基于学习的单幅彩色图像超分辨率重建

为了解决单幅彩色图像超分辨率重建中以最常用的欧几里得最小距离来选择最相似块进行重建可能带来的块效应等视觉效果不佳的问题以及由于训练样本较少可能出现最近邻块和待复原块之间误匹配问题,提出在重构算法中引入权值融合机制。通过最小化局部重建误差来计算K个最近邻域块的重建权数,可提升输出图像的高频成分,有效地改善了图像的全局效果,并且将重建后的高频细节作为训练样本再次进行重建以得到更高的图像峰值信噪比(PSNR)。实验结果表明,取4个最近邻块且加权系数依次分别为0.70、0.20、0.05、0.05时重建出的高频细节相对最完整和丰富。

Example-based Single Frame Color Image Super-resolution

In order to solve the problem of bad vision such as the block effect that may be brought by using the most common Euclidean minimum distance to select out the most similar patch to reconstruct and the problem of mismatching between the most similar patch and the patch to be recovered when the training samples are not enough in single color image super-resolution reconstruction,the weight fusion mechanism is proposed in the reconstruction algorithm. By minimizing the partial reconstruction error to calculate the weights of the K nearest patches,the high frequency details of output image and the global effects of the image are improved. What’s more,the reconstructed high frequency details are be used as the training samples to reconstruct again in order to yield a higher PSNR. The experiment results show that it can get the richest and most complete high frequency details while the K equals four and the weights respectively are 0. 70,0. 20,0. 05,0. 05.