Dimensionality reduction of SIFT descriptor using vector decomposition for image classification

Abstract

Image classification involves extraction of repeatable and robust feature points from images for classification. These extracted feature points are described in terms of feature vectors. Scale Invariant Feature Transform (SIFT) is one of the popular feature descriptor for obtaining the feature vectors from images for image classification, image retrieval and object recognition. However, the dimension of feature vector obtained using SIFT is high. Hence, the computation time to build decision model using SIFT for image classification, image retrieval and object recognition is high and also requires large memory. In this work, we have proposed SIFT-64 and SIFT-32 to reduce the dimension of the conventional SIFT for image classification. To check the efficacy of the proposed SIFT-64 and SIFT-32, experiments were performed on two well-known publicly available datasets namely, CalTech6 and Graz. The performance is evaluated in terms of classification accuracy, training and testing time. Experimental results demonstrate superior performance of the SIFT-64 and SIFT-32 in comparison to conventional SIFT in terms of both training and testing time without compromising much on classification accuracy. The proposed feature descriptor outperforms the PCA-SIFT descriptor in terms of classification accuracy, training and testing time.