Research On Multi-scale Face Detection Algorithm Based On Deep Learning

Face detection technology is an important research direction in computer vision with extensive research prospects that has been widly used in the field of face recognition,face key point detection,and expression recognition technology.However,in reality,it is still a chanllenging task due to the complicated shooting environment.The current mainstream face detection algorithms mainly draw on target detection algorithms,including two-stage detection algorithms represented by the RCNN series of algorithms and single-stage detection algorithms represented by the SSD algorithm.In contrast,two-stage detection algorithm has higher detection accuracy but lower detection speed,which can not meet the actual requirements.The detection accuracy of the single-stage algorithm is lower than two-stage detection algorithm,but it has high real-time performance that has broad market prospects.Therefore,how to improve the performance of single-stage detection algorithms is the research focus of current face detection algorithms.To deal with these existing problems,this thesis proposes a series of improved algorithms based on the SSD algorithm.The SSD algorithm extracts the feature maps of different convolutional layers,directly performs classification and regression without using Context information.In order to solve this problem,this thesis proposes a multi-layer feature fusion algorithm.From the top to down,different feature fusion algorithms are adopted for different convolution layers,including high-level feature enhancement algorithms and low-level feature stitching algorithms.On the high-level feature map,parallel networks and hole convolution operations are used to obtain more abstract features and enhance multi-scale information.On the low-level feature map,proportional stitching is adopted to retain more low-level features.This algorithm was tested on the WiderFace official data set.The average precision rates of the three subsets: easy,medium,and hard is 0.930?0.919?0.834.Based on RFBNet and Inception networks,this thesis proposes a receptive field network.The convolution kernel decomposition method is used to increase the depth and width of the network at the same time,and multiple dilated convolution stacks are used to simulate the change of eccentricity in human vision.Aiming at the problems of small-scale face detection,a random dense sampling strategy for small-scale anchors is proposed,and different numbers of anchors are set according to different convolutional layers.This algorithm was tested on the WiderFace official data set.The average precision rates of the three subsets: easy,medium,and hard is 0.942?0.921?0.85.