Algorithm Of Face Recognition Based On Deep Learning With Its Implementation On Embedded Platform

With the development of face recognition technology,face recognition is gradually moving towards commercialization and dailyization,bringing a lot of convenience to people's lives,such as person verification on train and online payment certification.Meanwhile face recognition is widely applied in security which makes great contribution.In order to accelerate the popularization of face recognition technology,the application on embedded devices becomes a hot issue.Therefore,this thesis deeply researches the face recognition technology based on embedded platform,and the main contents are as follows:1.Face detection algorithm.For real-time face detection on embedded platform,a more lightweight and fast face detection algorithm based on Retina Face is designed.The face algorithm performs the face classification and bounding box regression synchronously,and the overlapped bounding box is merged by non-maximum suppression.To achieve the design requirements,FPN is replaced with context enhanced module,moreover,depthwise separable convolution and channel reduction are employed.Finally,the algorithm can achieve faster speed with a small loss of accuracy.2.Face alignment algorithm.Current face alignment algorithm is generally based on landmarks,while landmark detection can be disturbed by emotion,pose and so on.To avoid the interferences,a fast face alignment algorithm which doesn't need to detect landmarks is designed,and the algorithm is based on spatial transform network(STN)that fits the affine transformation.Some improvements are employed in STN,for example,the 5×5 convolution is replaced with two layers of3×3 convolution,and two fully connected layers are replaced with global average pooling and 1×1convolution.Finally,the algorithm can get better alignment results and faster speed.3.Face recognition algorithm.For real-time face recognition on embedded platform,a face feature extraction algorithm is designed based on Mobile Face Net and Arc Face.In order to be more applicable on embedded platform,some methods are used to make Mobile Face Net more lightweigh.To make the feature extracted more expressive,attention model is employed in Mobile Face Net,which introduces fewer parameters and FLOPs.Arc Face is used for network training,after that,feature is extracted by the improved network.Then,cosine similarity is applied to do feature matching.Finally,the algorithm can achieve faster speed with a small loss of accuracy.4.Face recognition software based on embedded platform.Based on the entire face recognition algorithm,a face recognition software is developed,which runs on Jetson TX2.The software can realize face recognition in real time,and also it can display the real-time picture and recognition results on the interface.