A general algorithm for recognizing small, vague, and imager-alike objects in a nonuniformly illuminated medical diagnostic image

An algorithm for recognizing small, vague, imagery-alike objects in a nonuniformly illuminated image is presented. By small, we mean their size could be only a few pixels, by vague, we mean their boundary could be vaguely defined, and by imagery-alike, we mean their colors are similar and their shapes are overlapped when viewed in a thresholded binary image. The algorithm consists of four major steps: (1) perform an image normalization operation, (2) construct n global-thresholded binary images, (3) design and apply matched filters to the binary images for object recognition, and finally, (4) determine the object class by voting. Without loss of generality, nonuniformly illuminated retinal images are used as an example to describe the algorithm.