Date of Award

Summer 2010

Document Type


Degree Name

Doctor of Philosophy (PhD)


Electrical/Computer Engineering

Committee Director

Zia-ur Rahman

Committee Member

Frederic McKenzie

Committee Member

Jiang Li

Committee Member

Roland Lawrence


The unrealistic assumption that noise can be modeled as independent, additive and uniform can lead to problems when edge detection methods are applied to low signal-to-noise ratio (SNR) images. The main reason for this is because the filter scale and the threshold for the gradient are difficult to determine at a regional or local scale when the noise estimate is on a global scale. Therefore, in this dissertation, we attempt to solve these problems by using more than one filter to detect the edges and discarding the global thresholding method in the edge discrimination. The proposed multi-scale edge detection algorithms utilize the multi-scale description to detect and localize edges. Furthermore, instead of using the single default global threshold, a local dynamic threshold is introduced to discriminate between edges and non-edges. The proposed algorithms also perform connectivity analysis on edge maps to ensure that small, disconnected edges are removed. Experiments where the methods are applied to a sequence of images of the same scene with different SNRs show the methods to be robust to noise. Additionally, a new noise reduction algorithm based on the multi-scale edge analysis is proposed. In general, an edge—high frequency information in an image—would be filtered or suppressed after image smoothing. With the help of multi-scale edge detection algorithms, the overall edge structure of the original image could be preserved when only the isolated edge information that represents noise gets filtered out. Experimental results show that this method is robust to high levels of noise, correctly preserving the edges. We also propose a new method for evaluating the performance of edge detection algorithms. It is based on information-theoretic analysis of the edge detection algorithms in the context of an end-to-end visual communication channel. We use the information between the scene and the output of the edge-detection algorithm, ala Shannon, to evaluate the performance. An edge detection algorithm is considered to have high performance only if the information rate from the scene to the edge approaches the maximum possible. Therefore, this information-theoretic analysis becomes a new method to allow comparison between different edge detection operators for a given end-to-end image processing system.