基于密集连接的单目图像深度估计  

Monocular depth estimation based on dense connections

作　　者：王泉德[1] 程凯 WANG Quande;CHENG Kai(School of Electronic Information,Wuhan University,Wuhan 430072,China)

机构地区：[1]武汉大学电子信息学院,湖北武汉430072

出　　处：《华中科技大学学报（自然科学版）》2023年第11期75-82,共8页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61701351)。

摘　　要：为解决当前单目图像深度估计方法中重复简单的上采样操作导致深度图物体边缘模糊、图像质量较差的问题,提出一种密集连接的单目图像深度估计方法.该方法采用端到端的编码器和解码器结构,从单个RGB图像进行深度估计.编码器引入性能优良的卷积神经网络EfficientNet-B5,可以高效地提取图像全局上下文特征.解码器设计为密集连接的上采样特征金字塔结构,将全局上下文特征从低分辨率转移到高分辨率,以获得更高质量的深度图.此外,通过设计一种全分辨率多尺度损失函数进一步提升物体边缘的深度估计精度.在NYU Depth V2室内场景深度数据集和KITTI室外场景深度数据集上的训练和测试结果表明:本方法可以产生高精度的深度估计结果,预测的深度图边缘清晰、轮廓分明,所设计的消融实验充分验证了所提出方法各模块的合理性.To solve the problems of blurred edges of depth map objects and poor image quality caused by the repetition of simple upsampling operations in the current monocular depth estimation methods,a densely connected monocular depth estimation method was proposed.The method used an end-to-end encoder and decoder architecture for depth estimation from a single RGB image.The encoder introduced the high-performance convolutional neural network EfficientNet-B5,which could efficiently extract the global context features of images.The decoder was designed as a densely connected up-sampled feature pyramid structure to transfer global contextual features from low resolution to high resolution for higher quality depth maps.In addition,the depth estimation accuracy of object edges was further improved by designing a full-resolution multi-scale loss function.The training and testing results on the NYU Depth V2 indoor scene depth dataset and the KITTI outdoor scene depth dataset show that the proposed method can produce high-precision depth estimation results,and the predicted depth maps have clear edges and well-defined outlines.The designed ablation experiments could fully validate the reasonableness of the proposed method modules.

关 键 词：计算机视觉 深度学习 单目图像深度估计 编解码器结构 多尺度损失函数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]