Face Feature Analysis In Surveillance Video

The development of computer hardware has promoted the progress of deep learning technology in the field of image recognition.Based on the convolutional neural networks,the image classification,the face recognition method has achieved remarkable results.In recent years,with the rapid development of artificial intelligence,multimedia and other technologies,surveillance video has been widely used in home security,financial security,and intelligent travel.The face feature analysis in surveillance videos has become a hot topic for scholars.Most of the traditional research methods ignore the need of real-time performance in the surveillance video,and rarely achieve simultaneous research on multiple attributes and lacking further resolution of the few attribute data problems.Therefore,this paper mainly studies the real-time and accuracy of face feature analysis in the surveillance video,and solves the problem of small amount of data from the perspective of image generation,the target is further optimizing the model of attribute analysis.The main research content is as follows:Firstly,the article illustrates the significance of the face feature analysis in the surveillance video from the following aspacts,such as composition,characteristics,application fields and so on.Summarizing the method of selecting the network structure for face feature analysis in the surveillance video,and introducing the single network multi-target network structure of the Lighten CNN based on Multi-task.Several important evaluation indicators for face feature analysis are described.Secondly,this paper uses a method for generating face attribute data based on BEGAN,CycleGAN.The existing surveillance video data is relatively small and this has a great influence on the final accuracy of the model.In order to reduce the data amount influence on the final face feature analysis results,using the best performing network BEGAN to generate the face attribute data in the surveillance video.In this strategy,the method of merging the existing surveillance video data with the open source dataset CelebA is proposed to carry out the training of the BEGAN network,and an optimal iterative model is used to generate the face attribute data based on the input noise.The generated face data is added to the original surveillance video dataset,and the attribute model is optimized based on the extended data.At the same time,using the model of CycleGAN to generate face attribute data of wearing glasses and wearing a gauze mask.The experimental results show that the generated face attribute data has a high degree of authenticity in terms of clarity and detail,and based on the extended attribute dataset,the model is further optimized.Thirdly,a semi-automated data annotation strategy based on threshold is designed for model training and sample extension.Because the original manual annotation method brings heavy workload.The method uses a small amount of manually labeled data for training,through the model's prediction on unlabeled data,the probability values in each category are obtained.The threshold strategy allows the model to automatically label the samples when the predicted probability is greater than the specified threshold.Experimental results show that the threshold strategy can reduce the burden of manual annotation while ensuring the accuracy of data annotation,and optimize the recognition effect of the model.Finally,an optimized model is used to analyze the credibility of face sequence in the surveillance video.