Research On Robust Face Detection Based On YOLOv3 In Complex Scenes

In recent years,face detection has been widely used in various aspects such as campus access control,facial payment,video surveillance,and target tracking.At the same time,people are positioning and locating feature points based on faces in terms of real-time,security,and reliability.Related applications such as detection and recognition put forward higher requirements.In addition,face detection often fails to achieve the expected results in complex scenes,such as dense faces,different scales,high similarity to the background,high occlusion and so on.Therefore,technical challenges in face detection still exist,and robust face detection in complex scenes is still a hot research topic at home and abroad.In order to improve the applicability of face detection in complex scenes,aiming at the problems and challenges encountered in the detection,a robust face detection model based on YOLOv3 is designed and implemented.The specific work is as follows:(1)In complex scenes,face objects and image backgrounds are prone to misdetection when their shapes,colors are similar.Therefore,A face detection model based on attention mechanism,YOLOv3-attention,is proposed.Firstly,the attention mechanism based on the mixed domain is used to obtain the information of the face region that needs to be focused in the image by generating the feature weight,and to suppress the information of the irrelevant background region.In addition,the anchor used in the network is also an important part.In this paper,the size of the real face frame of the k-means++ algorithm is used for cluster analysis,and the anchor box with a smaller size is set to capture the target of small face.Experiments on the WIDERFACE dataset verify that the YOLOv3-attention algorithm has a significantly higher recall rate on human faces than the YOLOv3 algorithm,and the network detection accuracy is further improved.(2)The scales of faces in the images vary and the most model can obtain high detection accuracy on large and medium-sized faces,but the detection effect for small faces is not good.In order to solve the problem of missing detection of small faces in the scene,a detection model SR-YOLOv3-attention is designed to improve the detection performance of low-resolution and small faces.The Dark Net53 is used as the backbone network.Firstly,the mixed domain attention module is used to process the output features of blurry and small faces,and then the image super-resolution reconstruction module is integrated to enhance the data of lowresolution faces.In this way,the method can make up for the high-level semantic information that is not in the shallow features,and enhance the target texture information lost in the deep features.Tested on the WIDERFACE data set,the detection accuracy of SR-YOLOv3-attention on the three test subsets of easy,medium,and hard are 0.946,0.937,and 0.872,respectively.It is compared with other face detection algorithms MTCNN,CMS-RCNN,and HR.Compared with S3 FD,the measurement accuracy on the test subset Hard has been increased by 0.243,0.229,0.053,0.013,respectively.The experimental results verify that SR-YOLOv3-attention can make good use of face information,effectively detect difficult-to-detect faces in the image,and has good robustness.