Infrared star image denoising using regions with deep reinforcement learning

Efficient and accurate extraction and restoration of star targets in infrared star images with small number of frames is a growing need for optical adaptive image processing. Among the various noise in star images, mixed Poisson-Gaussian noise is difficult to be accurately suppressed due to its complicated distribution function. Aiming at obtaining the true value of star targets’ signal intensity in infrared images, a novel star target extraction and denoising model called regions with deep reinforcement learning (RDRL) is designed and developed in this study. This fully-automatic model contains two modules: (1) star region extraction module (SREM) that generates star regions within the image through an iterative algorithm based on geometric centroid method (GCM); (2) denoising module that performs an iterative denoising process on the star regions based on deep reinforcement learning. The denoising algorithm is tested on infrared star images, and the experiment results indicate that the proposed RDRL denoising model is able to achieve more accurate restoration with a smaller number of calculations than existing star image denoising methods.