Research On CNN-based Monocular Depth Estimation

The image depth information can be used to understand the geometric relationship of image scenes,and has important applications in robots,scene understanding,three-dimensional reconstruction and other fields.In recent years,with the rapid development of deep learning,a new approach has been opened up for the study of depth estimation of monocular images.In this paper,based on the deep learning network,two aspects of the depth image of the monocular image are studied:1.Because monocular visual cues have a correlation with depth features,this paper constructs a deep convolutional network from both the local features of the image and the global layout.By introducing multi-scale convolution kernels and increasing the network depth,Multi-scale features of multi-scale maps are introduced into the final depth prediction process.Experiments show that the improved algorithm is more prominent in the detailed representation of the depth-resulting image.2.This paper proposes a monocular image depth estimation algorithm based on multi-features fusion network.The algorithm consists of two partial networks: feature extraction network and fusion prediction network,similar to codec structure.The fusion forecasting network gradually merges the extracted features of each layer,and introduces a deconvolution layer during the fusion process,so that the resolution of the predicted image is the same as the network input image,and finally the predicted depth image is obtained.The fusion forecasting network gradually merges the extracted features of each layer,and introduces a deconvolution layer during the fusion process,so that the resolution of the predicted depth image is the same as the network input image,and finally the predicted depth image is obtained.This paper uses the NYU v2 data set for simulation experiments and uses the average relative error,root mean square error,and threshold accuracy to evaluate the results.Experiments show that the multi-feature fusion network based on full convolution can effectively reduce the network parameters and improve the accuracy of the depth result map.