MUMBAI, India, July 24 -- Intellectual Property India has published a patent application (202611062792 A) filed by Maharishi Markandeshwar Deemed To Be University on May 18, 2026, for A Hierarchical Attention Based Deep Learning System And Method For Offline Handwritten Signature Verification With Forgery-Independent Training.
Inventors include Ritika; and Dr. Dalip.
The application for the patent was published on July 17, 2026, under issue no. 29/2026.
Abstract: A HIERARCHICAL ATTENTION BASED DEEP LEARNING SYSTEM AND METHOD FOR OFFLINE HANDWRITTEN SIGNATURE VERIFICATION WITH FORGERY-INDEPENDENT TRAINING The present invention is related to a system and method for offline handwritten signature verification using deep learning techniques trained exclusively on genuine signature samples only. The system uses a Residual Neural Network (ResNet) with hierarchical feature aggregation and layer wise attention applied for better extraction of discriminative features. With the use of feature aggregation of the multiple layers of ResNet, the system becomes capable of extracting both the global overall shape of signature and local stroke level details from the signature images. The extracted features are refined with the attention module added on each layer for enabling the system to focus on more important features. Further, the system uses hybrid loss combining circle and cross entropy loss that provides a balance between metric learning and classification objectives to optimize feature embeddings. The system is capable of handling variations in signature styles across different writers and scripts. It includes mechanisms for using global threshold-based decisions for authentication purposes in both writer dependent and writer independent modes. The invention eliminates the need of forged samples. It provides high accuracy and lower error rates as compared to existing systems and thus enhances security and reliability in identity verification applications.
Disclaimer: Curated by HT Syndication.
