| With the rapid development of artificial intelligence and mobile Internet,face detection and recognition has become a hot research field in the field of computer vision.With the rapid development and widespread adoption of mobile smart phones in the past decade,face detection and recognition technologies have been applied to image processing and identity authentication applications on mobile phones.Thanks to the enormous performance gains of deep learning in the field of target detection and image recognition,the application of deep learning to face detection and recognition tasks has become a research direction for many researchers and has achieved excellent detection and recognition results.At present,the development trend of deep learning is to obtain higher accuracy by training deeper and more complex network models on GPUs.It is difficult to apply a large-scale,computationally time-consuming deep learning model to mobile smart phones due to the limited storage and computational capabilities of mobile-side hardware.Therefore,the use of small-scale,low-latency lightweight deep learning model for face detection and recognition becomes the main idea of this paper.In the face detection stage,this paper uses multitasking cascade convolutional neural network MTCNN to achieve.The MTCNN network is a cascade of three lightweight networks.It uses the intrinsic correlation between facial feature point positioning and boundary regression of face candidate frames to improve the performance of the network.Face detection and alignment can be performed in a coarse to fine manner.This paper uses face datasets WIDER FACE and CelebA to train the MTCNN network and perform test evaluation on the face detection benchmark dataset FDDB.When the number of false detections is 500,the face detection rate of MTCNN reaches 93.77%,which exceeds Many other face detection algorithms.The MTCNN is applied to the Android platform.On OnePlus5 mobile phone,single face images with the size of 480×640 are tested.The mean detection time is 133 ms,and the speed is very fast.On the basis of face detection,this paper uses lightweight deep neural network MobileNets to extract features for face verification and face recognition.The MobileNets model is built on a deep decomposable convolution.Deeply decomposable convolution is a form of decomposed convolution,which can be solved by decomposing a standard convolution into a deep convolution and a dot convolution(1 × 1 convolution kernel).Deep convolution applies each convolution kernel to each channel,and dot convolution is used to combine the output of channel convolutions.This decomposition has the effect of greatly reducing the size of the model and the amount of computation.In this paper,the MobileNets model is trained based on Softmax cross-entropy Loss and Triplet Loss on face dataset CASIA-WebFace,and the test is performed on the face recognition detection dataset LFW.The MobileNets model based on Triplet Loss training achieved 97.61.% of the verification recognition accuracy rate,the classification accuracy rate of 97.50% was obtained on the 10 types of face datasets collected in this paper,which exceeds many face recognition algorithms and human face recognition accuracy.In addition,this article reduces the size of the network and increases the speed of the network by adjusting the two hyperparameters of the MobileNets with very low loss of precision.The MobileNets model is applied to the Android platform.On the OnePlus5 mobile phone,single face images with a size of 480×640 are tested.When the verification accuracy rate is 95.03%,the verification process takes an average of 308 ms.When recognition accuracy rate is 96.27%,the process takes an average of 179 ms.The recognition speed is very fast and with no sense of delay.Now,the iterative update speed of smart phones is very fast,each generation will have a significant increase in performance over the previous generation.Therefore,the face recognition model proposed in this paper has a wide range of application prospects. |
PDF Full Text Request