Research On Image Restoration Based On Vibration Detection Using Fiber Optic Gyroscope

No matter in daily photography or high-resolution remote sensing, bad lighting condition means the increase of exposure time, the accompanied problem is, during the exposure time, the hand shake or platform vibration results in blurred images, which will greatly reduce the amount of information contained in the images. In mathematics, image blurring is modeled by convolving the clear image with a point spread function (PSF) plus some noise. Its inverse process is called image deconvolution, which is a typical ill-posed problem, the result is usually contaminated by amplified noise and ringing artifacts. However, regularization methods can be used to improve the performance of image deconvolution algorithm to achieve result of high quality. Image deconvolution algorithm is very useful, especially in the application of high-resolution remote sensing, which can effectively avoid the sharp rise of the cost caused by the design and manufacture of imaging systems with large aperture, long focal length and the corresponding stabilization equipments. Therefore, the thesis focuses on the image deconvolution algorithm and detailed discusses some key problems about it, which are summarized as follows.The first is about how to obtain the PSF. PSF plays a very important role in image deconvolution. There are various methods which can be adopted to achieve it, such as the point lighting source method, the knife edge method and the method based on motion trajectory which is proposed by S.K.Nayar and M.Ben-Ezra. According to the characteristics of vibration, we propose a PSF reconstruction method using the fiber optic gyroscope. The fiber optic gyroscope is a kind of angular velocity sensing device, which has the advantages of compact structure and high detection accuracy. We rigidly connect two fiber optic gyroscopes whose sensitive axes are perpendicular to each other to the optical axis of the camera, such that we can get the angle position of the vibration in two directions during the exposure time, then according to the relationship between the object and the image, we can obtain the motion trajectory of the image and reconstruct the PSF. In addition, we also improve the traditional PSF reconstruction method based on motion trajectory. In the improved version, with the method of setting a high sampling frequency, the complicated geometric and interpolation calculations are converted into a simple statistical calculation, and thus enhance the efficiency.According to whether the PSF is known, image deconvolution algorithms can be divided into two kinds, i.e., non-blind and blind image deconvolution. The thesis propose three non-blind image deconvolution algorithms, they are the piecewise local regularized Richardson-Lucy (RL) algorithm, the regularized RL algorithm based on natural image gradient prior and the algorithm based on Gaussian Scale Mixture Fields of Experts (GSM FoE) prior. The first two are derived from the model of Poisson noise, which are all improved versions of standard RL algorithm. In the piecewise local regularized RL algorithm, we design a regularization term based on the Gaussian Markov random field, and adopt a piecewise power function to adjust the smoothing strength to suppress the amplified noise and ringing artifacts in the restored image. Similarly, in the regularized RL algorithm based on natural image gradient prior, the model of Poisson noise is combined with the prior of image gradient, because the prior is in good agreement with the sparse probabilistic distribution of the gradient of nature images, the algorithm can efficiently improve the performance of standard RL algorithm and reach result of high quality. GSM FoE which is derived from Fields of Experts (FoE) is a very effective probabilistic model for natural image, in contrast to traditional priors, its advantage is that all the filters used to construct the GSM FoE prior are trained with a database, thus it can seize the characteristics of natural images more accurately. The thesis introduces this prior into non-blind image deconvolution and adopts the Split Bregman method to solve the resulted cost function. Experimental results show that its performance meets or exceeds some state of the art methods.The thesis also proposes three blind image deconvolution algorithms, i.e., the single image deconvolution algorithm based on Fields of Experts (FoE) prior, the single image deconvolution algorithm based on natural image gradient prior and the deconvolution algorithm with multiple blurred images. There are two innovations in the single image deconvolution algorithm based on FoE prior. Firstly, we introduce the FoE prior into blind image deconvolution. similar to the GSM FoE prior, all its parameters and filters are trained from natural images, thus it is of high accuracy. Meanwhile, a prior based on Student-t function is also used to regularize the PSF. Secondly, we improve the traditional alternating minimization (AM) algorithm which is usually used to solve the problem of blind image deconvolution. In each iteration of the improved algorithm, the mid-restored image obtained from the former iteration is used as a constraint, and thus it will not converge to the unwanted "blurry" result. The single image deconvolution algorithm based on natural image gradient prior adopts the proposed regularized RL algorithm with natural image gradient prior. The result is reached by the traditional AM algorithm. The innovation is that the Gaussian noise model is used to approximate the Poisson noise model to estimate the PSF, so the computation speed and accuracy are enhanced. We also propose a blind image deconvolution algorithm adopting multiple blurred images, the natural image gradient prior and a L1norm based prior is used to regularize the clear image and the PSF respectively. Since multiple frames contain more information about the clear image, it can achieve results of high quality.Finally, we also set up experimental system to verify the effectiveness of the proposed PSF reconstruction method and the image deconvolution algorithms. The results show that the image deconvolution method based on vibration detection using fiber optic gyroscope is robust, it can effectively enhance the quality of the obtained image. Meanwhile, the performances of the deconvolution algorithms proposed in this thesis are also good. Especially, the blind image deconvolution algorithm based on the FoE prior and the non-blind image deconvolution algorithm based on the GSM FoE prior are comparable with some state of the art methods.