Research On Visual Consistency In Augmented Reality Based On Deep Learning

Augmented reality technology is committed to superimposing computer-generated virtual objects into real scenes.In the superposition of virtual objects and real scenes,it is necessary to consider the spatial geometry and appearance consistency of the two in fusion.These visual consistency issues are enhanced An important and difficult problem in the study of real technology.The main problem of the traditional visual consistency research algorithm is that it depends on auxiliary equipment,the application has limitations,and the algorithm is too complicated,which affects the wide application of augmented reality technology.This paper studies the problem of virtual and real consistency based on deep learning,and explores effective geometric consistency strategies and illumination consistency mechanisms.First,the thesis studies a scene three-dimensional structure estimation and camera positioning algorithm based on multi-task learning.Using the design concept of Simultaneous Localization and Mapping(SLAM),an end-to-end multi-task learning network structure is designed.In order to effectively obtain the output of multi-tasking,the scene structure estimation sub-network and camera position parameter estimation sub-network,And adopts the geometric consistency of the three-dimensional space as the loss function,constrains the network performance,and obtains satisfactory results of the scene three-dimensional structure estimation and camera positioning,avoiding the problems of feature detection and accurate matching in traditional research,while enhancing The study of the consistency of the reality and the reality laid the foundation.Secondly,the thesis explores a virtual-real geometric consistency fusion strategy.In the research,the idea of scene semantic segmentation is integrated,and a virtual-real fusion framework including two stages of semantic information extraction and global map fusion is designed.In the semantic extraction stage,a multi-task network framework is designed to obtain more accurate camera parameters while acquiring scene semantic information,to ensure that the geometry of the virtual target is accurately positioned during the fusion of scene semantic information,To ensure the accuracy of map fusion.In addition,in order to improve the geometric positioning accuracy,a global fusion strategy of the target object is designed to optimize the geometric position of the virtual object and ensure the accuracy of the virtual object's relative geometric position.Further experiments and analysis show that the studied virtual-real geometric consistency fusion strategy can effectively achieve the virtual-real geometric consistency fusion,and can avoid the restrictions on the marker requirements in traditional problem research.The research stage of the thesis also explores a fusion algorithm of virtual and real illumination consistency.Design a deep learning network architecture based on unsupervised learning to estimate scene lighting parameters.The network topology includes two sub-networks:surface parameter estimation sub-network and light intensity estimation sub-network.In the surface parameter estimation sub-network,the surface normal vector and roughness of the scene are obtained,which solves the problem of data labeling in lighting parameter estimation.At the same time,in the light intensity estimation sub-network,unsupervised learning is used to design the surface consistency loss function to ensure the accuracy of the scene light intensity estimation.Further,the rendering result is used as a constraint of lighting consistency for optimization to ensure the accuracy of network lighting parameter estimation.Further experimental results show that the strategy explored can effectively achieve the integration of lighting consistency between virtual objects and real scenes,and provide an effective solution for augmented reality visual consistency.In a word,in the research of this paper,we systematically analyze and study the problem of geometric consistency and the problem of lighting consistency,and explore effective strategies.Through scheme design and analysis of experimental results,we have obtained effective solutions to problems and also enhanced The practical visual consistency problem proposes a feasible solution route,and the researched strategy can provide effectiye measures for practical application.