MUMBAI, India, June 26 -- Intellectual Property India has published a patent application (202621048858 A) filed by Beyondata Solutions Private Limited on April 16, 2026, for Method And System For Context-Aware Visual-Semantic Correction In Domain-Specific Ocr Pipelines.
Inventors include Nishant Singh Tomar; and Dipesh Prajapati.
The application for the patent was published on June 19, 2026, under issue no. 25/2026.
Abstract: The present disclosure relates to a system (102) and method for generating corrected textual output from a rasterized document representation using visual-semantic processing. The system (102) receives document data, transforms the data for machine interpretation, and processes the transformed version using a feature-encoding module (304) and a region-localization module (306) to generate hierarchical feature encodings and region proposals. An extraction module (308) generates preliminary textual tokens associated with positional information. A training module (310) trains a domain-specific small language model (312B) using noisy tokens, spatial features, and positional data. A multimodal representation generator (312A) combines visual and textual cues to form embeddings used by a correction module (312) to generate corrected textual tokens. A confidence scoring module (314) and an output module (316) produce structured corrected textual output linked to positional metadata. Further, the system (102) includes an image-cropping module (318) for handling image regions within the document. FIG. 1
Disclaimer: Curated by HT Syndication.
