Face Detection Based On Improved SSD Network

In recent years,with the rapid development of deep learning technology,face detection technology has made great progress and has become a popular research in the field of computer vision.However,face detection is susceptible to factors in scale,illumination and occlusion,which reduces the accuracy of face detection.Compared with traditional detection methods,SSD(Single Shot MultiBox Detector)network detection method has high efficiency detection speed and good detection effect,but there are also some shortcomings.Therefore,this paper proposes a face detection method based on improved SSD network,which has achieved good results in terms of detection accuracy and detection speed.The main research work of this article has the following two aspects:(1)For face detection in multi-scale and other environments,the detection effect of traditional SSD network has problems of missed detection and false detection.The SSD network based on ResNeXt network is proposed.The original vgg16 network is replaced by the ResNext-50 network in the basic network,using the ResNeXt-50 network structure can not only reduce the difficulty of network training but also reduce the redundancy of feature extraction,continuously enhance the ability to express facial features,and further improve the accuracy of face detection.Finally,the experimental results on the WiderFace verification set show that the SSD network based on ResNeXt network can improve the detection accuracy.(2)In order to solve the problem of small face detection accuracy in SSD network,the SSD network based on feature fusion method is proposed.First of all,deconvolution is used to learn the rich semantic information of high-level features for high-level feature maps,dilation convolution is used to learn the good location information of low-level features for low-level feature maps,and convolution is used to reduce the number of channels of features for middle-level feature maps,then the high-level feature maps,middle-level feature maps and low-level feature maps are connected and fused to obtain a new feature map.Finally,the residual layer is used in the prediction layer of the network,and the loss function layer uses a GIoU Loss replaces the original Smooth L1 Loss.The experimental results on the WiderFace verification set and the Pascal Face data set show that the SSD network based on the feature fusion method has better performance in detecting accuracy and speed of small faces.