Use of Wavelet-based Basis Functions to Extract Rotation Invariant Features for Automatic Image Recognition

In this paper we explore the use of orthogonal functions as generators of representative, compact descriptors of image content. In Image Analysis and Pattern Recognition such descriptors are referred to as image features, and there are some useful properties they should possess such as rotation invariance and the capacity to identify different instances of one class of images. We exemplify our algorithmic methodology using the family of Daubechies wavelets, since they form an orthogonal function set. We benchmark the quality of the image features generated by doing a comparative OCR experiment with three different sets of image features. Our algorithm can use a wide variety of orthogonal functions to generate rotation invariant features, thus providing the flexibility to identify sets of image features that are best suited for the recognition of different classes of images. Key-Words: Rotation Invariant Features, Zernike Moments, Haar Wavelets, Daubechies Wavelets, Orthogonal Functions, OCR