Robust and automated unimodal histogram thresholding and potential applications

In this paper, three new histogram-based algorithms are presented to segment images expressing unimodal intensity histograms. These algorithms are applied to laser scanning confocal microscope images (known to often exhibit unimodal histograms) to identify fluorescent signals, and other applications are also shown. The first algorithm facilitates linear diffusion to investigate dynamic histogram features in scale-space. The second algorithm is based on a histogram comparison between a reference area and the whole image at reduced scale. The third algorithm uses the maximisation of a between-class variance criterion applied to image histograms. Results obtained from automatic thresholding of confocal microscopy images show good agreement between the algorithms. Further applications to segment other images are also shown. (C) 2004 Pattern Recognition Society. Published by Elsevier Ltd. All rights reserved.