基于迁移学习的高效单目图像深度估计  

Efficient Monocular Image Depth Estimation Based on Transfer Learning

作　　者：刘佳涛 张亚萍[1] 杨雨薇 Liu Jiatao;Zhang Yaping;Yang Yuwei(School of Information Science and Technology,Yunnan Normal University,Kunming 650500,Yunnan,China;Nantong Institute of Technology,Nantong 226000,Jiangsu,China)

机构地区：[1]云南师范大学信息学院,云南昆明650500 [2]南通理工学院,江苏南通226000

出　　处：《激光与光电子学进展》2022年第16期226-234,共9页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61863037);云南省“万人计划”青年拔尖人才专项;南通市科技局项目(JC2019108)。

摘　　要：在进行三维重建、场景理解等计算机视觉任务时,从二维图像中恢复三维空间中的深度信息是一项基本的任务。当前使用深度学习完成该任务时,精确度较高的方法往往需要巨大的数据量,而这些数据的获取通常复杂且开销大。针对这个问题,提出了一种基于迁移学习的全局自注意力编解码网络。所提网络以单张图像作为输入,在编码时的每一个阶段都具有全局性的感受域,解码后把深度回归任务转化为一种分类任务,在保证模型精确度的前提下大大降低所需的训练数据量。实验结果表明,与当前先进的深度估计网络AdaBins和DPT-Hybrid相比,所提网络在均方根误差上降低了约2.2%和0.3%,在训练数据量上降低了约80%和99.6%。When performing computer vision tasks such as three-dimensional reconstruction and scene understanding, it is a basic task to recover depth information in three-dimensional space from two-dimensional images. When deep learning is currently used to complete this task, methods with higher accuracy often require a huge amount of data, and the acquisition of these data is usually complicated and expensive. In response to this problem, this paper based on transfer learning, and proposes a encoder-decoder network using global self-attention. It takes a single image as input and has a global receptive field at each stage of encoding. After decoding, the depth regression task is transformed into a classification task, greatly reducing the amount of training data required while ensuring the accuracy of the model. The experimental results show that compared with the current state-of-the-art depth estimation networks AdaBins and DPT-Hybrid, the designed model reduces the root mean square error by about 2. 2% and 0. 3%, and reduces the amount of training data by about 80% and99. 6%.

关 键 词：成像系统 迁移学习 单目视觉 深度估计 自注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]