MUMBAI, India, Sept. 22 -- Intellectual Property India has published a patent application (202631056142 A) filed by National Institute Of Technology Sikkim on May 03, 2026, for A System And Method For Image-Based Object Detection Using Deep Neural Network Architectures.
Inventors include Abhishek Rajan; Sayak Dutta; and Surabhi Hom Choudhury.
The application for the patent was published on September 18, 2026, under issue no. 38/2026.
Abstract: The present invention relates to a system and method for image-based object detection using deep neural network architectures. The present invention implements four- scale detection head that extends detection to a P6 level to directly model very large objects, an enlarged SPPF kernel at the deep P5 stage to increase local receptive field where large-object context is critical, asymmetric stage depth allocation that concentrates representational capacity in the P5 path for large-object features, and specialized downsampling modules that preserve complementary contextual and high-frequency information during spatial reduction. The combination of these features results in improvement in large-object localization, higher mean average precision on daatsets with substantial scale variation, increased recall for both small and large object classes, and retention of real-time through on modern graphical processing units.
Disclaimer: Curated by HT Syndication.
