Infrared scene-based non-uniformity correction based on deep learning model

Abstract

In this paper, an infrared scene-based non-uniformity correction method based on deep learning technology has been proposed. This method combines the scene-based infrared non-uniformity correction with the state-of-the-art deep learning technology. The traditional scene-based non-uniformity correction technologies generally face the problem that when the radiation of the scene is changed, the correction parameters may not converge anymore, and the correction process will rise again. Multiple times of correction will potentially increase the risk of the correction failure. The deep learning can help setting up a systematic correction parameter which self-adaptive to the thermal imager, which means that once the parameters are computed, the non-uniformity will be corrected according to the deep learning network by itself. Although the pre-calculation period of setting up the parameters take much time, the upcoming correction process is much easier than the traditional technologies. We use the feed-forward denoising convolutional network as the fundamental structure, and deploy the modified residual learning process as well as the batch normalization process onto it. Figures and charts show the priority of our method.