作者:张涛，(张晓利)，任彦
\n

Authors:ZHANG Tao, (ZHANG Xiao-li), REN Yan
摘要:摘要:针对单目视觉图像深度估计时存在精度低的问题，提出一种Transformer和CNN融合的单目图像深度估计方法。首先，采用ResNet-50作为编码器-解码器网络的主干网络对图像特征进行提取，同时在编码器-解码器网络中采用层级融合的方法，将编码器各层级特征进行融合作为解码器的输入，提升深度估计网络对多尺度特征信息的利用率。其次，采用Transformer网络对解码器的输出特征进行全局分析，Transformer网络中的多头注意力机制从解码器输出的深层特征中估计深度信息，提高深度估计网络对多尺度特征的提取能力进而提高深度图的精准度。在NYU Depth-v2数据集上完成模型有效性验证。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24.3%，在均方根误差指标上降低61.3%。证明其在单目图像深度估计的可行性。
Abstract:Abstract:Aiming at the problem of low accuracy in monocular vision image depth estimation, a monocular image depth estimation method based on Transformer and convolutional neural network is proposed. First, ResNet-50 is used as the backbone network of the encoder-decoder network to extract image features. At the same time, the encoder-decoder network adopts a level fusion method to fuse the features of each level of the encoder as the decoder to input to improve the utilization of multi-scale feature information by the depth estimation network. Secondly, the Transformer network is used to perform global analysis on the output features of the decoder. The multi-head attention mechanism in the Transformer network estimates the depth information from the deep features output by the decoder, which improves the depth estimation network′s ability to extract multi-scale features and thus improves the depth map accuracy. The validation of the model was completed on the NYU Depth-v2 dataset. Experimental results show that compared with multi-scale convolutional neural networks, this method is improved by 19.9% in δ＜1.25 and root mean square error is reduced by 49.9%. The feasibility of proposed method in estimating depth from a monocular image is proved.

PDF全文下载地址:

可免费Download/下载PDF全文