Author

Navid Serrano

Abstract

The advent and wide acceptance of digital imaging technology has motivated an upsurge in research focused on managing the ever-growing number of digital images. Current research in image manipulation represents a general shift in the field of computer vision from traditional image analysis based on low-level features (e.g. color and texture) to semantic scene understanding based on high-level features (e.g. grass and sky). One particular area of investigation is scene categorization, where the organization of a large number of images is treated as a classification problem. Generally, the classification involves mapping a set of traditional low-level features to semantically meaningful categories, such as indoor and outdoor scenes, using a classifier engine. Successful indoor/outdoor scene categorization is beneficial to a number of image manipulation applications, as indoor and outdoor scenes represent among the most general scene types. In content-based image retrieval, for example, a query for a scene containing a sunset can be restricted to images in the database pre-categorized as outdoor scenes. Also, in image enhancement, categorization of a scene as indoor vs. outdoor can lead to improved color balancing and tone reproduction. Prior research in scene classification has shown that high-level information can, in fact, be inferred from low-level image features. Classification rates of roughly 90% have been reported using low-level features to predict indoor scenes vs. outdoor scenes. However, the high classification rates are often achieved by using computationally expensive, high-dimensional feature sets, thus limiting the practical implementation of such systems. To address this problem, a low complexity, low-dimensional feature set was extracted in a variety of configurations in the work presented here. Due to their excellent generalization performance, Support Vector Machines (SVMs) were used to manage the tradeoff between reduced dimensionality and increased classification accuracy. It was determined that features extracted from image subblocks, as opposed to the full image, can yield better classification rates when combined in a second stage. In particular, applying SVMs in two stages led to an indoor/outdoor classification accuracy of 90.2% on a large database of consumer photographs provided by Kodak. Finally, it was also shown that low-level and semantic features can be integrated efficiently using Bayesian networks for increased accuracy. Specifically, the integration of grass and sky semantic features with color and texture low-level features increased the indoor/outdoor classification rate to 92.8% on the same database of images.

Library of Congress Subject Headings

Image processing--Digital techniques; Computer algorithms; Images, Photographic--Classification

Publication Date

5-1-2002

Document Type

Thesis

Department, Program, or Center

Electrical Engineering (KGCOE)

Advisor

Dianat, Sohail

Advisor/Committee Member

Rao, Raghuveer

Advisor/Committee Member

Luo, Jiebo

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: TA1637 .S477 2002

Campus

RIT – Main Campus

Share

COinS