计算机应用 ›› 2019, Vol. 39 ›› Issue (9): 2552-2557.DOI: 10.11772/j.issn.1001-9081.2019020373

• 人工智能 • 上一篇    下一篇

盲去模糊的多尺度编解码深度卷积网络

贾瑞明, 邱桢芝, 崔家礼, 王一丁   

  1. 北方工业大学 信息学院, 北京 100144
  • 收稿日期:2019-03-07 修回日期:2019-04-19 发布日期:2019-05-14 出版日期:2019-09-10
  • 通讯作者: 贾瑞明
  • 作者简介:贾瑞明(1978-),男,山东青岛人,助理研究员,博士,主要研究方向:计算机视觉、深度学习、模式识别;邱桢芝(1994-),女,山西长治人,硕士研究生,主要研究方向:计算机视觉、深度学习;崔家礼(1975-),男,山东枣庄人,助理研究员,博士,主要研究方向:图像处理、模式识别;王一丁(1967-),男,辽宁沈阳人,教授,博士,主要研究方向:图像处理、图像分析与识别。
  • 基金资助:

    国家自然科学基金面上项目(61673021)。

Deep multi-scale encoder-decoder convolutional network for blind deblurring

JIA Ruiming, QIU Zhenzhi, CUI Jiali, WANG Yiding   

  1. School of Information Science and Technology, North China University of Technology, Beijing 100144, China
  • Received:2019-03-07 Revised:2019-04-19 Online:2019-05-14 Published:2019-09-10
  • Supported by:

    This work is partially supported by the National Natural Science Foundation of China (61673021).

摘要:

针对拍摄场景中物体运动不一致所带来的非均匀模糊,为提高复杂运动场景中去模糊的效果,提出一种多尺度编解码深度卷积网络。该网络采用"从粗到细"的多尺度级联结构,在模糊核未知条件下,实现盲去模糊;其中,在该网络的编解码模块中,提出一种快速多尺度残差块,使用两个感受野不同的分支增强网络对多尺度特征的适应能力;此外,在编解码之间增加跳跃连接,丰富解码端信息。与2018年国际计算机视觉与模式识别会议(CVPR)上提出的多尺度循环网络相比,峰值信噪比(PSNR)高出0.06 dB;与2017年CVPR上提出的深度多尺度卷积网络相比,峰值信噪比和平均结构相似性(MSSIM)分别提高了1.4%和3.2%。实验结果表明,该网络能快速去除图像模糊,恢复出图像原有的边缘结构和纹理细节。

关键词: 盲去模糊, 多尺度结构, 跳跃连接, 编解码, 卷积神经网络

Abstract:

Aiming at the heterogeneous blur of images caused by inconsistent motion of objects in the shooting scene, a deep multi-scale encoder-decoder convolutional network was proposed to improve the deblurring effect in complex motion scenes. A multi-scale cascade structure named "from coarse to fine" was applied to this network, and blind deblurring was achieved with the blur kernel unknown. In the encoder-decoder module of the network, a fast multi-scale residual block was proposed, which used two branches with different receptive fields to enhance the adaptability of the network to multi-scale features. In addition, skip connections were added between the encoder and the decoder to enrich the information of the decoder. The Peak Signal-to-Noise Ratio (PSNR) value pf this network is 0.06 dB higher than that of the Scale-recurrent Network proposed on CVPR(Conference on Computer Vision and Pattern Recognition)2018; the PSNR and Mean Structural Similarity (MSSIM) values are increased by 1.4% and 3.2% respectively compared to those of the deep multi-scale convolution network proposed on CVPR2017. The experimental results show that the proposed network can deblur the image quickly and restore the edge structure and texture details of the image.

Key words: blind deblurring, multi-scale structure, skip connection, encoder-decoder, Convolutional Neural Network (CNN)

中图分类号: