Measurement and correction of defocus blur

Blur is a widespread and major source of degradation of imagery. This dissertation has developed a set of algorithms for the measurement and correction of the defocus blur which is modeled by a Gaussian function. The developed algorithms utilize the information contained in the edge neighborhoods and can be implemented using facet models. All the developed algorithms have been tested against real images and produced encouraging results.; We categorize the focus measurement problems into three classes: (a) Determination of the best focused image among a sequence of images; (b) Classification of multiple objects in one image into in-focus and out-of-focus ones; and (c) Blur parameter identification. A new focus criterion function, the energy of gradient of image Laplacian, has been proposed to solve Problem (a) above. When applied to the edge points of an image, this function measures the strength or sharpness of the edges. Therefore it relates defocus blur to edge sharpness. An algorithm has been developed to solve Problem (b) by computing the boundary sharpness of the object. Also, a closed form solution with its two varieties has been developed to solve Problem (c).; Two blur-correction related issues, image enhancement and image deblurring, are discussed in this dissertation. To develop blur correction algorithms, a new edge-sharpening algorithm has been developed using the graylevel gradient first, then extended from 1-D to 2-D. When the blur parameters is known or closely estimated, the blurred edge, which is modeled by a step edge convolved with a Gaussian function, can be reconstructed nearly to the original step edge. In this case the sharpening algorithm becomes a deblurring algorithm.; The facet model services as a useful tool for implementing the developed algorithms. We have proved a theorem stating that the selection of polynomial bases will not affect the least-squares fitting solution. According to this theorem, the facet models with different orders and window sizes can easily be generated and extended from 2-D to 3-D.