计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (9): 197-202.DOI: 10.3778/j.issn.1002-8331.1801-0072

• 图形图像处理 • 上一篇    下一篇

基于卷积神经网络的图像超分辨率重建

刘鹏飞1,2,3,4,赵怀慈1,3,4,刘明第1,3,4   

  1. 1.中国科学院 沈阳自动化研究所,沈阳 110016
    2.中国科学院大学,北京 100049
    3.中国科学院 光电信息处理重点实验室,沈阳 110016
    4.辽宁省图像理解与视觉计算重点实验室,沈阳 110016
  • 出版日期:2019-05-01 发布日期:2019-04-28

Image Super-Resolution Based on Convolutional Neural Network

LIU Pengfei1,2,3,4, ZHAO Huaici1,3,4, LIU Mingdi1,3,4   

  1. 1.Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang 110016, China
    2.University of Chinese Academy of Sciences, Beijing 100049, China
    3.Key Laboratory of Opto-Electronic Information Processing, Chinese Academy of Sciences, Shenyang 110016, China
    4.The Key Lab of Image Understanding and Computer Vision, Liaoning Province, Shenyang 110016, China
  • Online:2019-05-01 Published:2019-04-28

摘要: 单幅图像超分辨率(Super Resolution,SR)重建,是计算机视觉领域的一个经典问题,其目的在于从一个低分辨率图像得到一个高分辨率图像。目前的卷积神经网络重建算法只有三层结构,浅层结构在处理内部结构复杂的数据时,会出现表征能力不足的问题,因此提出了一个基于特征转移的八层卷积神经网络结构来实现图像超分辨率重建。针对不同的测试集,提出的卷积神经网络模型取得了更佳的超分辨率结果,不管是在主观视觉上还是在客观评价指标上均有明显改善,把数据集图像放大3倍时,对于不同算法的对比图像,该算法的峰值信噪比最高,而且在清晰度方面尤其是图像纹理边缘得到了增强。实验结果证明了基于迁移转移的八层卷积神经网络对图像超分辨率重建的有效性,且网络的收敛速度更快,在精细度方面具有更高的优势。

关键词: 图像超分辨率, 深度学习, 卷积神经网络, 特征转移

Abstract: Single Image Super-Rsolution(SISR), which aims at obtaining a high-resolution image from a single low-resolution image, is a classical problem in computer vision. At present the convolution neural network reconstruction algorithm only has three layers. When dealing with real word data and internal structure is complex, it will appear problem of insufficient characterization ability. In order to meet the requirements of higher super-resolution reconstruction precision, an eight-layer structure is proposed by feature transfer. For different test datasets, the model of Transfer Learning Super-Resolution using Convolutional Neural Network(TLSRCNN) achieves good super-resolution results, and the subjective visual effect and objective evaluation indices are both improved obviously. Peak signal-to-noise ratio of TLSRCNN is the highest in contrast to other algorithms. The image resolution and edge sharpness are enhanced. Experimental results demonstrate TLSRCNN effectiveness of image super-resolution reconstruction, and network convergence speed is faster, which can recover image texture details better in comparison with traditional method.

Key words: image super-resolution, deep learning, convolution neural network, feature transfer