Patents

A system for protecting sensitive video by automatically detecting and obscuring human skin tones. Redaction information is encrypted and stored with the video. When authorized personnel need to view unredacted content, they use unique decryption keys to reconstruct the original video, which then receives a personalized digital watermark identifying the specific user — enabling organizations to trace unauthorized leaks back to their source. Key application: law enforcement and government agencies protecting surveillance footage.

A system enabling restoration, enhancement, and segmentation of an input image at arbitrary resolution using a single trained ML model. The key innovation restructures image channels to match the model's input requirements — processing color channels separately for multiple restructurings, and together for single restructurings. This enables one model to handle images from 1 MP to 108+ MP, replacing multiple specialized networks and reducing device memory footprint.

A system that identifies blur types and intensity in photographs, distinguishing between unintentional blur (motion, camera shake, focus failure) and intentional artistic effects (bokeh, portrait mode). By analyzing entropy patterns across image regions and fusing probability scores, the device determines whether blur correction is needed — avoiding degradation of artistic effects. Operates on single frames without frequency transforms, suitable for mobile devices.

A neural network architecture employing cascaded Volterra filters in hierarchical layers. The system uses repeated lower-order (2nd or 3rd order) filters to approximate high-order filters, significantly reducing computational parameters compared to conventional deep learning approaches. Introduces non-linearities through controlled higher-order convolutions rather than standard activation functions. Demonstrated on action recognition (UCF-101, HMDB-51) and multi-modal data fusion with improved noise robustness.