Object Recognition Using SVM Based Bag of Combined Features

Abstract

The evolving and task-dependent nature of visual categories of problems prompts for an example based solutions involving machine learning approach. One such technique, ‘Bag of features’ approach has gained popularity in computer vision applications, including texture recognition, image classification and robot localization. Despite being quite newer, system based on bag of feature method has set performance standards on benchmark of image classification and has achieved drastic scalability in image retrieval. In this paper, a novel approach is being presented for object recognition using combined bag of features and multiclass Support Vector Machine (SVM). Proposed approach presents SVM based combined bag of features method and emphasizes on better recognition and more classification accuracy from images dataset. Although nearest neighbor classifiers have been employed in this area previously, but they suffer from high variance problem in case of limited sampling. Root-SIFT and SURF descriptors are combined for the construction of combined set of bag features, which has reasonable computational complexity yielding excellent results. A Caltech dataset; considered to be very challenging database because of objects are embedded in clutter background having different poses and scales; has been used for testing efficacy of approach. Comparison, made among state-of-the-art approaches shows promising result. Our experiments show state-of-the-art performance on benchmark dataset for object recognition. On Caltech dataset, we achieved a correct classification rate of 65 and 72 at 15 and 30 training images.