Region saliency as a measure for colour segmentation stability

Local Instability Problem of Image Segmentation Algorithms: Systematic Study and an Ensemble-Based Solution

“…A di®erent approach to analyze stability of segmentations is proposed by Heidemann, 17 where stability means that segmented regions remain undisturbed by di®erent kinds of degradation: noise, changes of viewpoint and changes of lighting. A goodness function based on color saliency is de¯ned and it is empirically shown that the goodness function predicts whether the segmentation is stable or not.…”

“…28 Although there has been a lot of improvement in the development of region-based segmentation methods in the last decade, 7,8,10,22,40 relatively little work is concerned with the stability of these algorithms. 17,26,38,40 For this study we choose three representative region-growing (RG) algorithms and the watershed transform, which can be classi¯ed as region-based segmentation algorithm. The reason for selecting these algorithms is that they are inherently local.…”

Colour saliency-based parameter optimisation for adaptive colour segmentation

“…The colour saliency (CS) [1] is calculated for every cluster R i (i [1,k] where k is the number of clusters) of the segmented image and is defined as the average colour difference with respect to the 4-connected neighbouring (N4) regions. In equation (1), NBP(R i ) is the total number of border pixels of the region R i , BP(R i ) is the set of pixels that defines the boundary of region R i and N diff_N4 (x, y) represents the number of pixels in the 4-connected neighbourhood (N4) of the current pixel (x, y) that belong to a region different than R i .…”

“…If this distance is smaller than ICV SOM , the node that has the largest weighted variance (confidence value) is eliminated. It has been quantitatively demonstrated in [1] that the maximisation of the border contrast (this implies the maximisation of the average saliency S avg ) leads to improved region stability. A high region stability in the context of image segmentation means that the resulting regions in the segmented data are not erroneously divided due to spurious texture or uneven illumination.…”