激光与光电子学进展, 2018, 55 (6): 061010, 网络出版: 2018-09-11  

基于帧间信息提取的单幅红外图像深度估计 下载: 1211次

Depth Estimation of Single Infrared Image Based on Interframe Information Extraction
作者单位
1 华东理工大学信息科学与工程学院, 上海 200237
2 东华大学信息科学与技术学院, 上海 201620
摘要
针对红外图像存在纹理信息不丰富和边缘信息较少导致深度估计精度难以提高的问题,本文设计一种深层神经网络估计红外图像的深度,该网络融合了一个二维(2D)残差神经网络和一个三维(3D)卷积神经网络。传统单幅红外图像的深度估计方法遗漏了帧间信息,容易出现物体轮廓模糊甚至丢失的情况。在2D和3D网络输入端分别加入稠密光流和前后帧图像。进一步将3D卷积网络提取的视频特征与2D残差网络的特征图做权值连接。不同于传统神经网络的全连接层,全卷积层突破了输入图片的尺寸限制。实验结果表明,本文提出的红外图像深度估计方法具有较高的精度,估计出的物体轮廓更清晰完整。
Abstract
In view of lacking of the texture information and the edge information in the infrared image, the accuracy of depth estimation is hard to be improved. We propose a deep neural network to estimate the depth of infrared images. The network combines a two-dimensional (2D) residual neural network and a three-dimensional(3D) convolution neural network. The traditional methods of estimating the depth of a single infrared image omits the interframe information and is prone to fuzzy or even missing object contour. The 2D and 3D network inputs are added dense optical flow and the frame before and after the image, respectively. Secondly, the feature map extracted from the 3D convolutional network is further connected to the feature maps of the 2D residual network. Unlike the fully connected layer of the traditional neural network, fully convolutional layer breaks through the size constraints of the input. The experimental results show that the accuracy of the proposed infrared image depth estimation method is improved, and the object contour estimated is clear and complete.

顾婷婷, 赵海涛, 孙韶媛. 基于帧间信息提取的单幅红外图像深度估计[J]. 激光与光电子学进展, 2018, 55(6): 061010. Tingting Gu, Haitao Zhao, Shaoyuan Sun. Depth Estimation of Single Infrared Image Based on Interframe Information Extraction[J]. Laser & Optoelectronics Progress, 2018, 55(6): 061010.

引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!