MUMBAI, India, June 30 -- Intellectual Property India has published a patent application (202641074362 A) filed by Vellore Institute Of Technology on June 16, 2026, for Frequency-Aware Transformer For Remote Sensing Image Classification.
Inventors include Avinash Chandra; and Shinde Vikrant Rajendra.
The application for the patent was published on June 26, 2026, under issue no. 26/2026.
Abstract: ABSTRACT FREQUENCY-AWARE TRANSFORMER FOR REMOTE SENSING IMAGE CLASSIFICATION A computer-implemented method for classifying remote sensing images comprises decomposing an input image into low-frequency, mid-frequency, and high-frequency band images using Non-Subsampled Contourlet Transform. Feature extraction streams (120) extract feature tensors from each band image through vision transformer backbones. A frequency-aware positional encoding module (130) applies encodings representing spatial position and frequency band identity. A hierarchical frequency gating module (140) controls information flow using a lateral inhibition circuit with a learnable matrix having positive diagonal entries for self-excitation and negative off-diagonal entries for cross-band suppression with row-wise softmax normalization, generating per-band gate values via sigmoid activation. An adaptive frequency band attention module (150) computes band-specific attention using hypernetwork-generated modulation vectors with shared projection matrices modulated through element-wise multiplication. A progressive frequency fusion pyramid module (160) fuses gated feature tensors bottom-up through gated cross-attention stages. A classification head (170) generates a classification output. (Fig. 1)
Disclaimer: Curated by HT Syndication.
