MUMBAI, India, Aug. 12 -- Intellectual Property India has published a patent application (202621074107 A) filed by Pathan Sarfaraz Iqbal Khan; and Dr. Sanjay Y. Azade on June 15, 2026, for Biologically Guided Roi-Aligned Spectral-Spatial Deep Learning System For Curcuminoid Concentration Prediction Using Hyperspectral Imaging.
Inventors include Pathan Sarfaraz Iqbal Khan; and Dr. Sanjay Y. Azade.
The application for the patent was published on August 07, 2026, under issue no. 32/2026.
Abstract: BIOLOGICALLY GUIDED ROI-ALIGNED SPECTRAL-SPATIAL DEEP LEARNING SYSTEM FOR CURCUMINOID CONCENTRATION PREDICTION USING HYPERSPECTRAL IMAGING Abstract The present invention discloses a biologically guided ROI-aligned spectral-spatial deep learning system for non- destructive prediction of curcuminoid concentration in turmeric rhizomes using hyperspectral imaging. Conventional hyperspectral prediction approaches typically assign laboratory-derived High Performance Liquid Chromatography (HPLC) measurements to entire hyperspectral images, resulting in spectral-label mismatch caused by biological heterogeneity and spatial variation in chemical composition. The proposed invention addresses this limitation through a Region of Interest (ROI)-HPLC alignment framework that establishes correspondence between laboratory-analyzed sample regions and hyperspectral image data used for model training and inference. The system comprises a hyperspectral image acquisition module, an ROI-HPLC alignment engine, a biological spectral prior generation module, spectral and spatial feature extraction networks, an Adaptive Spectral Prior Temperature (ASPT) mechanism, an adaptive gated feature fusion module, and a concentration prediction engine. Biological wavelength priors associated with curcuminoid-sensitive spectral regions are incorporated into attention-based learning to improve feature discrimination and model interpretability. Spectral representations are extracted through transformer-based learning modules, while spatial representations are learned using convolutional neural network architectures. The extracted features are dynamically integrated through an adaptive gating mechanism that optimally balances spectral and spatial contributions for each sample. The invention significantly reduces spectral-label mismatch, improves prediction robustness, enhances generalization across turmeric families and concentration regimes, and enables rapid, scalable, cloud-based deployment for agricultural quality assessment, breeding programs, nutraceutical manufacturing, and pharmaceutical raw material evaluation. The disclosed framework provides a technically improved system for biochemical concentration prediction from hyperspectral imagery through biologically informed learning and ROI-aligned supervision.
Disclaimer: Curated by HT Syndication.
