基于群稀疏系数估计的图像重构算法

基于稀疏表示的图像先验信息模型被广泛用于实现图像的重构中。针对稀疏表示中字典选择与系数估计的关键问题,提出了基于稀疏表示与非局部自相似性相结合的图像重构方法。首先通过欧氏距离的块匹配寻找相似图像块,并利用左右字典分别对相似图像块集合进行局部稀疏与非局部稀疏表示,以获得更稀疏准确的稀疏表示系数。进一步针对传统阈值收缩法对稀疏系数估计精度不足的问题,利用伯格曼迭代算法快速有效地求解重构模型,并采用线性最小均方误差估计准则(LMMSE)实现稀疏系数的估计,以保证对包含图像纹理细节信息的小系数的精确估计。实验结果表明,本文方法不仅在PSNR等客观指标上达到了目前先进水平,而且重构后图像拥有更为丰富的细节信息,整体视觉效果更加清晰。

Image reconstruction algorithm based on group sparse coefficient estimation

Sparse representation based image prior information model has been widely used in image reconstruction. Aiming at the key problems of dictionary selection and coefficient estimation in sparse representation, this paper proposes the image reconstruction method based on sparse representation combined with nonlocal self similarity. Firstly, the patch matching based on Euclidean distance is used to search the similar image patches; then, the local and nonlocal sparse representation of the similar image patch set is performed using left and right dictionaries respectively, so that the sparser and more accurate sparse representation coefficients are obtained. Next, aiming at the problem of the insufficient sparse coefficient estimation accuracy of the traditional threshold shrinkage method, this paper adopts Bregman iteration algorithm to solve the reconstruction model fast and efficiently; and the Linear Minimum Mean square Error (LMMSE) estimation criterion is adopted to achieve the sparse coefficient estimation, which can ensure the accurate estimation of the small coefficients containing the information of the image texture details. The experiment results demonstrate that the proposed method not only achieves the state of the art performance in the objective specifications such as peak signal to noise ratio (PSNR) and etc., but also makes the reconstructed image have richer detail information and the overall visual effect clearer.