11.45am,Lecture Hall 4
M.Sc Thesis Defence
Bispectrum invariants for learning images
Chennai Mathematical Institute.
In this thesis, we study the PhD work of Risi Kondor in which he demonstrates how bispectrum invariants can be used to classify translated and rotated images. The scope of this work is to gain the necessary theoretical background, repeat the experiments on MNIST dataset done by Kondor and perform some other experiments in order to get a better understanding of the features and if possible improve Kondor's results. The basic idea is to map the image in 2-dimensional space to a unit sphere. Then a translation or a rotation of the input image is given by an element of the group SO(3) acting on the (image wrapped about the) sphere. Kondor shows that the bispectrum of the image yields a translation and rotation invariant feature vector, which can be used to classify digits.