Research On Distributed Real Time Face Retrieval Technology

Face recognition has been an important subject in the field of computer vision research and engineering applications,scholars have been exploring how to identify the identity in a large number of face images.Nowadays surveillance cameras are widely used in the maintenance of public security and crime control.Face recognition based on surveillance video,can be used for key person tracking and large-scale face retrieval,has important significance for public security,security and customs etc.While the number of surveillance cameras is high,face recognition based on video stream in the need to ensure the recognition accuracy,also need to meet the requirements of high throughput and real-time response.The distributed real-time surveillance video stream processing platform,based on the stream computing framework Storm and distributed message queue Kafka,is for real-time analysis of customs video stream to identify "water person".In real-time aspect,we must first ensure that the algorithm including face detection and alignment,face feature extraction and face recognition to meet the real-time requirements.In the face detection and alignment algorithm,three cascade convolutional neural networks are used to detect human face and location the key points.In face feature extraction algorithm,The 32 layer convolutional neural network model based on residual block design has the accuracy of 99.42% for face verification and public data set LFW.And extracting a face feature requires only a few milliseconds,which can meets the real-time requirements.Statistical analysis shows that in the process of the video analysis,candidate selection for the subsequent accurate matching becomes the bottleneck.Fast hash face recognition method based on depth hashing can be a good solution to this bottleneck problem.Comparing the retrieval efficiency between the multiple segment hash index algorithm and the GPU accelerated hash search algorithm in different scenarios.The experimental results show that,when the number of bits is small or the data size is large,the multi index algorithm has higher query efficiency,otherwise,the GPU based accelerated hash search algorithm performs better.