In this paper, we develop a system to classify the outputs of image segmentation algorithms as perceptually relevant or perceptually irrelevant with respect to human perception. The work is aimed at figurative images. We previously investigated human visual perception of trademark images and established a body of ground truth data in the form of trademark images and their respective human segmentations. The work indicated that there is a core set of segmentations for each image that people perceive. Here we use this core set of segmentations to train a classifier to classify closed shapes output from an image segmentation algorithm so that the method returns the image segments that match those produced by people. We demonstrate that a perceptual relevance classifier is attainable and identify a good methodology to achieve this. The paper compares MLP, SVM, Bayes and regression classifiers for classifying shapes. MLPs perform best with an overall accuracy of 96.4%.
|Published - 2007