Oversegmentation reduction in watershed-based gray-level image segmentation

Gray-level image segmentation is the first task for any image analysis process, and is concerned with the identification of the objects of interest in a digital image. In this paper, we suggest a segmentation technique based on the use of watershed transformation to obtain a preliminary partition of the input gray-level image into regions, homogeneous with respect to a given property, and on the successive classification of the obtained regions in two classes (foreground and background). An important step of our process is related to the reduction of oversegmentation, which affects the watershed transform. We suggest two alternative criteria to reduce oversegmentation. The first criterion is based on the use of two suitable processes, called flooding and digging, and requires repeated application of the watershed transformation. The second criterion involves the use of multi-scale image representation. As for the classification of the regions of the obtained partition, our method is based on the locally maximal changes in gray-level among adjacent regions. This classification scheme works well for a wide class of gray-level images, e.g., a number of biological images, where the boundary between the foreground and the background is perceived wherever strong gray-level changes occur through the image.