基于多尺度特征融合的快速单目图像深度估计

Fast monocular depth estimation with multi-scale feature fusion

文章针对目前采用深度学习估计单目图像深度中存在推理时间长、物体边缘细节不清晰的问题,设计一种基于多尺度特征融合的快速单目图像深度估计网络。将GhostNet运用到单目图像深度估计网络的编码网络中,提高网络的编码速度;采用反卷积和双线性插值设计解码网络,并通过跨层连接将编码网络的特征与解码网络的特征融合增强深度图中物体的边缘细节。在通用数据集NYU Depth V2上训练和测试的结果表明,该文设计的网络模型得到的深度图细节保持较为完整,同时具有较高的推理速度。

Aiming at the long inference time and poor real-time performance and blurry object edge in depth map of most of the current deep learning methods to estimate the depth of monocular images,a fast model based on multi-scale feature fusion is designed.The GhostNet is applied to the encoding network of monocular depth estimation network to improve the encoding speed of the network.In order to improve the edge detail of objects in the depth map,a decoding network composed of deconvolution and bilinear interpolation is designed and the encoding features and decoding features are fused through cross-layer connection.Experiments in the recognized NYU Depth V2 database indicate that the proposed network shows a high inference speed and the depth map is more refined in the meantime.