Intriguing Findings of Frequency Selection for Image Deblurring

Blur was naturally analyzed in the frequency domain, by estimating latent sharp image and blur kernel given a blurry image. Recent progress on deblurring always designs end-to-end architectures aims at learning difference between pairs from pixel-level, which inevitably overlooks importance of kernels. This paper reveals an intriguing phenomenon that simply applying ReLU operation domain followed inverse Fourier transform, i.e., selection, provides faithful information about pattern (e.g., direction level, implicitly shows pattern). Based this observation, we attempt to leverage kernel-level for networks inserting operation, transform standard ResBlock. 1 × convolution is further added let network modulate flexible thresholds selection. We term our newly built block as Res FFT-ReLU Block, takes advantages both pixel-level features via frequency-spatial dual-domain representations. Extensive experiments are conducted acquire thorough analysis insights method. Moreover, after plugging proposed into NAFNet, can achieve 33.85 dB PSNR GoPro dataset. Our method noticeably improves backbone without introducing many parameters, while maintaining low computational complexity. Code available https://github.com/DeepMed-Lab/DeepRFT-AAAI2023.