Research On 1-Bit Compressive Sensing Signal Reconstruction Scheme Based On Bayesian Model

Compressive sensing(CS)is a sampling and reconstruction framework for sparse signals,and can carry out signal acquisition and compression simultaneously,hence breaks down the traditional Nyquist sampling theorem,and is considered as a great breakthrough in the field of signal processing.It has been widely used in information theory,image processing,pattern recognition,wireless communication,biomedical and other fields.With the continuous exploration of compression perception theory,there are various related research topics.1-bit compressive sensing(1-bit CS)is one of the focuses of recent study.Due to its simple hardware structure,high processing and transmission rate,1-bit CS has been extensively studied from its birth.It is the limit case of compressive sensing quantization wherein only the sign information of measurements are preserved,as a result,the reconstruction performance are heavily degraded and is much sensitive to the sign-flip noises.To overcome these issues,a research is made on the reconstruction algorithms of 1-bit CS,higher performance and stronger robustness are achieved through introducing the Bayesian inference into the CS framework.First,this thesis creatively proposes to a new scheme of adaptive thresholds,which makes use of the correlation among measurements,gradually estimates the signal by variational Bayesian inference and updates subsequent quantized measurements to improve the accuracy of reconstruction.Secondly,we use variational Bayesian inference to estimate distribution of noise and introduce weighting factors in signal reconstruction to suppress the influence of noise.The main contribution and findings of this thesis are list as follows:1.An adaptive threshold based 1-bit CS variational Bayesian inference reconstruction scheme is proposed.To overcome the issue of low latent efficiency of information utilization caused by fixed threshold value in the process of quantization,an adaptive scheme is adopted.Instead of constant threshold used in traditional quantization,this scheme tries to discover the hidden correlation among measurements,and adaptively adjusts the quantization thresholds of subsequent measurements according to the reconstructed signal from previous ones,thereby reducing the redundancy.In the reconstruction process,the Bayesian method is employed to model the signal and seek the solution.The experiments show that the scheme does not increase the complexity of the coding end in noise-free scenarios,and has a better signal reconstruction performance than those comparative methods mentioned in this thesis.2.A weighted 1-bit CS based variational Bayesian inference reconstruction scheme is proposed.Aim at the sign-flip problem caused by noises existed in practical transmission and quantization process,some weighting factors are assigned to the measurements in the reconstruction process to restrain those corrupted ones.Specifically,a variational Bayesian algorithm is developed to infer the distribution of noise signals,and the energies of the noises can be obtained.Accordingly,the weighting factor associated with each measurement can be set with reciprocal energy of the noise involved in that measurement,as such,the adverse effect of sign-flips can be suppressed and a better performance is achieved.Repeat the whole process until the signal converges.The experiments show that the proposed scheme can effectively suppress the effect of noise and improve the reconstruction performance dramatically,therefore,it is much suitable for scenarios where sign-flips occur,and displays a robustness in signal transmission.