Research On Indoor Scene Recognition Method Based On Graph Relation Network

As one of the basic tasks in the field of computer vision,indoor scene recognition is obviously different from object recognition and image recognition.Its image composition is relatively complex.In the process of recognition,not only the overall layout of the scene should be considered,but also the detailed information of the object is particularly critical.With the development of deep learning,this task plays a key role in image retrieval,intelligent driving,video surveillance and other applications.Therefore,further research on indoor scene recognition task is still of great significance.This paper mainly studies the recognition of indoor scene,which is divided into the following two parts:1.This paper studies an indoor scene recognition method combining object features and global features.It uses YOLOv3 as the object feature extraction network,extracts features from the object detection positions of large-scale objects,medium-scale objects and small-scale objects in the network,and fuses them as object features to supplement the global features of the scene,so as to improve the accuracy of indoor scene recognition.The experimental results show that the recognition model which combines the object feature and global feature improves the recognition accuracy of indoor scene obviously.2.Considering that some objects often appear at the same time in the real scene,this co-occurrence relationship can help to accurately identify the object,this paper improves on the basis of research content 1,and proposes an indoor scene recognition method based on graph relationship network to make use of the relationship between objects.In this method,the word embedding vector of the object class is used as the node feature,and the edge matrix is constructed by the co-occurrence probability between the objects,so as to construct the graph structure.A group of object class vectors are obtained by the Graph Convolution Network mapping,and the similarity between the object feature extracted by YOLOv3 and the group of object class vectors is measured,The semantic level feature representation reflecting the distribution of objects in the scene image is obtained,which can be used as a supplement to the object feature representation to further improve the accuracy of indoor scene recognition.The experimental results show that the scheme achieves a certain improvement in the recognition of indoor scenes.