| With the continuous increase of Internet video,the task of video moment retrieval has been widely used in fields such as video retrieval,media analysis,and intelligent monitoring.However,due to the massive and diverse nature of video data,traditional video moment retrieval methods require large-scale training,which poses new challenges in data privacy and model performance.To address the limitations of traditional methods,this paper proposes a new task,i.e.video moment retrieval task in the context of distributed data with privacy protection considerations,motivated by the advantages of distributed data and federated learning,to provide large-scale secure training for video moment retrieval.Federated learning refers to the training of a shared global model by multiple local devices(such as smartphones and computers)through encrypted communication without revealing data.Because federated learning can be trained without sharing data,it can ensure privacy and security.This thesis proposes two federated video moment retrieval methods for the proposed new task:federated video moment retrieval method based on grouped sequential federated learning and federated video moment retrieval method based on temporaldistribution-gap loss.The federated video moment retrieval method based on grouped sequential federated learning is designed to solve the problem of slow convergence and instability of traditional federated learning methods when used for video moment retrieval.The thesis divides the clients into multiple groups for sequential training within groups and parallel training between groups,which can improve model convergence speed,retrieval accuracy,and efficiency.The thesis conducted experiments on three different datasets to test retrieval accuracy and convergence speed,and conducted indepth visual analysis of retrieval efficiency on one of the datasets.The experimental results confirm our hypothesis and the effectiveness of our model.The federated video moment retrieval method based on temporal-distribution-gap loss is designed to solve the non-independent and identically distributed problem faced by the video moment retrieval task in the context of distributed data with privacy protection considerations.The thesis starts from the local training process of clients and the aggregation process of the server,using temporal distribution difference loss to regulate the local training process of clients,and then use the validation set among clients to provide selective aggregation methods for the server.By modeling the changes in data distribution,the robustness and generalization ability of the federated video moment retrieval model are ultimately improved.The thesis conducted experiments on two datasets,analyzed and discussed the results,and confirmed the improvement of the retrieval accuracy and the effectiveness of the two modules. |
