Automatic producing an image of a desired aspect ratio.

Archived from groups: sci.image.processing,comp.graphics.algorithms,rec.photo.digital (More info?)

I find information for my diploma work
about automatic scene-dependent image cropping, i.e., cropping
undesirable content from a picture and magnifying or zooming the
desired content to fill the entire ("crop to fill" ) photographic print.

While many digital cameras produce digital images of 2:3 aspect
ratio, a significant percentage of other cameras produce images of 3:4
aspect ratio. Because the most popular print format is 4x6, image
cropping has to occur when printing digital images of an original
aspect ratio of 3:4. A less than satisfactory alternative, called "crop
to fit" as opposed to "crop to fill", is to pad the 3:4 image with
white space to obtain the desired aspect ratio.
I need to provide intelligent image cropping according to an
automatic understanding of the image content and enforcement of
compositional rules so that
1. The main subject of the image is not cropped in part or in its
entirety.
2. Both smooth and textured background can be identified and removed in
part or in its entirety if necessary.
3. Common picture composition rules such as sufficient headroom can be
enforced.

Archived from groups: sci.image.processing,comp.graphics.algorithms,rec.photo.digital (More info?)

Joseph Meehan wrote:
> Egor Ivaniv wrote:
>
>> I find information for my diploma work
>>about automatic scene-dependent image cropping, i.e., cropping
>>undesirable content from a picture and magnifying or zooming the
>>desired content to fill the entire ("crop to fill" ) photographic
>>print.
>>
>> While many digital cameras produce digital images of 2:3 aspect
>>ratio, a significant percentage of other cameras produce images of 3:4
>>aspect ratio. Because the most popular print format is 4x6, image
>>cropping has to occur when printing digital images of an original
>>aspect ratio of 3:4. A less than satisfactory alternative, called
>>"crop to fit" as opposed to "crop to fill", is to pad the 3:4 image
>>with white space to obtain the desired aspect ratio.
>> I need to provide intelligent image cropping according to an
>>automatic understanding of the image content and enforcement of
>>compositional rules so that
>>1. The main subject of the image is not cropped in part or in its
>>entirety.
>>2. Both smooth and textured background can be identified and removed
>>in part or in its entirety if necessary.
>>3. Common picture composition rules such as sufficient headroom can be
>>enforced.
>
>
> It sounds like you want a machine to perform an artistic decision.
>
>

If you had a background of extremely high contrast, preferably white or
black, you could use variations of the PostScript infill operator to
crop for you. Based on a chroma key or luma key algorithm.

A somewhat related example with specific code appears at
http://www.tinaja.com/glib/bmfauto1.pdf

Otherwise, cropping can be done in a second or two using free
Imageview32. Tutorials at http://www.tinaja.com/auct01.asp

The general problem of image knockout remains unsolved.
Because intelligent decisions have to be made in murky shadow areas.

No knockout program I have tested works properly, and I do all my
knockouts by hand. Using plain old Paint. Followed by my automatic
backgrounder and auto vignetting routines.

It sounds like the sort of task that definitely should NOT be automated
in ANY manner. Disaster is certain.

--
Many thanks,

Don Lancaster
Synergetics 3860 West First Street Box 809 Thatcher, AZ 85552
voice: (928)428-4073 email: don@tinaja.com

Please visit my GURU's LAIR web site at http://www.tinaja.com

Archived from groups: sci.image.processing,comp.graphics.algorithms,rec.photo.digital (More info?)

A tough problem, but doable at the level of "produces something truly
dumb only once-in-a-while." E.g. the photo printing wizard in Windows
Exporer achieves this MS-esque level of perfection ;-). Unlike MS,
please give the user the option of overriding the automatic choice.

It seems this is a natural problem for a neural network. Your output is
just a single real number? Inputs will need much experimentation.
Normalize RGB magnitudes and then try various convolution filters with
both symmetric and oriented kernels (Laplacian, bandpass, etc.),
perhaps radially averaged power spectra and variance computed on
overlapping regions of varying sizes (Robert Ulichney on halftoning
comes to mind). As usual you will have to fiddle with input resolution,
layers and connectivity, discarding inputs with low weights, etc.
You'll need several thousand "good" croppings for a representative
training data set purturbed in logical ways to make the full training
set. It's a black art. A NN expert (especially a NN signal processing
expert) will have far more insight than me.