A Metric Learning Approach for Offline Writer Independent Signature Verification

An efficient offline system for writer independent signature verification is proposed in this work which is quite a difficult task in computer vision. The proposed system employs a novel global representation of signatures followed by the Mahalanobis distance based dissimilarity score to discriminate between the original signatures and their skilled forgeries. The global representation of an image containing signature is based on aggregation of local descriptors using a vocabulary. The global descriptors from the pair of images are then used to learn a low-rank distance metric which is not a trivial task owing to the high dimensionality of the descriptor. The experimental results are reported on two datasets namely CEDAR and BHSig260; both containing a sufficient number of writers and are used as benchmark datasets. A comparison with the state-of-the-art approaches demonstrates the efficiency of our proposed approach. Our proposed method achieved 100% accuracy on CEDAR dataset and outperformed all other methods on BHSig260 (Bengali) dataset. The results on BHSig260 (Hindi) dataset are also promising.