Principal curves is a new feature extraction method based on nonlinear transformation They are smooth self-consistent curves that passes through the “middle” of the distribution They perfectly reflect the structural features of the data The paper chooses principal curves to extract strokes of characters and segments numeral strings by grouping strokes based on the confidence of the classifiers The classifiers based on the segmented contour feature and the normalized template features are combined and experimental results indicate that the correlation of these two features is small The paper modifies the confidence of the combined classifier by posterior probabilities which are estimated by a novel class-conditional confidence transformation approach Experimental results indicate that the method is effective in the segmentation of numeral strings