units, intersect = 83.01%
- Gaussian Blur had enormous influence on gamut. Even very small radii reduce the gamut volume drastically (see attachment "GB_V(r).pdf"). Therefore it's difficult to find the right "dose". Too large radii cause serious clipping (and even r = 1 can be too large). The scaling of the slider in ImageTarget's UI should account for the exponential relation between r an V.
- Median filtering gives reasonable results, but the integer increments are somewhat coarse and not perceptually uniform.
- Bicubic downsampling gives very good results. It is easy to dose (the scale is roughly proportional to V^(1/3) and therefore perceptually nearly uniform). Moreover tiffgamut benefits from the smaller size of the downscaled images. IMHO it is the best of the tested methods to control the gamut mapping.
- Bicubic downsampling is strongly TRC dependent e.g. the effect of Downscaling to 20% with gamma 1 corresponds roughly to downscaling to 60% with L* TRC. 

- Gaussian blur is much less TRC dependent.

Klaus

Other related posts: