Image Restoration Methods And Its Application In Machine-Vision Based Surveillance

Nowadays, machine-vision based methods have been widely used. However, the results of the machine vision algorithms depend on the quality of the input image. In the engineering application, the images are often degraded, i.e. blurred and noisy. Thus, it is an effective way of improving the validity of the machine vision algorithms to restore the "idea image". In this thesis, based on the partial differential equation (PDE) and inverse problem theories, we propose three image restoration methods, i.e.1. We propose regularized backward heat diffusion method. By introducing the reg-ularization theory and methods, we propose an effective method of restraining the noise amplification, and then build the regularized backward heat diffusion method used for image restoration. In addition, we improve the corresponding theories and prove the effectiveness of the regularized backward heat diffusion method.2. We propose "learning—filtering" algorithm. We improve the traditional image restoration models which base on gradient operator, e.g. Tikhonov regularization method and the total variation (TV) model, by learning a set of effective filters to realize the feature mapping from the image and noise samples, rather than using the gradient operator directly. We also prove the validity and convergence of the "learning—filtering" algorithm.3. We propose the generalized motion blur recovery model. For a long time, the im-age motion blur is described as a convolutional process. We design a generalized motion blur model and the corresponding image restoration method. We imple-ment three subproblems:(1). passive navigation and optical flow estimation,(2) fast algorithm of generating the motion blur kernel; and (3) general motion-blur recovering model and the corresponding algorithm.Finally, these image restoration methods are integrated in the system used for railway-environment-surveillance, as a preprocess step for the machine vision algorithms.