Face Recognition Of Surveillance Video Based On The Supercomputing Cloud Platform

Video surveillance system now is widely used in the management of city,it has an important position in the field of traffic control,security management,and financial security.With the laying of a large number of surveillance cameras,massive surveillance videos have been produced everyday,through a conventional human means to monitor and watch the playback of surveillance has become increasingly powerless,there is a increasing need for intelligent processing techniques to free the people from this heavy work.CNN(Convolution neural networks)is the state-of-the-art in the field of face recognition.Through machine learning,CNN can achieve higher recognition rate than people,and it is the perfect automated processing technology to replace people.However,the current surveillance video processing still have many problems,such as construction dispersion,inconsistent standards,poor interconnecting.It has become increasingly urgent to build a unified,interconnected and efficient surveillance video processing platform in the current environment.The intelligent processing of mass surveillance videos require a large storage and processing platform,and provide convenient inquiry service for users via the Internet.Undoubtedly,the supercomputing cloud platform provides an excellent platform.To solve these problems,we have designed a recognition framework of surveillance video based on supercomputing cloud to storage and process the massive surveillance videos,and provide services through the cloud for users.After the completion of the design and implementation of the framework,it carried out a detailed experimental and evaluation.We mainly completed the following work:1,We design the surveillance video processing framework based on supercomputing cloud and use CNN to recognize the human face in the surveillance videos,provide inquiry service through the cloud for users.We also introduce the design of storage,frame processing and face recognition.2,We focus on the training of convolutional neural network and optimize the training methods.It takes long time to train the CNN,especially when the scale of the CNN is greater.By analyzing the CNN architecture and training principles defined,we use dichotomy method to find the convergence domain of learning rate.Based on the convergence domain,we have analyzed 6 training methods and proposed the optimized range of coefficients.3,Experiments and evaluation of training optimization.We select a typical CNN and the standard face recognition database for training,and verify the effectiveness of the training algorithm optimized through experiments.4,Experiments and evaluation of framework.By building the framework on the Tianhe-2 cloud platform,we complete the basic test environment deployment.We use videos data as the experiment data,and simulate the real using environment of the framework.Experimental results show that the framework is able to feedback to the users' inquiry within seconds.