Research On Image Restoration And Quality Assessment Based On Sparsity

Vision, as one of the most important channels via which human receive the external information, plays a key role for perceiving and understanding the physical world. With the rapid development of multimedia and sensor technologies, images have an important effect on our work and life. In the last decades, people can more conveniently record the multimedia information with the explosive growth of portable electronics. How-ever, as the devices and normal customers are non-professional, the captured images or videos are inevitable to be degraded by multiple distortions, which greatly influence the perception quality. In most time, people always prefer high-quality images. The goal of image restoration is to recover a high-quality image, which is free of distortion as much as possible. During the process of image restoration, one important issue is how to define the image quality. A high-performance image quality assessment (IQA) algorithm is expected to predict the image quality score, which is highly consistent with human perception quality. Although image restoration and image quality assessment are two different problems, they have strong relationship with each other. When an im-age is degraded by some distortions, image restoration is to filter out these distortions. However, IQA algorithm is to assess how the distortions influence the image quality. Therefore, better image quality is the goal of most image restoration algorithms, while a high-performance IQA scheme can provide useful guide information for designing an image restoration algorithm.In this dissertation, we select the sparsity property of images as the entry point and expand our discussion and research on image restoration and image quality assessment. In details, the selected topics include:license plate image deblurring from fast moving vehicles, noise reduction of video with mixture noise, and image blur/sharpness assess-ment. The main contribution and innovation can be summarized as the following three aspects:1. We propose a robust deblurring algorithm for license plate image from fast mov-ing vehicles. According to the mechanism of camera and the vehicles’motion path, the blur kernel that causes blurring can be approximated by linear uniform motion blur kernel. Therefore, the task of blur kernel estimation can be reduced to the estimation of two parameters:angle and length. Through sparse dictio-nary learning, the priori knowledge about sharp license plate images is implicitly expressed by the sparse dictionary. By analyzing the sparse representation coef-ficients of the recovered image, we determine the angle of the kernel based on the observation that the recovered image has the sparsest representation when the kernel angle corresponds to the genuine motion angle. Then, we estimate the length of the motion kernel with Radon transform in Fourier domain. Based on the above observation, we can estimate the blur kernel parameters robustly, which provides solid foundation for further license plate image recognition.2. We propose a new non-local video denoising scheme using low-rank representa-tion and total variation regularization. Through analyzing the distinct properties of video data and noise, we find that the video data exhibits strong self-similarity. In addition, video data is highly structured, of which the gradients follow some typical statistical law. In order to fully make use these two characters, different constraints are applied on video and noise. Then, the video data and noise can be separated by solving an optimization problem.3. We propose a no-reference image blur/sharpness assessment algorithm based on sparse representation. Image structure information plays a key role in the image perception quality. Therefore, it is very important to describe the image struc-ture information. In this dissertation, we observe that the atoms of sparse learned dictionary have clear structural meaning. Therefore, the sparse representation coefficients can convey substantial structural information about images. This ob-servation is the fundamental motivation of our proposed approach. In order to extract information across scales, we introduce the spatial pyramid framework into the pooling stage. Based on the known property of HVS, we obtain the fi-nal image sharpness assessment feature by max-pooling. Finally, a per-trained support vector regression is used to predict the image sharpness score.In a nutshell, we fully utilize the sparsity property of image and develop several novel and robust algorithms for image restoration and image quality assessment. Ex-perimental results demonstrate the superiority of our proposed approaches in term of effectiveness and robustness.