Date of Award

Fall 2014

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical and Computer Engineering

First Advisor

Edward J. Delp

Committee Chair

Edward J. Delp

Committee Member 1

Arif Ghafoor

Committee Member 2

Chih-Chun Wang

Committee Member 3

Jan P. Allebach

Abstract

Visual saliency is the perceptual process that makes attractive objects "stand out" from their surroundings in the low-level human visual system. Visual saliency has been modeled as a preprocessing step of the human visual system for selecting the important visual information from a scene. We investigate bottom-up visual saliency using spectral analysis approaches. We present separate and composite model families that generalize existing frequency domain visual saliency models. We propose several frequency domain visual saliency models to generate saliency maps using new spectrum processing methods and an entropy-based saliency map selection approach. A group of saliency map candidates are then obtained by inverse transform. A final saliency map is selected among the candidates by minimizing the entropy of the saliency map candidates. The proposed models based on the separate and composite model families are also extended to various color spaces. We develop an evaluation tool for benchmarking visual saliency models. Experimental results show that the proposed models are more accurate and efficient than most state-of-the-art visual saliency models in predicting eye fixation.^ We use the above visual saliency models to detect the location of hazardous material (hazmat) signs in complex scenes. We develop a hazmat sign location detection and content recognition system using visual saliency. Saliency maps are employed to extract salient regions that are likely to contain hazmat sign candidates and then use a Fourier descriptor based contour matching method to locate the border of hazmat signs in these regions. This visual saliency based approach is able to increase the accuracy of sign location detection, reduce the number of false positive objects, and speed up the overall image analysis process. We also propose a color recognition method to interpret the color inside the detected hazmat sign. Experimental results show that our proposed hazmat sign location detection method is capable of detecting and recognizing projective distorted, blurred, and shaded hazmat signs at various distances.^ In other work we investigate error concealment for scalable video coding (SVC). When video compressed with SVC is transmitted over loss-prone networks, the decompressed video can suffer severe visual degradation across multiple frames. In order to enhance the visual quality, we propose an inter-layer error concealment method using motion vector averaging and slice interleaving to deal with burst packet losses and error propagation. Experimental results show that the proposed error concealment methods outperform two existing methods.

Download

Share

COinS