Frequency-domain-guided deep Wiener deconvolution for optical sparse aperture image restoration.

Abstract

Optical sparse aperture (OSA) imaging systems synthesize a large effective aperture with reduced volume and weight, but their sparse pupil configurations attenuate the mid-frequency modulation transfer function and degrade image quality. Existing learning-based restoration methods mainly optimize spatial-domain reconstruction and provide limited explicit supervision for the frequency-dependent degradation characteristics of OSA systems. To address this issue, we propose a frequency-domain-guided deep Wiener deconvolution network (FDWDN), which incorporates the point spread function as an explicit physical prior and introduces complex-spectrum supervision into the reconstruction objective. This formulation enables the restoration process to be directly constrained in the frequency domain and naturally supports frequency-selective weighting. At an SNR of 40 dB, FDWDN improves PSNR from 25.16 to 33.43 dB, SSIM from 0.6381 to 0.8767, and mean mid-band Fourier ring correlation from 0.5376 to 0.7375. FDWDN also achieves better overall restoration performance than conventional methods and representative deep restoration networks. These results demonstrate that FDWDN provides accurate spatial reconstruction and improved mid-frequency spectral consistency for OSA image restoration.