结合支持向量回归和图像自相似的单幅图像超分辨率算法

目的 基于学习的单幅图像超分辨率算法是借助实例训练库由一幅低分辨率图像产生高分辨率图像。提出一种基于图像块自相似性和对非线性映射拟合较好的支持向量回归模型的单幅超分辨率方法，该方法不需使用外部图像训练库。方法 首先根据输入的低分辨率图像建立图像金字塔及包含低/高分辨率图像块对的集合；然后在低/高分辨率图像块对的集合中寻找与输入低分辨率图像块的相似块，利用支持向量回归模型学习这些低分辨率相似块和其对应的高分辨率图像块的中心像素之间的映射关系，进而得到未知高分辨率图像块的中心像素。结果 为了验证本文设计算法的有效性，选取结构和纹理不同的7幅彩色高分辨率图像，对其进行高斯模糊的2倍下采样后所得的低分辨率图像进行超分辨率重构，与双三次插值、基于稀疏表示及基于支持向量回归这3个超分辨率方法重建的高分辨率图像进行比较，峰值信噪比平均依次提升了2.37 dB、0.70 dB和0.57 dB。结论 实验结果表明，本文设计的算法能够很好地实现图像的超分辨率重构，特别是对纹理结构相似度高的图像具有更好的重构效果。

Single image super-resolution via support vector regression and image self-similarity

Objective The learning-based methods for single image super-resolution employ an instance training database to produce a high-resolution (HR) image from a single low-resolution (LR) input. In this study, we propose a new super-resolution framework without external training database. The proposed method is based on the self-similarity of images, which is a recurrence of image patches within an image or across image scales, and the support vector regression (SVR) model derives good fitting data via nonlinear mapping. Method First, the image pyramid of the input LR images is established, and the set of LR/HR image patch pairs is set. Second, we search similar image patches of input LR image patch in the set of LR/HR image patch pairs. Then, we use SVR to learn the map relationship between these similarity LR image patches and the pixel value of the center of the corresponding HR images. Finally, we can obtain the HR image patch through the aforementioned relationship and input LR image patch. Result We tested the proposed method on seven HR images with different textures and structures, which are downsampled by Gaussian blurring under a scalar factor of 2. The average PSNRs of our method are 2.37, 0.70, and 0.57 dB higher than the bicubic interpolation, the sparse representation-based super-resolution method, and the support vector regression-based super-resolution method, respectively. Conclusion Experimental results show that the proposed method can effectively achieve image super-resolution reconstruction, particularly for the image with a highly similar texture.