| In recent years,with the rapid development of artificial intelligence,face detection is one of the most popular research topics that has been applied to many practical scenarios in social life,such as mobile phone payment,intelligent security,and automatic driving.As soon as YOLO network came into existence,its real-time detection performance attracted much attention.Compared with the current practical detection network,YOLO has the advantage of detecting speed,but at the same time,it also sacrifices the detection accuracy.Until the emergence of the third generation of YOLO network,the network structure became deeper and the precision of the detection improve but at the same time reduced.With the rapid development of face detection network based on deep learning,how to apply to the real social life is facing great difficulties and challenges,such as too small and low pixel targets,too much storage space caused by too many model parameters,and the speed and accuracy of detection cannot be balanced.Aiming at the above problems,this paper proposes a new network model based on the regression thought of YOLO network.The main research contents of this paper are as follows:(1)For the face data set,the influence of different amount of anchors on the network model was studied.The unsupervised algorithm k-means++ was used to re-cluster the width and height of the face data set,so as to obtain the size and number of boundary boxes suitable for the face data set in this paper.(2)In this paper,a weighted non-maximal suppression algorithm is proposed to solve the problem of non-maximal suppression.The prediction box with a low score in the prediction box is not directly suppressed,but the score is lowered and compared with the set threshold value,and then suppressed below the threshold value,otherwise it is regarded as the output of the detection box.This method avoids the problem of missing detection of adjacent faces and improves the detection accuracy to some extent.(3)For face detection in complex scene and the size of the face percentage,there is give up small goal of face detection for big targets,this paper proposes a new loss function that the location prediction error of high contrast method used to calculate wide,and in view of the small target a single sample,use the focal loss as category forecast loss function,through the parameter adjusting the proportion of positive and negative samples,improve the weight of the small target,make small goal,effectively improve the detection results.(4)The traditional deep convolutional network for classification has the problem of too many parameters and too much storage space and too high computing cost,which makes it difficult to deploy the network model on mobile devices with limited computing power and internal storage space.In this paper,the optimization scheme of traditional convolutional neural network was analyzed.The depth convolutional network was optimized by using the depth separable convolutional layer,the linear bottleneck layer with inverse residual and the SE module,and a new lightweight detection network deft-yolo was formed by combining MobileNetv3 as classification network and YOLO layer.Experiments show that the network in this paper can effectively reduce the computational cost of the model and greatly reduce the storage space,and ensure the detection accuracy without loss. |
