基于压缩感知的视频重构方法研究

发布时间：2018-07-30 06:46

【摘要】：压缩感知理论提供了将模拟信号直接采样压缩为数字形式的有效途径,具有直接信息采样的特性。在该理论框架下,信号的采样和压缩同时以远低于奈奎斯特采样率的极低速率进行,显著地降低了数据采集、存储和传输代价,以及信号处理时间和计算成本,具有重要的军用和民用价值。基于压缩感知的视频采集与重构,所需测量值的数目远小于传统采样方法获得的数据量,因而可以大幅降低信号的采样和存储成本,从而降低对采集器件的要求和实现难度;同时可以在视频信号采集的同时实现压缩,大大降低了编码器复杂度,减少了对内存和运算资源的需求,使得在资源受限环境中可以实现低成本的(超)大分辨率视频采集和压缩。但是,直接将压缩感知理论应用于视频信号,往往会因为传统的正交变换系数的稀疏性很难达到压缩感知重构的要求而导致较差的重构质量;并且由于传统的压缩感知重构算法仅考虑了信号的稀疏性特征,并未考虑到信号自身的其他结构特征,而使其很难达到最优的视频重构质量。由于视频信号区别于一般信号最大的特点是存在大量空间/时间冗余,因此如何利用相关性是研究视频压缩感知理论的主要难题,目前在国内外仍处于研究的起步阶段。本文以压缩感知理论为基础,在保证视频采样终端低复杂度的前提下,以提高视频压缩感知重构质量为目的,围绕着视频信号的高效稀疏重构展开研究,旨在传统的压缩感知框架下通过引入视频信号空时稀疏的结构特征,将传统的非自适应结构变为自适应的视频压缩感知框架,以减少编码测量所需的采样率并提高视频重构质量。本文以国家自然科学基金、高等学校博士学科点专项科研基金等项目为研究平台,对视频压缩感知重构的关键技术进行研究。全文内容主要针对基于支撑集的压缩感知理论、视频压缩感知编码端速率控制、解码端高效重构算法以及分布式框架下视频压缩感知重构等四个方面展开研究,具体概括为:1)针对一般信号,研究有支撑集辅助的压缩感知理论及其在视频压缩感知中的应用。由于传统压缩感知理论的约束等距性很难在实际应用中被验证,故本文主要在相关性判别理论框架下研究有支撑集的压缩感知重构问题。本文在理论上证明了如果预测支撑集能够在精度与尺寸上满足一定条件,那么利用加权1l范数优化即可得到稳定的稀疏解;并且相比于没有支撑集的相关性判别条件,本文证明了利用支撑集可以得到更弱的充分条件以及更优的重构误差限。2)在视频压缩感知框架下研究速率控制算法,以实现自适应的视频采样,从而能够在不增加采样率的前提下进一步提高视频整体的重构质量。具体来说,由于采样终端无法得到视频信号像素域的结构特征,因而使得在视频压缩感知框架下研究速率控制备受挑战。本文首先提出了一种新颖的视频压缩感知失真模型;然后利用该模型设计了采样率与量化比特深度的联合优化算法,通过求解率失真优化问题实现率失真意义下最优的采样率与比特深度估计,进而能够在实现目标码率的同时得到最优的视频压缩感知重构质量。仿真实验结果表明,利用本文提出的速率控制算法可以较好地控制视频压缩测量的码率,并且相比于传统视频压缩感知系统可大幅提高重构的率失真性能。3)利用视频信号空时稀疏的结构特征,研究高效的视频压缩感知重构算法。具体来说,本文提出了一种正则化的加权基追踪去噪重构方法,通过预测视频信号的支撑集和像素值来辅助当前帧重构,并且基于交替方向乘子法构造了一种快速的迭代算法以实现该问题的求解。此外,本文通过分别构造视频信号在像素域与测量域的帧间相关模型,提出了一种基于最优相关模型的视频压缩感知重构方法,并构造了一种基于二阶Bregman分裂的迭代算法来实现该优化问题的求解。仿真结果表明,本文算法能够通过充分利用视频信号的结构特征实现高效重构,并且相比与传统方法能够提供更好的采样率-失真性能与主观图像质量。4)本文最后重点针对分布式视频压缩感知,研究讨论视频重构相关的问题。具体来说,本文首先在分布式视频压缩感知框架下研究当前帧与边信息帧的相关性,并构造了一种新颖的欠采样相关噪声模型。然后,在此基础上,本文提出了一种基于最大似然字典训练的分布式视频压缩感知系统,和一种基于字典学习和1l分析重构二者联合优化的系统,以及该框架下基于交替方向乘子法的迭代重构算法。仿真实验结果表明,本文算法相比于传统分布式视频压缩感知方法,均能够提供更好的重构质量。
[Abstract]:The compressed sensing theory provides an effective way to compress the direct sampling of analog signals into digital forms, with the characteristics of direct information sampling. In this theoretical framework, the sampling and compression of signals are far lower than the very low rate of Nyquist sampling rate, which significantly reduces the cost of data acquisition, storage and transmission, and the signal. Processing time and computational cost are of great military and civil value. Video acquisition and reconstruction based on compressed sensing are far less than the amount of data obtained from traditional sampling methods. Thus, the sampling and storage costs of the signals can be reduced greatly, thus the requirements and difficulties of the acquisition devices can be reduced; at the same time, it can be reduced. The simultaneous compression of video signal acquisition reduces the complexity of the encoder, reduces the demand for memory and computing resources, and makes low cost (super) high resolution video acquisition and compression in the resource constrained environment. However, the application of compressed sensing theory to video signals is often due to the traditional orthogonal design. The sparsity of the transform coefficients is difficult to achieve the requirements of the compressed sensing reconstruction and lead to the poor reconstruction quality. And because the traditional compression sensing reconstruction algorithm only takes into account the sparsity of the signal, it does not take into account the other structural features of the signal itself, and makes it difficult to achieve the best quality of the video reconfiguration. The largest characteristic of the general signal is the existence of a lot of space / time redundancy, so how to use the correlation is the main problem to study the video compression perception theory. At present, it is still at the beginning of the research at home and abroad. Based on the compression perception theory, this paper improves the video pressure on the premise of guaranteeing the low complexity of the video sampling terminal. In order to reduce the quality of perceptual reconstruction, this paper focuses on the efficient and sparse reconstruction of video signals. In the traditional compressed sensing framework, the traditional non adaptive structure is transformed into an adaptive video compression perception framework by introducing the spatial time sparsity of video signal, so as to reduce the sampling rate and improve the view of the coding measurement. In this paper, the key technology of video compression perception reconstruction is studied on the National