Design And Implementation Of Video And Image Semi-Automatic Annotation System

In supervised learning,data are constantly processed and annotated for algorithm training.Manual image annotation,though labor-intensive,consuming and inefficient,has long been the common practice in the past.Sometimes,it even requires a machine to deliver the results.Nowadays,the application of machine learning in varied fields have been proved a success.Hence,the optimization of image annotation system becomes also possible with machine learning.This thesis proposes a semi-automatic image annotation system,which features the application of focal point learning,video classification and 3D reconstruction of human bodies,thus allowing semi-automatic image annotation and manual tuning in the system.The thesis opens with introduction of research background,including briefing and reviews of related techniques in annotation systems.It then follows a detailed analysis of algorithms applied in the proposed system,namely,two-stream convolutional neural network for video classification and 3D reconstruction of human body for action recognition.The proposed semi-automatic image annotation system combines algorithms with manual tuning,thus considerably speeding up the annotation process with relatively reliable precision.The system focuses on the recognition of human body in motion with functions such as skeleton detection,human instance segmentation,motion stages identification and 3D pose annotation.This thesis presents the functional requirements and feasibility analysis of the system,and elaborates on the framework and modular design for different functions.It confirms availability of the system with testing of each functional module.Through the testing and contrast of varied algorithms,the accuracy of the applied algorithms is verified.Lastly,it ends with a summary of the performance and limitations of system,and puts forward prospects of the research.