| Face detection is a very basic problem in the field of computer vision and plays a great role in people's daily life.In recent years,with the rise of artificial intelligence,the related application of face detection technology has gradually increased.So far,face detection algorithm has been widely used in the field of security,entertainment,image retrieval,etc.It is developing at a very fast speed,affecting our lives in a variety of forms.Although the research of face detection algorithm has a relatively early origin and relatively rich technology accumulation,with the rapid rise of various applications,the current performance of face detection algorithm still cannot meet the needs of the existing society,especially in the actual scenes.It can be said that the research of face detection algorithm has a long way to go.In this paper,this article uses the training framework of convolutional neural network to study the face detection algorithm,aiming at improving the detection performance of the detector.Based on the single-step detector S3FD,this paper improves the two key steps of network structure and data processing,and proposes several improvement measures to improve the actual performance of the detector.In terms of network structure,since the feature fusion method can effectively utilize context information and improve the ability to express features,in order to improve the detection performance of smaller faces,this paper proposes the following improvements:(1)A low-level feature interleaving structure is used to combine the previous layer features,the current layer features,and the latter layer features to improve the expression of the current layer features.At the same time,for the high-level features are not suitable for the low-level features for fusion,this paper only selects the lower-level convolution layer for feature fusion.(2)This paper proposes a content fusion structure,which adopts the structure similar to Inception to better fuse the features with different sizes of receptive fields and improve the detection ability of faces at different scales.(3)In terms of data processing,this paper proposes a new data processing method.By adjusting the scale of the input face frame,the face frame and the anchor point can be easily matched together,thereby reducing the influence of background noise and further improving the performance of the network.In view of the above innovations,this paper proposes four enhanced network structures,namely VGG-16S,VGG-16S+,ResNet-50S and ResNet-50S＊,which can greatly improve the performance.In the actual test,the proposed network structures all show good performance,and achieve highly competitive results on FDDB and WIDER FACE test sets.Among them,this paper designed ResNet-50S network to reach the second place in the world in FDDB evaluation,while still maintaining speed above 10FPS. |
