Investigation On Restoration Of Rotatioanl Motion Blurred Image And Video Stabilization

Restoration of motion blurred image and video stabilization are both important subjects in the field of image processing. The random motion between the target and camera during the image exposure or video capture will cause the degradation of the image quality or the instability of the video. Such degradation or instability result in great difficulty for observing or succeeding job like image reorganization, so it's necessary to hold the process of image restoration and video stabilization. They are widely used in the field of television guidance, robot vision and video monitor, etc. In this paper, both of them are studied and innovations are promoted.For restoration of motion blurred image, current results focus on the common motion blur without nonlinear movement. For the rotational blur, it is much more complex because it is space variant. The former approach used to transfer the image to the polar coordinates through geometry transformation so that it can be dealt as the common one. But geometry transformation will not only bring great noise and computation but also gray-level interpolation error. On the other hand, most approaches, such as Weiner filtering or Kalman filtering, are used to correct the morbidity of restoration in frequency domain, which increase the computation complexity with twice Fourier transformations. The same situation happens on the constrained least squares filtering in space domain because of its computation of high order matrix.In this paper, a new approach for the restoration of rotational motion blurred image is presented by the means of gradient-loading. The main jobs of this part are as follows:1) The restoration method along the blur path is presented against the space-variant rotational blur. The degradation function model for the rotational motion blur is established according to the principle of image exposure. Through the deep study on the degradation matrix and precise deduction in spaces domain, the special qualities of degradation function matrix are proved and diagonal-loading in signal processing is introduced to image processing to correct the morbidity of restoration.2) The improved approach using gradient-loading is presented and proved through precise deduction in frequency domain to avoid the high-frequency-emphasis effect of diagonal-loading in image processing. Gradient-loading is a new filtering technique to correct the morbidity of degradation function matrix totally in space domain, and has little multiplication computation of high order matrix. Thus it is much simpler with low space and time complexity than traditional approaches. Compared with diagonal-loading, it makes use of the character of image processing and has the effect of suitable pixel-smoothening, which make the approach have better immunity for noise. The experimental results demonstrate the effectiveness, noise- resistibility, robustness and low complexity of the approach using gradient-loading.For video stabilization, current approaches are all based on the model of image motion in two-dimension plane. Such model describes the translation, defocusing and rotation of the image. For the current video stabilization technique, there're still some difficult points to be overcome. Firstly, it doesn't perform well in the case of serious rotation. That is mainly because the estimation of rotational motion is quite difficult. On the other hand, current image motion model is nonlinear which is very hard to be solved. Thus linear approximation is often made with the premise that the rotational motion is minor. And current approaches can't work in the case of multi-optical-axis transformation. That is because the model is based on the ideal situation to presume the viewpoint as infinitely far away from the target. So it doesn't include the information of projection plane changing.In this paper, a new approach for video stabilization based on Sift feature matching algorithm is presented. The main contributions of this part are as follows:1) A new motion model based on the theory of perspective projection is presented. This model is deducted from the coordinate's transformation in three-dimension world. It obeys the principle of perspective projection plane transformation. So this model is a much more exact description of actual motion includes translation, defocusing, especially the serious distortion caused by rotational motion and multi-optical-axis change. What's more, it is a totally linear model, which make it easy to be solved without any approximation.2) According to the model, four match points are needed to determine the whole transformation model, so points-matching are chosen as the method of motion estimation. Sift feature-matching algorithm is introduced to meet the requirement of rotation and multi-optical-axis change. Through deep study on the Sift algorithm, a simplified approach is introduced for the real-time environment.