Blur Removal In Natural Digital Images Using Self-reference Generative Networks

Abstract

The blur detection in a single image is considered as a pivotal issue in digital image processing applications, especially when the blurring is spatially-varying. This study introduces self-reference generative network (SR-GN) algorithm in order to deblur a digital natural blurred image. The proposed method divides the contaminated image into two main parts, which are the clear part and blur parts. The blur parts which caused by the shadowy channel are less sparse while the other parts have some dark patches, which can be cleaned by avenging them. In addition, due to the global performance of the proposed algorithm on public database, multicomponent loss function is utilized in order to perform further classification to the original patches of the blurred image to distinguish it from the contaminated counterparts as the experimental results demonstrated. The experimental results show that the proposed technique has an improvement in visual quality, and objective results as well in comparison to related advanced blur removal algorithms.