Motion blur is one of the most significant and common artifacts causing poor image quality in digital photography, in which many factors resulted. In imaging process, if the objects are moving quickly in the scene or the camera moves in the exposure interval, the image of the scene would blur along the direction of relative motion between the camera and the scene, e.g. camera shake, atmospheric turbulence. Recently, sparse representation model has been widely used in signal and image processing, which is an effective method to describe the natural images. In this article, a new deblurring approach based on sparse representation is proposed. An overcomplete dictionary learned from the trained image samples via the KSVD algorithm is designed to represent the latent image. The motion-blur kernel can be treated as a piece-wise smooth function in image domain, whose support is approximately a thin smooth curve, so we employed curvelet to represent the blur kernel. Both of overcomplete dictionary and curvelet system have high sparsity, which improves the robustness to the noise and more satisfies the observer's visual demand. With the two priors, we constructed restoration model of blurred images and succeeded to solve the optimization problem with the help of alternating minimization technique. The experiment results prove the method can preserve the texture of original images and suppress the ring artifacts effectively.
Aerial images captured using time delay and integration (TDI) charge-coupled devices (CCDs) could be blurred by three types of motion: forward image motion, turbulence disturbance and high frequency vibration. This work proposes a method to separately construct the three deterministic models by discerning or calculating the parameters from a single image blurred by all the three ones. Based on these models, we catch and separate the features existing in the power spectrum diagram, and select the methods with the best identification accuracy to the parameters. The results show the approach we mention can promise the accuracy of the determined parameters, which is helpful to improve the result of blind restoration algorithm.
To accurately discern the parameters of high frequency vibration blur model on a single TDI image, the research analyzes the imaging function when high frequency vibration occurs in TDI mode. The method of simplifying the vibration model is offered and verified, which promises the MTF will be only related with motion angle and vibration amplitude. Three algorithms for motion direction discerning are compared with one another, which are Radon transform, autocorrelation analysis and cepstral method. The conclusion reveals that cepstral method can measure the most accurate motion angle. Four algorithms for vibration amplitude discerning are compared, which are the quadratic Radon transform, cepstral analysis, autocorrelation analysis and direct analysis on frequency spectrum. It reveals that direct analysis on Log frequency spectrum is the most accurate for vibration amplitude. The research suggests that composition of cesptral method and direct analysis on log frequency spectrum could obtain the highly accurate parameters in high frequency vibration model.