The elastic net as visual category representation: Visualisation and classification

Abstract

In this paper we use the Elastic Net (EN) [9] as a visual category representation in feature space. We do this by training the EN on the high dimensional Pyramid Histogram of Visual Words (PHOW) features [2] often used in modern visual categorisation. By employing the topography preserving properties of the EN we visualise the features and draw some novel conclusions. We demonstrate how the EN can also be used as a Region of Interest detector [1]. Finally, inspired by biological vision we propose a new Visual Categorisation scheme that uses ENs as visual category representations. Our method shows promising results when tested on the Caltech101 [12] data set with several interesting future directions. © 2012 Springer-Verlag.