Title

Optimization and Performance of a Template and Histogram based Image Classifier

Presenter Information

James Smigaj

Document Type

Oral Presentation

Location

SURC 202

Start Date

17-5-2012

End Date

17-5-2012

Abstract

Within image classification tasks, many methods involve first extraction of low-level features and then use of these features to build a model that can be identified. This project evaluates one such classifier. The classifier uses the Harris Corner Detection algorithm to extract interest points within an image, and then uses these points to build a template that defines their spatial relationship. To provide additional information, multi-bin color and derivative histograms are used. This classifier contains several parameters that can be adjusted. The template uses a variable distance function; the histogram may vary in the number of bins, and these two components may be weighted differently in the final score. Therefore, appropriate methods are used to optimize these parameters. Finally, the optimized classifier is tested on images of different complexities.

Faculty Mentor(s)

Boris Kovalerchuk

Additional Mentoring Department

Computer Science

This document is currently not available here.

Share

COinS
 
May 17th, 9:10 AM May 17th, 9:30 AM

Optimization and Performance of a Template and Histogram based Image Classifier

SURC 202

Within image classification tasks, many methods involve first extraction of low-level features and then use of these features to build a model that can be identified. This project evaluates one such classifier. The classifier uses the Harris Corner Detection algorithm to extract interest points within an image, and then uses these points to build a template that defines their spatial relationship. To provide additional information, multi-bin color and derivative histograms are used. This classifier contains several parameters that can be adjusted. The template uses a variable distance function; the histogram may vary in the number of bins, and these two components may be weighted differently in the final score. Therefore, appropriate methods are used to optimize these parameters. Finally, the optimized classifier is tested on images of different complexities.