Potential Energy Distance based Image Retrieval
Open Access
- Author:
- Fang, Qi
- Graduate Program:
- Statistics
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- None
- Committee Members:
- Jia Li, Thesis Advisor/Co-Advisor
Le Bao, Thesis Advisor/Co-Advisor
David Russell Hunter, Thesis Advisor/Co-Advisor - Keywords:
- Image Retrieval
Potential Energy Distance - Abstract:
- Due to the large-scale use of digital cameras and easy access to the Internet, the number of images available on the Internet has exploded over the last twenty years and continues to grow. As a result, effective content-based image retrieval approaches are needed to help analyze and understand the large scale of images. Image retrieval approaches depend greatly on similarity measures. There are several popular similarity measures used for image retrieval, for example, Mallows distance and integrated region matching (IRM). Mallows distance is a metric, but IRM is much faster to compute. This thesis introduces a new measure - potential energy distance for image similarity computation. Potential energy distance is also a metric, but is much faster to compute. To evaluate the performance of potential energy distance, we conduct experiments to compare potential energy distance with Mallows distance and IRM. In our experiment, we use the MIR Flickr dataset that contains 25,000 images, and we evaluate different similarity measures from accuracy, speed, and robustness perspectives. Experiment results show that potential energy distance performs similarly to Mallows and IRM in accuracy, similarly to IRM but much faster than Mallows in speed, and fairly robust in image alternations.