Off-line Signature Verification through Machine Learning

Signature is a depiction of a person's name that is used as his/her identity proof, but it can be forged. Therefore, genuineness of a signature needs to be verified. In the present paper for verifying the o□ine signature, a small 3-layer deep convolutional neural network followed by a fully connected layer is proposed, whose trainable parameters are several orders of magnitude fewer than previously reported in literature. This network has been used in two dierent configurations: firstly as a feature extractor in hybrid classifier and secondly as an end to end classifier in Siamese network. In hybrid classifier scheme, support vector machine is used for verifying the genuineness of the signature. Siamese network contains two or more identical subnetworks joined by one or more fully connected layers. The proposed neural network is used as subnetwork in Siamese network. Both hybrid classifier and Siamese network are tested for both Writer independent (WI) and dependent (WD) verification on 3 datasets: CEDAR, GPDS Synthetic Signature, BHSig260. The proposed approach achieves percentage accuracy of 99.91, 92.28, 86.88, 90.58 in Hybrid classifier scheme, 99.87, 92.14, 86.33, 90.80 in Siamese network for WD verification and 80.26, 75.06, 89.33, 82.42 in Siamese network for WI verification in CEDAR, BENGALI, HINDI and GPDS respectively. Experimental results show that the verification accuracy for HINDI in WI outperforms the state of art approaches. The proposed approach has also achieved comparable performance in CEDAR and BENGALI for WD verification.