Siamese Network For Signature Verification On Various Distributions of Data
Arihant Jain, Ashish Ranjan Kumar · 2024
With the rapid growth of the human population as well as technology, an easy method of biometric verification is necessary and signatures are the simplest form of such methods, However the simplicity of signatures brings with it the prospect of forgery. Due to the sheer volume of verification that must be reliably and quickly carried out, an automated system to identify which signatures are genuine and which are forged is critical in the daily operation of most industries. Without the use of a pressure sensitive tablet, it is not possible to use pressure data for this purpose, however using the images of signatures we can devise a neural network to distinguish which signatures appear genuine and which appear forged. This method utilizes two convolutional neural networks to reduce images to a vector representation and compares them using Euclidean distance in order to classify the pair of images as matching or non-matching. Each image is preprocessed to remove noise and color as well as to reduce the size to an acceptable limit so that it can be run comfortably on the hardware being used. The results obtained show significant difference in performance of the model depending on what distribution is used to train the system, with mixed distributions outperforming the pure distributions by a large margin.