计算机工程 ›› 2018, Vol. 44 ›› Issue (9): 212-217.doi: 10.19678/j.issn.1000-3428.0047934

• 图形图像处理 • 上一篇    下一篇

基于稀疏表示与引导滤波的图像超分辨率重建

张万绪,史剑雄,陈晓璇,汪霖,赵明,周延,牛进平   

  1. 西北大学 信息科学与技术学院,西安 710127
  • 收稿日期:2017-07-12 出版日期:2018-09-15 发布日期:2018-09-15
  • 作者简介:张万绪(1965—),男,副教授,主研方向为计算机视觉、数字图像处理;史剑雄,硕士研究生;陈晓璇,讲师;汪霖,讲师、博士;赵明,硕士研究生;周延、牛进平,讲师、博士。
  • 基金项目:

    国家自然科学基金(61503300,61501372);中国博士后科学基金(2017M613186);西北大学科学研究基金(15NW30,14NW25);陕西省教育厅科学研究项目(17JK0766)。

Image Super-resolution Reconstruction Based on Sparse Representation and Guided Filtering

ZHANG Wanxu,SHI Jianxiong,CHEN Xiaoxuan,WANG Lin,ZHAO Ming,ZHOU Yan,NIU Jinping   

  1. School of Information Science and Technology,Northwest University,Xi’an 710127,China
  • Received:2017-07-12 Online:2018-09-15 Published:2018-09-15

摘要:

现有的图像超分辨率重建方法在图像处理和存储过程中不能有效地恢复更多的图像高频信息。为此,以稀疏表示的重建方法为基础,引入引导滤波,提升图像高频信息的重建效果。基于图像抠图技术,利用具有较好边缘保持特性的引导滤波将待重建的低分辨率图像分解为前 景色、背景色和边缘层。通过联合训练边缘层的高分辨率字典和低分辨率字典,采用稀疏表示的方法重建边缘层,而对于前景色和背景色则采用双三次插值的方法重建,并经过图像合成得到对应的高分辨率图像。实验结果表明,相较于基于稀疏表示的图像超分辨率重建方法, 基于引导滤波的重建图像在主客观评价指标上都有所提高。

关键词: 稀疏表示, 引导滤波, 图像抠图, 联合字典训练, 超分辨率重建

Abstract:

In view of the fact that the image super-resolution reconstruction method cannot effectively reconstruct more high frequency image information in the process of image processing and storage,this paper proposes an image super-resolution reconstruction method based on sparse representation combined with guided filtering.The high frequency image information can be recovered further during image super resolution.At the same time,image matting is introduced in this algorithm,so that the low-resolution image to be reconstructed is decomposed into foreground,background and edge layer by guided filtering.The clarity of the image is mainly determined by the edge layer,so the edge layer of high resolution image can be represented as a sparse linear combination in a dictionary trained from training images,while the foreground color and the background color are reconstructed by bicubic interpolation,then the final high-resolution image is obtained via image synthesis.Experimental results show that the peak signal-to-noise ratio and the structural similarity of the reconstructed image are improved,and the visual effects of the image are also improved,compared with image super-resolution reconstruction via sparse representation.

Key words: sparse representation, guided filtering, image matting, joint dictionary training, super-resolution reconstruction

中图分类号: