利用自监督卷积网络估计单图像深度信息

Estimating Depth from Single Image Using Unsupervised Convolutional Network

为了提高利用深度神经网络预测单图像深度信息的精确度,提出了一种采用自监督卷积神经网络进行单图像深度估计的方法.首先,该方法通过在编解码结构中引入残差结构、密集连接结构和跳跃连接等方式改进了单图像深度估计卷积神经网络,改善了网络的学习效率和性能,加快了网络的收敛速度;其次,通过结合灰度相似性、视差平滑和左右视差匹配等损失度量设计了一种更有效的损失函数,有效地降低了图像光照因素影响,遏制了图像深度的不连续性,并能保证左右视差的一致性,从而提高深度估计的鲁棒性;最后,采用立体图像作为训练数据,无需目标深度监督信息,实现了端到端的单幅图像深度估计.在TensorFlow框架下,用KITTI和Cityscapes数据集进行实验,结果表明,与目前的主流方法相比,该方法在预测深度的精确度方面有较大提升,拥有更好的深度预测性能.

To improve the accuracy of monocular depth estimation by deep learning,this paper proposes a method of unsupervised convolutional neural network for depth estimation from one single image.By introducing residual structure,dense connection structure and short-cut connection in the encode-decode network structure,the single image depth estimation convolutional neural network is improved and the learning efficiency and performance of the network are improved,and the convergence speed of the network is accelerated.Secondly,combined with the loss metrics such as gray similarity,disparity smoothing and left and right disparities matching,a more efficient loss function is designed,which effectively reduces the influence of image illumination factors,suppresses the discontinuity of image depth and ensures the consistency of left and right disparities.Thereby,the robustness of the depth estimation is improved.Finally,this method realizes the end-to-end depth estimation from single image,where stereo image sequences are used as the training data and the depth information in the single image scene can be estimated without the target depth supervision information.Experiments and comparisons on the KITTI and Cityscapes datasets with TensorFlow framework prove the effectiveness of the proposed method with higher accuracy and faster convergence.