摘要
目前基于深度学习的深度估计模型参数量大,难以适应移动端设备。针对此问题,提出一种可以部署在移动端的多尺度特征融合轻量级深度估计方法。首先,以MobileNetV2为主干,提取出4个尺度的特征。然后,通过构建编码器到解码器的跳跃连接路径,将4个尺度的特征进行融合,充分利用融合低层的位置信息和高层的语义信息。最后,融合后的特征通过卷积层得出高精度的深度图像。在NYU Depth Dataset V2数据集上进行了训练和测试,结果表明,该模型的参数量在仅有1.6×106的情况下,评估指标δ1高达0.812,在移动端的麒麟980 CPU上推理一幅图像仅需要0.094 s,具有实际应用价值。
The current depth estimation model based on depth learning has a large number of parameters,which is difficult to adapt to mobile devices.To address this issue,a lightweight depth estimation method with multi-scale feature fusion that can be deployed on mobile devices is proposed.Firstly,MobileNetV2 is used as the backbone to extract features of four scales.Then,by constructing skip connection paths from the encoder to the decoder,the features of the four scales are fused,fully utilizing the combined positional information from lower layers and semantic information from higher layers.Finally,the fused features are processed through convolutional layers to produce high-precision depth images.After training and testing on NYU Depth Dataset V2,the experimental results show that the proposed model achieves advanced performance with an evaluation index ofδ1 up to 0.812 while only having 1.6×106 parameters numbers.Additionally,it only takes 0.094 seconds to infer a single image on the Kirin 980 CPU of a mobile device,demonstrating its practical application value.
作者
陈磊
梁正友
孙宇
蔡俊民
CHEN Lei;LIANG Zheng-you;SUN Yu;CAI Jun-min(School of Computer and Electronics Information,Guangxi University,Nanning 530004;Guangxi Key Laboratory of Multimedia Communications and Network Technology,Nanning 530004,China)
出处
《计算机工程与科学》
CSCD
北大核心
2024年第9期1616-1624,共9页
Computer Engineering & Science
基金
国家自然科学基金(62171145)。
关键词
深度学习
深度估计
多尺度特征
轻量级网络
移动端模型
deep learning
depth estimation
multi-scale feature
lightweight network
mobile terminal model
作者简介
陈磊(1996),男,广东湛江人,硕士,研究方向为图像处理和人工智能。E-mail:leibnizchne@foxmail.com,通信地址:530004,广西南宁市广西大学计算机与电子信息学院;梁正友(1968),男,广西崇左人,博士,教授,CCF会员(16803M),研究方向为计算机视觉、点云处理和三维重建。E-mail:zhyliang@gxu.edu.cn;孙宇(1981),女,广西南宁人,博士,讲师,研究方向为智能计算、数据挖掘和深度学习。E-mail:sunyu@gxu.edu.cn;蔡俊民(1998),男,广西玉林人,硕士,研究方向为图像处理、点云处理和点云识别。E-mail:cjm982538@163.com。
