Research On Compressive Sensing Image Reconstruction

Compressive sensing (CS) is an emerging technology for signal processing. It breaks through the limitation of the Nyquist sampling theory and shifts the complexity burden of the encoder to the decoder by obtaining the compressed signal directly. In addition, it has a good performance both in robustness and confidentiality. Being applied in image processing for wireless multimedia sensor network (WMSN), CS can realize an uncomplicated and efficient signal processing mode at the encoder.Compressive sensing can reconstruct a signal accurately from a small number of measurements by utilizing the prior information. However, because of the complexity and variety in image signal structure, most of existing methods cannot accurately describe the structural characteristics of an image in different transform domains, which results in a poor performance. In order to improve the reconstruction performance, multiple regularization constraints model for CS image reconstruction (MRCS) is proposed in the paper. It exploits structural redundancy information of the image effectively-the sparsity of wavelet coefficients, the sparsity of gradient coefficient and the nonlocal similarity. The parameters of the prior regularization terms are adjusted adaptively to balance the influence of each corresponding structure characteristic on the model. The texture and geometry structure can be recovered clearly while sharp edge can be kept well in this way. An alternating iterative method based on variable-splitting technique is applied to solve the model. Moreover, in order to reduce the complexity, the calculations are accelerated by the approximate substitutions. In addition, a two-step solution is proposed. Firstly, by picking parts of constraints of the model, an approximate image can be reconstructed rapidly. Secondly, enhanced reconstruction based on it can be achieved by adding the constraint of nonlocal mean. Experimental results show that compared with the ALO method, MRCS improves PSNR by 2dB averagely at a time cost of 1/10, and it can also achieve a more pleasant visual quality. The results verify that the proposed method can realize a fast high-quality reconstruction at low sampling rate.