Local Jet Feature Space Framework for Image Processing and Representation

We present a unified framework for processing and representing images using a feature space related to local similarity. The visual data is represented by the versatile multiscale local jet feature space, possibly reduced by vector quantisation and/or represented by data structures enabling efficient nearest neighbours search (e.g. kd-trees). We demonstrate the interest of the local jet feature space processing through three fundamental low level tasks: noise reduction, motion estimation and background modelling/subtraction. We also show the potential of the framework in terms of higher level visual representation (e.g. recognition/retrieval).