Object Identification and matching in Images using SIFT features for Social Networking applications

Abstract

Our project aims at merging image processing and machine learning techniques to distinguish, in particular, famous buildings and landmarks of the world from one another. The challenge we faced was that intrinsically buildings belong to a single class so it is difficult to differentiate between them due to less intra class variability. The flow of our project is that first we use SIFT (Scale Invariant feature Transform) algorithm to extract low level features of large number of buildings of the world each forming a different class. After extracting features we use Bag of words technique which involves K mean clustering in order to find clusters and making a vocabulary file (dictionary). This file is used to construct histograms for each image of every class by assigning each feature to its nearest cluster and finding numerical value for occurrence of each cluster. These Histograms are used to train our system using SVM (support vector machine) technique .Our system predicts the class of a new input image, which is not from training images, by extracting its features and constructing a histogram. The classifier will finally indicate which class the unknown image falls into once the histogram is plotted on “n dimensional plane