Design And Implementation Of Face Detection Algorithm For Mobile Platform

Face detection plays an important role in face recognition technology.With the rapid development of mobile terminal market,how to develop and apply an efficient face detection algorithm in mobile terminal has become a research hotspot.However,the current face detection algorithm has some problems in the application of mobile terminal,such as low detection accuracy,slow running speed and poor portability,which hinder the widespread application of face recognition technology in mobile terminal.In response to the above problems,this paper proposes a more efficient lightweight detection algorithm named FaceYoLo based on Darknet-53 network,and then constructs a mobile terminal application system.The main work are as follows:(1)Design a mobile-oriented detection model with high precision.In order to improve the robustness of the model,we designed a multi-scale network module to capture the features of faces with different scales in the depth model,thus reducing the error rate caused by inputting faces with different sizes.In addition,we also proposed the central loss strategy,which makes the model more focus on the intra-class gap in the learning process and improves the precision of the model.(2)Improve the running efficiency of the algorithm in mobile terminal.Based on the characteristics that the larger convolution kernel can reduce the input spatial scale quickly,and the activation function CReLU can reduce the output channels,we designed a fast input module to cut the characteristic parameters of the input network,thus accelerating the running speed of the model in mobile terminal.(3)Optimize the model's transplanted memory space in mobile terminal.Based on the characteristics that group convolution which can process features in a distributed manner and the principle of quadratic expansion of 1x1 convolution,we designed a variable group module.This module can effectively reduce the output parameters of the fully connected layer while ensuring the detection accuracy,making the output model more lightweight,thereby improving the portability of the model in mobile terminal.(4)Develop a face recognition system for mobile platform based on FaceYoLoalgorithm.Compared with the MobileNet algorithm in mobile terminal,our algorithm nearly doubled in speed and successfully delivered applications.