A real-time narcissus correction algorithm based on Gaussian mixture models

Abstract

Cooled infrared optical systems often suffer from narcissus artifacts, which can severely degrade their imaging quality. Although the narcissus effects on image quality can be reduced using appropriate optical designs, the narcissus artifacts cannot be fully eliminated by this approach alone. Other approaches use image processing schemes to treat these artifacts as nonuniform noise and apply nonuniformity correction techniques to eliminate the artifacts. However, as the intensity and form of the narcissus artifacts depend on external factors, such as ambient temperature and focal length of the infrared imaging equipment, target-based nonuniformity correction techniques must frequently be calibrated against a standard blackbody radiation source. Moreover, scene-based nonuniformity corrections often produce ghosting artifacts owing to their overreliance on scene motions. In the present work, narcissus artifacts are treated as a type of additive low-frequency noise. Based on the characteristics of these artifacts and the requirements for real-time performance, we propose a new fitting method and a modified two-dimensional Gaussian mixture model (2D-GMM) for the low-frequency noise associated with narcissus artifacts. This allows the model parameters to be calculated rapidly without an iterative process. The resulting estimates of the narcissus signal can then be directly subtracted from the raw image to eliminate the narcissus artifact. The proposed correction algorithm was experimentally demonstrated to be capable of eliminating narcissus artifacts; furthermore, the fitting and correction processes of the algorithm are sufficiently fast to satisfy the requirements of real-time applications.