Recently we announced Treadco – bespoke machine learning.
It looks like software in general will be our product.
My … collaborator has a rather rough way of expressing himself. It made me blush, but he said it was time to pull something out of somewhere. As I said, he’s coarse and uncouth.
Anyway, our first open source teaser product will sharpen images. We’ve discovered the joys of numpy and python – Fourier transforms that use the GPU and handle normalization like a breeze. It means we can easily, in something like 30 lines of code, implement a regularized Jacobi solver to find the image and point spread function that provides an optimal sharpening for a degraded image.
The application we’re looking at is in cryo-EM, but my collaborator mutters about stupid biologists and confocal microscopy. (I’ll have to get him drunk and find out what he’s pissed about. It must be something – he’s usually straightforward about these things.)
For those of us who remember our numerical analysis (a rapidly vanishing group of computer scientists) Jacobi iterations are an easy way to solve large matrix problems. It’s not quite as efficient as Gauss-Siedel and SOR, but it can be implemented with the Fourier transform. Technically, regularization is a successive under-relaxation rather than over-relaxation. Most importantly, it’s a block iteration and can be implemented with high efficiency on all sorts of computer hardware. Careful regularization of the iterations, using noise levels in the image (the noise estimates don’t have to be stationary) results in an excellent and stable sharpening of the image. Without all that dangerous messing about with unsharp masks and inverse filters.
Anyway, Treadco has opened a github repository (empty for the moment) and is starting to produce.
By the way, there is no connection, none, absolutely none, with any Georgia State University resources.