|
Hi,
Actually my neatness metrics is an unsuccessful attempt, that I have coded very (very) quickly, to measure/catch the neatness of an image, originally to try to objectively prove that the NHW Project has more neatness.But you can forget about it, it is not really efficient!...
A real neatness metrics is way more complex and difficult to design (at least for me...).However I can give you quick details of my neatness metrics attempt, maybe there could be possible interesting parts in it: it is really very simple (too simple), again I coded it very very quickly.For each pixel of a decoded image, I compute the laplacian kernel of the pixel with its 8 neighbours, if this pixel is considered as neat (has a high score), then I quickly study its 8 neighbours to be sure this pixel is not an isolated point for example and rather belongs to a contour/edge.Then with this selected pixel in the decoded image, I compare it with the same position pixel in the original image, if this same position pixel in the original image is also a rather neat pixel (to be sure we're not on an artifact), then I compare if the neat pixel in the decoded image is "more neat/sharp" than the pixel in the original image, if the decoded neat pixel is above a threshold compared to the original one then I consider that this pixel gives more neatness/sharpness than the original image...
But this algorithm fails to catch all the neatness of an image, neatness is more complex actually.
Would you know if a good neatness metric algorithm exists?
Cheers,
Raphael
|