基于回归函数结合局部自相似的单帧图像超分辨率算法

A Single Frame Super Resolution Algorithm Based on Fusion of Regression Function and Local Self-similarity

摘要:

启发于过完备字典中稀疏线性组合的高分辨率图像的块与其对应的低分辨率局部块能很好地匹配,提出一种回归函数结合局部自相似的单帧图像超分辨率算法;该算法结合了实例图像块的学习和局部自相似图像块的学习,实例图像块的局部回归避免了从低分辨率到高分辨率图像块映射的病态性问题;通过局部自相似实例图像块学习获得非线性映射函数的一阶近似,从而获得低分辨率图像块相对应的高分辨率图像块,克服了实例图像块算法不足的问题；实验采用峰值信噪比(Peak Signal to Noise Ratio,PSNR)和均方误差(Root-mean-square error,RMSE)比较各算法效果;从实验结果数据可以看出,大多数情况下,提出的算法具有最高的峰值信噪比和最低的均方根误差,从实验结果图可以看出,提出的算法的纹理保留的最好,图像自然性最好,且运行时间也少于其他几种较新的算法,表明提出的算法更适合用于解决实际问题。

Abstract:

Inspired by that high resolution patches,which is the sparse linear combining of the over complete dictionary,can well match the low resolution patches. A new super resolution algorithm based on regression function and local self-similarity is proposed. Compared with the general single-frame super resolution algorithm,the advantage of the proposed algorithm combines the learning of example patches and local self-similarity. Local regression of example patches can avoid the ill-posed problem of LR patches to HR patches. Another aspect,the first-order approximation of the nonlinear mapping function obtained by local self-similar example patches can get the pair of the LR patches corresponding to the HR patches. That overcomes the shortcomings of examples patch algorithms. In this paper,Peak Signal to Noise Ratio (PSNR) and Root-mean-square error (RMSE) is used to compare the results of the algorithms. The data results show that the proposed algorithm has the highest PSNR value and the lowest MSE value in most cases. The experimental graph results show that the image’s texture can be reserved best,the nature of the images is very good. And the execution time of proposed algorithm is less than several other algorithms,which indicates the proposed algorithm is more suitable for applying to settle practical issue.