利用双通道卷积神经网络的图像超分辨率算法

发布时间：2019-06-11 09:21

【摘要】：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。
[Abstract]:Objective Image super-resolution algorithm has a wide range of requirements and research in practical application. However, traditional sample-based super-resolution algorithms use simple image gradient features to represent low-resolution image blocks, which are difficult to effectively distinguish different low-resolution image blocks. In order to solve this problem, based on the traditional sample super-resolution algorithm, a two-channel convolution neural network algorithm is proposed to learn the similarity between low-resolution and high-resolution image blocks for image super-resolution. Methods the effective similarity measurement between low resolution and high resolution image blocks was obtained by using deep convolution neural network learning. Then the corresponding high-resolution image block is reconstructed according to the similarity between the input low-resolution image block and the high-resolution image block dictionary primitive. Results the average peak signal-to-noise ratio (PSNR) of Set5 and Set14 datasets was 32. 53 dB and 29. 17 dB, respectively. Conclusion the algorithm solves the super-resolution problem from the point of view of similarity learning between low resolution and high resolution image blocks, which can better maintain the edge information in the resulting image and weaken the ringing phenomenon in the result. The algorithm in this paper can be well applied to the task of super-resolution enhancement of natural scene images.
【作者单位】：中国科学院自动化研究所智能感知与计算研究中心;
【基金】：国家重点基础研究发展计划(973)基金项目(2012CB316302) 国家自然科学基金项目(61322209,61175007)~~
【分类号】：TP391.41;TP183

【相似文献】