Evaluating the Effectiveness of an Image Segmentation method

Many image segmentation algorithms have been proposed to partition an image into foreground regions of interest and background regions to be ignored. We focus on examples related to images of botanical seeds presented to evaluate, from a statistical perspective, the effectiveness of the results provided by several image segmentation methods. More precisely, we assume that the separation of background pixels from foreground ones operated by a segmentation method needs to be further validated since, particularly for complex, it is very difficult to distinguish between the two categories even by a human eye or by powerful zooming. In this respect, the idea is to use a classification method, or classifier, in order to assess the degree of reliability of the separation between background and foreground pixels obtained from a standard segmentation image method. To this end, the comparison is made by evaluating, through the use of different types of classifiers, the accuracy of an image segmentation process. The statistical analysis involves many different settings in which each specific pre-processing method is, in turn, considered as the reference pre-processing method in the image segmentation process and the output of the different approaches proposed for image segmentation is used as response variable. In practice, in each setting the response variable is binary and corresponds, for each individual pixel, to the background/foreground assignment deriving from a specific segmentation method. The classification task is to ask a classifier to predict in the most accurate way the pixel category on the basis of the RGB intensities deriving form a specific pre- processing method. If a classifier is able to correctly predict all the available pixels, the relative segmentation method is 100% reliable. Thus, the more accurate is a classifier the more reliable is the pre-processing method at hand.