Parameterizing SIFT and Sparse Dictionary for SVM Based Multi-class Object Classification

Abstract

To build a general-purpose object recognizer, capable to recognize many different classes of objects is the most challenging task. Spatial pyramid matching (SPM), an extension and revision of bag-of-feature (BoF) computes histograms of native features at various levels of resolution. The support vector machine (SVM) using SPM gained large popularity in object classification. However, its uses are trivial due to the large computational cost and limited classification accuracy for very large classes of objects. This paper presents an extension of SPM based on tuneable SIFT sparse code using multi-class SVM to enhance the performance by encoding the SIFT features into sparse code and tuning its parameters. The various experimental studies are presented in the paper to investigate the effect of SIFT parameters to achieve better sparsity in the dictionary. During the number of experimentation, parameter tuning in SIFT feature extraction with linear-kernel SVM has improved the recognition accuracy for the datasets having a large number of classes and a huge number of images for each class. The comparison with state-of-art concludes that optimum tuning of the parameters of SIFT can minimize the feature vector size reducing the computational cost and achieving higher classification accuracy.