I am looking for an algorithm that could classify (in an unsupervised setting) handwritten symbols (or similar objects) in any orientation or translation. Simply put: if I throw a billion poker cards all over a casino, what would be the best way to sort them if you were a camera attached to a computer?
This would be easy with known alignments of the cards but I seek rotation/translation invariant solutions. I am running short on algorithms. I tried: a) all-to-all-rotation-2D-cross-correlation (works great but it takes more time than the age of the universe), b) sub-sampling by recognizing some features (like density of ink, etc), c) Monte Carlo this and that (and I suspect solution may be somewhere there).
Please do not provide extreme deep learning algorithms (unless they can run on almost a single CPU/GPU or server side & for free [open source] & are fully scalable).
Other than that any solution will be very welcome and properly credited.
Hi !
One algorithm that is doing that is called Temporal Population Code.
You can give it a try :
Article A wavelet-based neural model to optimize and read out a temp...
typically, you would use PCA to fit a parsimonious model to each of the group. This model allows you to compute the distance of an incoming observations to each of the existing groups. Then, the new data is assigned to that cluster to which it is closer. This is, in essence, the SIMCA approach to high dimension classification. It is essentially a PCA based approach.
You would also want to use a procedure that can handle outliers in the data.
we compare a bunch of PCA algorithms that can handle outliers here:
https://www.researchgate.net/publication/260211852_Finding_High-dimensional_Outliers_with_FastHCS
the algorithm that woks best in the bunch is also embarrassingly parallel (we actually ran the heavy simulation study in amazon's EC2 cloud). A free, c++ open source implementation of the algorithm is available via CRAN (we distribute it through the R infrastructure but it is essentially stand alone code)
Article The FastHCS algorithm for robust PCA
You might try the following:
A biologically plausible transform for visual recognition that is invariant to translation, scale, and rotation. Pavel Sountsov, David M. Santucci and John E. Lisman, Frontiers in Computational Neuroscience, 5(53), Nov 2011, DOI: 10.3389/fncom.2011.00053
- Badges
- Science topic