Research On Compressed Sensing Of Image And Video

In the field of signal processing with the basic framework of Shannon-Nyquist sampling theorem, the high-bandwidth image and video require usually a larger Nyquist sampling rate, which increases the costs of sampling and compressing image or video and proposes a tough challenge for the application of image and video when the acquisition time, energy consumption, or computing power is limited, and therefore it is necessary to find the new technology of sampling and coding image or video which can break restriction of Shannon-Nyquist sampling theorem. The theorem of Compressed Sensing(CS), which can also called the Compressive Sampling, states that the signal having sparse or compressive representation can be exactly reconstructed by using incoherent measurements acquired at the sub-Nyquist rate, and it has established the theoretical foundation for low-cost image or video acquisition and coding. However, the CS theorem regards the one-dimensional vector as its research subject, so that some new challenges will be produced when the CS theorem is applied into the two-dimensional image and three-dimensional video, such as the problem of large memory in the measuring process and high computational complexity of reconstruction resulting from the high dimensionality of image and video, the poor quality of reconstructed image or video resulting from the non-stationary statistics property and so on. In this dissertation, the solutions to these challenges of image-and-video CS are investigated systematically, and some novel conclusions and algorithms are achieved.After reviewing the application prospect, theoretical basis and research progress of image-and-video CS in detail, the system design of image CS is firstly analyzed and researched. Then, in basis of the contents on system design, the reconstruction models are constructed for image CS and Distributed Compressed Video Sensing(DCVS) respectively, and the corresponding solving algorithms are also designed. At last, considering the fact that the side information prediction is the key factor to improve the performance of DCVS reconstruction, and therefore side information prediction is further researched independently.The main contributions of this dissertation are summarized as follows:1. The three different block-by-block measurement based system of image CS are proposed. Firstly, the system of block-by-block measurement and global reconstruction is proposed, the decoder of this system reconstructs the whole image one time by introducing the reordering operator, and therefore it can avoid the blocking artifacts resulting from block-by-block reconstruction. Then, based on the system of block-by-block measurement and global reconstruction, the edge feature based block-by-block adaptive measurement is proposed to guarantee the more efficient measurement for further improving the performance of system. Considering that the encoder realized by the compressive imaging device cannot use the original discrete image, a block-by-block adaptive measurement is realized in the measurement domain, and this method can directly use the CS measurements to estimate the sample variance of each image block and then determinate the measurement rate of each block by using the sample variance to reveal the structural complexity of each block.2. The two reconstruction algorithms of image CS are proposed. In order to reducing the computational complexity of reconstruction, the best linear estimate based fast image CS reconstruction algorithm is proposed, and this algorithm replaces lots of nonlinear iterations in traditional CS reconstruction algorithm with linear projection, and therefore it shortens the time of recovering image. In order to improve the quality of reconstructed image, the PCA-based smoothed projected image CS reconstruction algorithm is proposed, and the method uses PCA technology to train the sparse representation matrix adapting to image structures for the hard thresholding, thus the reconstructed image quality is improved.3. The three DCVS reconstruction algorithms are proposed. At first, in basis of the traditional Wyner-Ziv(WZ) video codec system, the WZ-based DCVS reconstruction algorithm is proposed by replacing the channel coding based codec with the CS measurement and reconstruction. To solve the problem of inaccurate estimation to the parameters of virtual channel, the smoothed projected DCVS reconstruction algorithm is proposed. The method removes the virtual channel and directly uses CS measurements to evaluate the quality of side information depending on the Restricted Isometry Property of measurement matrix, and then the measurement rate of each image block is adaptively assigned by using the quality of side information and the edge features of image block. In order to completely eliminate the dependence of decoder on the encoder, the DCVS reconstruction algorithm combined with temporal-spatial characteristics is proposed. The method further deletes the feedback channel based on removing the virtual channel, thus the task of reconstruction is entirely shifted to decoder, and the performance of joint reconstruction is improved depending only on exploiting the temporal-spatial characteristics of video.4. The four side information prediction algorithms are proposed. For the situation of side information extrapolation, the Motion-Aligned Auto Regressive(MAAR) model based side information prediction algorithm is proposed. The method uses the Tikhonov regularization and overlapped interpolation to overcome the over-fitting problem existing in the MAAR model, thus the better quality of side information is obtained. For the situation of side information interpolation, the joint motion compensation based side information prediction algorithm is firstly proposed, and this method improves the fault-tolerance of motion compensation so as to enhance the quality of side information. Then, the hybrid motion estimation based side information prediction algorithm is proposed, and this method fully uses the advantages of unidirectional and bidirectional motion estimation to reduce the edge blurring of interpolated frame and the distortion resulting from occlusion problem. At last, the multi-resolution motion estimation based side information prediction algorithm is proposed, and this method applies the wavelet pyramid based multi-resolution search into the bidirectional motion estimation so as to suppress the quality degradation of interpolated frame resulting from no temporal symmetry and mismatch in the conventional bidirectional motion estimation.Finally, the main results of the dissertation are concluded and some issues for future research are proposed.