James Kang

Abstract

Prior research on novelty detection has primarily focused on algorithms to "detect" novelty for a given application domain. Effective storage, indexing and retrieval of novel events (beyond detection) are largely ignored as a problem in itself. In light of the recent advances in counter-terrorism efforts and link discovery initiatives, the need for effective data management of novel events assumes apparent importance. Automatically detecting novel events in video data streams is an extremely challenging task. The aim of this thesis is to provide evidence to the fact that the notion of novelty in video as perceived by a human is extremely subjective and therefore algorithmically illdefined. Though it comes as no surprise that current machine-based parametric learning systems to accurately mimic human novelty perception are far from perfect such systems have recently been very successful in exhaustively capturing novelty in video once the novelty function is well-defined by a human expert. So, how truly effective are these machine based novelty detection systems as compared to human novelty detection? In this paper we outline an experimental evaluation of the human vs machine based novelty systems in terms of qualitative performance. We then quantify this evaluation using a variety of metrics based on location of novel events, number of novel events found in the video, etc. We begin by describing a machine-based system for detecting novel events in video data streams. We then discuss the issues of designing an indexing-strategy or "Manga" (comic-book representation is termed as "manga" in Japanese) to effectively determine the "most-representative" novel frames for a video sequence. We then evaluate the performance of machine-based novelty detection system against human novelty detection and present the results. The distance metrics we suggest for novelty comparison may eventually aide a variety of end-users to effectively drive the indexing, retrieval and analysis of large video databases. It should also be noted that the techniques we describe in this paper are based on low-level features extracted from video such as color, intensity and focus of attention. The video processing component does not include any semantic processing such as object detection in video for this framework. We conjecture that such advances, though beyond the scope of this particular paper, would undoubtedly benefit the machine-based novelty detection systems and experimentally validate this. We believe that developing a novelty detection system that works in conjunction with the human expert will lead to a more user-centered data mining approach for such domains. JPEG 2000 is a new method of compressing images better than other image formats such as JPEG, GIF, PNG, etc. The main reason this format is in need for investigation is it allows metadata to be embedded within the image itself. The types of data can essentially be anything such as text, audio, video, images, etc. Currently image annotations are stored and collected side by side. Even though this method is very common, it brings up a lot of risks and flaws. Imagine if medical images were annotated by doctors to describe a tumor within the brain, then suddenly some of the annotations are lost. Without these annotations, the images itself would be useless. By embedding these annotations within the image will guarentee that the description and the image will never be seperated. The metadata embedded within the image has no influence to the image iteself. In this thesis we initially develop a metric to index novelty by comparing it to traditional indexing techniques and to human perception. In the second phase of this thesis, we will investigate the new emerging technology of JPEG 2000 and show that novelty stored in this format will outperform traditional image structures. One of the contributions this thesis is making is to develop metrics to measure the performance and quality between the query results of JPEG 2000 and traditional image formats. Since JPEG 2000 is a new technology, there are no existing metrics to measure this type of performance with traditional images.

Library of Congress Subject Headings

Data mining; Information storage and retrieval systems; Pattern recognition systems; Video recordings--Data processing; JPEG (Image coding standard)

Publication Date

2005

Document Type

Thesis

Department, Program, or Center

Computer Science (GCCIS)

Advisor

Teredesai, Ankur - Chair

Advisor/Committee Member

Misic, Vladimir

Advisor/Committee Member

Gaborski, Roger

Comments

Note: imported from RIT’s Digital Media Library running on DSpace to RIT Scholar Works. Physical copy available through RIT's The Wallace Library at: QA76.9.D343 K26 2005

Campus

RIT – Main Campus

Download

Share

COinS