Research On Blind Motion Deblurring Of Mobile Phone Photographs

Photographs taken with mobile phone cameras often suffer from motion blur,mainly due to the natural tremor of the photographer's hand and the moving objects in the scene.Lately,blind motion deblurring is becoming an important problem in digital image processing.Multiple images can provide more useful information for blind motion deblurring.Based on the random nature of camera shakes in the successive frames,multi-frame Fourier domain aggregation can effectively cope with motion blur due to camera shake.While leveraging the correlation between inter-frame motion and intra-frame object motion,method based on local piece-wise linear motion model approximation shows advantage in object motion deblurring.Based on the displacement between forward and backward optical flows,this thesis proposed a blur correction aware multi-frame fourier domain local aggregation deblurring algorithm through an improved consistency checking and handling method.Experiments on real and synthesized blurry sequences demonstrate that the proposed algorithm can effectively cope with the motion blur caused by camera shake and moving objects.In addition,we revisit in detail the classical spatially uniform single-image blind deblurring problem.We ask two important questions based on a thorough understanding of the recently successful maximum a posteriori（MAP）based blind deblurring approaches.They are the kernel estimation quality evaluation problem and the dual kernels problem,which we try to answer through theoretical and experimental analysis.We proposed a L₀ energy function based kernel estimation quality metric and claimed that in the dual kernel case,direct kernel estimation result tends to be close to one of the dual kernels.For non-uniform motion blur caused by real camera shake,the convolution model is not suitable.We assume locally uniform blur,and estimate multiple blur kernels from different local regions of the blurred image using existent blur kernel estimation method and use them to generate multiple deblurring images.Then a novel L₀ energy function based kernel estimation quality evaluation metric is used to automatically select the optimal local blur kernel and the corresponding deblurring patch for each overlapping blurred patch,and yield final deblurring result through weighted aggregation.Experiments show that this conceptually simple algorithm can effectively remove non-uniform motion blur caused by camera shake.Using the estimated local optimal blur kernels as a good initialization,we adopt the so-called Projective Motion Path Blur（PMPB）model to constrain the blur caused by camera shake to the camera geometric motion model.Considering two commonly used 3D approximated camera pose subspaces respectively,we restraining the possible camera poses in a low-dimensional subspace and iteratively estimate the weight for each pose.Experiments demonstrate that this algorithm can remove the non-uniform camera shake more effectively.We also analyze the differences and relationships between the two used 3D approximated camera pose subspaces with PMPB model,and briefly summarize their effectiveness in different situations.Lastly,we address a common problem scenario where both object motion blur and camera shake are involved:objects in the scene move rapidly in front of the background.In this case,the foreground object and the background undergo significantly different motion blur process and the background is partially occluded by the moving objects.We explicitly model the occlusion and blending effect in the boundary of foreground moving objects,and an iterative procedure is performed to solve an optimization problem that includes both natural image matting constraints and blur constraints to update the latent sharp images and the mask of foreground moving objects.Experiments show that the deblurring result of this method does not produce significant artifacts at the boundary between the foreground moving objects and the background,and can effectively remove motion blur.