基于多通道卷积神经网络的单幅图像深度估计  

SINGLE IMAGE DEPTH ESTIMATION BASED ON MULTI-CHANNEL CONVOLUTIONAL NEURAL NETWORK

作　　者：朱丙丽[1] 高晓琴[2] 阮玲英 Zhu Bingli;Gao Xiaoqing;Ruan Lingying(College of Computer Science and Engineering,Chongqing Three Gorges University,Chongqing 404100,China;Department of Information Engineering,Sichuan Vocational College of Industry and Commerce,Dujiangyan 611830,Sichuan,China)

机构地区：[1]重庆三峡学院计算机科学与工程学院,重庆404100 [2]四川工商职业技术学院信息工程系,四川都江堰611830

出　　处：《计算机应用与软件》2020年第6期230-233,241,共5页Computer Applications and Software

基　　金：国家自然科学基金项目(61602072);重庆市基础研究与前沿探索项目(cstc2016jcyjA0063,cstc2018jcyjAX0502);重庆市教委科技项目(KJ1710248);重庆三峡学院重点实验室“先进网络与智能技术”开放项目。

摘　　要：为了实现单幅图像的深度估计,提出基于多通道卷积神经网络的深度学习算法。基于ResNet框架提出一种新的网络框架及深度平衡欧几里得损失函数;对输入图像进行各种比例的裁剪,生成多通道卷积神经网络的输入对象;经过多通道卷积神经网络后,得到多个深度图候选对象,并将这些深度图候选对象映射合并为一个深度映射候选对象;通过傅里叶反变换生成最终估计深度图。实验结果表明,该算法在深度估计比较实验中取得了较好的结果。In order to realize the depth estimation of a single image,this paper proposes a deep learning algorithm based on multi-channel convolutional neural network(MCCNN).A new network framework and depth balanced Euclidean loss function based on the ResNet framework was proposed;the input image was clipped with various clipping ratios to generate the input object of the multi-channel convolutional neural network;through the multi-channel convolutional neural network,multiple depth map candidates were obtained,and these depth map candidates were merged into one depth map candidate;the final estimated depth map was generated by inverse Fourier transform.The experimental results show that our algorithm achieves good results in the comparison experiment of depth estimation.

关 键 词：卷积神经网络 单幅图像 深度估计 傅里叶分析 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]