MULTI-CLASS SEGMENTATION OF HETEROGENEOUS AREAS IN BIOMEDICAL AND ENVIRONMENTAL IMAGES BASED ON THE ASSESSMENT OF LOCAL EDGE DENSITY

Abstract. Imaging techniques employed in biomedical and ecological applications typically require complex equipment and experimental approaches, including sophisticated multispectral cameras, as well as physical markup of samples, altogether limiting their broad availability. Accordingly, computerized methods allowing to obtain similar information from images obtained in visible light spectrum with reasonable accuracy are of considerable interest. Edge detection methods are commonly used to find discriminating curves in image segmentation. Here we follow an alternative route and employ edge detection results as a separate metric characterizing local structural properties of the image. In turn, their characteristics such as density or orientation averaged in a gliding window are used as a virtual channel substituting multispectral imaging and/or physical markup of samples, and the following image segmentation procedures are performed by thresholding. In complex segmentation scenarios, a single fixed threshold often appears sufficient, and thus relevant adaptive multi-threshold algorithms are of interest, with slope difference distribution (SDD) thresholding algorithm representing a prominent example.