MUMBAI, India, Sept. 22 -- Intellectual Property India has published a patent application (202611093564 A) filed by Tanya Mall; Gaurav Singh; Anjani Kumar; Atul Anand; Shashi Kant; and Ravikant Kumar Nirala on August 02, 2026, for An Intelligent Deep Learning Framework For Unified Image Reconstruction From Raw Sensor Data Through Joint Demosaicking, Noise Suppression, And Resolution Enhancement.
Inventors include Tanya Mall; Gaurav Singh; Anjani Kumar; Atul Anand; Shashi Kant; and Ravikant Kumar Nirala.
The application for the patent was published on September 18, 2026, under issue no. 38/2026.
Abstract: An intelligent deep learning framework is presented for unified enhancement of raw digital images by joint image demosaicking, noise removal and accounting for image super-resolution. By building a single end-to-end network for convolutional neural networks (CNNs), the invention overcome the problems of limitations of conventional sequential image enhancement pipelines by incorporating multiple reconstruction operations into a single network. The proposed framework is used to input raw image data, as well as an indexing matrix in which represents one or more colour filter array cfa configurations. Multi-path feature extraction modules create complementary spatial and contextual representations which are further refined using attention based residual learning blocks and concentration modules. Efficient preservation of structural information with minimal reconstruction artifacts by hierarchical feature fusion. The improved feature representations are passed to the generative reconstruction modules designed for estimating the missing colour components, reject sensor noise and synthesize high resolution images with high perceptual quality. An adaptive feature aggregation and a residual learning contribute to increase the convergence stability, keep the fine textures, edge continuity and colour consistency. In addition, the invention allows for multiple CFCs to be supported with a single NN, thus avoiding the need for multiple NNs for each sensor layout configuration. This greatly decreases the quantity of computation, memory consumption and deployment overhead among a disparate imaging platform. The disclosed framework can be implemented in a smartphone, a digital camera, an autonomous vehicle, an industrial inspection system, a robotic vision system, a satellite imaging system, medical imaging equipment, a surveillance system, an embedded vision processor, and an edge-computing device. The invention can simultaneously optimize for multiple image restoration applications in a single computational structure and can offer more reconstruction accuracy, better visual quality, less propagation of artefacts and more efficient operation than the traditional image enhancement techniques.
Disclaimer: Curated by HT Syndication.
