基于稀疏特性的图像恢复和质量评价研究

发布时间：2018-11-05 11:19

【摘要】：视觉最为人类获取外部信息最主要的渠道之一,对人们感知和理解外部世界起到十分重要的作用。随着多媒体技术和传感器技术的飞速发展,图像对人们的生产、生活等产生越来越重要的影响。近十年以智能手机为代表的便携智能设备变得日益普及,人们可以比历史上任何时期更方便地记录视觉信息。然而,由非专业的人士或非专业设备获得的图像不可避免地存在各种各样的失真,从而造成人视觉感知体验的下降,甚至导致图像语义信息的破坏。在绝大多数时间,人们总是倾向于得到一个清晰、锋利、无噪声的高质量图像。图像恢复旨在滤除失真图像中的失真部分,从而达到提升图像质量的目标。在图像恢复的过程中,一个很重要的问题就是如何定义图像的感知质量。图像质量评价算法旨在通过计算机算法来模拟人眼视觉系统(Human Visual System, HVS)对图像质量的感知,实现与人类感知一致的图像质量评价。图像恢复和图像质量评价既相互联系又相互区别。当图像遭受失真污染,如果尝试将该失真滤除,就是图像恢复问题：如果尝试评估该失真给人眼造成的质量感知变化,就是图像质量评价问题。因此,绝大多数图像恢复算法需要以图像质量为风向标,而优秀的图像质量评价算法可以为图像恢复算法提供十分有效的指导信息。本课题以图像自有的稀疏特性为切入点,针对图像恢复和图像质量评价展开深入研究,具体包括：高速运动车牌去模糊,视频混合噪声去噪和图像模糊／锋利程度评价。本论文的主要工作和创新点可以总结为以下几点：1.论文提出了一种针对汽车高速运动导致的车牌模糊的鲁棒去模糊算法。首先,根据相机的成像原理及汽车的运动规律,将造成车牌模糊的卷积核简化为线性卷积核。从而,卷积核的估计问题可以简化为参数估计问题。通过稀疏字典学习,将清晰车牌图像的先验信息融合在稀疏字典中,发现去卷积结果的稀疏表达系数与某些卷积核参数之间存在着拟凸关系,利用此性质可以比较鲁棒地估计出卷积核参数,从而得到较好的车牌去卷积效果,为后面的车牌识别奠定基础。2.本文提出了一种针对视频混合噪声去噪的非局部算法。通过分析视频数据及噪声数据不同的特性,利用了视频数据在当前帧和周围数帧之间很强的自相关性。此外,视频数据有清晰的结构信息,其梯度分布符合一定的统计规律。本文从这两个不同的特性入手,对视频数据和噪声数据施加不同的特性约束,利用优化理论及方法,通过求解优化问题实现视频数据和噪声数据的分离,从而达到去噪的效果。3.本文提出了一种基于稀疏表达的图像模糊／锋利程度评价算法。图像的结构信息对于人的视觉质量感知起着十分重要的作用,因此如何描述图像结构信息是图像质量评价中的一个重要问题。通过稀疏字典学习,得到的稀疏字典项具有清晰的结构信息,这为使用稀疏表达进行图像质量评价奠定了基础。此外,通过构建多层金字塔,克服稀疏表达无法捕获跨尺度信息的缺点,利用最大化池化压缩稀疏表达系数的维度,从而实现图像模糊／锋利程度预测。本文充分利用图像(视频)数据的自有稀疏特性,针对图像恢复和图像质量评价等多个典型问题,设计了更加有效的图像恢复和质量评价算法,并深入了分析图像的稀疏特性,大量的实验结果表明了本文所提算法的有效性。
[Abstract]:Vision is one of the most important sources of external information and plays a very important role in people's perception and understanding of the outside world. With the rapid development of multimedia technology and sensor technology, the image has a more and more important influence on people's production and life. portable smart devices represented by smart phones have become increasingly popular in the past decade, and visual information can be recorded more conveniently than in any period of history. However, images obtained by non-professional or non-professional devices inevitably suffer from a variety of distortions, resulting in a decrease in human visual perception experience, or even the destruction of image semantic information. In most of the time, people tend to get a clear, sharp, noiseless high-quality image. The image restoration is intended to filter out distortion parts in the distorted image, thereby achieving the goal of improving image quality. In the process of image restoration, an important problem is how to define the perceived quality of an image. The image quality evaluation algorithm is designed to simulate human visual system (HVS) perception of image quality by computer algorithm to realize image quality evaluation consistent with human perception. Image restoration and image quality evaluation are both interrelated and different from each other. When the image is subjected to distortion pollution, if the distortion is attempted to be filtered out, it is an image restoration problem: if the quality perception change caused by the distortion to the human eye is attempted, the image quality evaluation problem is solved. Therefore, most of the image restoration algorithms need to be based on the image quality, and the excellent image quality evaluation algorithm can provide very effective guidance information for the image restoration algorithm. Aiming at image restoration and image quality evaluation, this paper focuses on image restoration and image quality evaluation, including: high speed motion license plate deblurring, video mixed noise de-noising and image blur/ sharpness evaluation. The main work and innovation points of this thesis can be summarized as follows: 1. In this paper, a fuzzy algorithm is proposed to blur the license plate blur caused by high-speed motor vehicle movement. First, according to the imaging principle of the camera and the motion law of the automobile, the convolution kernel which causes the license plate blur is simplified into a linear convolution kernel. Therefore, the estimation problem of the convolution kernel can be simplified to the parameter estimation problem. Through the sparse dictionary learning, the prior information of the clear license plate image is fused in the sparse dictionary, the quasi-convex relation exists between the sparse expression coefficient of the deconvolution result and some convolution kernel parameters, the convolution kernel parameter can be estimated by using the property, so that a better license plate deconvolution effect is obtained, and a foundation is laid for the following license plate identification. This paper presents a non-local algorithm to de-noising video mixed noise. By analyzing the different characteristics of the video data and the noise data, the self-correlation between the current frame and the surrounding number frame is very strong. In addition, the video data has clear structural information, and its gradient distribution is consistent with certain statistical rules. Based on these two different characteristics, different characteristic constraints are applied to the video data and the noise data, and the optimization theory and method are used to realize the separation of video data and noise data by solving the optimization problem so as to achieve the de-noising effect. An image fuzzy/ sharpness evaluation algorithm based on sparse expression is proposed in this paper. Image structure information plays an important role in human visual quality perception, so how to describe image structure information is an important problem in image quality evaluation. With sparse dictionary learning, the obtained sparse dictionary items have clear structural information, which lays a foundation for image quality evaluation using sparse expression. In addition, by constructing multi-layer pyramid, overcoming the shortcoming that sparse expression can't capture cross-scale information, utilizing the dimension of maximizing the pool to compress sparse expression coefficient, the image blur/ sharpness prediction is realized. This paper makes full use of the self-sparse characteristics of image (video) data, designs a more effective algorithm for image restoration and quality evaluation for many typical problems, such as image restoration and image quality evaluation, and deeply analyzes the sparse characteristics of image. A large number of experiments show the validity of the proposed algorithm.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】