Watershed-based segmentation of rock scenes and Proximity-based classfication of watershed regions under uncontrolled lighting conditions

A watershed-based segmentation approach that uses iterative bilateral smoothing with an adapting photometric similarity parameter for pre-filtering is adopted for the segmentation of rock scenes. But the resultant segmented images also contain non-rock watershed regions which are not desired for measuring rock sizes. A proximity-based classifier is applied for the removal of the latter using features that can be divided into rock shape, edge strength and region intensity characteristics. Subset feature selection based on Thornton’s seperability index is used to remove redundant and irrelevant features. We achieve final classification rates of 89.91% using a simple k-nearest neighbor Classifier.