Signature warping and greedy approach based offline signature verification

Offline signature verification paves way to an effective automation of authorizations needed in various real world applications. The use of a neural network in the automatic identification of signatures supports a faster validation that involves lowering labor costs as well as eliminating any form of bias. The models presently in use perform an image embedding based comparison between the test signature’s extracted features and those in the dataset. This often requires the original signee to go through the cumbersome task of producing signatures multiple times, which is not applicable in real world scenarios. The novelty of the proposed work in offline signature verification involves generating a dataset from a very minimalistic number of original signatures. Each of these has its own set of variations produced by augmenting the original image over composite functions. The most similar amongst these alternatives at any given point of time is selected via a greedy approach, thereby reducing the computation required over the Siamese Network. The model was tested on standard datasets as well as those that were locally generated. As a whole, the model is able to classify signatures with an accuracy of 97% and an F1 score of 0.97 on average.