Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical Engineering and Computer Science


Pramod K. Varshney


Computer Vision, Human Visual System, Image Processing, Medical Imaging, Performance Limits, Stochastic Resonance Noise

Subject Categories

Electrical and Computer Engineering


This dissertation investigates the problem of image processing based on stochastic resonance (SR) noise and human visual system (HVS) properties, where several novel frameworks and algorithms for object detection in images, image enhancement and image segmentation as well as the method to estimate the performance limit of image segmentation algorithms are developed.

Object detection in images is a fundamental problem whose goal is to make a decision if the object of interest is present or absent in a given image. We develop a framework and algorithm to enhance the detection performance of suboptimal detectors using SR noise, where we add a suitable dose of noise into the original image data and obtain the performance improvement. Micro-calcification detection is employed in this dissertation as an illustrative example. The comparative experiments with a large number of images verify the efficiency of the presented approach.

Image enhancement plays an important role and is widely used in various vision tasks. We develop two image enhancement approaches. One is based on SR noise, HVS-driven image quality evaluation metrics and the constrained multi-objective optimization (MOO) technique, which aims at refining the existing suboptimal image enhancement methods. Another is based on the selective enhancement framework, under which we develop several image enhancement algorithms. The two approaches are applied to many low quality images, and they outperform many existing enhancement algorithms.

Image segmentation is critical to image analysis. We present two segmentation algorithms driven by HVS properties, where we incorporate the human visual perception factors into the segmentation procedure and encode the prior expectation on the segmentation results into the objective functions through Markov random fields (MRF). Our experimental results show that the presented algorithms achieve higher segmentation accuracy than many representative segmentation and clustering algorithms available in the literature.

Performance limit, or performance bound, is very useful to evaluate different image segmentation algorithms and to analyze the segmentability of the given image content. We formulate image segmentation as a parameter estimation problem and derive a lower bound on the segmentation error, i.e., the mean square error (MSE) of the pixel labels considered in our work, using a modified Cramér-Rao bound (CRB). The derivation is based on the biased estimator assumption, whose reasonability is verified in this dissertation. Experimental results demonstrate the validity of the derived bound.


Open Access