| introduction |
computational process for me suggests, again an interventionless
solution to this 'problem'. This, coupled with golan's express desire for
me to illustrate just one feature in my work led me to an algorithmic
solution.
An obvious may to emphasize the background is to make it bigger. the intractable problem of image segmentation aside, simply trying to scale the background will leave one with a hole in the forground. an alternative is to 'make more' of the background and this takes us into the realm of texture synthesis. |
"Choose on of the scenes and describe in text how you would intend to make the background amplified through some computational process" |
| texture synthesis |
given a picture of some grass, a texture synthesis algorithm tries
to generate a lawn. this requires some assumtions about the way that the
human visaul system works and about images in the real world. the approach
taken here would be based on de Bonnet's work, if there hadn't been
mistakes in his paper. the approach now is a hybridised version of his
work, with greater flexibity. what the applet does (and it certainly
doesn't do it in real time) is take a wavelet transform of the patch
of the image to be resynthesised and then samples each level of the wavelet
pyramid from the distribution of pixels at that level. this sampling is
further constrained by all the parent structures of those pixels, and
by the parent structures taken from possible other images / filter banks.
the aim of this is to maintain the structure of the image while creating a 'new' image - i.e. the blades of grass remain blades, but they get put in different positions. |
|
| edge of abstract |
I'm aiming for late heron position meets klimt palette. heron's edge of
abstract landscapes and portraiture meeting klimts edge of impressionistic
colour fields. while it was originally indended for the forground to be
manually masked out (e.g. eliminated for the purposes of resampling) I
didn't have time to mask out the tree in the image, hence resynthesis takes
place on right hand 128x128 square of the image (reduced to 128 tall) and
the sampling thresholds increase towards the right of the image - as we let
the algorithm become more and more creative in its reoganisation of the
freqency levels of the image.
what I didn't achieve was performance of the algorithm. the resolution in the images effectively only 64x64 (there was literally not enough time last night before class to sample all the way down to 256x256) and this has adversly affected the quality of the images. I'll be running the algorithm overnight this week to create really large images. colour-space is obviously an issue here, and rgb (or hsv) is probebly a bad choice - perhaps I might do some stuff yuv. |
|
| the images | original image |
|
| threshold 0.4 image (large) | (large) | |
| final image |
