Image parsing using local features of superpixels and genetic algorithm
Issued Date
2014
Issued Date (B.E.)
2557
Available Date
Copyright Date
Resource Type
Series
Edition
Language
eng
File Type
application/pdf
No. of Pages/File Size
91 leaves
ISBN
ISSN
eISSN
Other identifier(s)
b192204
Identifier(s)
Access Rights
Access Status
Rights
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Rights Holder(s)
Physical Location
National Institute of Development Administration. Library and Information Center
Bibliographic Citation
Citation
Joseph, Ferdin Joe John (2014). Image parsing using local features of superpixels and genetic algorithm. Retrieved from: http://repository.nida.ac.th/handle/662723737/4127.
Title
Image parsing using local features of superpixels and genetic algorithm
Alternative Title(s)
Author(s)
Advisor(s)
Editor(s)
item.page.dc.contrubutor.advisor
Advisor's email
Contributor(s)
Contributor(s)
Abstract
Image porsing is a new avenue in the field of Computer Vision and Machine Learning. Similar to parsing of text into meaningful tokens and classifying them based on the user discretions, images are also parsed to understand the contents of the image. In this regard, there lies a need for image parsing systems to parse the images and understand the contents of the images in an efficient way. In the recent past various image parsing methodologies are proposed with huge fcaturc sct and ambiguity in performance based on class and pixels detected. Many existing methodologies usc thc parumeters connected with the features of pairs of superpixels, a standard set of features extracted from a superpixel. But the accuracy of images parsed and understood with respect to class and pixels is yet to be improved. This gives a need to device a new methodology to improve the classification accuracy with respect to class and pixels
In this reaseach, a new feature representation is proposed to better representan object in an image and so enhance the classification accuracy for the object. Each superpixel in the object area is assigned with a visual word id based on its superpixel colors and textures. All two grams of visual word ids of neighborings of superpixels in clockwise direction are counted to create an histrogram of the two grams for the object. This histrogram can represent the unique composition of superpixels within the object. Using the histograms and the shape features of training objects, object classification models based on Support Vector Machines are built for each class of objects
To parse an image, the image is first segmented into regions and a multiobjective genetic algorithm is used to do a search for the set of the image regions that would best form an object instance based on two objectives, i.e. the SVM object classification score and the size of the object. The size is included to avoid the premature convergence to local optima that represent only parts of the object Crowding is also incorporated into the genetic algorithm in order to simultaneously search for multiple solutions, cach represents an instance of same class objects present in the image. The experimental results in numerals as well as the quantitative results show that the proposed methodology is well ahead in accuracy than most of the existing methodologies. This methodology is also suitable to adapt for any real time applications which need image parsing
In this reaseach, a new feature representation is proposed to better representan object in an image and so enhance the classification accuracy for the object. Each superpixel in the object area is assigned with a visual word id based on its superpixel colors and textures. All two grams of visual word ids of neighborings of superpixels in clockwise direction are counted to create an histrogram of the two grams for the object. This histrogram can represent the unique composition of superpixels within the object. Using the histograms and the shape features of training objects, object classification models based on Support Vector Machines are built for each class of objects
To parse an image, the image is first segmented into regions and a multiobjective genetic algorithm is used to do a search for the set of the image regions that would best form an object instance based on two objectives, i.e. the SVM object classification score and the size of the object. The size is included to avoid the premature convergence to local optima that represent only parts of the object Crowding is also incorporated into the genetic algorithm in order to simultaneously search for multiple solutions, cach represents an instance of same class objects present in the image. The experimental results in numerals as well as the quantitative results show that the proposed methodology is well ahead in accuracy than most of the existing methodologies. This methodology is also suitable to adapt for any real time applications which need image parsing
Table of contents
Description
Dissertation (Ph.D. (Computer Science and Information Systems))--National Institute of Development Administration, 2014