HINRDNet: A half instance normalization residual dense network for passive millimetre wave image restoration

Abstract

During the imaging process of passive millimeter wave (PMMW), due to the long wavelength of millimeter wave, the small aperture of the antenna, and the introduction of noises, the obtained PMMW image are degraded by blur and noise simultaneously. To address these problems, we propose a passive millimeter-wave image restoration network called as HINRDNet which includes a denoising module (DNM), deblurring and denoising U-Net (DBNU-Net), and an image reconstruction module (RCM). In order to fully utilize local features, the proposed network is designed based on Half Instance Normalization Block, and is mainly composed of the densely connected sum dense HINBlock (SDHB) and the concatenate dense HINBlock (CDHB) block from low step to high step. In addition, the higher level features information is combined with low level features for global feature cross fusion. Finally, we construct remote sensing and infrared simulation datasets to train the network, and test it on simulated datasets and real PMMW degraded images. The experimental results for the simulated degraded images and the real PMMW degraded images demonstrate that our proposed method outperforms the existing state-of-the-art methods in terms of visual perception and evaluation metrics.