On discrimination functions for image feature extraction
Suresh K. Hungenahally · Griffith Research Online · 1997
Despite significant advances in Computer Vision during the last few decades, our understanding of the underpinnings of visual perception remains primitive. The research on computational studies of visual perception has been primarily based on the design and development of edge and feature detection techniques. These techniques are crucial to higher level algorithms for object recognition and feature extraction. Inspired by the neuro-morphology of visual receptive fields, a number of filtering techniques have been proposed in the literature on image processing notably the most influential work by Marr [1980,82], Levine [1985] and Canny [1986]. However, none of the previously proposed techniques provide a satisfactory solution to problems such as detecting blurred edges. It is possible to increase edge gradients but the existing techniques usually do not produce step edges, processing noise sensitive image data is extremely difficult, and the problem with edge maps that the context or background of the edges is lost-also pointed out in the scale -space filtering problem by Witkin, still remains broadly unsolved. An attempt has been made in this thesis to address these problems by defining a family of discrimination functions and implementing them as filters which selectively extract local information from the image while preserving its global features. A method of image decomposition and contextual filtering by treating the original image as an embedded set of images processed by a family of three -parameter (shape, size and order) functions is proposed. The functions developed in this thesis perform global scale tuning in conjunction with local feature enhancement thus enabling identification and localization of visual information in a natural fashion. Further, we introduce the notion of fractional discrimination functions by defining a continuous order of discrimination. Precisely, the contributions of this thesis are: Definition of a generalized framework for a class of functions and operators to perform robust (in the presence of noise), selective (band limited) and contextual feature enhancement. Definition of a family of multi -dimensional discrimination and fractional discrimination functions which provide uniform treatment of the existing functions and operators for edge and feature enhancement. Implementation of discrimination functions in the KHOROS visual programming environment using "C", and testing on simulated images with artificially added noise. Experimental analysis indicates that our approach is robust under noisy conditions. Testing of properties of selective feature extraction on texture images and contextual feature enhancement has been tested on medical images. The test yielded good texture detection in the presence of noise and texture variation and successful enhancement of bones, tissue matter and fluids and cavities for medical images.