| In recent years,with the widespread use of convolutional neural networks in computer vision tasks,it has led to the rapid development of artificial intelligence techniques,which includes the development of face detection tasks.Face detection is used in all aspects of our lives,such as security systems,face payment and humancomputer interaction,among others.At the same time,face detection serves as a critical step for face verification,face recognition,face alignment and expression analysis problems,making it one of the most important tasks in computer vision.Although there has been great progress in the task of face detection,and various excellent and advanced face detectors have emerged,and face detection has become more and more accurate,there are still many problems that need to be solved in face detection at this stage,for example,the detection of low-resolution faces has been a major challenge in the field of face detection.In addition,various portable embedded devices such as cell phones and surveillance devices need efficient face detection algorithms to support their necessary functions,and efficient and lightweight face detection algorithms need to be researched and developed.Therefore,in this paper,we propose two different algorithms to solve the challenges of low-resolution face detection and lightweight face detection algorithms.First,for the problem of low accuracy of face detection at Video Graphics Arra(VGA)resolution(640×480)at this stage,this paper selects the YOLOX algorithm as the main framework of the model in this paper and proposes a YOLOX-based lowresolution face detection algorithm.In this paper,by making corresponding improvements to the Backbone,Neck and Head parts of the model respectively,the corresponding network structure is proposed to make the model pay more attention to the features of small faces by fusing the attention mechanism to different parts of the model,and then improve the efficiency of small face detection.The proposed improved algorithm is tested at VGA resolution and achieves good results on WIDER FACE.Secondly,in response to the high demand for lightweight face detection algorithms in various portable embedded devices at this stage,this paper proposes a lightweight face detector based on RetinaFace,by introducing five face keypoints as additional supervised signals to provide model learning,adding deformable convolution(DCN)to Backbone for use,and proposing a deformable convolution-based The design uses Recurrent Convolutional Feature Fusion Network(RECFPN)and context module to obtain more semantic information and improve feature representation,which in turn improves the efficiency of the model in detecting obscured faces and difficult faces;Focal Loss is used to make the distribution of positive and negative samples more balanced during model training.The proposed algorithm is tested on WIDER FACE and good results are obtained. |
