Colloquium "Feature descriptors based on Local Pattern Spectra (LPS)"

Abstract:

Global shape-size pattern spectra can be seen as 2D histograms of the image component distribution over a number of shape and size classes. Prompted by their previous successful application to image retrieval as global descriptors, a local variant is introduced, suited for calculation on arbitrary connected regions. Pattern spectra can be efficiently calculated on a mathematical morphology structure called the max-tree, using a technique called granulometry. For computational efficiency, the MSER detector which can also be implemented on a max-tree structure is used. The local variant of the pattern spectra is then combined with normalized region moments and used as a Local Pattern Spectra (LPS) descriptor for the MSER regions. The presentation will cover the max-tree construction on the image, as well as its use for MSER detection and calculating the granulometries, including the attribute choice for quantifying shape and size classes. Challenges faced when adapting the global variant for local calculation and the parameter choices for LPS are discussed. Preliminary validation is done using a simple image retrieval system performing exact search (using no aggregation scheme) on the UCID database, and the performance compared to that of SIFT descriptors.

CV:
Petra Bosilj received her B.Sc. and M.Sc. degrees from the Faculty of Electrical Engineering and Computing, University of Zagreb, Croatia. The focus of her thesis was in computer vision and image processing, more specifically achieving coarse robot localization using content- based image retrieval techniques within a system for robot navigation. She is currently pursuing her Ph.D. degree at University of South Brittany, France, working with both with OBELIX team associated with IRISA Vannes and LinkMedia (previously TEXMEX) team of IRISA/INRIA Rennes. A portion of the research towards her thesis was done during a collaborative visit to the Johann Bernoulli Institute of the University of Groningen, the Netherlands. The goal of her Ph.D. project is applying hierarchical image models from mathematical morphology to image retrieval problems.