Ranklets: a family of multiscale orentation selective rank features

F. Smeraldi
Queen Mary, University of London


Rank features have long been known to the pattern recognition community as a robust tool suitable for conditions of high noise, low resolution or extreme variability. The popularity of rank features has however been limited by their reputation for being relatively coarse. Arguably one of the main drawbacks of most existing descriptors is the lack of orientation selectivity.

In this talk I introduce a family of rank features named Ranklets, that provide an orientation selective, multiscale representation of the image similar to that obtained with Haar wavelets. Ranklets are defined starting from the Mann-Whitney statistics and admit an intuitive combinatorial interpretation in terms of pairwise comparison of pixel values.
The algorithm can easily be extended to hexagonal pixel lattices for application in embedded systems such as some digital cameras. A recently developed extension of Ranklets based on the Siegel-Tukey statistics allows detection of second-order stimuli; this has been shown to be useful for the processing of visual textures.
Ranklets have been successfully applied to point tracking, face detection, the processing of mammographic images and texture classification among others; I will discuss some of these applications in my talk.

Part of the presentation will be on joint work with Mr George Azzopardi.

back to the list of talks