GAN for Blind Image Deblurring Based on Latent Image Extraction and Blur Kernel Estimation

We propose a GAN for image deblurring based on latent image extraction and blur kernel estimation, with which the single image deblurring assignment is successfully completed. We introduce the FFT to the image of the latent image extraction because the FFT can convert image transformation from spatial to frequency domain; this is a good solution for the convolutional neural network for partial frequency domain knowledge learning and gives a sharper picture; We also apply the cross-scale reproducibility of natural images to the extraction of blur kernel. By adding regularization constraints to the kernel to enhance estimation precision, the estimated kernel of the image is generated via fusing of local kernels after numerous iterations. Meanwhile, a multi-scale discriminator structure combining RSGAN and PatchGAN is used. RSGAN is applied as a global discriminator with more accurate classification criteria, while PatchGAN is applied as a local discriminator to determine the accuracy of local blur kernel. The experiment proves that our work is effective.