constant em=1 (3138 stipples)

Now it's working.

k=0.01, 10321

k=0.1, 4885

k=1,2619

Change em

(f)3080

(e)17003

A fast and automatic screening method

Based on our CAH:

1. segment the image by mean shift method.

2. Based on segmentation, assign different exclusion-based masks to each segmented regions.

3. Do CAH with exclusion assignment.

Total time: size 433*800, 55 seconds for segmentation and 15 seconds for CAH.

Next, I want to do something on the choice of the masks. Relate it to the content.

Still cannot get Fourier power spectrum like their paper

Here is what I got.

Matlab code I am using as follows,

 mag = fftshift(abs(fft2(double(X))));  % Amplitude spectrum

X is image.

Then I scale them into [0,255] and show them. It's not what their paper looks like.

Wang tile:

Penrose:

Ours: 

m1 and m2

Compare with 3, 9 and 9, 3. If m1 = 3 and m2 = 9, it will have more stipples for light area and fewer stipples for dark area. If m1 = 9 and m2 = 3, it will have more stipples for dark area and fewer stipples for light area. I think m1 and m2 can adjust the number of stipples for light area and dark area respectively. It also means if we want a more precise density control, we have to test the content of the image such as histogram.

m1, m2

3, 3    

3, 9

9, 3

9, 9 

For secord's results: Measurement change with different number

Tone:
            40k         |30k         |20k         |10k         |5k
baby:    21.8400  |22.2324  |23.0338  |24.4905  |24.0452
balloon: 18.5688 |19.2340  | 20.4272 |24.0532  |25.8144
woman: 21.4464 |22.2887  |24.6317  |28.7024  |26.9925
                           |               |              |               |
MSSIM:
baby:     0.1182 |0.1057    | 0.0969    |0.0930  | 0.0947
balloon: 0.5620 |0.5403    | 0.5225    | 0.5045 | 0.4916   
woman: 0.1383 |0.1146    | 0.0964    |0.0777  | 0.0711

Tone improved by Secord's calculation

1:

MSSIM:

PSNR:  

Tables

Stipple sizes affect the measurement too. Based on those two tables, it seems we should get better tone in some way. I am not sure if the adjustment can compete with Secord's method at very low budget. But it looks it should have some peak tone matching for our method. I am going to try different size calculation to see if we can improve the tone in some way.

MSSIM: Both of ours have better matching than Kim's method. Our ETF method is using very dense version. P1(0,0,0,7) and P2(5,3,0,7). Both our ETF and basic methods is using averge intensity (previously is using the original intensity) to adjust the stipple size when doing postprocess.

PSNR: Secord's method has the best matching. Both ours are better than others.(Old matlab codes for psnr is using unsigned int. Here is using double. Fix up the 'inf' problem.)

Increasing Size

It seems the tone is not fixed up by the postprocessing.

A searching map for nearest stipples:

3085 stipples with P(10,10,0.1,15):

size = 0.5 * d *(255-avg)/255-- d is the distance between the nearest stipples.

size = MIN+(Max-min)*(255-avg)/255

Secord's result

Extreme low budget (955 stipples)

Postprocess

Size effects

Our method prefers to small stipples. Here is an example to show double the stipple sizes. It reduces the number of stipples, but it is not good for structure. P(5,5,7), 6678 stipples (previously 26834 stipples with smaller sizes).