基于通道注意力的自监督深度估计方法  

Channel Attentive Self-supervised Network for Monocular Depth Estimation

作　　者：吴俊贤 何元烈[1] Wu Jun-xian;He Yuan-lie(School of Computer Science and Technology,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学计算机学院,广东广州510006

出　　处：《广东工业大学学报》2023年第2期22-29,共8页Journal of Guangdong University of Technology

基　　金：国家自然科学基金资助项目(61876043)。

摘　　要：提出了一种基于自监督深度学习和通道注意力的深度估计方法。虽然以往的方法已经能够生成高精度的深度图,但是它们忽略了图像中的通道信息。对通道之间的依赖关系进行显式建模,并根据建模结果重新校准通道权重能有效地提高网络性能,从而提高深度估计的精度。本文从两个方面引入通道注意力机制以增强网络模型的能力:在网络中插入SE(Suqeeze-and-Excitation)模块以提高网络模型获得特征图中通道间关系的能力;设计了一个多尺度融合通道注意力模块,实现融合多尺度像素特征和重新校准通道权重的功能。通过在KITTI数据集上的实验验证,所提方法在精准度、误差和深度图的具体效果上都优于现有的基于自监督深度学习的深度估计方法。A new method is proposed for self-supervised monocular depth estimation.Although recent methods have been able to produce high-precision depth maps,previous work has ignored channel-wise information in the image.To solve this problem,Channel Attention is introduced and improvements made in two aspects of the network structure:(a)a Squeeze-and-Excitation(SE)block is injected into the corresponding networks to capture the channel relation in feature map;(b)a Channel Attention Dense Connection(CADC)block is applied to combine multi-scale features and recalibrate channel-wise feature.Experiments on the KITTI dataset show the effectiveness of the proposed approach,which outperforms quantitatively and qualitatively the state-of-the-art Self-Supervised Depth Estimation methods.

关 键 词：单目深度估计 注意力机制 自监督深度学习 

分 类 号：TP249[自动化与计算机技术—检测技术与自动化装置]