Study On Motion Blurred Video Images' Restoration

Target on ringing artifacts caused by Richardson Lucy,that uniform deblurring are not good for local blur image and that inaccurate optical flow will affects deblurring,this paper proposes two novel methods.(1)We start from imaging theory,model the whole blur process using projective geometry.Considering that guide filter's strength of preserving edge and suppressing noise could overcome the drawback of Richardson Lucy algorithm's ringing artifacts,we combine guide filter with Richardson Lucy algorithm,proposing a novel algorithm,namely,guide filter regularization Richardson Lucy algorithm.This modified model shows its superiority in strong contrast edges' deconvolution.The edges with low contrast are smoothed out,unfortunately.We modify this model again by adding multiscale detail layers' deconvlution at each iteration of guide filter regularization Richardson Lucy.The experiments turn out to be great.This modified guide filter regularization Richardson Lucy algorithm which is incorporated with multi-scale details' deconvolution does reduce ringing artifacts when compared with standard Richardson Lucy and preserve more details than bilateral regularization Richardson Lucy in [17],showing its superiority obviously.To better illustrate deblurring results,we evaluate the deblurring quality with RMSE,PSNR and MSSIM.Results indicate that our method is the best compared with standard Richardson Lucy and bilateral regularization Richardson Lucy in [14].(2)The deblurring technique proposed in this part is based on blur detection.Targeting on aforementioned model's weakness of the local blur image's deblurring,we,firstly,refine it and propose an improved Kurtosis-feature-based blur detection algorithm.Mend the Kurtosis feature's selection mode.Comibining the mended Kurtosis feature with gradient feature,average spectrum and local learned filter,we use na?ve Bayesian classifier to learn the posterior for these feature sets.After that,we build a multi-scale relationship of these features in different scales.The final score represents blur detection result.This update model has improved blur detection accuracy especially for direction blur motion images.Incorporate this blur detection model with aforementioned multiscale guide filter regularization Richardson Lucy algorithm,the local blur image could be deblurred better.Leading this new blur detection algorithm into video deblurring,we mend it again and propose a motion-detection-based video deblurring.Firstly go into object detection using GMM and Three Frame Differencing,label them,do blur detection and calculate their blur levels(mean of the normalized local blur features)respectively.Secondly,obtain each labeled area's optical flow by optical flow algorithm of LK.Thirdly,we build a coarse to fine vetor selection model to generate blur kernel of each label.Finally,we deblur the blur area(the labeled area with high blur level)using the proposed multiscale guide filter regularization Richardson Lucy model.The deblurring results tell us that the choosen vector which is used to generate blur kernel agrees with motion objects'.Besides,the dedeblurring results are obviously better than uniform deblurring.