Size variations

Screen shots

1-2

1-3

1-4

Fourier analysis

The link for Wang tile:

http://johanneskopf.de/publications/blue_noise/spectrum/index.html

Their RAPSD

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

Ours:

I = 120

RaPsd

I = 191(0.75)

RaPsd

Using the same number of stipples for Secord's method and high packing density for Martin's method

Some changes:

1. In previous table, our methods are using m1,m2 = (3,2). This table is using m1, m2 = (5,3). 

2. Secord's method is using the same number of stipples as our ETF method, which improves the values a lot.

3. Martin's method is using high packing density value(4.0).

Discussion:

1. It seems CPSNR cannot say anything. It might be explained as the contrast for CAH is different for the contrast for stippling, since stipples are two dimensions. Contrast is not based on density only. The sizes of stipples and the directionality affect the contrast too. I guess it is not proper for our measurement.

2. As for MSSIM, our ETF method higher than our without etf method and Secord method, but lower than Kang's method. It makes sense. But I cannot explain why FSED and Ostromoukhov methods have higher scores in this comparisons. They always have lower number of stipples than our other three methods.

3.The high PSNR scores for our methods make me worry about the conversion validation.Some confidence comes from the values of Secord's method since it uses the same transition and has lower scores.

4. Except that the woman4 image uses more than 35000 stipples, the number for other images is around 20000 stipples. Secord's method is using the same number of stipples as the ETF method, always higher than other three methods. Kang said the number of stipples is around 8000 to 14000 without counting the pixels or lines used for the ETF edges.Based on my codes, the number of pixels on ETF edges is larger than 14000 or 20000. I guess the number of stipples for testing is fair in some way.

CPSNR

MSSIM

PSNR

Introducing texture as a feature field

It is not easy to have an interesting merging.It depends on both image contents.

One layer: 

Two layers:

Three Layers:Another two examples 

Using straight lines

Something about our stippling

1. Our stippling method works for both FS error diffusion and Ostromoukhov method by adjusting errors for each stipple based on the sizes of stipples. We can say it is a general transition from error diffusion to stippling.

2. Conversion issues(svg to png)

http://www.fileformat.info/convert/image/svg2raster.htm

3. Here are tables for comparisons among current stipplings I can get.

It surprises me Kang's method got the highest scores.

CPSNR

MSSIM

PSNR

Examples for styles

(1)Stippling:

(2)Abstract

Intermediate result

Final result

(3) Line

(4) Stroke-based

Density

Threshold the gradients to select three kinds of field

A particle with smallest gradients uses arbitrary field; the middle group uses smoother field; the particles in highest group use the real field.Hard thresholds cannot make sure the good segmentation.

I am working on the calculation of textureness.