作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• 图形图像处理 • 上一篇    下一篇

基于深度学习的深度图超分辨率采样

王晓晖,盛斌,申瑞民   

  1. (上海交通大学 计算机科学与工程系,上海 201100)
  • 收稿日期:2016-11-16 出版日期:2017-11-15 发布日期:2017-11-15
  • 作者简介:王晓晖(1992—),男,硕士,主研方向为计算机视觉、机器学习;盛斌(通信作者),副教授;申瑞民,教授。
  • 基金资助:
    国家自然科学基金(61671290)。

Deep Depth Graph Super Resolution Sampling Based on Depth Learning

WANG Xiaohui,SHENG Bin,SHEN Ruimin   

  1. (Department of Computer Science and Engineering,Shanghai Jiaotong University,Shanghai 201100,China)
  • Received:2016-11-16 Online:2017-11-15 Published:2017-11-15

摘要: 在深度图像采集场景下,为利用场景高分辨色彩图进行超分辨率上采样,提出一种采用卷积神经网络自适应学习局部滤波器核的算法,通过同时应用稠密/高分辨率颜色信息和稀疏/低分辨率深度信息全面提取场景信息。在Middlebury和ToFMark数据集上的实验结果表明,与传统深度超分辨率算法相比,提出的算法能够取得较好的超分辨率结果,尤其在颜色和深度的边缘、纹理不匹配区域,具有更好的鲁棒性。

关键词: 深度超分辨率, 上采样, 滤波, 深度学习, 卷积神经网络, 立体视觉

Abstract: In the scene of depth image acquisition,in order to use the high-resolution color map of the scene for superresolution upper sampling,this paper proposes an adaptive learning algorithm for local filter kernels using convolutional neural network.It utilizes both the dense/high-resolution color and the sparse/low-resolution depth information to extract the scene informationentirely.Experimental results on the Middlebury and ToFMark datasets show that,compared with traditional depth superresolution algorithms,the proposed algorithm is capable of obtaining the best super-resolution results.Especially in the color and depth edge as well as the texture mismatch region,it has better robustness.

Key words: Depth Super Resolution(DSR), upsampling, filtering, depth learning, Convolutional Neural Network(CNN), stereo vision

中图分类号: