Quantitative evaluation of different thresholding methods using automatic reference image creation via PCA

摘要

This article proposes a principal component analysis-based automatic approach to generate reference image for evaluating different thresholding techniques. Twenty one thresholding methods have been considered for reference image creation and evaluated using five standard performance indices. Literature suggests a few performance measures, among which F-measure, modified Hausdorff distance, edge mismatch error, relative area error and object level consistency error are popular. However, correlation analysis of these metrics reveal that only F-measure, modified Hausdorff distance and edge mismatch error retain non-redundant information. Thus the best thresholding method can be determined from these three indices for different types of images, automatically. Experimental results demonstrate the potential of different thresholding methods and select the best binary segmentation technique for particular type of image set.