基于局部结构相似与协同表示的超分辨率图像重建

Super-Resolution Image Reconstruction Based on Local Structural Similarity and Collaborative Representation

提出一种基于局部几何结构相似性和协同表示的超分辨率图像重建算法.该算法利用l2范数正则化的协同表示和局部几何相似约束模型求解低分辨率图像块在低分辨率字典下的线性表示系数,并利用这一系数重构出高分辨率图像块.文中基于l2范数的系数求解模型可得到解析解而不涉及局部最小解,相较于l1稀疏性约束具有较低的复杂度.实验结果表明,该算法对小尺寸超分辨率图像重建可行且有效,并在重构效果上具有明显的优越性.进一步研究表明,在放大因子增大和存在噪声的情况下,该算法较传统算法重构效果也有显著提高.

An approach for super-resolution image reconstruction is presented based on local structural similarity and collaborative representation. The collaborative representation l2-norm regularization and local similarity constraint are employed to seek a linear combination for a patch of low-resolution input image with respect to the low-resolution dictionary. Then,the high-resolution image patch is generated by virtue of the coefficients of this combination and the high-resolution dictionary. In addition,the l2-norm based objective function implies an analytical solution and it does not involve local minima. Hence,it performs at a lower complexity compared to l1-sparsity constraint model. The experimental results demonstrate that the proposed method is feasible and effective for small super-resolution image reconstruction and outperforms the bicubic interpolation method and sparse representation super-resolution model on both visual effect and numerical results. Further research shows that the proposed method also performs well for large magnification factors and noisy data.