Motion Blur Removal of Digital Photographs

Abstract

Motion blur is a common artifact that produces disappointing blurry images with inevitable information loss. It is caused by the nature of imaging sensors that accumulate incoming lights for an amount of time to produce an image. During exposure, if the camera sensor moves or objects move, a motion blurred image will be obtained.Deblurring is to find the latent sharp image from a given blurred image. Formally, motion blur has been often modeled using a convolution based blur model, where a blurred image is the convolution result between a latent sharp image and a blur kernel. Then, deblurring becomes a deconvolution problem. In non-blind deconvolution, the motion blur kernel is given, and the problem is to recover the latent image from a blurry version using the kernel. In blind deconvolution, the kernel is unknown and the problem is to estimate the blur kernel as well as the latent image.While deblurring has been studied extensively my many researchers, it is still very challenging. The major difficulties can be summarized as follows: first there is missing information. Motion blur causes irreversible information loss. Moreover, the blur kernel is unknown in most cases, which makes the problem more difficult. Second, while the convolution based blur model has been widely used, it often does not hold in practice. In real cases, images are usually non-uniformly blurred, e.g., due to object motions and rotational camera shakes. Noise and outliers, e.g., saturated pixels and sensor defects, also severely degrade the performance of deblurring methods.In this thesis, we present software-based solutions to overcome the difficulties mentioned above. Specifically, the thesis includes the following topics:Fast uniform motion deblurring from a single blurred image: Previous blind deblurring methods need a huge computation time. In this work, we propose a fast blind deconvolution method, which produces a deblurring result in only a few seconds. The experimental results show that our method is not only faster but also more reliable than previous methods.Handling outliers in non-blind image deconvolution: Outliers, such as sensor defects and saturated pixels, cause severe artifacts in previous methods. In this work, we explicitly model outliers and derive an Expectation-Maximization based non-blind deconvolution method.Blind deconvolution methods for non-uniform motion blur: Although the uniform blur model has been widely used, real-photographed images usually have non-uniform blur caused by object motions and rotational camera shakes. In this work, we propose two methods for handling non-uniform motion blur.Video motion deblurring: Motion blurred video can significantly degrades the performance of computer vision algorithms. We use unblurred frames to achieve very reliable video deblurring results.