Deep dual de-noising network for single corrupted infrared image

Abstract

Infrared images that can detect radiant energy emitted from objects and people are receiving more attention than ever. However, the infrared image is essentially corrupted by stripe pattern noise (FPN) due to the detector's signal readout properties. Recently, various researches based on deep learning have been proposed for FPN de-noising. However, existing deep learning-based methods still have performance limitations such as lack of robustness and discrimination. In particular, since FPN has the property of appearing in the vertical direction, image information in the vertical direction could be corrupted inevitably during a de-noising process. Most of the existing methods did not sufficiently consider the lost details recovery mechanism. In this paper, we propose a novel deep dual de-noising network based on a residual learning strategy to tackle the above limitations. To this end, first, we designed the stripe pattern de-noising network based on a two-level encoding-decoding structure capable of symmetrically combining feature maps, and FPN was estimated from corrupted images through residual learning. At this time, in order to extract the stripe noise effectively, the functional module composed of the densely connected residual layers was applied. Secondly, detailed information recovering network was designed to restore lost details in the de-noising process. The coarsely estimate FPN is decomposed into four sub-bands using the Haar wavelet transformation, and through residual learning, the lost information is directly estimated using the details extraction module from the low-frequency components and horizontal high-frequency components including both FPN and lost information. Finally, from the learning result of the stripe pattern noise de-noising network, and the lost information extracted by the detailed information recovering network, we reconstructed a noise-free infrared image without FPN while maintaining image details. Through comparative experiments using corrupted synthetic infrared images and real infrared images, we showed qualitatively and quantitatively that the proposed method achieved outstanding performance in terms of the de-noising robustness against FPN intensities, the improved de-noising discrimination for different FPN, and the recovering ability of lost information