MUMBAI, India, Aug. 10 -- Intellectual Property India has published a patent application (202641089403 A) filed by Vellore Institute Of Technology on July 22, 2026, for Rgb-Only Specular Highlight Removal Framework Using Physically Grounded Synthetic Supervision And Latent Feature Inpainting.
Inventors include Sidharth Gautam; Om Shinde; and Shivam Singh.
The application for the patent was published on July 31, 2026, under issue no. 31/2026.
Abstract: The present invention relates to a deep learning-based framework for removing specular highlights from a single RGB image without requiring paired reflection-free training data, specialized imaging hardware, or domain-specific adaptation. The framework incorporates a Virtual Highlight Synthesis (VHS) pipeline that generates physically plausible synthetic training pairs using monocular scene geometry estimation, Fresnel-aware reflectance modeling, and randomized illumination sampling. Hierarchical feature representations extracted by a frozen pre-trained encoder are processed to localize highlight regions, followed by latent feature-space inpainting that reconstructs corrupted features using contextual priors and self-attention mechanisms. A multi-scale decoder generates a reflection-free diffuse image with preserved structural and semantic consistency. The framework generalizes across natural and endoscopic imaging domains within a single model and enhances downstream computer vision tasks, including stereo reconstruction, camera pose estimation, optical flow, and feature correspondence, by reducing photometric inconsistencies caused by specular reflections.
Disclaimer: Curated by HT Syndication.
