Content-Based Image Retrieval via Subspace-Projected Salient Features

In this paper we present a novel image representation method which treats images as frequency histograms of salient features. The histograms are computed making use of linear discriminant analysis (LDA). The method employs saliency feature extraction and image binarisation. Then subspace-projected features are extracted. Using the saliency maps as the positive and negative labels, the image features are mapped onto a lower-dimensional space using LDA. This enables us to construct a 3D-histogram by direct binning on the feature space. This gives rise to a "cube" of image features which have been projected onto a ower-dimensional space so as to maximise the separability of the salient regions with respect to the background. Image retrieval can be performed by computing the distances between the histograms for the query image and the images in the database. We demonstrate our algorithm on a real world database and compare our results to those yielded by codebook representation.