High fidelity single image blind deblur via GAN.

To reconstruct the high resolution image makes sense from single image with low resolution. Most conventional methods assume that the blur kernel is known, however the blur kernel within single blurring image is always unknown, then it is necessary to generate the kernel dynamically during deblurring. The paper proposes single image high-fidelity blind deblurring method based on GAN. The degradation of super resolution networks is first used to synthesize high resolution images. A blur kernel discriminator is then trained to analyze the generated high-resolution images and errors that occur when the prediction generator provides incorrect blur kernel. Thus, the blur kernel provided by the generator is closer to the actual image. Finally, through the redefinition and optimization of loss function, the perception loss is replaced by the per-pixel loss to obtain better visual effects. Experiments show that the proposed method can achieve high fidelity deblurring results 1.2% higher than those of traditional methods. [ABSTRACT FROM AUTHOR]