MUMBAI, India, Aug. 12 -- Intellectual Property India has published a patent application (202611069608 A) filed by Manipal University Jaipur on June 03, 2026, for Ai Framework For Reliable Vehicle Damage Assessment Using Vision-Language Models And Epistemic Uncertainty.
Inventors include Bagesh Kumar; Vishal Pandey; Sahil Kumar; Harsh Markam; Prakhar Shukla; and Shivansh Mishra.
The application for the patent was published on August 07, 2026, under issue no. 32/2026.
Abstract: The present invention relates to an artificial intelligence (AI)-based framework for reliable vehicle damage assessment using vision-language models (VLMs) and epistemic uncertainty estimation. The framework includes a VLM context engine configured to identify vehicle type, viewing angle, and surface conditions from raw images prior to segmentation. A nine-point quality screening module detects issues including glare, blur, reflections, and mud, and selectively applies corrective enhancement. A DINOv2 backbone with feature pyramid adaptation extracts multi-scale features for a dual-head Mask2Former decoder that simultaneously performs vehicle anatomy recognition and damage segmentation through a shared backbone in a single forward pass, thereby reducing latency and GPU memory consumption. Monte Carlo dropout generates confidence-aware predictions for automated approval or human review. A post-processing pipeline localizes damage on vehicle parts, detects missing components through void analysis, evaluates severity, and classifies damage as cosmetic or functional. The framework outputs structured reports, annotated images, and uncertainty heatmaps for efficient insurance assessment and claim processing.
Disclaimer: Curated by HT Syndication.
