TY - GEN
T1 - GaN-based image deblurring using DCT discriminator
AU - Tomosada, Hiroki
AU - Kudo, Takahiro
AU - Fujisawa, Takanori
AU - Ikehara, Masaaki
N1 - Publisher Copyright:
© 2020 IEEE
PY - 2020
Y1 - 2020
N2 - In this paper, we propose high quality image debluring by using discrete cosine transform (DCT) with less computational complexity. Recently, Convolutional Neural Network (CNN) and Generative Adversarial Network (GAN) based algorithms have been proposed for image deblurring. Moreover, multi-scale architecture of CNN restores blurred image cleary and suppresses more ringing artifacts or block noise, but it takes much time to process. To solve these problems, we propose a method that preserves texture and suppresses ringing artifacts in the restored image without multi-scale architecture using DCT based loss named “DeblurDCTGAN.”. It compares frequency domain of the images made from deblurred image and ground truth image by using DCT. Hereby, DeblurDCTGAN can reduce block noise or ringing artifacts while maintaining deblurring performance. Our experimental results show that DeblurDCTGAN gets the highest performances on both PSNR and SSIM comparing with other conventional methods in GoPro, DVD, NFS and HIDE test Dataset. Also, the running time per pair of DeblurDCTGAN is faster than others.
AB - In this paper, we propose high quality image debluring by using discrete cosine transform (DCT) with less computational complexity. Recently, Convolutional Neural Network (CNN) and Generative Adversarial Network (GAN) based algorithms have been proposed for image deblurring. Moreover, multi-scale architecture of CNN restores blurred image cleary and suppresses more ringing artifacts or block noise, but it takes much time to process. To solve these problems, we propose a method that preserves texture and suppresses ringing artifacts in the restored image without multi-scale architecture using DCT based loss named “DeblurDCTGAN.”. It compares frequency domain of the images made from deblurred image and ground truth image by using DCT. Hereby, DeblurDCTGAN can reduce block noise or ringing artifacts while maintaining deblurring performance. Our experimental results show that DeblurDCTGAN gets the highest performances on both PSNR and SSIM comparing with other conventional methods in GoPro, DVD, NFS and HIDE test Dataset. Also, the running time per pair of DeblurDCTGAN is faster than others.
KW - Blind deconvolution
KW - DCT(Discrete cosine transform)
KW - GAN
KW - Image deblurring
UR - http://www.scopus.com/inward/record.url?scp=85110441024&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85110441024&partnerID=8YFLogxK
U2 - 10.1109/ICPR48806.2021.9412584
DO - 10.1109/ICPR48806.2021.9412584
M3 - Conference contribution
AN - SCOPUS:85110441024
T3 - Proceedings - International Conference on Pattern Recognition
SP - 3675
EP - 3681
BT - Proceedings of ICPR 2020 - 25th International Conference on Pattern Recognition
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 25th International Conference on Pattern Recognition, ICPR 2020
Y2 - 10 January 2021 through 15 January 2021
ER -