Improved KNN Matting For Motion Blurred Objects

Motion blur may happen when camera or objects move during photographing, however, existing matting research mainly aims for clear image. In this thesis, we study matting of motion blurred image. Especially, we are interested in KNN matting used in the clear image, which don’t need assume the linear color model and apply a complex sampling strategy. We study the properties of motion blurred image and propose improvements for extending KNN matting to the motion blurred image. The main work can be summarized as following:First, the status quo of motion blur image matting and the basic principles of matting as well as motion blur are discussed. This thesis reviews the popular matting algorithms for both clear image and blur image. The basic theories of motion blurring are also reviewed.Second, an improved KNN matting based on blur process modeling is proposed. In KNN matting, the pixel matching error will be introduced when image is motion blurred because of the K-nearest neighbor matching algorithm. To overcome this shortcoming, the improved KNN matting incorporates blur kernel information is proposed.Third, an improved KNN matting based on blur feature extraction is proposed. To overcome the instability of blur kernel estimation, a blur measurement feature incorporated KNN matting is proposed. Firstly, the blur measurement feature using singular values is extracted and, consequently, a new feature composed of color, position and blur information is introduced. Then, the matting Laplacian matrix is constructed by combining local smooth prior and nonlocal principle. At last, additional local gradient smoothness constraint is introduced into the model.