论文部分内容阅读
目的图像超分辨率算法在实际应用中有着较为广泛的需求和研究。然而传统基于样本的超分辨率算法均使用简单的图像梯度特征表征低分辨率图像块,这些特征难以有效地区分不同的低分辨率图像块。针对此问题,在传统基于样本超分辨率算法的基础上,提出双通道卷积神经网络学习低分辨率与高分辨率图像块相似度进行图像超分辨率的算法。方法首先利用深度卷积神经网络学习得到有效的低分辨率与高分辨率图像块之间相似性度量,然后根据输入低分辨率图像块与高分辨率图像块字典基元的相似度重构出对应的高分辨率图像块。结果本文算法在Set5和Set14数据集上放大3倍情况下分别取得了平均峰值信噪比(PSNR)为32.53 d B与29.17 d B的效果。结论本文算法从低分辨率与高分辨率图像块相似度学习角度解决图像超分辨率问题,可以更好地保持结果图像中的边缘信息,减弱结果中的振铃现象。本文算法可以很好地适用于自然场景图像的超分辨率增强任务。
The purpose of the image super-resolution algorithm in the practical application has a more extensive needs and research. However, the traditional sample-based super-resolution algorithms use simple image gradient features to characterize low-resolution image blocks, which are difficult to effectively distinguish between different low-resolution image blocks. In view of this problem, based on the traditional sample super-resolution algorithm, this paper proposes an algorithm of dual-channel convolutional neural network to learn the image super-resolution by learning the similarity between low-resolution and high-resolution image blocks. Methods Firstly, the similarity measure between effective low-resolution and high-resolution image blocks is obtained by using deep convolutional neural network, and then reconstructed according to the similarity of input primitives of low-resolution image blocks and high-resolution image blocks Corresponding high-resolution image blocks. Results The proposed algorithm achieves the average peak signal-to-noise ratio (PSNR) of 32.53 d B and 29.17 d B, respectively, when magnified three times over the Set5 and Set14 datasets. Conclusion This algorithm solves the problem of super-resolution of images from the perspective of similarity between low-resolution and high-resolution image blocks, and can better preserve the edge information in the result image and weaken the ringing phenomenon in the result. The proposed algorithm is well suited to the super-resolution enhancement tasks of natural scene images.