Analysis of SVM kernels for content based image retrieval system

The Content Based Image Retrieval (CBIR) system uses knowledge of computer vision for the research. Nowadays, a lot of images are being generated due to extensive use of smart phone and being spreaded worldwide due to the popularity of social media and high speed internet services. To remove the language ambiguity problem, images are required sophisticated analysis instead of simple textual base analysis and hence CBIR is developed. It is a new way to retrieve images based on color, texture, and shape features of the images. The major challenges of the CBIR systems are the retrieval accuracy and the computational complexity. In this paper, color moments are used due to its small feature vector which will reduce computational complexity. Gray Level Co-Occurrence Matrix is used as texture feature to extract a repetitive pattern of the image. Support Vector Machine (SVM) is used as a classifier and it will remove irrelevant images and hence improve retrieval accuracy. The linear and non-linear SVM classifier is used to predict category of the query images and filter out the irrelevant images. In non-linear SVM classifier, three different kernels: Polynomial, Radial Bases Function (RBF) and Sigmoidal function are used. It is proved that RBF performs good due to its exponential kernel and hence resolve the problem in infinite dimensions. The results of CBIR are compared for linear and nonlinear SVM classifier and also for different fusion techniques in different color space using average precision rate.