基于近似消息传递与卡通纹理模型的图像重构

Image reconstruction based on approximate message passing and cartoon-texture model

近似消息传递(approximate message passing,AMP)的高相变性能与低计算复杂度使其非常适用于图像重构等大数据量应用领域。如何充分利用图像的结构化稀疏先验是基于AMP研究图像重构的一个关键问题。将卡通-纹理模型引入AMP图像重构,根据迭代滤波中待处理图像卡通、纹理成分的不同特点,设计基于双树复数小波变换与全变差的层次化AMP滤波算子,进而分析AMP迭代次数对滤波对象结构特征与滤波算子性能的影响,研究AMP的阶段化滤波操作,提出一种基于卡通-纹理模型与分段滤波的AMP图像重构算法。实验表明,该算法能够更好地保留图像轮廓与纹理信息,提高图像的重构质量。

Due to its high phase transition performance and low computational complexity, approximate message passing （AMP） becomes a powerful reconstruction scheme in large-scale applications such as image reconstruction. How to employ the image structured sparsity priors is a key subject for the study of AMP image reconstruction. The cartoon-texture model is introduced into the AMP image reconstruction. According to the different characters of the cartoon and texture component, a hierarchical filtering operator is designed for AMP image reconstruction based on dual-tree complex wavelet transform and total variation. Effects of the iterations on the structural characteristics of the to-be denoised images and the performance of the filtering operators are analyzed. An AMP image reconstruction algorithm based on the cartoon-texture model and staging filtering is proposed. Experimental results show that this algorithm can preserve more contour and texture information, and improve the quality of the reconstructed images.