| 基于张量低秩和TV正则化的图像超分辨率重建 |
| |
| 引用本文： | 刘小花,唐贵进.基于张量低秩和TV正则化的图像超分辨率重建[J].教育技术导刊,2019,18(12):187-191. |
| |
| 作者姓名： | 刘小花 唐贵进 |
| |
| 作者单位： | 1. 南京邮电大学 电子与光学工程学院 微电子学院；2. 南京邮电大学 江苏省图像处理与图像通信重点实验室，江苏 南京 210003 |
| |
| 基金项目： | 江苏省研究生科研创新计划项目（KYCX17_0776）；南京邮电大学科研基金项目（NY218089，NY219076） |
| |
| 摘 要： | 由于低秩先验能够有效地学习图像数据的冗余和数据的全局结构，因此低秩约束在矩阵填充中得到广泛应用。以往的研究表明，低秩约束对张量恢复具有显著影响，这些工作往往通过Tucker秩解决，然而Tucker秩不能捕获张量的内在相关性。提出一种新的基于张量链秩1（Tensor-Train Rank-1，TTR1）分解的逼近张量核范数的邻近算子。低秩约束能够很好地捕获数据的全局结构，但不能利用可视化数据的局部平滑性，因此提出将张量低秩和全变分（total variation，TV）正则化相结合的超分辨率（super-resolution， SR）重建方法，充分利用图像冗余性、全局结构信息和图像局部平滑性，实现图像的SR重建。实验结果表明，相比于Tucker低秩和TV正则化模型（LRTV_SR），该方法在峰值信噪比指标上平均提高了0.2dB，充分验证了基于TTR1分解的张量低秩约束在超分辨率重建中更能保留彩色图像的全局结构特性。
|
| 关 键 词： | 张量低秩 全变分 超分辨率重建 |
| 收稿时间： | 2019-08-06 |
Image Super-resolution Reconstruction Based on Tensor Low-rank and TV Regularization |
| |
| LIU Xiao-hua,TANG Gui-jin.Image Super-resolution Reconstruction Based on Tensor Low-rank and TV Regularization[J].Introduction of Educational Technology,2019,18(12):187-191. |
| |
| Authors: | LIU Xiao-hua TANG Gui-jin |
| |
| Institution: | 1. College of Electronic and Optical Engineering & College of Microelectronics， Nanjing University of Posts and Telecommunications；2. Jiangsu Key Laboratory of Image Processing & Image Communication， Nanjing University of Posts and Telecommunications，Nanjing 210003， China |
| |
| Abstract: | Since low-rank prior can effectively learn the image data redundancies and global structures， low-rank constraint has been widely used in matrix completion. Previous works have shown that low-rank constraint has significant effects on tensor recovery， and these tasks are often solved by Tucker rank. However， Tucker rank does not capture the intrinsic correlation of a tensor. Therefore， we propose a new proximity operator based on tensor train rank-1 （TTR1） decomposition to approximate the tensor nuclear norm. In addition， low-rank constraint can capture the global structure of data well， but it cannot take advantage of the local smoothness of visual data. In order to make full use of image redundancy， global structure information and image local smoothness， we propose a super resolution （SR） method by combining tensor low-rank and total variation （TV） regularizations. Experimental results show that the proposed algorithm achieves an improvement of 0.2dB averagely compared with LRTV_SR. It is fully verified that the proposed low-rank tensor constraint based on TTR1 decomposition can retain the global structure characteristics of color images in super-resolution reconstruction. |
| |
| Keywords: | tensor low-rank total variation super-resolution reconstruction |
|
| 点击此处可从《教育技术导刊》浏览原始摘要信息 |
| 点击此处可从《教育技术导刊》下载免费的PDF全文 |