Natural Science Foundation and the special scientific research fund of the doctoral discipline point of the University. The full text is mainly based on the compression perception theory based on the support set, the rate control of video compression perceptual coding end, and the efficient reconstruction of the decoder The algorithm and the video compression perception reconstruction under the distributed framework are studied in four aspects, which are summarized as follows: 1) the compression perception theory with support set assisted and its application in video compression perception are studied for the general signal. The constraint ISO distance of the traditional compression theory is difficult to be verified in practical applications. It is necessary to study the problem of compressed sensing reconstruction with support set under the framework of correlation discriminant theory. This paper theoretically proves that if the predictive support set can satisfy a certain condition on the precision and size, then the stable sparse solution can be obtained by using the weighted 1L norm optimization, and compared to the correlation criterion of the unsupported set, this method can be obtained. It is proved that using the support set can obtain more weak sufficient conditions and better reconstruction error limit.2), the rate control algorithm is studied under the video compression perception framework to realize adaptive video sampling, which can further improve the quality of the reconstruction of the whole video without increasing the sampling rate. It is impossible to obtain the structural features of the pixel domain of the video signal, which makes it very challenging to study the rate control in the video compression perception framework. In this paper, a novel video compression perception distortion model is proposed first. Then, a joint optimization algorithm of sampling rate and quantization bit depth is designed by using the model, and the rate distortion is optimized by solving the rate distortion. The optimization problem realizes the optimal sampling rate and bit depth estimation in the sense of rate distortion, and then can achieve the optimal video compression perception reconstruction quality at the same time that the target rate is realized. The simulation experiment results show that the rate control algorithm proposed in this paper can better control the rate of the optical frequency compression measurement, and compared to the traditional view. The frequency compression perception system can greatly improve the rate distortion performance of the reconstructed.3) using the spatial time sparse structure feature of the video signal to study the efficient video compression perception reconstruction algorithm. In this paper, a regularized weighted base tracking denoising method is proposed, which is used to assist the current video signal support set and pixel value. A fast iterative algorithm is constructed based on the alternating direction multiplier method to solve the problem. In addition, a video compression sensing reconstruction method based on the optimal correlation model is proposed by constructing the inter frame correlation model of the video signal in the pixel domain and the measurement domain, and a new method based on the two order is constructed. The simulation results show that the algorithm can make full use of the structural features of video signals to achieve efficient reconstruction, and can provide better sampling rate distortion performance and subjective image quality.4 compared with traditional methods. Finally, this paper focuses on distributed video compression. In this paper, we first study the correlation between the current frame and the edge information frame in the framework of distributed video compression, and construct a novel model of the less sampling correlation noise. Then, on this basis, a distributed maximum likelihood dictionary training is proposed. The video compression perception system, and a system based on dictionary learning and 1L analysis reconstructing the two joint optimization, and the iterative reconstruction algorithm based on the alternating direction multiplier method. The simulation results show that the proposed algorithm can provide better reconstruction quality compared with the traditional distributed video compression perception method.
【学位授予单位】：西安电子科技大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TN919.81

【相似文献】