Lightweight And Multi-pose Face Recognition Method Based On Deep Learning

With the rapid development of deep learning,more and more technologies in the field of computer vision are being applied to people's daily lives.As a branch of computer vision,face recognition,while deep learning is developing rapidly,face recognition technology based on deep learning methods has also achieved great achievements.Most of the face images in the existing public datasets are single pose(front face)faces.In actual applications,the face images that need to be recognized may be multi-pose.A single pose dataset cannot achieve good results in multi-pose recognition tasks.In order to solve this problem,800 volunteers were selected to build a multi-pose dataset.The dataset contains a total of 4000 images of 800 people,each person collects 5 poses(front face,left face,right face,head down,head up),the scene of each person at the time of collection is different,and all people are in different age groups.The distance of the collected face image relative to the camera is also different.The face detection algorithm used in this paper is the Multi-task Cascaded Convolutional Networks(MTCNN)algorithm,and its network structure is composed of three cascaded convolutional neural networks PNet,RNet,and ONet.The size of the input image is larger,the number of layers of the image pyramid is higher,resulting in more time-consuming PNet.In this paper,the number of image pyramid layers is fixed by dynamically adjusting the minimum face parameters.Regardless of the size of the input image,the image pyramid is fixed at 8.In addition,the more faces in the input image,the more time-consuming RNet and ONet.This paper uses the depthwise convolution and channel shuffle operations to replace the standard convolutional of the original MTCNN network to optimize the second problem.Due to insufficient computing power on the mobile terminal and limited computing resources,large-scale deep learning models cannot be successfully applied on it.In response to this problem,this paper proposes an efficient lightweight network.During training,in order to improve the contribution of the multi-pose dataset to the entire feature extraction network,the sample selection method was modified.During recognition,in order to improve the accuracy of face recognition in multi-pose tasks,the current face pose is judged based on the key points of the face,and then the current face features are compared with all the features of the corresponding pose in the face feature database to obtain Face matching results.The experimental results show that the lightweight network proposed in this paper has an accuracy of 99.34% on the LFW dataset,and a complete set of face recognition processes on the mobile phone with Qualcomm Snapdragon 820 processor has a derivation time of 49 ms,which is close to real-time.At the same time,the multi-pose data set and multi-pose recognition method proposed in this paper have also improved the recognition accuracy.