“…DTW can tolerate spatial variations unlike other methods such as XOR, Euclidean Distance Mapping, Sum of Squared Differences [47]. Alternative to the methods matching words based on whole images or profile-based features [46], recently other features are also experimented, including word contours [57], gradients [34,50,55,65], shape context descriptors [35], Harris corner detector outputs [51], line segments [12,13], and interest points [7]. In [11,21,49,57], the problem of writing style variations in multi-writer datasets is tackled, but these studies generally require isolated words.…”