Offline Signature Verification Using Pre-Trained Deep Convolution Neural Network: SqueezeNet
Bhimraj Prasai Chetry, Biswajit Kar · International Journal of Electronics and Communication Engineering · 2025
Offline Signature Verification is a very important research area because signatures evolve throughout a person’s life and have many applications such as person authentication, verification in financial transactions, institute certifications, legal documentation, etc. It has been socially, legally, and culturally accepted as a behavioural biometric for centuries. So, it is more prone to forgery than any other biometrics. So, in order to counteract forgery and accept genuine signatures, we have proposed an offline signature verification system using a pre-trained deep convolutional neural network called “SqueezeNet v1.0” to enhance the verification accuracy of the system. Here, the use of a pretrained SqueezeNet model is an effective approach, especially when we need a lightweight model that can perform well with fast inference in resource-constrained environments like signature verification. Signature verification is challenging work because of large intra-class diversity and small inter-class distinction while considering forgeries. Despite the progress made with traditional methods, these techniques often face challenges related to feature engineering and performance under noisy conditions, making them less effective compared to modern deep learning-based approaches. With the progress of deep learning, offline signature verification has seen significant improvements, particularly Convolutional Neural Networks (CNNs), which are able to self learn hierarchical feature representations from raw signature images, eliminating the need for manual feature extraction. Here, skilled forgery signatures of each user are used for training and testing purposes to make the system robust and more accurate. Our system is trained and tested on the CEDAR database for all fifty-five users having different types of signature information, yielding average testing accuracy of 98.98% using random forgeries and 98.07% using skilled forgeries. Testing accuracy of random forgeries lies between 93.75%-100% and testing accuracy of skilled forgeries lies between 72.92%-100%.