| With the rapid development of computer hardware level and deep learning,computer vision has achieved rapid development.Face detection is an important branch in the field of computer vision.Face detection systems based on deep learning have been widely used in real life.In some special environments,the accuracy of face detection systems even exceeds that of the human eye.2019 outbreak of novel coronavirus pneumonia COVID-19,individuals have been wearing masks to anticipate the spread of the infection beneath social distancing.In the case of wearing a mask correctly,the facial features from the nose to the chin will be blocked by the mask,which will reduce the accuracy of the face detection system.In daily life,some decorations of people will also block some features of the human face.For example,wearing sunglasses will block the features of the human eyes;wearing a hat will have a certain impact on the facial features due to the occlusion of light.Therefore,how to design a face detection system suitable for occlusion scenes is an urgent problem to be solved.When the facial features change due to factors such as illumination and occlusion,the current general object detection model based on deep learning is not ideal in terms of detection accuracy and speed in the field of face detection.Therefore,based on the general object detection model,conducts the following research on the problem of accuracy degradation caused by occlusion:(1)Since occluded face detection needs to pay attention to the category information of the object and the spatial position information of the object at the same time,and a large amount of facial feature information will be destroyed when the face is occluded by the mask,so on the basis of the YOLOv4 model,a new method based on The face detection model that combines the attention mechanism introduces the DCA(Dilated Convolutional Attention)attention mechanism module that takes into account both channel attention and spatial attention,and uses a variety of data enhancement methods to expand the data to prevent the model from overfitting.Improve the model’s ability to handle occluded faces.The experimental results show that compared with the original model,the YOLO-DCA model proposed improves m AP by 4.66%.(2)In order to solve the problem of performance degradation due to object occlusion and complex background in face detection,and the difficulty of deploying low-performance devices due to the large convolutional neural network model,self-deconvolution SD(SelfDe Convolution)is introduced on the basis of YOLOv4-Tiny.module,which proposes the YOLO-SD-Tiny network model.The SD module with larger receptive field and lighter weight is introduced into the feature pyramid network,Mish activation function is used in some network layers,and the loss function adopts CIOU bounding box regression loss and GHM(Gradient Harmonizing Mechanism)classification loss.The experimental results show that: compared with the original YOLOv4-Tiny,the proposed YOLO-SD-Tiny has a slight reduction in the number of parameters for face detection in occlusion scenes,an increase of 4.89% in accuracy,and an increase in detection speed.9.64%. |
