Tiny Face Detection Based On Deep Learning

In unconstrained scenes,the traditional methods of face detection are implemented by matching features of facial contours,facial key points,skin color,etc.,which greatly affects the accuracy of face detection because the effective features of tiny faces are seriously lost.In order to improve the accuracy of tiny face detection in unconstrained scenes,in this thesis,we analyze the reasons for low recall and precision of tiny face detection and research on deep neural network algorithm for the detection of tiny face.The specific work and results are as follows:First of all,we summarize the traditional methods and deep learning methods in the field of face detection,and introduce the basic algorithms of deep learning.In order to solve the problem of the loss of effective information of tiny faces caused by the neural network pooling layers,we propose a feature fusion network of feature reuse of adjacent detection layers,which provides more features for face classification and coordinate regression,thus improving the recall of the tiny face detection.Secondly,in order to effectively detect faces of muti-scales and further improve the recall of the tiny face detection,this thesis analyzes the advantages and disadvantages of several classic object detection networks,and constructs a Feature Fusion Single Shot Scale-invariant Face Detector(FS~3FD)for tiny face detection.It proves the effectiveness of muti-scale pyramid network for face detection.In addition,in order to assist the detection of tiny faces,we also introduce a context module consisting of dilated convolutions with dilation rates of 1,2,4,and 8 in this network,so as to increase the receptive field of the feature map without reducing the image resolution.It proves the context information for tiny face has a significant effect on improving detection accuracy.Then,in response to the misdetection of tiny faces caused by the interference of background information,this thesis introduces an attention module based on mixed domain to obtain the key areas of interest,and suppress areas that are similar to the shape and color of tiny faces by generating feature weights in the face image.The experimental results show that the multi-scale network based on attention mechanism is effective in detecting tiny faces,and adding attention module to FS~3FD can further improve the detection accuracy of the network.Finally,the methods proposed in this thesis are verified on the WIDER FACE dataset.Compared with the mainstream face detection algorithms,it achieves a better average precision in all level faces,i.e.0.948(Easy),0.940(Medium)and 0.893(Hard)for validation set,which demonstrates the feasibility and effectiveness of these methods in the tiny face detection task.