A HYBRID METHOD OF FEATURE EXTRACTION FOR SIGNATURES VERIFICATION USING CNN AND HOG A MULTI-CLASSIFICATION APPROACH

Abstract

The feature extraction stage of the offline signature verification system is considered crucial and has a
substantial influence on the system's performance. This is because the quantity and calibration of the extracted
features determine the system's ability to differentiate between genuine and forged signatures. This paper presents a
novel approach for extracting features from signature photos, using a combination of a Convolutional Neural
Network (CNN) and Histogram of Oriented Gradients (HOG). The extracted features are then subjected to a feature
selection technique (Decision Trees) to determine the most important ones. Ultimately, the CNN and HOG
techniques were merged. The hybrid method's effectiveness was evaluated using three classifiers: long short-term
memory, support vector machine, and K-nearest Neighbor. The experimental results demonstrated that our proposed
model performed well in terms of efficiency and predictive capability, achieving high accuracies using the CEDAR
dataset. The accuracy of our assessment is considered very significant, especially considering that we examined
expertly produced signatures, which are more difficult to detect compared to other types of fabricated signatures
such as basic or reverse forgery. The research incorporates many modifications, including the use of Xception for
feature extraction using HOG-RFE and a Voting Classifier for dataset analysis. Through this approach, we achieved
a remarkable accuracy of 100% in the improved verification of signatures using CNN and HOG, employing a multiclassification
methodology. A Flask framework that is easy to use and includes SQLite integration allows for user
registration and signin, making it viable for usability testing in cybersecurity apps.