A content-based image retrieval using PCA and SOM

Abstract

Image search engines have progressed to allow an efficient retrieval. A common trend consists in the construction of a visual vocabulary, in order to apply the BOW model for image indexing. In this paper, we proposed an approach to build an efficient visual vocabulary: First, the feature space composed of SIFT descriptors is transformed into a lower-dimensional space using the Principal Component Analysis (PCA). Second, the resulting feature space is clustered using the Self Organising Map (SOM) and it results in a map of visual words. The proposed model, called PCA-SOM, is evaluated using a dataset of vehicle images from Pascal VOC 2007 benchmark and the experiments show encouraging results.