Agent-based Performance Optimization For Real-time Distributed Face Recognition

At present,face recognition system has been widely used in various fields.With the increasing scale of the processed videos in face recognition system,the traditional centralized face recognition methods have been unable to meet the needs of scalability.Abstract:Since the current face recognition system can't handle multiple videos at the same time,this paper proposes an agent based distributed face recognition model,which has multiple agents and one server.The agent can perform face detection,tracking and feature extraction of many videos simultaneously,and the server can perform face recognition of many videos.The proposed model effectively improves the efficiency of face recognition,and increase the extensibility.However,it will lead to processing time delay if the quantity of tasks which will be processed by agents not load balance,and the CPU utilization of the agent which processes the most tasks will be explosion.In this paper,we designed a solution for performance optimization by loading balance the tasks of agents.For load balancing problem,an improved particle swarm optimization algorithm and a new type of genetic algorithm are proposed.For convergence of particle swarm optimization algorithm,a convergence factor is introduced and the inertia weight is removed.In addition,a perturbation strategy is introduced to guarantee the algorithm to get the optimal solution.In view of the traditional genetic algorithm on the issue of population diversity and the number of iterations,an improved genetic algorithm based on the number of samples and the length of the chromosome is proposed.Through a lot of simulation experiments,the results show that the load balance for agents can effectively improve the video processing times,and significantly reduces the burden heavier proxy CPU utilization.In addition,in view of the problem that the tasks processed by agents unevenly resulted from the quantity of tasks are dynamic change as time goes on.This paper presents a dynamic optimization allocation method to keep load balanced for agents.In order to ensure the accuracy of the dynamic optimization allocation,this paper proposes a forecasting mechanism to predict the number of pedestrian in video at a certain moment.For predicting mechanism,an extreme learning machine algorithm is proposed to predict the number of pedestrian.According to the forecast data,the dynamic optimal allocation can effectively reduce the error caused by the abrupt change of the task quantity.Finally,the load balance algorithm for agent,prediction algorithm for the number of pedestrians and dynamic optimization allocation method have been applied to the distributed real-time face recognition model based on agent,and a lot of simulation experiments and performance analysis are carried out in a real network environment.Experimental results show that the optimization method proposed in this paper can effectively improve the efficiency of the distributed real-time face recognition model based on agent,and improve the scalability.