Research On Object Detection Algorithm Based On Object Relationship And Scene Context Constraints

Object detection is a fundamental and very important problem in the field of computer vision.Accurate and efficient detection algorithms are requisite for the application of objection detection.In recent years,although significant progress has been made in the research of object detection algorithms,these algorithms often rely only on the local area features of each object in the image and ignore the relationships between objects and the scene information,making the detection performance poor.In order to use the relationships between objects and scene information,a post-correction algorithm based on prior co-occurrence between objects and a feature enhancement method based on scene-object relationships are designed to improve the detection effect,and experimental verification is carried out.Aiming at decreasing errors in the existing object detection methods,a postcorrection algorithm based on prior information is proposed.Through statistics theory,the co-occurrence relationships between each category in the dataset is obtained,and the information is stored through the co-occurrence matrix.The detection results obtained by the existing object detection algorithm are combined with the co-occurrence relationships between the categories to establish a label-to-label graph,which is the input of the graph convolutional neural network,and the output of the network is the correction vector.The result of existing object detection is corrected by the correction vector to improve the detection accuracy.A comparative experiment is conducted on PASCAL VOC,and the experimental results show that the algorithm can improve the accuracy of object detection.In order to make full use of the relationships between objects and the relationships between the scene and the object,an object detection method that combines the object relationships and the scene information is proposed.By representing regional features as nodes and using neural networks to generate edges with prior information,a regionto-region graph is established.The graph convolutional neural network is introduced to extract the relationships between the nodes in the graph,and feature enhancement based on the object relationships is realized.Statistics on the dataset then is used to obtain the prior co-occurance information between the scene and the object,adaptively generate the coefficients containing the prior co-occurance of the scene and the object through the neural network,and then utilize the coefficients to screen the scene features to achieve feature enhancement based on the scene and object relationships.The enhanced features are aggregated and fed into the subsequent network for detection.The method is verified experimentally on the PASCAL VOC dataset and the COCO dataset,and the experimental results show that the proposed method can improve the accuracy of the object detection.