Font Size: a A A

Block Adaptive Compressed Sensing With Coevolution Optimization And Image Prior

Posted on:2015-04-13Degree:DoctorType:Dissertation
Country:ChinaCandidate:R F WangFull Text:PDF
GTID:1108330464968883Subject:Circuits and Systems
Abstract/Summary:PDF Full Text Request
When applying to 2D images, the compressed sensing(CS) theory faces several challenges, including the high computing complexity of reconstruction algorithm and excessive memory consumptions in the implementation of the sensing matrixes. The block compressed sensing(BCS) technique can well solve these problems with the advantages such as low storage and complexity, good real-time performance and easy hardware implementation.However, the existing BCS methods utilizing convex and greedy algorithms cannot exploit image prior in CS measurement and image reconstruction. In this thesis, a novel block adaptive CS algorithm exploiting the natural computation techniques and image priors is proposed. This thesis mainly focuses on the extraction of the prior knowledge of natural images and its utilization for developing effective BCS algorithms. The main contributions of this thesis are summarized as follows:(1) A universal self-adaptive dynamic control strategy of population size is proposed. This strategy can be easily combined with various evolution computation methods because its implementation is independent of the evolutionary operation details. Based on the strategy, a method which can vary the number of increase/decrease on the basis of the logistic model is proposed. We also designed an increase operator giving consideration to the effectiveness and diversity adaptively, as well as a decrease operator with the diversity. The strategy is applied to two different nature computation methods. Experimental evaluation is conducted on both a set of standard test functions and a new set of benchmark functions CEC05. The results show that the new algorithms with proposed strategy outperform the original algorithms on both the precision and convergence rate.(2) To increase the effectiveness of the traditional evolutionary algorithm for CS image reconstruction, an adaptvie coevolution multiscale compressed sensing algorithm(ACE-MCS) is proposed for nonconvex CS, which is developed by introducing cooperative coevolution to the above Sa DCPS evolution computation methods. The CS fitness function is defined, and some strategies are designed, including random creatinginitial population and dynamic grouping based on the reisdual correlation, as well as global debiasing. Experimental results on simulated sparse image demonstrated the effectiveness of the ACE-MCS algorithm, which solves the nonconvex CS reconstruction problem using the techniques of nature computation.(3) For our ACE-MCS performance decreases sharply with the increase of image complexity, an edge-based adaptive measurement method(EAM) is proposed. In addition, based on the EAM, an improved algorithm called ECEA-MCS(Edge-based Coevolution Adaptive Multiscale CS) is proposed. In EAM algorithm, to achieve the adaptive CS measurement in wavelet domain, the energy differences between the coefficients blocks of the same and different scales were firstly analysised and then the edge information extracted from the low-frequence coefficients was used to compute the weights for different scales. The adaptability of ECEA-MCS lie in adaptive measurement with EAM, and edge-based adaptive reconstrution. The medical image and nature image experimental results show that EAM can significantly improve existing CS algorithms on both the reconstructed image quality and the visual effect. Meanwhile, ECEA-MCS can provide the good reconstruction accuracy and remain less computing time comparing with ACE-MCS.(4) Conventional block based CS method that uses non-adaptive measurement and reconstruction algorithm fail to exploit the prior knowledge of natural images, leading to poor reconstruction performance. To attack this problem, two algorithms of blocked adaptive CS based on texture information and blocked adaptive CS based on visual saliency are proposed. In this two methods, the number of measurements were adaptively assigned to the image blocks according to the complexity of the blocks. Thus, the reconstruction qaulity can be improved without the increase of the total number of measurements. Furthermore, to better preserve the image details and supppress noise, the filtering threshold is adaptively computed based on the prior knowledge of natural images. Experimental results on medical images, natural images and the SAR images show that both the BACS algorithm and the VS-BACS algorithm outperform conventional wavelet-based CS algorithms in terms of visual quality and objective metrics(e.g., PSNR).(5) The rich repetitive patterns of natural images have been proved to be very effectivefor image reconstruction. To exploit the nonlocal self-similarity of natural images, the nonlocal block adaptive CS algorithm(NBACS) and the nonlocal block adaptive multi-scale CS algorithm(NBA-MCS) are proposed. In NBACS algorithm, the visual saliency detection is combined with the nonlocal reconstruction, while in NBA-MCS algorithm the adaptive measurement of the edge information is integrated with the nonlocal based wavelet CS algorithm. The proposed CS algorithms can conveniently integrate the existing nonlocal image denoiser into the CS image reconstruction algorithms. Experimental results on medical images, natural images, SAR images and noisy natural images demonstrate that the effectiveness of the proposed algorithms with comparison to other CS algorithms proposed in this thesis.
Keywords/Search Tags:block compressed sensing, coevolution, wavelet transform, image prior, adaptive measurement, adaptive reconstruction
PDF Full Text Request
Related items