Enhancement Of Face Detection Accuracy Based On Deep Learning

Face recognition process is divided into face detection,key point location and recognition.Face detection can detect faces and locate human faces through image;Key points are located through the process of locating features(eyes,mouths,etc.)of faces;Identify through the process of comparing the data collected by the device with the database.At present,there are several problems in face detection,such as variability,occlusion,uneven influence of illumination and smaller target objects that are difficult to be detected by machine.Variability of face is influential because of the color of skin,facial expression may change with age grows.The difference between the images collected and the images entered in the database is about partially occluded objects,which will effect on the subsequent training and testing results.The influence of illumination is serious,when the light is not uniform,the equipment cannot detect the face in the picture.There are smaller faces,fewer features of the pictures can be obtained,it is easy for computer to detect face with errors.In view of the fact that the target object is smaller,it leads to the reduction of accuracy.So a study was done in this paper.To a certain extent,it reduces the detection error and optimizes the face detection process.In this paper,we use the knowledge of convolution neural network in deep learning to compare the thought methods of target detection.Based on the framework idea of SSD paper,the framework is divided into two parts,the process of feature selection and the process of target detection,I make experiments on face data set WIDERFACE,try to solve the problem of face detector based on the anchors.As the target object becomes smaller,the general problem of detection accuracy is seriously reduced.The main research contents and results of this article are as follows:(1)The designed detection network can target different size target objects.Different layers of the detection network design different candidate frame sizes to ensure different target object detection.(2)Design the match strategy between the candidate box and the real box to improve the recall rate of the small face detection.The face detector of this paper based on the Titan X(Pascal)test platform has an accuracy of 0.97 on the published face detection benchmark data set FDDB.