Face Image Detection Based On Feature Enhanced YOLOv4 Network

Face detection technology has a high application value due to its wide range of applications,so it has been a hot topic in the field of computer vision research has always attracted attention.However,in the use of actual complex scenes,face detection is susceptible to factors such as the distance from the imaging device,the imaging angle,the changes in light and shade of the external environment,and the varying degrees of occlusion,which causes the detection accuracy to fail to reach the standard.At the same time,face detection applications need to have a faster detection speed to meet actual application requirements.The YOLOv4 network detection method has both good detection accuracy and faster detection speed.It is currently one of the recognized general target detection algorithms.It can be applied to face detection to achieve good results,but there are also some false detections.Problems such as missed detection of small-scale faces.The face detection method based on the improved YOLOv4 algorithm proposed in this paper is correspondingly studied for the above problems.The main research work is as follows:(1)Propose a new backbone network structure for feature extraction.Firstly,the feature enhancement module(Feature Enhance Module,FEM)is used for feature enhancement in the multi-scale feature fusion PANet architecture.This module not only integrates the feature information of different levels in the context,but also obtains expanded receptive fields through the fusion of expanded convolutional subnets once,twice,and three times,and a feature map with enhanced features,so that the entire network can obtain more There are more recognizable and robust features,which improve the ability to extract and express facial features,thereby improving the accuracy of model face detection.In addition,improvements were made on the backbone network CSPDarknet53 of YOLOv4,and the first Cross Stage Partial(CSP)was replaced with the original Darknet residual layer.Finally,the Spatial Pyramid Pooling(SPP)in the Path Aggregation Network(PANet)is CSP-ized.(2)Based on the confidence loss function of the YOLOv4 network,a new loss function is proposed in combination with the focal loss function(Focal Loss).The focus loss function is mainly used to solve the problem of the imbalance of the positive and negative sample ratio in the single-stage target detection network.Through the parameter setting,the proportion of a large number of simple face samples in the training can be reduced,and the training focus on small-scale difficult face samples can be improved.,so as to finally achieve the goal of improving the detection effect of small-scale difficult face samples.(3)On the basis of the above improvements,a lightweight design was carried out,and the deep separable convolution was used to replace the ordinary convolution in the multi-scale feature fusion PANet system in the original network,and the main feature extraction network feature extraction ability was not lost at the same time,To a certain extent,reduces the amount of parameters of the entire model,and accelerates the model’s face detection and inference speed.Finally,the method proposed in this paper has been extensively verified on the WIDER FACE face database.The experimental results show that the face detection network proposed in this paper to improve the YOLOv4 algorithm not only effectively reduces the false detection rate of faces,but also at small scale faces It has good performance in detection accuracy.