A model-independent method for local blur estimation and its application to edge detection

Knowledge is hidden in images in form of objects, structures, patterns and their relationships, which are acquired through devices associated with various artifacts including blurring and noise. This paper presents a model-independent method for local blur-scale estimation based on a novel hypothesis that gradients inside a blur-scale region follow a Gaussian distribution with non-zero mean. New statistical test criteria involving maximal likelihood functions are presented to test the hypothesis and applied for blur-scale estimation. Also, the applications of blur-scale for scale-based gradient and edge computation are presented. In the context of scale-based edge computation, new methods are introduced to suppress false gradient maxima avoiding double edging artifacts. New methods are examined on computer-generated as well as real-life images with varying blur and noise. Experimental results show that computed blur-scale using the new algorithm is accurate (r = 0.95) and scale-based gradients are visually satisfactory at both sharp as well as blurred edge locations. Performance of the new edge detection algorithm is quantitatively examined and compared with two popular methods, and the results show that, at various contrast-to-noise ratio, the new method is superior to the others in terms of overall accuracy (92 to 96%), true edge detection (96 to 98%), and false edge reduction (93 to 100%).