As the Internet of Things matures the burden of carrying a device for access to information (knowledge) resources will be eliminated.
Instead of making devices carried by a person aware of the location, we are working to make the location aware of the person.
This is a natural progression of the technology, and is need to provide a safe and secure place for diverse populations to interact.
In the modeling of images we may make convenient simplifications (looks symmetric, smoother is better, etc).
Remember: "noise" may just be signals we have yet to identify, and apparent "symmetry" might be an artifact of low resolution and low sample size.
When reviewing papers may be hard to discover the "tuning" parameters and assumptions embedded in the proposed techniques which render the approach of little value.
Joint Image Reconstruction and Segmentation Using the Potts Model. M. Storath. et. al.
1) Embeds a model (in this case piecewise constant) and counts dissimilar neighboring pixels to select which reconstructed image is the best "match" to the given data (radon transform)
2) Creates nice piecewise constant images. Throws out all of the texture (tissue) information which the scanners acquire.
I am completing an approach that uses simple linear algebra to solve (or reduce the number of unknowns) in constraint satisfaction problems. Scilab code for 6x6 Sudoku will be posted soon (9x9 in a bit).
Applications include classification in addition to classical CFP.
Here I will be posting all of my code I have written on wavelets, solving constraint satisfaction problems, and numerical methods. All code is free to use.
I will be posting techniques to for image segmentation. Some things to keep in mind:
- Voxels or Pixels may span a region which intersects multiple classes, thus mixes the signature of this classes and represent a weighted average of the classes.
- Multi-layer segmentation or classification approaches can be used to improve the results
- Atlas based techniques can be used. For example, if you know where there is a "pure" voxel of CSF in the image, then use the sample value and compare to the predicted value returned by the imager.
- For vector valued classifiers, remember, the features which are best for representation are not necessarily the ones best for segmentation.
Started a page on image registration. Need to be a bit careful, many approaches corrupt the dataset right at the beginning, complicating the interpretation of the results.
I updated the Fourier Shape Descriptors abit. It was pointed out that it can be a bit opaque, Thanks Ian!!.
Things to I like to keep in mind: we typically are working with finite samples; and transforms often reduce to matrices being applied to the vector of samples. If the samples of the data are in columns, it is a good idea to look at the rows of the transform. These are the "shapes" you are matching (taking the inner products).
And, the transform you are taking, often assumes something at the boundaries, for example, the data repeats periodically....
Put up the category for Imagery Resources. If you wish to contribute to the Imagery Resources , send me a link to the resources's main page where we determine the restriction on the data use.
I am using Linked In as the contributor public profile page.