Intriguing Findings of Frequency Selection for Image Deblurring

23 Nov 2021  ·  Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, Yan Wang ·

Blur was naturally analyzed in the frequency domain, by estimating the latent sharp image and the blur kernel given a blurry image. Recent progress on image deblurring always designs end-to-end architectures and aims at learning the difference between blurry and sharp image pairs from pixel-level, which inevitably overlooks the importance of blur kernels. This paper reveals an intriguing phenomenon that simply applying ReLU operation on the frequency domain of a blur image followed by inverse Fourier transform, i.e., frequency selection, provides faithful information about the blur pattern (e.g., the blur direction and blur level, implicitly shows the kernel pattern). Based on this observation, we attempt to leverage kernel-level information for image deblurring networks by inserting Fourier transform, ReLU operation, and inverse Fourier transform to the standard ResBlock. 1x1 convolution is further added to let the network modulate flexible thresholds for frequency selection. We term our newly built block as Res FFT-ReLU Block, which takes advantages of both kernel-level and pixel-level features via learning frequency-spatial dual-domain representations. Extensive experiments are conducted to acquire a thorough analysis on the insights of the method. Moreover, after plugging the proposed block into NAFNet, we can achieve 33.85 dB in PSNR on GoPro dataset. Our method noticeably improves backbone architectures without introducing many parameters, while maintaining low computational complexity. Code is available at

PDF Abstract

Results from the Paper

Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Image Deblurring GoPro DeepRFT+ PSNR 33.23 # 16
SSIM 0.963 # 15
Deblurring GoPro DeepRFT+ PSNR 33.23 # 15
SSIM 0.963 # 14
Deblurring HIDE (trained on GOPRO) DeepRFT+ PSNR (sRGB) 31.42 # 5
SSIM (sRGB) 0.944 # 6
Deblurring MSU BASED Deeprft (REDS) SSIM 0.94479 # 8
PSNR 31.32349 # 6
VMAF 66.46811 # 9
LPIPS 0.08139 # 2
ERQAv2.0 0.74339 # 6
Subjective 0.4622 # 7
Deblurring MSU BASED Deeprft (GoPro) SSIM 0.94484 # 7
PSNR 31.57612 # 3
VMAF 66.55057 # 8
LPIPS 0.08326 # 8
ERQAv2.0 0.74323 # 7
Subjective 0.5354 # 6
Deblurring RealBlur-J DeepRFT+ SSIM (sRGB) 0.931 # 3
PSNR (sRGB) 32.19 # 4
Deblurring RealBlur-J (trained on GoPro) DeepRFT+ PSNR (sRGB) 28.97 # 4
SSIM (sRGB) 0.884 # 3
Deblurring RealBlur-R DeepRFT+ PSNR (sRGB) 39.84 # 2
SSIM (sRGB) 0.972 # 3
Deblurring RealBlur-R (trained on GoPro) DeepRFT PSNR (sRGB) 36.06 # 6
SSIM (sRGB) 0.954 # 6