CFF-Net: cross-frequency fusion network for aero-degraded image restoration

During high-speed flight, aero-optical effects are a key factor for the image degradation (including image distortion, offset, intensity reduction and image blur) of aircraft with optical imaging systems. Effective solution for neutralizing aero-optical effects is essential to improve the performance of detection, identification and tracking of the aircraft. Conventional algorithms naturally analyse blur in the frequency domain by estimating the blur kernel of the target image and the given blurred image, which is a cumbersome process. Deep learning-based aero-degraded image restoration method aims to learn the nonlinear relationship between ideal sharp image and degraded images at the pixel level through an end-to-end architecture, which usually ignores the frequency domain information. In this paper, we propose a Cross-Frequency Fusion network (CFF-Net) for aero-degraded image restoration. Simulation results show that the network is capable of achieving high-performance aero-degraded image restoration, and the average PSNR of the restored image is improved by 8.33 dB and the average SSIM is improved by 30.70% compared to the degraded image.