Image restoration for optical synthetic aperture system via patched maximum–minimum intensity prior and unsupervised DenoiseNet

Optical synthetic aperture system improves spatial resolution significantly. However, the image will be degraded and blurred because of the sparse sub-mirror arrangement. Therefore, it is of great significance to exploit the restoration method to obtain an ideal latent image. A method combining maximum a posterior (MAP) framework model optimization with deep learning is presented to enhance imaging quality. The model is optimized by patched maximum–minimum intensity (PMMI) to alleviate insufficient representation of existing priors, while DenoiseNet is designed to train the device and eliminate noise interference. Unsupervised learning is applied to solve the acquisition difficulty of original image labels dispensed with clean samples as training set. When a noise of 30 dB is added, the PSNR of the remote sensing test set increases from 20.90 dB to 23.01 dB. SSIM ameliorates from 0.345 to 0.651 by 88.6% and MS-SSIM from 0.781 to 0.883 by 13.1%. The mid-frequency MTF also has a significant improvement.