Date of Award:


Document Type:


Degree Name:

Doctor of Philosophy (PhD)


Computer Science

Committee Chair(s)

Heng-Da Cheng


Heng-Da Cheng


Lie Zhu


Haitao Wang


Curtis Dyreson


Vicki Allan


The human visual system can automatically emphasize some parts of the image and ignore the other parts when seeing an image or a scene. Visual Saliency Estimation (VSE) aims to imitate this functionality of the human visual system to estimate the degree of human attention attracted by different image regions and locate the salient object. The study of VSE will help us explore the way human visual systems extract objects from an image. It has wide applications, such as robot navigation, video surveillance, object tracking, self-driving, etc.

The current VSE approaches on natural images models generic visual stimuli based on lower-level image features, e.g., locations, local/global contrast, and feature correlation. However, existing models still suffered from some drawbacks. First, these methods fail in the cases when the objects are near the image borders. Second, due to imperfect model assumptions, many methods cannot achieve good results when the images have complicated backgrounds. In this work, I focuses on solving these challenges on the natural images by proposing a new framework with more robust task-related priors, and I apply the framework to low-quality biomedical images.

The new framework formulates VSE on natural images as a quadratic program (QP) problem. It proposes an adaptive center-based bias hypothesis to replace the most common image center-based center-bias, which is much more robust even when the objects are far away from the image center. Second, it models a new smoothness term to force similar color having similar saliency statistics, which is more robust than that based on region dissimilarity when the image has a complicated background or low contrast. The new approach achieves the best performance among 11 latest methods on three public datasets. Three approaches based on the framework by integrating both high-level domain-knowledge and robust low-level saliency assumptions are utilized to imitate the radiologists' attention to detect breast tumors from breast ultrasound images.