基于编解码卷积神经网络的单张图像深度估计

摘 要：针对传统方法在单目视觉图像深度估计时存在鲁棒性差、精度低等问题，提出一 种基于卷积神经网络(CNN)的单张图像深度估计方法。首先，提出层级融合编码器-解码器网络， 该网络是对端到端的编码器-解码器网络结构的一种改进。编码器端引入层级融合模块，并通过 对多层级特征进行融合，提升网络对多尺度信息的利用率。其次，提出多感受野残差模块，其 作为解码器的主要组成部分，负责从高级语义信息中估计深度信息。同时，多感受野残差模块 可灵活地调整网络感受野大小，提高网络对多尺度特征的提取能力。在 NYUD v2 数据集上完 成网络模型有效性验证。实验结果表明，与多尺度卷积神经网络相比，该方法在精度 δ<1.25 上 提高约 4.4%，在平均相对误差指标上降低约 8.2%。证明其在单张图像深度估计的可行性。

Single Image Depth Estimation Based on Encoder-Decoder Convolution Neural Network

Abstract: Focusing on the poor robustness and lower accuracy in traditional methods of estimating depth in monocular vision, a method based on convolution neural network (CNN) is proposed for predicting depth from a single image. At first, fused-layers encoder-decoder network is presented. This network is an improvement of the end-to-end encoder-decoder network structure. Fused-layers block is added to encoder network, and the network utilization of multi-scale information is improved by this block with fusing multi-layers feature. Then, a multi-receptive field res-block is proposed, which is the main component of the decoder and used for estimating depth from high-level semantic information. Meanwhile, the network capacity of multi-scale feature extraction is enhanced because the size of receptive field is flexible to change in multi-receptive field res-block. The validation of proposed network is conducted on NYUD v2 dataset, and compared with multi-scale convolution neural network, experimental results show that the accuracy of proposed method is improved by about 4.4% in δ<1.25 and average relative error is reduced by about 8.2%. The feasibility of proposed method in estimating depth from a single image is proved.