| 基于深度反卷积神经网络的图像超分辨率算法 |
| |
| 引用本文： | 彭亚丽,张鲁,张钰,刘侍刚,郭敏.基于深度反卷积神经网络的图像超分辨率算法[J].软件学报,2018,29(4):926-934. |
| |
| 作者姓名： | 彭亚丽 张鲁 张钰 刘侍刚 郭敏 |
| |
| 作者单位： | 现代教学技术教育部重点实验室(陕西师范大学), 陕西 西安 710062;陕西省教学信息技术工程实验室(陕西师范大学), 陕西 西安 710119,现代教学技术教育部重点实验室(陕西师范大学), 陕西 西安 710062;陕西省教学信息技术工程实验室(陕西师范大学), 陕西 西安 710119,现代教学技术教育部重点实验室(陕西师范大学), 陕西 西安 710062;陕西省教学信息技术工程实验室(陕西师范大学), 陕西 西安 710119,现代教学技术教育部重点实验室(陕西师范大学), 陕西 西安 710062;陕西省教学信息技术工程实验室(陕西师范大学), 陕西 西安 710119,陕西省教学信息技术工程实验室(陕西师范大学), 陕西 西安 710119;陕西师范大学计算机科学学院, 陕西 西安 710119 |
| |
| 基金项目： | 国家自然科学基金（No.61672333，61402274，61772325）；陕西省工业科技攻关项目（No.2016GY-081）；教育部高等教育司产学合作协同育人项目（No.201701023062）；陕西自然科学基金（No.2017JQ6074）；陕西省农业攻关项目（No.2016NY-176）；陕西省重点科技创新团队计划项目（No.2014KTC-18）；教育部科技发展中心“云数融合科教创新”基金（No.2017A07053）；陕西师范大学学习科学交叉学科培育计划资助；中央高校基本科研业务费专项资金资助（No.2017CSY024，GK201603091，GK201703054） |
| |
| 摘 要： | 图像超分辨率一直是底层视觉领域的研究热点，现有基于卷积神经网络的方法直接利用传统网络模型，未对图像超分辨率属于回归问题这一本质进行优化，其网络学习能力较弱，训练时间较长，重建图像的质量仍有提升空间。针对这些问题，本文提出了基于深度反卷积神经网络的图像超分辨率算法，该算法利用反卷积层对低分辨率图像进行上采样处理，再经深度映射消除由反卷积层造成的噪声和伪影现象，使用残差学习降低网络复杂度，同时避免了因网络过深导致的网络退化问题。在Set5、Set14等测试集中，本文算法的PSNR、SSIM、IFC三项评价指标皆优于FSRCNN，重建图像的视觉效果同样验证了本文算法出色的性能。
|
| 关 键 词： | 卷积神经网络 图像超分辨率 深度映射 上采样 |
| 收稿时间： | 2017/4/29 0:00:00 |
| 修稿时间： | 2017/6/26 0:00:00 |
Deep Deconvolution Neural Network for Image Super-Resolution |
| |
| PENG Ya-Li,ZHANG Lu,ZHANG Yu,LIU Shi-Gang and GUO Min.Deep Deconvolution Neural Network for Image Super-Resolution[J].Journal of Software,2018,29(4):926-934. |
| |
| Authors: | PENG Ya-Li ZHANG Lu ZHANG Yu LIU Shi-Gang and GUO Min |
| |
| Affiliation: | Key Laboratory of Modern Teaching Technology, Ministry of Education(Shaanxi Normal University), Xi''an 710062, China;Engineering Laboratory of Teaching Information Technology of Shaanxi Province(Shaanxi Normal University), Xi''an 710119, China,Key Laboratory of Modern Teaching Technology, Ministry of Education(Shaanxi Normal University), Xi''an 710062, China;Engineering Laboratory of Teaching Information Technology of Shaanxi Province(Shaanxi Normal University), Xi''an 710119, China,Key Laboratory of Modern Teaching Technology, Ministry of Education(Shaanxi Normal University), Xi''an 710062, China;Engineering Laboratory of Teaching Information Technology of Shaanxi Province(Shaanxi Normal University), Xi''an 710119, China,Key Laboratory of Modern Teaching Technology, Ministry of Education(Shaanxi Normal University), Xi''an 710062, China;Engineering Laboratory of Teaching Information Technology of Shaanxi Province(Shaanxi Normal University), Xi''an 710119, China and Engineering Laboratory of Teaching Information Technology of Shaanxi Province(Shaanxi Normal University), Xi''an 710119, China;School of Computer Science, Shaanxi Normal University, Xi''an 710119, China |
| |
| Abstract: | Image super resolution is a research hotpot in the field of low level vision. The existing methods based on convolutional neural network do not optimize the image super resolution as a regression problem. These methods is weak in learning ability and require too much time in training step, and the quality of the reconstruction image still has room for improvement. To solve above mentioned problems, we propose a method based on deep deconvolution neural network, which upsamples low resolution image by deconvolution layer firstly, and then uses deep mapping to eliminate the noise and artifacts caused by deconvolution layer. The residual learning reduces the network complexity and avoids the network degradation caused by the depth network. In Set5, Set14 and other datasets, our method performs better than FSRCNN in PSNR, SSIM, IFC and visual. |
| |
| Keywords: | convolutional neural network image super resolution deep mapping deconvolution |
|
| 点击此处可从《软件学报》浏览原始摘要信息 |
|
点击此处可从《软件学报》下载全文 |
|
