Offline signature verification using lightweight deep learning

The biometric approach to signature verification serves as the foundational method predominantly employed for personal purposes in validating significant documents and legal transactions within both private and governmental sectors. Detecting counterfeit signatures is of paramount importance; however, manual methodologies reliant on subject matter experts entail substantial time, effort, and expense, making them susceptible to errors and requiring significant expertise. Addressing these challenges, this thesis presents advancements in offline signature detection and verification, considering devices with limited resources and inadequate samples of collected signatures.The proposed model utilizes lightweight deep-learning techniques to classify signatures as genuine or forged, achieving a delicate balance between high accuracy and resource constraints. In this study, four lightweight deep learning models were employed: MobileNet V3, MobileNet V2, Mobile Net V1, and EfficientNetB0. These models were evaluated on two datasets: GPDS, the largest signature database in the field, and FHD, which was specifically compiled for research purposes as a challenge and incentive. Results indicated that MobileNet V3 outperformed its predecessors and EfficientNetB0 in terms of accuracy and practicality, achieving an accuracy rate of 87.48% on the GPDS dataset and 86.75% on the FHD dataset.The relevance of this research lies in providing an effective and reliable tool to aid decision-making in the field of biometrics and signature verification, crucial across all sectors requiring individual authentication and verification. This contributes significantly to overall management and the attainment of satisfactory quality standards.