摘要
为了获取信息完整的深度图以提高预测深度图的质量,解决单目深度估计模型中特征融合的问题,提出一种融合多尺度和不同层特征的双流神经网络模型。该模型采用ResNet-50残差网络结构提取深度特征信息,利用金字塔结构融合不同层次的图像特征,实现低层、中层和高层的特征融合,保证不同层次特征的有效互补,改善多层间特征信息的传递,在一定程度上避免了信息的遗漏和缺失。在KITTI(Karlsruhe Institute of Technology and Toyota Technological Institute)数据集上进行试验,结果表明,该模型的均方根误差为2.3704,对数均方根误差为0.229,平均对数误差为0.118,阈值精度分别为0.686、0.951、0.977,实现了较好的评测结果。
In order to obtain the depth map with complete information,boost the quality of prediction depth map,and solve the problem of feature fusion in monocular depth estimation model,a dual-stream neural network model was proposed integrating multi-scale and multi-layer features.In this model,ResNet-50 residual network structure was used to extract depth feature information,and pyramid structure was used to fuse image features of different levels to realize feature fusion of low,middle and high levels,so as to ensure the effective complementarity of features at different levels,improve the transmission of feature information among multiple levels,and to a certain extent,avoid the omission and lack of information.The experimental results on the KITTI(Karlsruhe Institute of Technology and Toyota Technological Institute)dataset show that the root mean square error of the model is 2.3704,the root mean squared log error is 0.229,the average log10 error is 0.118,and the threshold accuracy is 0.686,0.951 and 0.977,respectively,which achieves sound evaluation results.
作者
叶绿
段婷
朱家懿
Nwobodo Samuel Chuwkuebuka
Annor Arnold Antwi
YE Lü;DUAN Ting;ZHU Jiayi;Nwodobo Samuel Chuwkuebuka;Annor Arnold Antwi(School of Information and Electronic Engineering,Zhejiang Universitity of Science and Technology,Hangzhou 310023,Zhejiang,China;School of Mechanical and Engineering,Zhejiang Universitity of Science and Technology,Hangzhou 310023,Zhejiang,China)
出处
《浙江科技学院学报》
CAS
2020年第4期257-263,共7页
Journal of Zhejiang University of Science and Technology
关键词
特征融合
双流神经网络
金字塔结构
feature fusion
dual-stream neural network
pyramid structure