Framework for extracting rotation invariant features for image classification and an application using haar wavelets

We propose a general framework for extracting rotation invariant features from images for the tasks of image analysis and classification. Our framework is inspired in the form of the Zernike set of orthogonal functions. It provides a way to use a set of one-dimensional functions to form an orthogonal set over the unit disk by non-linearly scaling its domain, and then associating it an exponential term. When the images are projected into the subspace created with the proposed framework, the rotations in the image affect only the exponential term while the value of the orthogonal functions serve as rotation invariant features. We exemplify our framework using the Haar wavelet functions to extract features from several thousand images of symbols. We then use the features in an OCR experiment to demonstrate the robustness of the method.