I have a segmentation task that could benefit from texture analysis. At
the moment, I can solve it using color (separating out one channel) and
thresholding), but the most obvious distinguishing characteristic to the
human eye is a pattern of particles which have a well defined size . The
spacing between particles is less well defined. Most of the particles are
well separated.
I have considered template matching using a disc of the right size, but I’m
looking for other options.
Are there any well established texture operators available in ImageJ (or
FIJI)?
Original images are about 1000x2000, RGB color. Reducing this to gray scale
by selecting only one channel is easy. I could cut the resolution -
perhaps by a factor of 4 in each dimension. I think I would like a texture
operator which operates on about a 25x25 window and produces a probability
that the window is filled with the desired pattern.
High speed is not a concern. I have only a few 10’s if images to analyze.
The ultimate goal is to combine the color, intensity, and texture
information to provide a probability at each pixel.
I don’t have enough data to consider training a general segmented (but I
might try that) - if only because by the time we produced the samples we
would have solved the original problem.
Similarly, I’m not sure that Fourier analysis is cost effective - but it’s
been 30 years since I’ve tried that - so it’s on my list of thing to try.
I’m looking for a classical texture operator approach using a small number
of kernels, so I can build a n-dimensional descriptor at each point and
classify based on those n-vectors. Of course, I’d prefer to use something
that’s already implemented - before I re-re-re-invent the wheel.
The kernels would need to be about 25x25 at the original resolution, but I
think I can work with something as small as 8x8.
--
ImageJ mailing list:
http://imagej.nih.gov/ij/list.html
Locked
