Research Of Deblurring From A Single Motion Image Based On Detection And Segmentation

With the vigorous development of digital photography technology,more and more branches appear in image processing,and the restoration of blurred images is one of the important research directions.In many cases of image blurring,the blurring caused by the relative motion between the camera and the objects is called motion blurring.What's more,the motion of the camera will make the whole image blurred,while the motion of the objects in the scene will lead to the local blurring of the motion image.According to the mathematical model of motion blurred image,the core of its restoration is the blur kernel estimation.For global motion blurred images,it is usually assumed that there is only one blur kernel in the whole image.For local motion blurred images,the blur kernel varies with the spatial variation,and the mathematical model is more complex.Restoration algorithms for local motion blurred images mostly depend on multiple images,or assume that blurred images are caused by uniform linear motion.The method of depending on deep learning to solve the problem of non-uniform motion blurring,often takes a lot of time due to its huge amount of computation,and generally requires hardware support.Therefore,after studying the previous work in the field of blurred image restoration,this thesis will focus on the research of local motion blurred image restoration,and propose some improved algorithms to improve the quality of image restoration.Firstly,in view of the complexity and high time-consuming characteristics of previous algorithms,this paper proposes a restoration framework of single local motion blurred image.The framework improves the utilization ratio between the results of each step and also improves the experimental efficiency.The central idea of this framework is as follows:After detecting and segmenting the motion blurred regions,the blurred region is restored to obtain a complete clear image.Secondly,an algorithm of blurred region detection is proposed in this thesis.Considering the relationship between the singular value and the blur model,and the positive effect of multi-scale model on improving the accuracy of region detection,and also the effective correction of matting segmentation for the mapping of the blurred region.By combining the image singular value with multi-scale perception and closed-Form matting algorithm,a two-step blurred region detection is realized.Under the condition of reducing the computational complexity,the detection accuracy of the blurred region is effectively improved.The experimental results show that the improved algorithm has better detection results and higher applicability.Finally,an improved method of restoring and splitting the segmented local motion blurred image is proposed as a result of the edge problem between the restored blurred region and the clear region.Using the blur kernel of the detected blurred region,to restore the images obtained in the process of region segmentation.Finally,remove the redundant part,and split it with the clear region.This method solves the problem of splicing edge of the restored blurred region and the clear region.The experimental results show that the improved method can effectively use the detection results of the previous step,solve the problem of image stitching after restoration,and avoid the boundary effect.