Cardiff University | Prifysgol Caerdydd ORCA
Online Research @ Cardiff 
WelshClear Cookie - decide language by browser settings

Multi-task feature selection for advancing performance of image segmentation

Liu, Han and Zhao, Huihuang 2018. Multi-task feature selection for advancing performance of image segmentation. Presented at: International Conference on Machine Learning and Cybernetics, Chengdu, China, 15-18 July 2018. 2018 International Conference on Wavelet Analysis and Pattern Recognition (ICWAPR). Piscataway, New Jersey: IEEE, pp. 244-249. 10.1109/ICWAPR.2018.8521328

This is the latest version of this item.

[img]
Preview
PDF - Accepted Post-Print Version
Download (147kB) | Preview

Abstract

Image segmentation is a popular application area of machine learning. In this context, each target region drawn from an image is defined as a class towards recognition of instances that belong to this region (class). In order to train classifiers that recognize the target region to which an instance belongs, it is important to extract and select features relevant to the region. In traditional machine learning, all features extracted from different regions are simply used together to form a single feature set for training classifiers, and feature selection is usually designed to evaluate the capability of each feature or feature subset in discriminating one class from other classes. However, it is possible that some features are only relevant to one class but irrelevant to all the other classes. From this point of view, it is necessary to undertake feature selection for each specific class, i.e, a relevant feature subset is selected for each specific class. In this paper, we propose the so-called multi-task feature selection approach for identifying features relevant to each target region towards effective image segmentation. This way of feature selection requires to transform a multi-class classification task into $n$ binary classification tasks, where $n$ is the number of classes. In particular, the Prism algorithm is used to produce a set of rules for class specific feature selection and the K nearest neighbour algorithm is used for training a classifier on a feature subset selected for each class. The experimental results show that the multi-task feature selection approach leads to an significant improvement of classification performance comparing with traditional feature selection approaches.

Item Type: Conference or Workshop Item (Paper)
Date Type: Publication
Status: Published
Schools: Computer Science & Informatics
Publisher: IEEE
ISBN: 978-1-5386-5218-3
ISSN: 2158-5709
Date of First Compliant Deposit: 8 May 2018
Date of Acceptance: 8 May 2018
Last Modified: 15 Oct 2020 01:34
URI: http://orca-mwe.cf.ac.uk/id/eprint/119822

Available Versions of this Item

Actions (repository staff only)

Edit Item Edit Item

Downloads

Downloads per month over past year

View more statistics