Blurred Image Restoration And Image Quality Assessment Algorithm Based On Deep Neural Networks

Image blurring makes the loss of key information such as license plates,military targets,criminal suspects,etc.;With the development of deep learning and hardware devices,deep learning has achieved very good results in current high-level image processing tasks such as detection,identification,and other tasks.However,in the low-level image processing tasks,there are still some problems.This article focuses on existing problems in image restoration and quality assessment algorithms based on deep learning,and the combination of image restoration and evaluation in order to use the evaluation to improve the restoration.Firstly,an end-to-end multi-scale image restoration network is designed to meet the real-time requirements of the restoration algorithm.The network takes blurred images as input,and estimates the residuals between the clear images and blurred images at multiple scales.Compared to the other restoration methods based on deep learning,the proposed algorithm firstly remove the estimation error of the previous scale,and then estimate the error on the next scale which provides the network with the ability to handle different scales of blurring.Secondly,we research the combination of the commonly used image evaluation algorithms and the image restoration based on the deep learning.For the traditional image quality assessment algorithm SSIM can not be well combined with the deep learning based image restoration network,th is paper reanalyzes the factors considered by the SSIM algorithm,and simplifi es it to make it easier so that it can be used in the gradient descent method;This paper also introduce the regular terms in the traditional image restoration algorithm into the image restoration network,and the image restoration effect is improved;For the problem that the MSE loss function does not meet the image restoration task,we proposed a weighted MSE method to improve the image restoration effect.Then,by analyzing the common image blurring model,an image blur kernel estimate network is designed.The network can estimates the size of the image blur kernel which is related to the blur,different with the traditional method which first estimate the blur kernel and then use the deconvolution method,the blur kernel estimate network is used as a quality evaluation network,which can avoid the ringing caused by blur kernel estimation errors and is more efficient.Finally,aiming at the problems existing in the current deep learning-based image restoration algorithms that cannot be trained with the real blurred image,we propose a method combining the non-reference quality assessment algorithm and the deblurring network which can be trained with the real images.