DeepPSF: On-Orbit Point Spread Function Estimation of Space Camera With Deep Learning

Obtaining an on-orbit space camera’s point spread function (PSF) is challenging but necessary for remote sensing image (RSI) restoration. Current methods rely on regular ground targets, causing determining the PSF at any position on the camera sensor to involve significant human effort and be highly inefficient. To reduce the difficulty and improve the precision of estimating the PSFs of an on-orbit space camera, this paper proposes DeepPSF, a novel PSF prediction method based on deep learning and Fourier transformation. DeepPSF employs a dual-stream convolutional neural network to extract multi-scale features from blurred and reference images, introduces a channel-wise Wiener filtering block for PSF feature calculation in frequency domain, and reconstructs high-precision PSF through a CNN network. Experiments demonstrate: 1) On synthetic datasets, DeepPSF achieves PSF prediction with 58.2dB PSNR (SSIM> 0.64), significantly outperforming Wiener filtering and phase-only image based kernel estimation method (POI); 2) When combined with the non-blind deblurring algorithm DWDN, it delivers 26.1dB restoration PSNR, surpassing comparative methods; 3) Real remote sensing image tests validate its adaptability to complex scenarios. This method provides an efficient solution for full field-of-view PSF modeling of on-orbit cameras.