Monocular Depth Estimation Based On Deep Learning

Monocular depth prediction is a hot research topic in the field of computer vision in recent years.Obtaining depth information from monocular images is an important method to understand the geometric relationship of the scene,and it is also a key technology for 3D reconstruction and viewpoint synthesis.Compared with the multi-factor supervision of binocular and multi-viewpoint image depth prediction,monocular depth prediction is more challenging.In order to improve the accuracy and generalization ability of in-depth prediction,three aspects of research work have been carried out by using the convolutional deep learning networks:(1)To address the problems of inaccurate geometry and blurred edges in predicted depth images,a monocular depth estimation algorithm based on multi-scale structural similarity and gradient matching is proposed to achieve monocular depth prediction using a joint loss consisting of multis-cale structural similarity loss and scale-invariant gradient matching loss to rank the relative depth point pairs.Numerical experiments and subjective evaluation results show that the obtained monocular depth maps have more accurate geometry,clearer edges at depth discontinuities,more robust results,and some generalization performance.(2)To address the problem that the existing monocular depth prediction network lacks interconnection between feature channels,resulting in inadequate and inaccurate capture of depth information in images,a monocular depth network estimation algorithm based on splitting attention mechanism is proposed to improve the connection between feature maps across groups by adding an attention mechanism to the feature map channels to further enhance the interaction between feature maps and the representation of feature maps,so as to improve the The ability to capture depth information further improves the accuracy and generalization of monocular image depth estimation.(3)Spherical panoramas have the potential to produce more accurate,complete and proportionally consistent scene reconstructions as they have a complete view of the environment and provide a relatively complete description of the scene.This paper utilises the previous algorithms for engineering applications,based on the optimised depth prediction network in Chapter 4,to perform sub-regional depth prediction of monocular spherical panoramas and post-process the generated rough depth prediction results to obtain accurate and undistorted spherical panoramic depth prediction maps.