Research On Face Detection Technology For Label Assignment And Multi-scale Feature Fusion

Face detection technology refers to the technology by which all faces are recognized and the corresponding face bounding boxes are returned from a given image,which is one of the hot researches in the current computer vision field.At present,face detection technology has been widely used in smart phones such as beauty photos,face payment,and community security access control.In the last decade,although face detection technology has developed rapidly,its performance can still be improved in unconstrained environments,and its performance and detection speed on mobile devices are not satisfactory.Therefore,face detection technology still needs to be deeply studied,which has academic and engineering application value.In recent years,the image recognition and target detection of deep convolutional neural network technology have developed rapidly,and breakthroughs have also been made in face detection based on this technology.Compared with traditional hand-crafted feature techniques,the deep convolutional neural network can learn more robust features,which are used in different face detection tasks to achieve superior performance.Although deep convolutional neural network technology has been adopted for face detection with excellent performance,there are still challenges in different scenarios.On the one hand,faces are easily affected by scale,pose,facial expression,occlusion,and illumination in natural environments,especially small faces have few available features,resulting in low detection accuracy;on the other hand,mobile devices have limited storage space and computing power,so the size and computing consumption of the network are limited.Faced with the above problems,this thesis designs three different face detection networks from the perspectives of label assignment and multiscale feature fusion,and applies them to the scene tasks of detecting tiny faces in dense crowds,detecting faces at different scales in a complex environment and detecting faces in real time.The research contents are as follows:(1)A high performance face detection network for tiny faces has been constructed.The designed model solves the problem of low accuracy face detection from two perspectives.Based on insufficient positive samples for tiny faces,an improved anchor compensation strategy is introduced to allocate enough high-quality positive samples for tiny faces;based on enhancing the features for detecting tiny faces,a multi-scale contextual feature selection model is designed to obtain features with richer contextual information.(2)The face detector with effective training sample selection and multi-scale feature learning is proposed,which improves the detection performance of faces.The core idea is to automatically select positive and negative samples through an improved training sample selection strategy,so that faces of different scales can be trained.Secondly,a residual feature pyramid fusion module is introduced to generate more powerful and robust multi-scale facial features.(3)A lightweight anchor-free face detection model has been proposed.Based on a singlestage lightweight anchor-free object detection framework,an improved dynamic soft label assigner has been designed,which can ensure that enough high-quality samples are assigned to faces of different scales before training and the optimal label assignment results are obtained through network learning.In addition,an improved assign guidance module is introduced to obtain features more suitable for classification and regression tasks,which can further improve the detection accuracy of the model without affecting the inference speed.The research conducts experimental analysis and comparison of two commonly used data sets for face detection tasks,WiderFace and FDDB,it is proved that the proposed method is effective.