Author(s)

Shi Zhibo

Corresponding Author

Shi Zhibo

Abstract

The acquisition of depth information is a very important part of scene analysis, which is mainly divided into two methods: sensor acquisition and image processing. The technology of sensor technology has very high demands on the circumference, so the image processing is a more general method. The traditional method uses binocular stereo calibration to get the depth by geometric relation, but it is still restrained because of the circumstance. Therefore, as the closest to the actual situation of the square method, monocular image depth estimator has a very large research value; For this reason, a method for depth estimation of monocular image based on DenseNet is proposed, which is based on multi-scale convolutional neural network. The DenseNet structure is added to optimize the feature acquisition process by using the characteristics of DenseNet strong feature transfer and feature reuse. The results show that the average relative error is 0.119. The root mean square error is 0.547. And the logarithmic mean space error is 0.052.

Keywords

Depth calculation, Depth estimation, Convolution neural network, Multi-scale, DenseNet