空间自适应正则化的图像超分重建算法

Image Super-resolution Reconstruction Algorithm Based on Spatial Adaptive Regularization

为提高稀疏表示系数的精度和图像的分辨率,提出一种基于稀疏表示和正则化技术的超分重建算法.首先引入自回归正则化项,通过样本图像来训练出描述图像局部结构的自回归模型,每个图像块自适应选择一个自回归模型用以调节解空间,实现图像局部的自适应性控制.然后,引入非局部相似正则化项作为自回归正则化项的补充,用于保持图像边缘清晰度.从而,完整构造出一种基于自回归正则化和非局部相似正则化的稀疏编码目标函数.为了进一步恢复图像,实现图像去噪、去模糊,利用总变分正则化实现全局优化.实验结果表明,与L1SR、SISR、ANR、NE+LS、NE+NNLS、NE+LLE和A+(16 atoms)等算法相比,无论在主观视觉效果还是客观评价指标上,提出的算法都取得了更好的超分重建效果.

In order to improve the accuracy of sparse representation coefficients and the resolution of the image, a novel super reconstruction algorithm based on sparse representation and regularization technique is proposed. First, the auto-regressive （ AR ） regularization term is introduced in sparse coding objective function. The AR model which describes the local structure of the image can be trained by using the sample images. And each image patch adaptively selects an AR model to adjust the solution space and realize the image local adaptive control. Then, the non-local （NL） similarity regularization term is introduced as a complement to the AR regularization term, which is used to preserve the edge sharpness of the image. Therefore, the sparse coding objective function is constructed based on the AR regularization and NL similarity regularization. In order to restore the image and improve the performance of image denoising and deblurring further, the total-variation regularization is adopted to realize the global optimization. Experimental results validate that compared with L1SR, SISR, ANR, NE ＋ LS, NE ＋ NNLS, NE ＋ LLE and A ＋ （ 16 atoms）methods, the proposed approach achieves better super-resolution reconstruction effects in both subjective visual effects and objective evaluation criteria.