Design And Implementation Of Image Quality Assessment For City-level Video Surveillance

With the continuous advancement of smart security construction in smart cities,the number of surveillance cameras owned by smart security in a second-tier or above city usually approaches or exceeds one million.Therefore,for the intelligent monitoring system,not only the clarity of the monitoring video is required,but also the ability to obtain effective information from the monitoring video.As for such a huge scale of cameras,it has become impractical to detect and evaluate of the quality of each camera manually.Accordingly,the scientific video quality management of city-scale cameras has become one of the research hotspots of smart security.In order to achieve scientific video quality management of city-scale cameras,this paper studies the design and implementation of the quality assessment system for video images in this senario.Firstly,due to the three-level organizational structure of city bureau,sub-bureau and police station in the public security video surveillance private network,distributed deployment is adopted to implement the proposed surveillance video quality analysis and management system.In the same time,considering the convenience of users,B/S architecture is selected in this system.According to the development principle of the public security system,users are divides into two categories in this paper: administrators and visitors.In addition,function modules are designed separately for specific users.And the part I am responsible for in this project is the design of the web UI and front and rear interfaces of the system.In terms of technology selection,frameworks and component libraries such as Vue and Antdesign Vue are used for UI and interaction with users,and Axios is used for front and rear data interaction.Then,the residual network Res Ne Xt50 is selected for monitoring video quality detection.By supplementing surveillance images in the public TID2013 distortion dataset,expanding image types and then performing data enhancement.Furthermore,network training is optimized by introducing Leaky-Re LU activation function and cross-entropy loss function.In view of the low classification accuracy of Res Ne Xt50 on the expanded dataset,the c SE module is introduced to improve the feature extraction ability.Finally,the overall test of the above system is carried out,including functional test of each module,UI test and security test of the whole,and experiments in actual scenarios.The results show that the functions of the system meet the design requirements.