A Face Detection Algorithm Based On Light-weight Convolutional Neural Network

The subject of this thesis is to implement a face detection algorithm based on light-weight convolutional neural network.The last two years there have been many face detection algorithm based on convolutional neural networks,and their accuracy is much higher than face detection algorithm based on traditional machine learning methods.But their networks are relatively large,and it's not suitable for planting on the mobile device.Our goal is to compress the network model as well as get a faster face detection model with keeping the accuracy of algorithm.At first,this thesis designed a light-weight face detection algorithm based on convolutional neural network.it's the first time to combine the existing networks,SqueezeNet and RPN network,to implement a dedicated network for face detection.This network have an advantage in network speed,model size than the networks based on VGG16.Secondly,we fused features from SqueezeNet's different convolution layers.Reference currently available algorithm about features fusion,we used this method for designing lightweight face detection network,so that we can get better feature in the convolution layers.Finally,we use two methods to compress the detection network.One way is to use smaller connection layer compression,and the other is to use SVD decomposition.Through these two methods,we make lightweight face detection network model smaller and detecting the face faster.