Research On The Small Scale Face Detection Algorithm Based On Convolutional Neural Network

Face detection has important application prospects in the fields of human-computer interaction,intelligent security and criminal investigation and detection.With the rise of deep learning,face detection technology has achieved better results than traditional methods by virtue of the powerful feature extraction ability of convolutional neural network,but there are still problems of low accuracy in small-scale face detection.In this paper,we study and analyze the task of small-scale face detection based on convolutional neural network.The main work is as follows:Firstly,taking advantage of the fast speed of YOLO object detection algorithm and aiming at the problem of low accuracy of small-scale face detection,a small-scale face detection method FusionYOLO based on improved YOLO is proposed.At first,considering the small proportion of small-scale face in the image and the aspect ratio is close to 1,an improved k-means clustering algorithm is used to select the appropriate candidate region scale.Then,the multi-level feature map is fused with fine-grained features to enhance the dependence of feature context information and improve the representation ability of smallscale face features.At last,combined with the suitable candidate regions in the clustering results,the YOLO prediction layer is redesigned to form a network structure suitable for small-scale face detection.The experimental results on WIDER FACE dataset show that compared with YOLO,the precision and recall of FusionYOLO are increased by 14.9% and 22.7% respectively,while maintaining good detection speed.Secondly,in order to further improve the accuracy of small-scale face detection,a small-scale face detection algorithm AttFusionYOLO is proposed.The algorithm consists of two parts: At first,A data enhancement method DenseGAN based on GAN improvement is proposed,GAN model is used to learn the mapping relationship between random noise and small-scale face images,which can be used to generate new images with similar samples,at the same time,dense connection blocks and network structure are used to improve the common problem of generator mode collapse in GAN algorithm training process,which can reduce the cost of manual labeling and increase the diversity of data;At last,on the basis of FusionYOLO,channel attention mechanism is introduced to model the correlation between each channel of feature map,and the weight of the model is weighted with the information of the original feature channel to achieve the re-calibration of small-scale face features.The experimental results show that the precision and recall of AttFusionYOLO are improved by 0.3% and 0.6% respectively compared with FusionYOLO,which proves the validity of the method.