Research On Spatial Variant Motion Deblurring Technologies

With the wide spread of digital photography, image quality enhancementhas gradually become one of the heated topics in the area of digitalmultimedia. And the achievements in this area have significant meaning andgreat application value in the development of digital content industries, suchas video post-production, animation creation, interactive entertainment and soon. As one of the most important part of image quality enhancementtechnologies, deblurring can be categorized into handling two kinds of blur,namely defocus blur and motion blur. Defocus blur usually takes place whenthe focal length is not matched. While motion blur is often caused by therelative motion between camera lens and the objects shot, which usuallyoccurs if the camera shakes or objects move quickly during the exposure time.We mainly discuss motion blur in this paper rather than defocus blur.Generally, the mathematical model of motion blur can be defined as theconvolution of the latent image and the blur kernel, plus random noise. As aconsequence, the problem of motion deblurring can be simplified as adeconvolution process based on the estimation of blur kernel.Motion deblurring methods can usually be divided into spatial invariantmethods and spatial variant methods, according to the properties of the blurkernel (also known as point spread function). Depending on whether the blurkernel is different for different pixels, the mathematical model can be singlekernel or multi kernel. Spatial invariant algorithms, or single kernelalgorithms, assume that the blur kernel remains the same for all the pixels inan image, which is a simplification of motion deblurring problems. Spatialvariant algorithms calculate different blur kernels for each pixel in an image, thus are much more complicated and difficult.The aims of this paper are to first study and analyze existing motiondeblurring methods, then combine all the heated-discussed and difficultproblems in the motion deblur area, and finally focus on spatial variantmotion deblurring methods based on single image. How to perform effectiveimage segmentation according to the similarity of blur kernel and how toapply spatial invariant motion deblurring methods to different imagesegments are the main points discussed in this paper.On the whole, this paper proposes a new spatial variant motiondeblurring method based on image segmentation, which includes two majorsteps:(1) estimate the blur kernel of different segmented regions andoverlapping regions according to the assumption that the blur kernels are thesame within each region and are similar between adjacent regions;(2) putdifferent regions back together by use of overlapping regions to make thefinal result smooth and natural. Experimental results showed that thisalgorithm is better than those methods based on single kernel.