Single Stage Face Detection Algorithm Based On Anchors Optimization

Face detection is the premise of many related applications,and its performance will directly affect the subsequent operation.At present,face detection algorithms have made great progress,but there are still omissions in the detection of large angle faces and blurred faces.At the same time,the detection speed needs to be further improved.Therefore,this paper proposes a single stage face detection algorithm based on anchors optimization.The main contents are as follows:Firstly,based on the SSD detection framework,we design a lightweight convolution neural network,which is mainly composed of standard convolution,depthwise separable convolution,Inception module and so on.The method of maxout background label is used to reduce the false positive rate and to reduce the imbalance between positive and negative samples to a certain extent.At the same time,the robustness of the algorithm to the size of face is improved by designing a large range of layers which are related to anchors.The experimental results show that the designed network structure is reasonable,and its detection performance is better than other basic network structures,and the number of model parameters is less.Then the algorithm is further optimized based on this network structure.It mainly includes the following four items:(1)We simulate human visual mechanism and add atrous convolution into Inception module.The atrous convolution with different expansion rates corresponds to different centrifugal rates,which makes the position closer to the center have higher weight,and then enhances the semantic information of feature map.(2)We design feature hierarchy structure.Firstly,we return the high-level semantic information from the bottom-up network structure;Then we combine it with the top-down feature map to enhance the semantic information and location information of the feature map.So that all feature maps with different sizes in the feature pyramid have rich semantic information,thereby improving the detection performance of detectors.(3)Based on the feature hierarchy structure,the anchors cascade optimization method is used to initialize the anchors of the next level with the anchors whose position is refined.At the same time,different IOU thresholds are used for anchors matching at different levels to gradually improve the quality of anchors,thereby improving the detection performance of detectors.(4)Negative sample screening mechanism is used to screen out the negative samples which are easy to classify.To some extent,the imbalance problem caused by the negative samples which are easy to classify is alleviated.Finally,the algorithm is validated in several face databases.The experimental results show that the proposed face detection algorithm is robust to small size faces,large angle faces,blurred faces and so on and has good detection speed.On the FDDB database,when the number of false positives is 2000,the recall rate is 96.50%,and the average accuracy rates on AFW database and PASCAL FACE database are respectively 99.01%and 97.8%.The test speed on Tesla K80 and cuDNN v5.1.10 is 14 FPS.