Research On Face Detection Algorithm Based On SSD Convolutional Network

Face detection has always been a popular research field,especially in recent years,deep learning has made great progress.Since 2014,the birth of a large number of access control,security and other fields face startups have greatly promoted the research and development of face detection.The traditional face detection algorithm is based on SVM and other machine learning algorithms.It adopts cascading and AdaBoost strategies,which has relatively small amount of computation,and has obvious advantages in speed.The drawback is that the detection accuracy is not good.At present,face detection algorithm based on deep learning greatly improves the performance of face detection.However,the huge amount of computation of convolution neural network makes these methods need to improve in real-time.SSD is a new face detection framework of convolutional neural network,its face detection mechanism is different from candidate box mechanism.Different layer in convolution neural network has different receptive fields and different dimensions,the receptive field of neurons in the superficial network layers is small,but the output feature map is large.The receptive fields of neurons in deep layers is big,but its output feature map is small.Therefore,in the SSD convolutional network,different level feature maps not only extract features as the input of the next convolutional,but also output the face class score and locate the face box accurately using detection and prediction network.This have greatly improved the utilization efficiency of network characteristics,saved the network forward propagation overhead of selecting the candidate frame,and speeded up the detection speed.The research of SSD network shows that increasing the size of input images can improve the accuracy of detection,but the large output feature map will inevitably lead to the increase of convolutional computation and reduce the detection speed.Using the method of model compression,the network parameters can be reduced and the running speed of the model can be accelerated.In this paper,we use deformable convolution instead of ordinary convolution layers to enlarge the receptive field of neurons,increase the adaptability of convolution to shape,and improve the accuracy.By transforming model parameters to ternary weights,the size of the model is reduced,the speed of convolution operation is accelerated and the balance of accuracy and efficiency is maintained.