Generative Adversarial Networks for Blind Image Deblurring

Abstract

In imaging applications, blurred images are of significant challenge and when the issue is addressed without the prior knowledge of blur kernel i-e blind image deblurring, it become more complex. A novel approach for blind image deblurring using Generative Adversarial Network (GANs) for the restoration of sharpness of images is presented in this thesis. A special type of attention module is added in the generator model of GANs which is helpful for finding corelation among the pixels of image. A generator and discriminator network in GANs are trained in adversarial manner and the mapping between blurry and sharp image is learned in unsupervised manner in the proposed framework. Unlike, other traditional deblurring methods that are based on the estimation of blur kernel this method by-pass the need of blur kernel estimation and learns to generate visually pleasing images from the blurred input directly. Extensive experiments are carried out on GoPro benchmark dataset determines that GAN-based deblurring method out-performs the existing traditional state-of-the-art methods in terms of visual quality and objective metrics such as Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index (SSIM). The GAN-based approach shows superior results in complex scenarios when the blur patterns are complex. The ability to surpass the blur kernel makes this method more applicable in real world scenario as the accurate knowledge is challenging to achieve. The contribution in the field of blind image deblurring is made by the introduction of approach that addresses the traditional methods limitations and the results out-performs from the previous techniques. In the proposed frame work the potential of GANs is showcased in tackling the challenge of Blind Image Deblurring. Further avenues are opened for the advancement in the field of Image Processing.