Patent · US Active

Scene-based nonuniformity correction using a convolutional recurrent neural network

US10929955B2 · kind B2 · utility

0Cited by
22References
19Claims
0Family size

Assignee

Inventor

Key dates

Filing dateFeb 14, 2019
Grant dateFeb 23, 2021
Priority date
Expiry dateJun 30, 2039

Classification

  • Technology area (CPC G)Physics
  • CPC primaryG06T2207/20084
  • WIPO fieldAudio-visual technology
  • WIPO sectorElectrical engineering

Abstract

A method and apparatus for correcting nonuniformity noise in thermal images. The method comprises receiving a current image being part of a stream of thermal images; concatenating the current image from the stream of thermal images with hidden state images; processing, by a first convolutional neural network, the concatenated image to extract a number of feature channels; generating based on the feature channels at least a first multiplicative mask; processing, by a second convolutional neural network, a masked concatenated image to compute a weighting parameter, wherein the masked concatenated image is resulted by applying the first multiplicative mask on the concatenated image; and simulating, using the weighting parameter, an infinite impulse response (IIR)-style updating scheme to estimate the nonuniformity noise in the current image.

Source: USPTO / EPO open patent data. Objective bibliographic and citation counts.