Research On Signal Reconstruction Algorithm Based On Compressed Sensing

In recent years, compressed sensing theory received extensive attention of the researchers at home and abroad, the compressed sensing theory breakthrough the limitation of the Nyquist sampling rate, which makes it possible to compress the compressible signal when it is sampling. This technique breaks through the conventional Nyqusit sampling theory, thus has promising applications in signal capturing, compressing, storage, disposal and transmission.As compressed sensing reconstruction algorithm is the core issue of compressed sensing theory, this paper mainly studies the compression perception reconstruction algorithm. Meanwhile, with the application of compressed sensing theory as the background, we mainly studies speech signal processing technology based on compression perception reconstruction algorithm. The combination of this novel technology and speech signal processing technology of the urgent need for a breakthrough brings hope to improve the performance of speech signal processing system, which is of great theoretical significance and practical application value.This paper is mainly on the following several aspects:Firstly, we studied the compression perception theory and its application in the field of pattern recognition, and analyzed the relation and distinction between sparse representation theory and compression perception. Main compression perception reconstruction algorithm are introduced, and the advantages and disadvantages of the algorithm and the improved direction are analyzed.Secondly, we proposed the improved subspace tracking algorithm based on the wavelet tree model. Analyze the drawback of the traditional SP algorithms, which only use the simple prior knowledge that the signal becoming sparse or compressible with a certain dictionary. Then put the signal wavelet tree model into SP algorithm was put forward to improve the refactoring efficiency of the algorithm, and improved subspace tracking algorithm based on the wavelet tree model was proposed. Through simulation experiments, it proves that with the prior knowledge, the signal wavelet tree model, the SP algorithm based on wavelet tree model proposed in this paper has higher efficiency of refactoring when compared with the traditional SP algorithm.Thirdly, we put forward voice signal compression and reconstruction method based on the adaptive subgradient projection. Because the adoption of the new iteration process and simple projection expression, and the adaptive adjustment coefficient of expansion in the iteration process, make the proposed algorithm full use of the advantages of convergence speed and steady state on different stages, the compressed sensing signal will be effectively reconstruct. Compared to the same category algorithm, the computational complexity of this algorithm also has a certain advantage.Then, a kind of speaker recognition technology based on GMM mean vector sparse decomposition was put forward. Apply compressed perception theory into speaker recognition, and make the GMM super vector as sparse matrix of compression perception. The experiments has proved that this kind of compressed sensing is feasible to be applied into speaker recognition, even under the circumstances of less speaker corpora, not under the general background to the case, it also can get good recognition effect.Finally, the paper makes a summary on the overallwork and achievements, and prospect the future work.