A Face Detection Algorithm Based On One-Stage And Multi-Fusion Network

While face plays a great role as one of the most important biometric identification features,as an initial part of associated applications,face detection has a massive impact on subsequent steps and final result.There are still many challenges in face detection,many false positives will be occurred when faces in under poor lighting conditions,with very low illumination or resolution,and so on.Some of methods cooperated with specific domains for a futher speed improvement to satisfying a need on real-time performance.In this paper,we proposed a face detection algorithm based on one-stage network and multiple aspects fusion,and the main contents of this paper are summarized as follows:We proposed a new method of one-stage convolutional neural network consisting of Simple Block Layers(SBL)and Dense Multi-Fusion Layers(DMFL).In SBL,there are only simple convolution layers and pooling layers.In order to obtain a light-weight network,we just compress SBL on the number of output channels.Experiments have shown that the improved SBL compressing the model effectively,with reducing the parameters multiplied and achieving the purpose of fast filtering the images.Considering more discriminative features,this paper designs a multi-resolution fusion,multiscale overlapping receptive filed fusion and anchor matching strategy in DMFL.Experiments have shown that features extracted from DMFL used to powerful on the face representation.To make a balance between hard and easy samples,we used a semi-focal loss as classification task loss.To obtain features about abnormal pose or deformable state of faces in a further learning,deformable convolution is used in deeper layers regarded as a sampling optimization.Experimental results show that semi-focal loss can reduce the weight loss of easy samples,and the network tends to be sensitive to difficult samples learning,sampling optimization concentrates on the sampling around the face on face adapting to geometric deformation as well.In the post-processing stage,an improved NMS method is propose.The candidate boxes are scored twice and then to remove the higher overlap rate associated with current box.When IoU is greater than a given threshold,a new score is produced by multiplied by the score attenuation factor instead of being set to zero directly.Experiments have shown that the improved NMS method is effective on accuracy in the testing process.Based on all above designs and improvements,the proposed face detection model from our one-stage and multi-fusion network is not only a light-weight but also a discriminative model which can get the strong discriminant features of complex facial in an unconstrained environment.Finally,experimental result demonstrates the proposed algorithm achieves an accuracy of 96.5%on FDDB dataset and 99.4%on AFW dataset.Also the average test speed of detection achieves 18fps using CPU.