摘要
针对现有单目图像深度估计方法中存在估计效果较差等问题,为提升单目图像深度估计的有效性,设计一种基于自监督学习的单目图像深度估计算法。首先获取单目图像绝对深度特征、相对深度特征以及位置特征,采用深度差的对数作为损失函数,得到图像点对点的关联特征;然后采用表面法线损失函数处理深度图表面法线,降低图像表面波动,计算图像光度损失函数与两个图像之间的相似度;最后将单目图像输入到自监督学习网络,计算深度幅值,完成基于自监督学习的单目图像深度估计。实验主要分为客观评价和主观评价,在客观评价部分验证了本文估计算法深度预测的准确性达到了91.2%,深度图边缘误差仅为3%,该方法具备较高的准确性;在主观评价中验证了本文算法能够真实预测图形特征,有效提高图像深度估计效果。
In order to improve the effectiveness of monocular image depth estimation,a monocular image depth estimation algorithm based on self supervised learning is designed.The absolute depth feature,relative depth feature and position feature of monocular image are obtained,and the logarithm of depth difference is used as the loss function to obtain the point-to-point correlation feature of image;The surface normal loss function is used to deal with the surface normal of the depth map to reduce the surface fluctuation of the image,and the similarity between the image photometric loss function and the two images is calculated;The monocular image is input into the self supervised learning network,the depth amplitude is calculated,and the monocular image depth estimation based on self supervised learning is completed.The experiment is mainly divided into objective evaluation and subjective evaluation.In the objective evaluation part,the accuracy of depth prediction of the proposed estimation algorithm reaches 91.2%,and the edge error of depth map is only 3%.This method has high accuracy.In subjective evaluation,the proposed algorithm is verified to be able to predict the image features and improve the image depth estimation effect effectively.
作者
白琳
刘林军
李轩昂
吴沙
刘汝庆
BAI Lin;LIU Lin-jun;LI Xuan-ang;WU Sha;LIU Ru-qing(School of Computer,Electronics and Information,Guangxi University,Nanning 530004,China;College of Electronic Engineering,Guangxi Normal University,Guilin 541004,China)
出处
《吉林大学学报（工学版）》
EI
CAS
CSCD
北大核心
2023年第4期1139-1145,共7页
Journal of Jilin University:Engineering and Technology Edition
基金
国家自然科学基金项目(61966003)
广西自然科学基金项目(2020GXNSFAA159171)。
关键词
软件工程
自监督学习
单目图像
深度估计
特征提取
损失函数
software engineering
self-supervised learning
monocular image
depth estimation
feature extraction
loss function
