Research On One-Stage Complex Face Detection Methods

In recent years,face detection has achieved great progress with the development of deep learning.However,the high accuracy of existing methods still highly depends on complex models and computing resources due to complex faces such as low resolution,severe occlusion,and background interference,and hence it is limited when applied into real-lift scenarios.So,singlestage and lightweight complex face detection methods have drawn much attention.In this dissertation,based on single-stage detection frameworks,face detection methods have been proposed by introducing some tricks including attention mechanism,lightweight model design and others to improve the detection performance of complex faces.The proposed methods have been evaluated on the public datasets,achieving good balance of detection accuracy,reasoning rate and computing resources.In addition,face detection software with related algorithms is designed.The main contents are as follows:(1)An improved YOLOv3-attention model has been proposed by introducing the attention mechanism and data augmentation to improve the feature perception ability of the deep network,and the model's ability to learn complex facial features,respectively.The improved model achieves the accuracy of 0.942,0.920 and 0.821 with the speed of 28 FPS on the Wider Face dataset,respectively,and also obtains the highest accuracy improvement(5.3%)on the complex subset,outperforming some advanced algorithms.In addition,an accuracy of 0.965 has been achieved on the FDDB dataset.It shows that the proposed model is an effective model with high accuracy and efficiency in complex face detection.(2)An improved lightweight complex face detection model based on Retina Face has been proposed by using MobileNetV2 as backbone,by introducing lightweight attention mechanism and deformable convolution to improve the ability of learning effective features and multi-scale information.The pruning optimization in the loss function is employed to improve the detection efficiency,and Soft-NMS replace the non-maximum suppression in the test phase to improve the accuracy.The improved model achieves the detection accuracy of 0.942,0.926,and 0.832 with the speed of 149 FPS on Wider Face datasets,where an improvement of 8.7% on hard subset is achieved.Besides,0.979 detection accuracy on the FDDB dataset is obtained.The proposed method shows a good balance between the accuracy and efficiency of complex face detection.(3)Based on the images of MAFA and Wider Face,we unify the data annotation format and add the mask wearing classification labels to construct the self-made dataset called Mix Face.Using MobileNetV2 as the backbone,and the joint pooling method is adopted to improve the feature utilization rate.The proposed mask wearing detector can detect the face and monitor whether wearing a mask at the same time.The model achieves the average accuracy of 80% and79.4% on the Mix Face and MAFA data sets,respectively,with an inference rate of 46 FPS for MAFA data.The results show that the proposed model has achieved high accuracy and real time detection in the detection of mask wearing conditions.