Learning a Legibility Classifier for Deformed Letters

We introduce an approach for classifying a deformed letter from the English alphabet based on its legibility. Given an image of a deformed letter d, our method proposes the use of the filtered medial axis (i.e. strokes) as its simplified representation, from which we can compute a feature vector. Based on the anatomy of the letter d we identify the strokes as: ascender, descender and counter. Feature correspondence across deformations is ambiguous; thus we propose using a kernel-based multi-instance classifier model. Our classifier is trained using legibility score labels collected through a user-study. Our approach obtains an accuracy of 71.1% and outperforms a classic boundary-based method. Our findings suggest that, although our features are meaningful descriptors of the letter forms, a robust methodology for acquisition of legibility scores is required to ensure accuracy.