Research On Scene Similarity Calculation And Labeling Method Based On Faster R-CNN

With the continuous exploration of scientific researchers,the research problems have become more and more complex.Target detection only obtains the category information and location information of the target in the scene,while ignoring the internal connections of the targets in the scene.In real life,the scenes that need to be processed are more complex,and the dependencies between targets are closer.How to efficiently analyze the spatial layout and semantic relations of each target in a complex scene,and calculate the similarity between complex scenes will become more and more important.In order to solve the above problems,this paper proposes a similarity calculation method for complex scenes based on the Faster R-CNN target detection network,and produces and annotates a complex scene dataset,namely the CSBOGM(Complex Scene Based On Graph Model)dataset.The research work is as follows:(1)A method of scene similarity calculation based on Faster R-CNN is proposed.It mainly consists of five parts:a)Proposing a data preprocessing method based on Faster R-CNN,its purpose is to realize the target detection of the image,the de-redundancy of the target detection frame and the unique number of target detection frame;b)Propose a scene the target tree construction method,realizes the division and merging of targets of different sizes in complex scenes,at the same time,provides node attribute information for constructing the scene graph model;c)Propose an algorithm for analyzing the relative position relationship between targets,which is used to process the semantic connection between targets in complex scenes,and provide edge attribute information for constructing scene graph models;d)Method based on above b and c,two scene graph model construction methods are proposed,and their graph models are named OT(Object Tree)graph model and PGAOT(Primary Goal And Object Tree)graph model;e)Based on the OT scene graph model and the PGAOT scene graph model,two scene similarity calculation methods are proposed.The experimental results on the Corel-1K dataset and Caltech101 dataset show:the method proposed in this article has achieved good results in visual perception and objective evaluation,which proves the effectiveness and practicability of the method proposed in this article.(2)A complex scene dataset was produced and annotated,and it was named the CSBOGM dataset.Considering that real scenes are often more complicated,and the existing image retrieval datasets have a single scene target,these datasets cannot effectively replace real scenes,and cannot prove the effect of existing image retrieval method in complex scenes.Therefore,this paper produced the CSBOGM dataset,which contains 10 complex scenes and a total of 6200 images.The experimental results on the CSBOGM dataset further prove the effectiveness of the method proposed in this paper,at the same time,the two comparison methods have also achieved satisfactory results on the CSBOGM dataset,which reflects the rationality of the CSBOGM dataset produced in this article from the side.In summary,based on the deep learning target detection method,this article proposes a method suitable for complex scene analysis and scene similarity calculation.At the same time,the CSBOGM dataset produced in this article is more complex than the existing image retrieval dataset.Experiments show:the method proposed in this paper has high interpretability and practicability,and the relative rationality of making the dataset.