Research On Restoration Algorithms Of Motion Blurred Image

With the development of camera technology, the resolution of image which obtained from either aerospace camera or industrial camera is getting higher and higher, the noise suppression is getting better and better. However, many cases will produce motion-blurred images, such as hand-held camera shake within the integration time, the movement of the target in the scene, the relative motion between the satellite along the orbital flight and the Earth’s rotation during space camera shooting. In order to recover the fine definition, high-resolution images from the degraded inputs, the restoration of motion-blurred images is required.Space-invariant motion-blurred image restoration is the core research content, and the theory and methods of motion-blurred image restoration is research base, the main works and contributions are as follows:Based on the study of the relations between the image gradient and the motion vector, it is deduced the constraint between the gradient of image transparency and motion vector, reducing unknown value. Using the method of spectral matting, alpha components are extracted automatically and its gradient is solved. The motion vector is found by the expectation-maximization algorithm based on RANSAC, and the motion-blurred kernel is estimated using this motion vector. Finally, the latent image is restored by the total variation regularization method. Experimental results show that the blur scale estimation error is less than one pixel using this method on blurred images of2-20pixels, the blur angle estimation error is less than1°. Linear space-invariant motion blurred images or segmentation of moving targets can be restored effectively.Based on the study of the relationship between straight line edges of the motion-blurred image and the Radon transform, it is discovered that motion-blurred kernel can be estimated by detecting the edge of straight lines on blurred image. Taking1°as unit, find the orthogonal slices of different direction straight lines, that is Radon projection of blur kernel, and blur kernel is estimated iteratively using this priori constraint. Experimental results show that the straight line edges exist in the blurred image, the method can estimate the blur kernel of any motion form, and get a better restoration.The iterative restoration method is studied based on Bayesian framework. First, the initial value of the blur kernel is estimated quickly based on the Gaussian priori constraint. Then, the estimated blur kernel converges adaptively to the true blur kernel using iterative support detection method. Second, in order to reduce window truncation influence by the sensor obtaining the image, the latent image is restored by pyramid layers. The latent image is composed by smooth region image and detail region image. The current level smooth region image is come from the interpolation of last level restoration result, the current level only restore detail region image, and use the high-pass filter to remove the ringing artifacts, restoring the sharp, content-rich image. Experimental results show that the method can restore the blurred images successfully which blur scale is larger more than15pixels, ringing artifacts are suppressed effectively, and the details of image is recovered clearly.it was designed a dual CMOS imaging system which acquired the same scene images simultaneously, one CMOS sensor acquired high frame rate, low spatial resolution image sequences, another CMOS sensor acquired low frame rate, high spatial resolution image. The global motion path was obtained by computing the high frame rate, low resolution CMOS sensor acquired image sequences using optical flow method. Under the constraint of energy conservation and the constraint of energy should be proportional to integration interval, initial motion blur kernel was estimated. Motion blur kernel was refined using alternating iterative method by Bayesiari criterion. Finally, a sharp image was restored quickly and effectively from the low frame rate, high spatial resolution of CMOS sensor acquired blur image using TV-L1algorithm. Simulation and experiment results indicate that there are over38%of the simulation result images which error ratios are less than2, restoration images are less affected by noise and less ringing artifacts, the space-invariance motion blurred photographs are deblurred effectively.All of the algorithms are verified by experiments with the actual photography, and have good application value. This research is the most important part of the Eleventh Five pre-research project "CMOS camera electronics system development", the project has been delivered and accepted in September2011.