Research And Implementation Of Crowdsourcing Annotation System For Still Image Visualactivity

At present,image recognition and understanding algorithms based on high-level semantics have been widely used in artificial intelligence systems.As a hot research topic in the field of computer vision,human visual activity recognition has positive significance for the semantic understanding of image content.In the past,the research on human activity recognition mainly focused on the analysis of spatiotemporal sequence data collected by mobile sensors such as video data and smart phones.Moreover,the international mainstream image data sets lack the data of activity label,the annotation tools at home and abroad do not have the annotation function specifically targeted for image activity semantics.Therefore,the lack of relevant experimental data sets makes the activity recognition task based on still image face great challenges.In this thesis,the problem of lack of annotation data in the semantic recognition of still image visual activity is studied.Under the background of the popularization of mobile devices,the "crowd-sourcing" idea is used to research and develop a still image visual activity manual annotation system based on Android platform.Based on the reference of Chinese and foreign literatures,this thesis first introduces the related technologies and overall design schemes of thesystem,then elaborates the implementation scheme of key functional modules,and finally illustrates the test results of the system.The thesis' s main tasks are as follows:1.Designing and developing a WEB system for manual annotating of visual activity of images;2.Adopting a crowd-sourcing task automatic allocation algorithm based on pricing mechanism to improve the allocation efficiency of image annotation tasks;3.Using web crawler technology to collect and organize image-assisted text labels to enrich the semantics of images further;4.Utilizing Word2 vec training word embedding model,convert qualified annotation information into word embedding and store it in the database.The application in the actual deployment system shows that the operation of the system is simple and smooth,the function modules of each algorithm are stable and efficient,and the advantages of mobile terminals are fully utilized to collect and organize the data of image activity annotation.