Research On Face Detection In Large Scenes

Nowadays,face recognition is regarded as an important means of identity verification,personnel search and traffic statistics in many fields.Cameras are installed in many scenes such as stations,airports,ticket windows and bustling streets to monitor and recognize faces.In order to realize intelligent recognition of faces,it is necessary to detect faces from video images with many faces before further face recognition.Therefore,face detection is the prerequisite for face recognition.At present,there are many face detection methods,which can be roughly divided into traditional face detection methods and face detection algorithms based on deep learning.Due to the large number of people and different positions in large scenes,the face posture and expression are not controlled,and there are also differences in environmental conditions such as light,so choosing an accurate and efficient face detection algorithm faces challenges.This paper mainly studies the related technology of face detection and uses the students' classroom scenes as experimental scenes.Firstly,the face detection methods based on skin color and Adaboost are introduced.Secondly,the MTCNN(Multi-task Cascaded Convolutional Networks)algorithm,YOLO algorithm and Faster RCNN algorithm based on deep learning are analyzed and studied.By comparing different algorithms,the MTCNN,YOLO,and Adaboost algorithms,which have relatively good results,are finally applied to the face detection of students in the classroom,and the Adaboost,YOLO,and MTCNN algorithms with significant effects are improved respectively.For the Adaboost algorithm,this article employs a new weight update strategy,and reduces the number of features to be extracted.At the same time,the skin color region is pre-extracted for the image to be detected.The improved algorithm effectively shortens the detection time and improves the accuracy.For the YOLO algorithm,it is mainly improved for YOLOv2.It adjusts the network structure,increases the fusion of low-level and high-level feature maps,and uses a new loss function to improve the detection rate of small-sized faces.For the MTCNN algorithm,this paper uses a new activation function and non-maximum suppression method,adding a convolution layer and a global average pooling layer,which improves the detection effect of occluded faces and accelerates training.The above improvements effectively improve theperformance of each algorithm.On the training samples,this article puts real classroom student pictures into the training sample set for training.The Adaboost method,YOLO algorithm and MTCNN algorithm before and after the improvement were respectively used in the face detection experiment of students in classroom scene and graduation photo scene,and the experimental results were compared and analyzed.Finally,the face timing detection strategy of students was proposed.The experimental test results show that the face detection method based on MTCNN achieves best effect in the detection accuracy,and can be well applied to face detection in college classrooms,providing important reference value for the selection of face detection schemes in other large scenes.