Enhancement of object classification via diversifying features

• Main Author: Tan, Yang Zhi.
• Other Authors: Teoh Eam Khwang
• Format: Final Year Project
• Language: English
• Published: 2011
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/46006
• Institution: Nanyang Technological University

Object classification has been an active research field for many decades due to its various applications in computer vision. However, only recently there has been an interest in classifying objects that belong to the same family. Normal methods of feature extraction and classification are proven to have high rates of error. The need to improve the performance of family object classification arises due to the fact that no single feature representation can truly represent all the family traits of an object. The object referring to in this project is a car, and its family is the brand of the car. The objective of this project is to focus on the fusion of classifier outputs similar to a mixture of experts system to increase the performance of modern day family object classification. This final year project proposes the use of various feature representations to capture the family traits of a car. Amongst the various types of features that are available, Gabor, SIFT and LBP features are chosen to capture the salient family traits that are present in the car image. Gabor features were chosen because they are robust against illumination and slight pose changes. SIFT features capture the shape, contour details of the car. Lastly LBP proves to be useful in representing textures. Classifiers such as SVM and AdaBoost are used for the classification of the features. The robustness of this aforementioned system was tested on 3 sets of brands, which are BMW, Audi and Mercedes. It was shown that fusion of 18 segments proved superior to 27 segments, whilst reducing the FNR by 20%.