Research On Face Detection In Complex Scenes Based On YOLOv3

Face detection has always been one of the important research directions in the field of computer vision.It mainly uses machine learning and image processing techniques to extract facial features and build effective classifiers to obtain face information.In recent years,with the continuous advancement of face detection technology,its application industries have gradually expanded,such as banks,large commercial centers,railway traffic safety and other fields.However,the most commonly used algorithm is the traditional face detection algorithm based on VJ(Viola-Jones),which is mainly aimed at face images with strong constraints.This type of face image has a simple background and a single scene.At the same time,the face size is similar and most of the faces are frontal faces.But,in actual application scenarios,the constraint conditions are weak,and the detection accuracy will be affected by many factors,such as scale changes,lighting conditions and face angle changes,occlusion,blurring,etc.,this makes the accuracy of traditional face detection algorithms drop sharply.It has been unable to meet the needs of the actual situation.With the development of deep learning,the current research on face detection algorithms has gradually turned to deep learning.With the powerful feature extraction function of neural networks,certain breakthroughs have been made in face detection in complex scenes.This paper aims to improve the effect of face detection in complex scenes,and studies the excellent object detection algorithm YOLOv3.Through the improvement of the network and the optimization of parameters,it can fully extract the facial features in complex scenes and improving the detection accuracy.For clarity,the main contributions of this work can be summarized as three fold:(1)The YOLOv3 algorithm is applied to face detection in a complex background.According to the clustering results of the complex face data set WIDER FACE,adjust the number and dimensions of a priori frames,optimize anchor point parameters,and improve the effectiveness of the YOLOv3 network for face detection,and adjust the confidence threshold to find the appropriate value in complex scenarios.Experiments show that the performance of the optimized algorithm has been greatly improved;compared with VJ,DMP and other algorithms,its performance has been significantly improved,and the average accuracy is more than 10% higher than that of ACF,Tow-stage CNN and Multiscale Cascade CNN.(2)In order to further improve the sensitivity of the detector model to small-scale facial features,the FPN structure adopted in YOLOv3 is improved.The improved FPN further uses the high-level semantic features of different scales in the backbone structure to construct new feature maps.Experiments show that the improvement of the FPN network structure can improved the model's small-scale face detection accuracy in complex scenes.(3)For the backbone network of YOLOv3,when convolution operation and feature extraction are performed on the input image,there will be a problem that the intermediate residual module has fewer channels,which limits the learning ability and the feature information cannot be better circulated in each layer.This paper introduces the improved residual neural module(Improved Residual Networks,IRes Net)module to rebuild the spatial structure of the entire model network.Experiments show that the fusion of Darknet-53 and IRes Net can greatly improve the overall performance of the YOLOv3 network model;especially the improvement is very significant in the detection of small object.