为了有效地利用图像的特征作为指导重建的先验知识, 解决常规超分辨率算法对边缘与结构等细节恢复不足的问题, 提出一种改进的超分辨率算法. 对待重建图像进行低秩分解, 得到不同特征的低秩子图像和稀疏子图像;对于低秩子图像, 提出采用基于正则化技术的稀疏表示超分辨率算法进行重建, 先通过在低秩子图像中寻找相似图像块构造非局部相似正则化项, 得到图像的非局部冗余, 以保持边缘信息; 再通过局部线性嵌入方法构造流形学习正则化项, 获得图像的结构先验知识, 以增强结构信息. 对于稀疏子图像则不参与稀疏表示超分辨率重建, 而是采用双三次插值法进行重建. 实验结果表明, 与其他算法相比, 无论在主观视觉效果上, 还是在峰值信噪比和结构相似性指标上, 文中算法都有显著的提高.%To effectively utilize the image features as priori knowledge for guiding reconstruction and solve the problem that conventional super-resolution algorithms suffer from insufficient recovery of details such as edges and structures, an improved super-resolution algorithm is proposed. The image to be reconstructed was decom-posed into a low-rank sub-image and a sparse sub-image with different features by the low-rank matrix recovery. The low-rank sub-image was reconstructed with the proposed super-resolution algorithm based on sparse repre-sentation and regularization technique. Firstly, the non-local similarity regularization term was constructed by us-ing similar image patches found in the low-rank sub-image to obtain the non-local redundancy of the image to preserve the edge information. Then the manifold learning regularization term was constructed by applying the locally linear embedding method to get prior knowledge of the image structure to enhance the structural informa-tion. The sparse sub-image was not involved in super-resolution algorithm based on sparse representation, and in-stead it was reconstructed with bi-cubic interpolation. Experimental results demonstrate that the proposed algo-rithm has significant improvement over other algorithms in terms of subjective visual effects, peak signal-to-noise ratio, and structural similarity measure.
展开▼
