Research On Motion Blur Image Deblurring Technology Based On Computational Photography

Motion blur image restoration problem has been known in the field of image processing. With the development of microelectronic, sensor, computer and other technologies, people began to use these new technologies to improve all kinds of problems and limitations in tranditional photography, and gradually a new research field - computational photography formed. Computational photography applies multiple advanced technologies of different fields into the restoration of motion-blurred images, and skillfully solves the traditional image restoration which needs to depend on complex mathematical models, which has currently become a hot research.This paper focuses on the theme of restoration of motion-blurred images. Hybrid camera and coded exposure techniques are used to estimate the point spread functions of motion-blurred images and make the ill-posed deconvolution problem well-posed. An image sequence is captured simultaneously by high-speed low-resolution image sensor in the exposure time of the high-resolution low-speed sensor. The motion parameters during exposure of high-resolution sensor are obtained by visual measurement techniques using the image sequence. The point spread function is estimated by modeling the motion blur using the motion parameters. Coded exposure photography which changes the imaging process of camera is used to capture the motion-blurred image by high-resolution sensor. The coded exposure motion-blurred image can be deblurred using direct deconvolution which is not ill-posed. This paper's research contents and innovations are as follows:Motion-blurred images Modeling. The traditional study of motion-blurred images restoration which uses a model of two-dimensional clear image to two-dimensional blur image is difficult to accurately describe the process of motion blur. This paper starts from the source of image acquisition, the camera imaging process into the analysis of the process of motion blur, established a motion blur model of three-dimensional scene to two-dimensional image. Algorithms for the point spread function estimation of linear and complex motion-blurred images based on imaging model are proposed which laid a theoretical foundation for later research in this paper.Hybrid vision system calibrating. The high-speed sensor which is used for estimating the point spread function of the motion-blurred image has a different spatial resolution from the high-resolution sensor. A calibration model of hybrid vision system is established which uses spatial scale as the standard metrics. Based on three-dimensional geometric constraints, a relative external parameters estimation algorithm for the hybird visual system is proposed. Comparative experiments between a recognized fine calibration algorithm and our calibration algorithm show that the alogrithm has higher calibration accuracy and calibration stability. In the application of detection of intersection points of chessboard's black and black grids, SUSAN algoritm has a problem of differentiating the edge points and corrner points. A chessboard corner detection algorithm based on SUSAN which combines multi-direction restriction of symmetry and uniformity of the corners is proposed. The real corners of the chessboard are confirmed by computing the restriction of symmetry and uniformity in the result of SUSAN algorithm. The experiments results show that the proposed algorithm has a better validity and precision.Three-dimensional motion measurement based on binocular stereo vision. The motion measurement mission needs to detect and match the features in four-dimensional space (3D space + 1 time dimension). The matching constraints of binocular stereo vision and motion are taking into account at the same time, a stereo motion measurment algorithm based on stereo-motion constraints SIFT matching is proposed. In the matching process, the epipolar constraint is used to reduce the search space and the motion constraint is used to eliminate false matches. For local motion-blurred image caused by moving target, we present an accurate moving target detection algorithm based on color"background Subtraction"and SUSAN. Firstly, a color"background Subtraction"method is used for rough target detection. Secondly, an enhanced SUSAN method is used for accurate target edge estimation. After filling the holes in the edge, accurate target area is estimated. The experiments results show that the target detection algorithm can detect accurate moving target area and the motion measurement algorithm has a better time efficiency and measurement accuracy.Image restoration based on coded exposure photography. Coded exposure changed exposure mode of traditational photography and make the point spread function invertible. Aiming at the problems of coded exposure such as motion blur on texture background, bad time efficiency and resolution losing, we present a local motion blur model based on"image matting", a fast code seach method and a point spread function estimation algorithm based on coded exposure. Combining the hybrid vision system and coded exposure, an image restoration method based on hybrid coded exposure is porposed to break the limitation of coded exposure method. Our method can get the motion distant by motion measurement automatically and can handle the local motion blur images in texture background and accelerated motion.Experiments which compared our method with other well-known algorithms show that our method can deblur the motion blur images in many situations. Our method has a better effectiveness than the coded exposure method and a better time efficiency than the tranditational restoration methods which is applicable to real time applications after programing optimization.