MUMBAI, India, June 30 -- Intellectual Property India has published a patent application (202641075418 A) filed by Vellore Institute Of Technology on June 18, 2026, for Handwritten Numeral Recognition Using Hybrid Neural Network With Dynamic Gating.
Inventors include Bharanidhara N N; Aadarsh Pranav Sureshbabu; Nitheesh Kumar Vinayagam; and K C Ghoraav.
The application for the patent was published on June 26, 2026, under issue no. 26/2026.
Abstract: ABSTRACT HANDWRITTEN NUMERAL RECOGNITION USING HYBRID NEURAL NETWORK WITH DYNAMIC GATING A method for handwritten numeral recognition comprises receiving, by a processor (12), an input image (100) containing a handwritten numeral and generating a stroke descriptor map (200) comprising a skeleton channel (202) encoding a morphological skeleton, a curvature channel (204) encoding a gradient magnitude, and a direction channel (206) encoding a stroke orientation. The method comprises extracting a first feature vector (302) using a convolutional neural network branch (300), extracting a second feature vector (402) using a transformer branch (400), and encoding the stroke descriptor map (200) into a third feature vector (502). The method comprises generating per-sample gate values (602) comprising first and second gate weights using a gating network (600), computing a fused feature vector (700) by combining the feature vectors weighted by the gate weights and a fixed residual weight, and classifying the handwritten numeral based on the fused feature vector (700).
Disclaimer: Curated by HT Syndication.
