| We present a method to automatically construct features capable of putting images in correspondence with each other (i.e. registration) without requiring control points or landmarks on the images in question. It is clear that a technique is required for analyzing the dependence structure inside an image. Typically images have a high degree of correlation, thus having redundant information within the image set. Transforming the set of images into its principal components results in a data set where the resulting information is uncorrelated, thus eliminating the redundancies. Our method is based primarily on Principal Component Analysis (PCA) and a suitable image representation, The feature set is a collection of weight vectors and the correspondence mapping is done using the distances between those vectors. Only a small number of vectors is needed. While we illustrate the approach in the context of image registration, other applications of this method are possible, like distributed sensor networks specifically when a sensor network is build up observing devices, like camcorders, CCDs, etc. Most notably, the method could be used for image matching and retrieval, being insensitive to rotation. |
