Single Image Depth Estimation Based on Improved U-NET

In response to the problems of low accuracy and complex network structure in existing deep learning based monocular image depth estimation algorithms, an improved U-NET [1] monocular image depth estimation method based on dense skip connections is proposed. The traditional U-NET network has the problem of depth distortion when processing depth estimation tasks. In response to this problem, this algorithm proposes skip connections to achieve multi-scale feature fusion from encoder to decoder; And introduce attention mechanism into the decoder to further improve the accuracy of depth estimation. The training and verification were carried out on the indoor data set KITTI [2] and the outdoor data set NYU Depth V2 respectively. The experimental results show that compared with the advanced algorithm DORN, all indicators have been improved to varying degrees, in which the root-mean-square error has been reduced by 24.1% and the square relative error has been reduced by 15.2%.