CLASSIFICATION OF REAL AND FORGED SIGNATURE USINGVGG-16 ALGORITHM

Abstract

Signature authentication is of paramount importance in applications such as banking, legal transactions, and security systems. This study introduces a novel approach to automatic signature verification by leveraging a modified VGG-16 convolutional neural network (CNN), distinct from traditional image classification tasks. While VGG-16 is widely recognized for its effectiveness in general object classification, our work adapts and optimizes the model specifically for the challenges of signature verification. A key innovation in modified VGG-16 model is the incorporation of an Attention Mechanism using the Squeeze-and-Excitation (SE) block, which enhances the feature map by adaptively recalibrating channel-wise feature responses. This attention mechanism allows the model to focus on the most relevant features of the signature, such as specific strokes or pen pressure, improving the accuracy of classification and making the model more sensitive to small but significant differences between genuine and forged signatures. Additionally, custom Dense Layers are added to the VGG-16 architecture,

which include a combination of fully connected layers and Dropout to prevent overfitting. These layers work synergistically with the SE block, enabling the model to learn complex patterns and make accurate binary decisions (authentic or forged signatures) based on the enhanced features. In comparing the performance of ANN, CNN, and VGG-16 models for signature verification, significant differences are observed in key metrics such as accuracy, precision, recall, and F1-Score. The ANN model has the lowest accuracy at 74.16%, with a precision of 83%, recall of 74%, and an F1-Score of 72%, indicating that it struggles to correctly classify signatures. The CNN model performs better, achieving an accuracy of 85.97%, with 87% precision, 86% recall, and an F1-Score of 86%, showing its improved capability in detecting forgeries. However, the VGG-16 model outperforms both, with a high accuracy of 96.88%, and precision, recall, and F1-Score of 97%, making it the most reliable model due to its deep learning capacity and ability to generalize well to various signature styles. Key contributions of this research include the introduction of aggressive data augmentation techniques such as rotation, width/height shift, shear, zoom, and horizontal flipping, which improve model robustness and generalization, particularly when dealing with small datasets typical of signature verification tasks. Additionally, we designed custom classification layers on top of the VGG-16 base, tailored for the binary classification of authentic versus forged signatures, with the integration of dropout layers to reduce overfitting. Unlike existing approaches that apply VGG-16 for general classification, our model incorporates real-time metric monitoring during training, tracking precision, recall, F1-score, and confusion matrices, providing a continuous, in-depth understanding of model performance. We also introduce an in-depth feature extraction and visualization process to illustrate how the model learns distinctive features from signature images, distinguishing small yet significant differences between genuine and forged signatures. Through these contributions, this study significantly advances signature verification technology, offering an effective, high-precision model suitable for real-world applications such as banking and security systems. By leveraging the VGG-16 model with task-specific modifications, including attention mechanisms and dense layers, our approach addresses the unique challenges of signature verification and enhances the reliability of automated authentication systems