Object Detection Fused With Knowledge Graph

Recently,object detection technology has developed rapidly and been widely used in many fields,including automatic driving,underwater navigation,and security monitoring.Currently,advanced object detectors can highly-quality represent the features of each region,but cannot reason with common sense.In large-scale complex scenes,traditional object detectors cannot accurately identify occluded or small-sized objects.In response to this problem,this thesis constructs the knowledge graph through the Concept Net dataset and proposes anobject detection model fused with knowledge graph.The model can not only utilize the features of images but also capture relational features between objects.The main research work of this thesis is as follows:(1)By constructing the knowledge graph,this thesis designs the ODFKG(Object Detection Fused with Knowledge Graph)model.The model aims to use the image’s features and capture the relationship between objects to improve performance.The ODFKG model uses basic detection network to collect the weights of the original classification,integrating the high-level semantic representation of each category.At the same time,the model uses the RWR(Random Walk with Restart)algorithm to explore the overall structure of the knowledge graph iteratively,obtaining a semantic consistency matrix to enhance the features of each object.The ODFKG model obtains good detection results on the COCO2017 validation set,indicating that the ODFKG model can effectively improve the detection performance of the object detector.(2)This thesis proposes the ODFKG-A model(Object Detection Fused with a Knowledge Graph Based on Attention)to better capture different information in the network and enhance the feature representation ability of the network.The attention mechanism recalibrates features,emphasizing relationships between channels while suppressing noise information.The detection accuracy of the model on the VOC2007 test set and COCO2017 validation set is 80.8% and 56%,respectively.(3)In the NMS(Non-Maximum Suppression)algorithm,the Io U(Intersection over Union)only considers the overlapping area between two detection boxes,which often causes wrong suppression.By improving the NMS algorithm,this thesis proposes the Soft CIo U-NMS algorithm.To enhance the ODFKG model’s ability to capture global context features,this thesis also proposes the ODFKG-G(Object Detection Fused with a Knowledge Graph Based on Global Context Features).While focusing on the features of images themselves,the model can capture the relationship information between different spatial locations as well,and then strengthen the original features by aggregating the same global features.The model achieves 89.5% on the COCO128 validation set,which is 0.6%higher than the ODFKG model.