Effects of Image Retrieval from Image Database using Linear Kernel and Hellinger Kernel Mapping of SVM

In this paper basically we have compared the efficiency of image retrieval using the most efficient type of classifiers i.e. Support Vector Machine (SVM) with linear kernel mapping and the Hellinger kernel mapping applied to various classes of images and also varied representation of the corresponding image classes using Matlab R2009a.The results obtained from simulation show that Hellinger kernel mapping yields improved performance as compared to the linear kernel mapping. The database consists of a collection of images of different classes whose feature vectors are calculated using Dense Scale Invariant Feature transform (SIFT) and are quantized to visual words whose frequency is recorded in a histogram for each spatial tile of the image. Then the resultant feature vectors are used to train both the linear kernel and Hellinger kernel for different class of images and varied representation of them. The resulting precision and recall graphs and Average Precision (AP) gives us the performance efficiency of various classes of images with varied representation and the classifier mapping used. It is observed from the graphs and AP values that efficiency of the system is increased with Hellinger kernel given more positive images are contained in the database. KeywordsHellinger Kernel, Image Classification, Image Retrieval, Kernel Functions, Linear kernel, SIFT, Support Vector Machine