Dimentionality Reduction and Scene Classification Using Multi Kernel Support Vector Machine

Abstract

A scene can be classified by comparing its important features with features stored in the database. It plays an important role in automated surveillance applications including pedestrian detection, indoor positioning, categorizing in semantic classes, etc. The classification of images into semantic categories is tough nowadays; especially for very large categories of the scene, the accuracy of classification is very poor. Therefore, the objective of the proposed work in the thesis is to increase the scene classification accuracy amongst a large number of classes. newlineOne of the challenges in a large class is to handle a huge database and a correspondingly large number of feature sets. Sparse coding has emerged as a powerful tool for feature extraction and classification. Sparse representation has attracted much attention from researchers in fields of signal processing, image processing, and computer vision and pattern recognition. As a lot of elements are zero, it reduces the total computational time taken for operations and memory space required to store the data. The dictionary is a generalized concept to form the basis function in linear algebra. For a dictionary, better localization of transformation provides better sparsity. There are many algorithms to achieve sparse representation. However classic sparse representation algorithms for classification fail to integrate the label information for training images. newlineThe Machine learning is an alternative powerful algorithm used for classification. The binary classifiers are really annoying to deal with when you have too many classes. The Multi-class classifiers work efficiently when having really many classes. There are several classifiers available for this purpose, i.e. Support Vector Machine (SVM), Artificial Neural Network (ANN), etc. The each classifier has its own merits and demerits. The SVM has an L2 Regularization feature. So, it has good generalization capabilities that prevent it from over-fitting and can efficiently handle non-linear data using the Kernel