Object Entity Alignment Technology Based On Knowledge Graph

In recent years,with the development of knowledge mapping(KGS),entity alignment has developed rapidly.The objective of entity alignment in knowledge graph is to find entity pairs pointing to the same object in heterogeneous graph.In the field of computer vision,the task of finding the same object in two different images should also be realized,so as to facilitate the application in target tracking and target recognition.Since there are few research methods of image alignment at present,this paper attempts to make a breakthrough from the perspective of knowledge graph.Based on the analysis of entity alignment technology of knowledge graph,this paper proposes the solution ideas and methods of image alignment through knowledge transfer,so as to achieve the alignment between different images.Therefore,this paper proposes a technology of object entity alignment based on knowledge graph.Because the entity alignment work of knowledge graph is usually divided into two parts: knowledge extraction and entity alignment,the image alignment work of this paper will be discussed in detail from these two parts.For the part of knowledge extraction,this paper uses the visual relationship prediction model to extract image information,including the objects in the image and the relationship between the objects.The visual relationship prediction model based on language priori is used as the reference model.Firstly,there are too many random combinations of entity pairs in the two images,resulting in redundant calculation that does not conform to the real relationship.This paper adds an entity-pair suggestion algorithm to the benchmark model.This algorithm can filter entity pairs before visual relationship prediction,remove some entity pairs that do not conform to the real situation,and reduce the computational complexity of relationship prediction model.Secondly,the image target detection of the benchmark model uses Region based CNN(RCNN).In order to improve the speed and performance of target detection,this paper replaces RCNN with Fast RCNN.For the part of entity alignment,TransE is used to embed image information into low dimensional vector space.Since different images will be embedded in different vector Spaces,the learning method of joint vector space is used to learn the union of two vector Spaces according to partially aligned seed pairs,so as to judge whether the target entities are aligned according to the distance between the targets in the same vector space.Finally,the improved algorithm is evaluated on VRD and VG-based data sets,and recall rate and accuracy rate are used as evaluation criteria.Experimental results show that the proposed method is feasible in image entity alignment.