Research On Scene Graph Generation Algorithm Based On Causality

Scene graph generation is an important research direction in the field of computer vision.Many researchers continue to explore and study problems in the field of image deep understanding.To deepen the research on image understanding,based on the re-search foundation of target recognition,image subtitles,visual question answering,and natural language processing,the task of scene graph generation has gradually become a hot research topic in the current field of computer vision and natural language processing.Scene graph generation tasks require rich and advanced semantic understanding,which is a huge challenge for machines.To improve the quality of the scene graph generation model,this paper proposes a Scene Fusion Module(SFM)and a Causality Reasoning Module(CRM)for the insufficient utilization of scene information and the long-tail dis-tribution of predicates in the data set.The main research contents are as follows:Because the standard information without scene tags in the current data set and the currently existing models only use the object-level context fusion method,ignoring a lot of environmental information and other issues,this paper proposes the SFM module to build a scene graph generation model based on scene fusion.The model uses a combination of parts of the image to model the entire scene,and obtains the feature representation of the dimensions of the object feature through the layered convolutional network module,and shares its scene information.Compared with existing models on the One Shot dataset,SFM can effectively promote the fusion of image scene information and object features,which proves the effectiveness of SFM in the task of generating scene graphs.Aiming at the long-tail distribution of relational predicates in the current data set,overlapping of relational predicates,and insufficient fine-grained relational predicates,this paper introduces a CRM module to build a scene graph generation model based on causal reasoning.This model constructs a cause-effect graph based on the scene graph task through the CRM module,counts the object features under the guidance of the cause-effect graph,and further integrates it with the relational features.Compared with the existing model on the One Shot data set,the CRM module can observe the main influence of the predicate on the formation of relational triples in the case of the long-tail distribution of the relational predicate,which proves that the model can improve the predicate long-tail problem and Distinguish the size of some predicates.Finally,to make reasonable use of scene information and alleviate the problem of long-tail distribution,this paper proposes a scene graph generation model for causal rea-soning based on scene fusion.The model uses the feature representation learning module of the multi-head attention mechanism in the basic network to perform feature fusion.The effectiveness of SFM and CRM is verified through ablation experiments and multi-ple comparative experiments,and it is proved that the model can effectively improve the quality of scene graph generation.