1. UKM Learning and Research Repository
2. Master Thesis / Tesis Master
3. Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Title:	Ensembles of filter based classifiers for visual object categorization
Authors:	Inas A.A. Yosif (P50083)
Supervisor:	Azizi Abdullah, Dr.
Keywords:	Image processing -- Digital techniques Pattern recognition systems Computer vision Image processing -- Mathematics Universiti Kebangsaan Malaysia -- Dissertations Dissertations, Academic -- Malaysia
Issue Date:	13-Jun-2012
Description:	Object recognition systems need effective descriptor classifiers to achieve good performance levels. Based on literature, edge and orientation information are important and widely used in describing objects. They are represented by a histogram of size N bins and computed by using color or/and intensity signals changes in a certain angular range. One problem in using a single filter is the stronger filter responds to edge structures the more sensitive it to orientation. The filters that are sensitive to orientation tend to respond to non-edge structures. Thus, a set of different filters are needed to efficiently describe images such as compass filters namely Robinson filters. The filters contain several basic filters that allow measuring of different local edge orientation strength. Another possible problem in combining the different filter based descriptors using the naive approach is it can lead to increase over-fitting and prevent generalization performance. For these reasons, the research is to propose the ensemble learning algorithms of Support Vector Machines (SVMs) by combining multiple edge and orientation descriptors based on several basic primitive filters. The methodology consists of two main parts. The first part provides a low-level description for describing images. In this part, the proposed method starts with constructing several feature maps of the input image. The maps are constructed by convolving the image data using a set of M different filters namely Robinson filters resulting M different feature maps. After that, the edge and orientation descriptors namely edge histogram, histograms of oriented gradients and Scale Invariant Feature Transform on fixed grid are used to describe the maps for indexing. The second part provides the image classification task. In this part, the support vector machine algorithm is used to train models of different filter based descriptors. After that, all models are used to compute the right output class by using the class probabilities of all trained support vector models. In this case, single or individual classifier, naive classifier and ensemble learning classifiers namely product rule, mean rule and majority voting rule are used to measure the recognition performance. Experimental results on 20 classes of the Caltech-101 object dataset show that the ensemble methods outperform single and naive approaches at about 71.23% of accuracy rates.,“Certification of Master’s/Doctoral Thesis” is not available,Master of Information Technology
Pages:	94
Call Number:	TA1637.5.Y647 2012 3 tesis
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.