Research On Image Retrieval Algorithm Based On AR Cloud System

With the development of computing power of mobile devices,it has been possible to implement complex image processing such as feature detection and matching of augmented reality(AR)in mobile devices.However,there are still many problems in mobile AR applications.For example,the matching speed between Mark image and background image database is slow,the query efficiency is low,and so on.In order to reduce the burden of image matching,we unload the heavy computing tasks to the cloud environment.Because the existing research and technology can not meet the needs of dynamically adding and deleting new image data,we propose a new architecture to meet their own needs and improve the matching and retrieval speed of background images.Firstly,based on hierarchical clustering,the vocabulary tree is improved and the binary tree of high-dimensional space is constructed by using the random projection forest approximate nearest neighbor search method for reference.The addition and deletion of the image do not need to reconstruct the structure,which greatly reduces the burden of the background database.At the same time,the binary tree structure combined with the inverted vocabulary tree index system reduces the time complexity and improves the speed of image retrieval.Then,we improve the TF-IDF algorithm,inspired by VLAD to calculate the relevance of words,so that the weight of visual word bag vocabulary is more reasonable,so as to improve the accuracy and efficiency.In our evaluations,the mobile AR application built with the CloudAR framework runs at 30 frames per second(FPS)on average with precise tracking.Our results also show that CloudAR outperforms one of the leading commercial AR framework in several performance metrics.Experiments verify that the architecture proposed in this paper can accurately match the searched images.On the other hand,comparing with other methods,we find that our architecture can effectively improve the retrieval speed compared with other methods,and it is suitable for most of the current image descriptors.