压缩感知图像的块子带自适应稀疏表示规则化重构

Image Reconstruction Regularized by Patch Bandwise Adaptive Sparse Representation for Compressive Sensing

针对自然图像信号的非平稳特性和不同图像块的变换域系数的分布差异较大,基于分块图像子带自适应稀疏表示规则化,提出了一种新的压缩感知图像重构方法.先利用非局部相似块组估计每个分块图像变换域各子带系数的均值和标准差,再将图像块各子带系数进行去均值并关于标准差归一化,最后将去均值归一化处理的子带系数的l_1范数表示用于规则化压缩感知重构.由于块子带自适应稀疏表示更加合理地表达了稀疏系数的重要性,使得重构图像能够更好地保留纹理、边缘等细节信息.大量的实验结果表明:相比组稀疏表示的压缩感知重构算法,该方法重构图像的峰值信噪比平均提高了0.69 dB.

In view of the non-stationary characteristics of natural image and the large difference of the transformation coefficients in different image patches,a new image reconstruction method for compressive sensing(CS)was presented in this paper based on patch bandwise adaptive sparse representation(PBASR)regularization.Specifically,both the expectation and the variance parameters of the distribution of image patches subband coefficients were first adaptively estimated based nonlocal similar patch group.Then,the subband coefficients subtracted by the expectation values were normalized by the variance.Lastly,the l 1 norm of the resulting coefficients were used to regularize CS image reconstruction.Since the subband adaptive sparse representation reasonably reflected the importance of sparse coefficients,the reconstructed image well retained the details such as the textures and edges.Extensive experiments demonstrated that compared with compressed sensing reconstruction algorithm with group sparse representation,the peak signal-to-noise ratio of reconstructed image was improved by 0.69 dB.