I=
Pakistan Research Repository Home
 

Title of Thesis

Feature Selection for Image Classification

Author(s)

Nadeem Abbas Zaidi

Institute/University/Department Details
Department of Electrical Engineering / University of Engineering & Technology, Lahore
Session
2009
Subject
Electrical Engineering
Number of Pages
150
Keywords (Extracted from title, table of contents and abstract of thesis)
Feature, Selection, Image, Classification

Abstract
Image classification has earned enormous attention due to the advent of modem day applications involving image base information and now an extensive research has been carried out in this field. Developing computationally efficient algorithms for image classification without compromising the classification accuracy is of primary importance. A typical approach to the image classification problem consists of pre-processing. feature extraction. feature selection and classifier design. Pre- processing includes segmentation of an image to separate individual objects from an image, removing of any erroneous noise from an image, image cropping, image normalization and thinning etc. Feature extraction means finding out various attributes as well as characteristics of objects found within an image. These characteristics, usually called features are used to describe an object. Selection of features is probably the most important factor in achieving high recognition rite. The extracted features must be invariant to the distortions and variations such as translational, rotational and scale of the objects. These selected features from the objects are used as an input to the classifier.
Efficient feature selection for image classification is one such component of a classification algorithm that has received considerable attention. In this dissertation, we propose to adopt a framework of sparse representations to address the problem of feature selection for image classification. A new cost function for sparse feature selection is constructed using two major components namely: sparsity and Fisher's discrimination power. Three algorithms are proposed for the solution of sparse representation problem. The algorithm I is inspired by Orthogonal Matching Pursuit in which we choose the best basis function in each iteration from the overcomplete dictionary that maximized the cost function. Algorithm 2 is the generalized form of algorithm I which gives the global solution of the problem. The solution of proposed cost function is presented in algorithm 3. The performance of the algorithms are quantified both in terms of the Fisher's discrimination power and classification accuracy for different number of selected features. LibSVM classifier is used for classification. Efficiency and robustness of the developed algorithm is demonstrated through experiments with COIL-20 damsel with different noise and occlusion levels. A comparison between the cost function with and without reconstruction error is also carried out to illustrate the importance of discrimination power with respect to reconstruction error in classification applications.
In recent years, filter bank approaches including wavelets and overcomplete multiresolution dictionaries have been used extensively in image classification. The key issue in filter bank based classification methods is the selection of the best and the moist compact representation out of the feature set generated by the overcomplete dictionary. In feature selection for classification, two important factors are separability between the different classes and the compactness or sparsity of the selected feature set. The formulation of the proposed sparse representation for image classification method is further improved by using the separability for the measure of discrimination. In this dissertation, we propose a new approach for feature selection in the framework of sparse representations by combining the separability and sparseness into a single cost function. Orthogonal Matching Pursuit algorithm is employed to extract a sparse set of discriminative features and LibSVM classifier is used for the consequent classification. Efficiency and robustness of the developed algorithm is demonstrated through experiments with different image databases and different noise and occlusion levels. The proposed algorithm is compared Shih Fu Chang algorithm A comparison between the proposed cost function and a conventional cost function that only considers the energy of the selected features is also presented to illustrate the improvement in performance. The proposed cost function is offered high accuracy with a small number of features.

Download Full Thesis
7,637 KB
S. No. Chapter Title of the Chapters Page Size (KB)
1 0 CONTENTS

 

ii
100 KB
2

1

INTRODUCTION 1
750 KB
3 2 BACKGROUND 13
794 KB
4 3 SPARSE AND DISCRIMINATIVE REPRESENTATION FOR IMAGE CLASSIFICATION 28
2,323 KB
5 4 SEPARABILITY BASED FEATURE SELECTION FOR IMAGE CLASSIFICATION 73
1,838 KB
6 5 CONCLUSION AND FUTURE WORK 109
416 KB
7 6 BIBLIOGRAPHY 116
686 KB