计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (20): 138-145.DOI: 10.3778/j.issn.1002-8331.1908-0002

• 模式识别与人工智能 • 上一篇    下一篇

双路融合的深度估计神经网络方法研究

刘春,吴一珩   

  1. 湖北工业大学 计算机学院,武汉 430068
  • 出版日期:2020-10-15 发布日期:2020-10-13

Research on Neural Network Method for Depth Estimation Based on Two-Way Fusion

LIU Chun, WU Yiheng   

  1. School of Computer Science, Hubei University of Technology, Wuhan 430068, China
  • Online:2020-10-15 Published:2020-10-13

摘要:

从单目视觉中恢复深度信息是计算机视觉领域的经典问题,结合传统算法的深度学习方法是近年来的研究热点,但在神经网络的算法融合、参照物标定和应用场景上还有限制。提出了一种双路融合深度估计神经网络结构,分别基于深度与深度梯度的语义信息进行网络训练,对特征融合后再次训练得到最终的细节特征,并通过单次标定的方法解决真实参照物标定工作量大的问题。该网络结构能根据单张RGB图片推测出富有细节的深度信息,网络模型基于KITTI的深度图数据集训练,实验包括KITTI测试集和部分实际场景图集,结果表明该方法在深度信息细节的重建上优于对比深度估计方案,在大视场场景下的鲁棒性优良。

关键词: 深度估计, 单目视觉, 人工智能, 神经网络

Abstract:

Recovering depth information from monocular vision is a classical problem in the field of computer vision. The deep learning method combined with traditional algorithms is a hot research topic in recent years, but there are still limitations in algorithm fusion of neural networks, reference calibration and application scenarios. This paper proposes a two-way fusion depth estimation neural network structure, which trains the network based on the semantic information of depth gradient and depth gradient respectively, then trains the final detailed features again after feature fusion, and solves the problem of heavy workload of real reference calibration by single calibration method. The network structure can infer detailed depth information from a single RGB image. The network model is based on KITTI depth map data set training. Experiments include KITTI test set and some actual scene atlas. The results show that the method is superior to the contrast depth estimation scheme in depth information detail reconstruction and in large field of view scenario with good robustness.

Key words: depth estimation, monocular vision, artificial intelligence, neural network