Tensor-based representations and Boosted Kernel Learning for Content-Based Image Categorisation

When?

Tuesday 11 December 2012, 14:00 to 15:00

Where?

CVSSP Seminar Room (40bAB05)

Open to:

Public, Staff, Students

Speaker:

Prof Philippe Gosselin, ENSEA, Paris, France

Abstract:

This talk presents methods for content-based image indexing and learning techniques for image categorisation in databases. The first part of the presentation will address methods for the indexing of images thanks to their visual content. After a short history of matching-based techniques of low-level visual descriptors, we will present a method based on tensor called Vectors of Locally Aggregated Tensors (VLAT). This method is an extension of a method based on visual dictionaries called VLAD. In the case of VLAT, we also consider visual dictionaries, however the metric to compare distribution of descriptors in each cluster is a linearisation of Lyu's kernel function on bags. This leads to better retrieval results, but also to very large signatures. In order to reduce the size of these signatures, we will present two solutions that, in addition to the dimension reduction, also increase retrieval results. Experimental results will be presented and compared to state-of-the-art in dictionary-based methods. For similarity search we will consider a one million database, and for categorisation we will consider the VOC2007 benchmark.

The second part of the presentation will address the learning of categories thanks to existing image signatures, like the ones we present in the first part. More specifically, we will present a method that combines two popular learning frameworks in image indexing : Boosting and Kernel Machines/SVM. The aim of this method is to first learn a linear combination of Mercer's kernel function, and then use this combination with SVM classifiers to learn the categories. However, on the contrary to Multiple Kernel Learning (MKL) techniques, we learn the combination thanks to a Boosting approach. In this scope, we don't consider a few complex kernel functions, but a large number of "weak kernels" : basic but easy to design kernel functions. Furthermore, we propose an iterative learning algorithm with a complexity linear with the size of the training database, and show a proof of its convergence. Experimental results will be presented and compared to state-of-the-art in MKL thanks to the Oxford Flowers databases.