| Under the trend of the deep integration of AI and safety prevention and control,the efficient,convenient and reliable biometric recognition technology represented by face recognition is widely used in the industrial technology upgrading of automobile intelligent security.In recent years,face recognition driven by deep learning has also made a leap forward development in keeping with the development pace of artificial intelligence.However,limited by vehicle computing resources,artificial intelligence technology represented by deep neural network is difficult to be applied in vehicle terminals on a large scale.At present,deep learning algorithm is only implemented by cloud platform or industrial computer,but there are few researches on how to break through the bottleneck of computing power to realize deep learning algorithm.In this study,the driver in the vehicle environment with multi factor interference is taken as the object.Through the comparative study of a variety of face recognition algorithms under deep learning,a face recognition method with strong real-time,high accuracy and suitable for algorithm transplantation in the vehicle environment is proposed.The main work of this paper includes:1.Designing and building the deep convolution neural network model,it is based on the Googlenet network,aiming at the application requirements of high accuracy and strong real-time of face recognition in the vehicle environment,combined with the characteristics of the target triplet to minimize the distance between the same classes and maximize the distance between the different classes,the Cap＿conv neural network model for face recognition in the vehicle environment is improved The convolution and pooling’s layer of the network model is used as the feature extraction layer of the face’s area,and the capsule network is used to reconstruct the feature vector and realize the final face feature vector representation.Through the deep neural network model built by training and testing,the facial image feature information collected by the vehicle camera is extracted and mapped to another new feature space to achieve the comparison between the target to be tested and the feature face space in the collected data set.Through the index detection of the model,it is found that Cap＿conv neural network has good accuracy and real-time performance,but there is a high proportion of computing resources.2.In view of the high computational resource consumption rate of Cap＿conv neural network,which makes it difficult to apply it in the resource constrained environment,an optimization strategy of face recognition network model suitable for vehicle environment is introduced.On the basis of Cap＿conv nerual network,the original convolution module is disassembled by lightweight neural network,and the residual module is connected with the shortcut.Combined with the fitting operation of the network,the task of face recognition under the interference of multiple factors is realized quickly and accurately.According to the test results,the optimized Cap＿conv nerual network still has high real-time performance and accuracy under the embedded transplantation of vehicle system.The experimental results show that the optimized Cap＿conv neural network has98.06% higher accuracy,38% less computing resources and 35 ms faster processing speed for face recognition.It is more suitable for face recognition in vehicle environment and has strong practical application value. |
