【作者】 李国安；

【导师】 杨铀；

【作者基本信息】 华中科技大学 ， 信息与通信工程， 2023， 硕士

【摘要】 单目图像深度估计旨在从单张图像中恢复出像素级别的深度信息,其在机器人导航和增强现实等领域有着广阔的应用前景。然而,要让单目图像深度估计真正适用于现实的应用,现有方法仍然存在如下问题:首先,现有基于深度神经网络的方法通过深层网络进行特征提取,但是在深层网络的编码和解码过程中难以保留和利用边缘细节信息,从而产生模糊的边缘。其次,现有基于卷积神经网络或Transformer的方法受限于有限感受野和空间归纳偏置,导致现有方法在远距离区域的深度估计会出现尺度错误和细节丢失的问题。本文针对上述问题展开研究,形成以下成果:(1)针对单目图像深度估计中的边缘模糊问题,提出了一种基于多尺度语义图神经网络的深度估计算法。为了在深层网络的编码和解码过程中保留和利用边缘信息,使用跳跃连接将边缘细节信息丰富的浅层特征连接到解码过程中。为了充分挖掘上述特征中的边缘信息,设计了一个语义图神经网络模块和聚合、更新机制来进行浅层特征的融合。此外,为了解决固定尺寸的卷积核导致的跨边缘插值问题,利用语义分割图来引导图神经网络在同一语义段内自适应地进行聚合和更新。实验表明,与同类算法相比,在客观指标上均取得最佳结果,在RMSE(Root Mean Square Error)上取得了最高24.9%的提升,且所得到的深度图具有更加锐利的边缘。(2)针对远距离区域会出现尺度错误和细节丢失的问题,提出了一种基于双分支编码的深度估计算法。为了解决卷积运算符难以有效建模长程相关性导致远距离区域出现尺度错误的问题,设计了一个能利用全局感受野来建模长程相关性的Transformer分支来进行特征提取,并引入卷积神经网络分支来进行特征提取以解决远距离区域丢失细节的问题。为了解决特征之间存在异构性的问题,提出了一个多尺度交叉注意力解码器,通过逐像素交互的方式提取互补信息,实现了两者的优势互补。实验结果表明,与同类算法相比,在客观指标上均取得最佳结果,RMSE从0.344下降到0.329,在远距离区域的深度估计尺度更准确且更好地保留了细节信息。更多还原

【Abstract】 Monocular depth estimation aims to restore pixel-wise depth information from a single image,demonstrating broad application prospects in domains such as robotic navigation and augmented reality.However,for monocular depth estimation to be truly applicable in practical applications,existing methods still exhibit the following issues: Firstly,current deep neural network-based approaches employ deep networks for feature extraction,but struggle to preserve and utilize edge detail information during the encoding and decoding processes,resulting in blurred edges.Secondly,existing convolutional neural network or Transformer-based techniques are constrained by limited receptive fields and spatial inductive biases,leading to scale inaccuracies and detail loss in depth estimation for distant regions.This thesis investigates the aforementioned problems and presents the following results:(1)To address edge blurring in monocular depth estimation,a depth estimation algorithm based on a multi-scale semantic graph neural network is proposed.To preserve and utilize edge information in the encoding and decoding processes of deep networks,a skip connection is employed to connect shallow features rich in edge details to the decoding process.In order to fully exploit the edge information in these features,a semantic graph neural network module is designed along with aggregation and update mechanisms to fuse shallow features.Furthermore,to overcome the problem of edge interpolation caused by fixed-size convolutional kernels,a semantic segmentation map is utilized to guide the adaptive aggregation and update of the graph neural network within the same semantic segment.Experimental results demonstrate that the proposed method achieves the best performance in objective metrics and obtains a remarkable 24.9% improvement in RMSE(Root Mean Square Error),while producing depth maps with sharper edges.(2)To address scale inaccuracies and detail loss in depth estimation for distant regions,a dual-branch encoding-based depth estimation algorithm is proposed.In order to resolve the issue of ineffective modeling of long-range correlations by convolution operators leading to scale errors in distant regions,a Transformer branch is designed to leverage global receptive fields for modeling long-range correlations in feature extraction.Additionally,a convolutional neural network branch is introduced to address the problem of detail loss in distant regions during feature extraction.To tackle the heterogeneity between features,a multi-scale cross-attention decoder is proposed,which extracts complementary information through pixel-wise interactions,thereby leveraging the strengths of both branches.Experimental results demonstrate that the proposed method achieves the best performance in objective metrics when compared to similar algorithms.The RMSE is reduced from 0.344 to 0.329,resulting in more accurate scale estimation and better preservation of detail information in long-range depth estimation regions.更多还原