Multilingual signature-verification by generalized combined segmentation verification

We propose a combined segmentation-verification technique for Multilingual signature verification. One limitation of original segmentation verification method is that it is not applicable for such signatures which are difficult in segmentation like Latin-scripts signatures. To overcome this limitation, we employ signature segmentation using the position of gravity center of whole signature strokes instead of an interval space between names in the signature image. Three grayscale-gradient features are extracted from whole signature image and two segmented signature images, left-hand and right-hand side and evaluated the Mahalanobis distances from reference samples. The on-line feature based technique employs dynamic programming (DP) matching for time series data of the two segmented and one whole signatures. Three resultant distance values from off-line verification and three resultant dissimilarity values are input to SVM to make final decision of genuine or forgery. We evaluated the performance of the proposed method on SigComp2011 dataset which consists of Chinese and Dutch signatures. In the results of evaluation, the proposed technique achieved 1.02% and 4.29% EER(Equal Error Rate)for Chinese and Dutch signatures respectively, which are significantly lower than and comparable to those of the best performances in SigComp2011 competition. These results confirm that the proposed generalized combined segmentation-verification by gravity center is effective for accuracy improvement of multi-script signature verification.