Design And Implementation Of A Scenic Spot Imageand Text Recording System Based On Human-Object Interaction

In order to meet the needs of tourists to write travel notes,the scenic area provides functions for tourists to take photos and push.If the interaction between tourists and scenic facilities in tourist photos can be identified,and personalized graphic and textual records can be generated by integrating tourist photos and interactive action text descriptions,it will better meet the travel needs of tourists.The human object interaction(HOI)recognition technology is used in the image and text recording system to detect the interaction between tourists and objects.In response to the problems of low detection accuracy and high computational overhead in traditional HOI algorithms,this dissertation investigates the effective learning and long tail distribution of human and object features.The main research work is as follows:1.Design a HOI algorithm based on instance aware attention moduleIn order to fully learn the characteristics of humans and objects,this dissertation designs an instance aware attention module between the human object pair decoder and the interactive decoder,promoting information aggregation between the two detection branches and fully learning the characteristics of humans and objects.Experimental results have shown that compared to the cascaded basic model,the HOI algorithm based on instance aware attention module has increased mAP by 1.62 and 1.56 on the V-COCO dataset,and increased mAP by 0.24 and 0.37 on the detection of complete and non rare categories on the HICO-DET dataset,respectively.However,it has decreased by 0.11 on rare categories.2.Design a dynamic reweighting training mechanism for HOI algorithmIn response to the long tail problem caused by the combination properties of HOI samples,this dissertation proposes a dynamic reweighting mechanism as a training strategy to capture real-time weights of samples and alleviate the long tail effect.The experimental results show that compared with the cascaded basic model,the HOI algorithm based on instance-aware attention module and dynamic reweighting mechanism proposed in this dissertation has improved mAP by 1.81 and 1.73 on the V-COCO dataset,and 0.71、0.67 and 0.74 on the HICO-DET dataset,respectively.The parameter quantity is 46.7M,and the inference speed is 70ms,which has good practical application value.3.Design and implement a graphic and text recording system based on human-object interaction recognitionUsing the proposed HOI algorithm,this dissertation designs a graphic and text recording system in scenic spots,and completes the design and implementation of a graphic and text recording business platform and a graphic and text recording management platform based on the SpringMVC framework.