Variable Sampling Methods Of Compressed Sensing Exploiting Image Features

Compressed Sensing(CS) was developed from the theory of sparse representation. It's a theory of accurate reconstruction after reduce dimensionality of high-dimension signals which can be represented sparsely. CS breaks through the bottleneck of the speed of sensing and bandwidth, and combines the sampling and compression into one step, which can reduce the waste of hardware resources and acquisition time. That makes the CS widely applied in data acquisition system of many fields. As the sparse representation of signals keeps image features well, the CS framework has a significance advantage in the applications such as matrix completion, data sepration, super-resolution and Pan-sharpening.Data acquisition is one of the most important steps in CS, and plays a critical role in compression and reconstruction. The distribution of projection coefficients of images in orthogonal basis shows a property of “cluster”. But in the step of acquisition in traditional CS framework, the fixed measurement matrix is employed, without considering the influence of the distribution of sparse coefficients. That makes the results of reconstruction are not quite well. We analyze the drawbacks brought by fixed measurement matrix, and proposed two algorithms to allot the sampling rate by combining features in spacial and sparse domains.1. An improved CS algorithm based on the Discrete Cosine Transform(DCT) fan-shaped segmentation is proposed. In order to improve the quality of the reconstructed image and reduce the time-consuming, the low-pass coefficients in the form of fan-shaped segmentation is preserved, then the middle/high-pass coefficients in different sampling rate are measured. Finally, the orthogonal matching pursuit(OMP) algorithm was used to recover the middle/high-pass DCT coefficients. So, the image could be reconstructed by the inverse DCT transform. The experiment results showed that when the compressing ratio is lower, the PSNR was improved distinctly compared with the single layer wavelet transform and so on.2. An adaptive sampling algorithm of block-divided compressed sensing for images based on textural feature is proposed. Firstly, the spacial frequency is utilized to extract the textural features of image blocks; Secondly, each block is categorized into the smooth blocks or the textual blocks based on the textual features, and the basic sampling rate is obtained simultaneously; Thirdly, adjusting the subrate of subbands using the statistical characteristics of the coefficients in wavelet domain based on basic sample rate for the textural blocks. Finally, the smooth projected Landweber is employed to reconstruct images. The experiment results show that when the compressing ratio is modest, the quality of reconstructed images can be improved greatly by the proposed algorithm comparing with other block-based compressed sensing algorithms from the aspects of objective indicator and visual effect.