DeepSign lightweight ShuffleNet with transfer learning for signature verification

Signature verification confirms the authenticity of a handwritten signature that matches it to the claimed identity. This process analyses dynamic features like pressure, speed and stroke order as well as static features such as shape and structure. Existing systems often face challenges with variations in signature styles and inconsistencies across sessions that affect accuracy. Moreover, many models demand large datasets and heavy computational power that limits their real-world applications. So, this paper proposes a Transfer Learning based Deep Signature Verification (TL-DSV) that can handle a broader range of signature variations and is better at distinguishing between forged and genuine signatures. The TL-DSV begins with the input of a signature image that undergoes preprocessing techniques like Multistage Medav Filtering Approach (MMFA) to enhance clarity and remove noise. Image augmentation techniques such as rotation and flipping are applied to generate variations of the signature that helps to generalize the model better. The image is passed through feature extraction such that stroke extraction, global features, Enhanced LGTrP based Local features (EL-Lf) and deep features are extracted. For verification, a Transfer Learning integrated Attention Ghost module based ShuffleNet (TL-AGShN) is used to categorize the signature as authentic or forged through leveraging pre-trained networks to fine-tune on the specific dataset and improve accuracy. The system ultimately outputs a decision regarding the signature's authenticity that marks it as either genuine or forged based on the learned features and patterns. The TL-AGShN approach has introduced a new framework for carrying out the verification proposes to resulted the betterment with the accuracy of 97.2% in 90%.