Research On The Optimization Of Iterative Deblurring Algorithm Based On Image Quality-aware Features

Image acquisition,transmission,storage and other processes are very vulnerable to factors such as imaging equipment,external environment,this generates poor image quality and then a variety of interference information,subsequently image processing task cannot be carried out smoothly.Therefore,image restoration algorithms appear especially important.Among these algorithms,image deblurring algorithm is an important category in image restoration algorithm.The regularization method is widely used in the field of image deblurring due to its excellent performance.It usually solves the optimal estimation of the objective function for the latent image in an iterative manner.However,due to lack of effective iterative stopping criterion,such methods generally set iterative parameter according to experience,so that these algorithms are short of flexibility and efficiency,and even cause the restored image to be destroyed.To solve the shortcomings of iterative deblurring algorithm,this thesis proposes an iterative stopping optimization strategy.Specifically,it is achieved by iterative estimation and adaptive stopping.(1)In the initial iterative point estimation step,an iterative estimation based on image quality-aware features(IE-IQF)algorithm is proposed.The algorithm is based on image quality-aware features and local mean estimation(LME)method.Firstly,the different types and degrees of distortion of blur are added to a set of typical natural images to obtain blurred images.The feature sample database is extracted by the natural scene statistical method under the spatial domain to extract the quality-aware features of the blurred image,and then the iterative operation of deblurring is based on the nonlocally centralized sparse representation(NCSR)algorithm,and determines the optimal iterative point by recording the quality similarity index(FSIM)after each iteration of the NCSR algorithm and using the optimal iterative point Calibrate the corresponding feature vector in the sample library;Finally,use the LME method to estimate the number of iterative steps.Through the above steps,the estimation of the initial iterative monitoring point is completed,so that the subsequent adaptive iterative stopping task has a more purposeful monitoring of the deblurring metric,and the fast and effective algorithm optimization is completed.(2)Adaptive iterative stopping step.Firstly,a deblurring measure(DM)is designed through the statistical characteristics of the residual image(the difference between the intermediate estimated image obtained during the iterative process and the convolution after the blur kernel and the blurred image).Monitoring was performed within a certain range before the iterative step number,and an adaptive iterative stopping condition(adaptive ISC,AISC)was designed using this DM metric,and then the adaptive stopping of the NCSR algorithm was realized.Extensive experiments show that,the IE-IQF algorithm can roughly determine the number of iterative deblurring steps required for a blurred image.The average estimation error is about 60 steps.After two steps of the above optimization strategy,the performance of the improved NCSR algorithm for restoration of the image on the PSNR,SSIM,and FSIM image quality values is similar to the original algorithm.Specifically,the difference range is-0.0465 to 0.3868,-0.0004 to 0.0219,-0.0011 to 0.0079.And the improved NCSR algorithm performs very efficiently,and the average execution efficiency is improved by 31.3% compared with the original algorithm.To prove the versatility of AISC,AISC is applied in other iterative deblurring algorithms after reasonably adjusting the relevant parameters.The experimental results show that,the improved other iteration deblurring algorithm can improve the efficiency of the algorithm under the premise of the restoration of image quality.Take the Lena image with a size of 256?256 as an example,the average difference between the PSNR,SSIM,and FSIM before and after improvement is-0.0917,-0.0033,and-0.0025,and the average execution efficiency increases by 12.7%.