基于自监督学习的热红外图像景深估计方法  

Depth estimation of thermal infrared images based on selfsupervised learning

作　　者：丁萌[1] 关松[2] 李帅 于快快 徐一鸣[1] DING Meng;GUAN Song;LI Shuai;YU Kuai-Kuai;XU Yi-Ming(College of Civil Aviation,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China;Science and Technology on Electro-Optical Information Security Control Laboratory,Tianjin 300308,China)

机构地区：[1]南京航空航天大学民航学院,江苏南京211106 [2]光电信息控制和安全技术重点实验室,天津300308

出　　处：《红外与毫米波学报》2023年第6期907-916,共10页Journal of Infrared and Millimeter Waves

基　　金：光电信息控制和安全技术重点实验室开放基金(JCKY2022210C005);国家自然科学基金(U2033201);航空科学基金(20220058052001)。

摘　　要：从热红外图像对比度低、细节信息不足等特点出发,提出了一种面向热红外图像的景深估计方法。首先,设计了一种红外特征聚合模块,提高了对目标物边缘和小目标的全方位深度信息获取能力;其次,在特征融合模块中引入了通道注意力机制,进一步融合通道间的交互信息;在此基础上,建立了一种深度估计网络,实现热红外图像的像素级景深估计。消融实验与对比实验的结果表明,该方法在热红外图像像素级景深估计中性能优于其他代表性方法。Depth estimation based on unsupervised learning is one of the important issues in the field of computer vision.However,existing algorithms of depth estimation are mainly designed based on visible images.Compared with visible images,thermal infrared images have the disadvantages of low contrast and insufficient detailed information.To this end,a depth estimation network is constructed and an unsupervised depth estimation method is proposed for thermal infrared images according to their characteristics.The network consists of three parts:feature extraction module,feature aggregation module,and feature fusion module.Firstly,a feature aggregation module is designed to improve network ability to acquire the edge information of target objects and the small object information of the image.Secondly,the channel attention mechanism is introduced in feature fusion module to effectively capture the interaction relationship between different channels.Finally,a depth estimation network for thermal infrared images is established.In this network,the model parameters are trained by thermal infrared sequence images to achieve the pixel-level depth estimation of a single thermal infrared image.The results of ablation studies and comparative experiments fully demonstrate that the performance of the proposed method in pixel-level depth estimation of thermal infrared image outperforms other representative methods.

关 键 词：红外图像 无监督学习 单目深度估计 特征聚合 通道注意力机制 

分 类 号：TP29[自动化与计算机技术—检测技术与自动化装置]