Research On Face Detection And Alignment Algorithm For Embedded System Implementation

Face detection and alignment are key steps in face recognition.Face detection is the premise of face alignment,and its duty is to find all the face in the image and draw them out,in order to show where they are.Based on the detected face,face alignment can locate facial landmarks such as eyes?nose?mouth and so on.We can extract the features of different facial landmarks which contains lots of information,thus face alignment is widely used in face recognition,facial expression recognition,virtual reality,and augmented reality.With the widely use of embedded devices such as mobile phones,the efficiency of face detection and alignment algorithms has become a hot topic in embedded devices.This paper proposes a face detection and alignment algorithm for embedded devices.Currently widely used face detection algorithms are based on deep learning methods including YOLO?SSD and so on,and the alignment algorithms including SDM,FPS3000,and others.Due to the lack of computing power in embedded devices and limited hardware resources,the detection based on deep learning method cannot run at real-time and the corresponding model is big.Besides,the regression alignment algorithm based on one feature,can lead to the problem of insufficient detection accuracy.Based on the above problems,this paper proposes different optimization and improvement methods,and to realizes real-time face detection and alignment algorithm on embedded systems.In this paper,the main research work is as follows:(1)Face detection: Focusing on how to optimize the detection speed and model size,aiming at the problem of slow detection speed and large model in detection neural network in the YOLO algorithm.From the perspectives of network structure optimization and network design,different neural network structures are proposed.Among them,the face detection network based on SqueezeNet design has 2.2% detection accuracy loss and 2.6% overlap rate.At the same time,the detection speed has increased by 41.2%,and the model size has been reduced by 29.5%.(2)Face alignment: FPS3000 alignment algorithm is very fast,but the alignment effect is poor.In order to improve the accuracy of alignment,this paper optimizes both the alignment features and the selection of feature points.The result of the random forest calculation will be regenerated by using the concatenated code as the new feature,and the selection of feature points dynamically changes the search range according to the number of regressions.Through the experimental comparison,using the optimized alignment algorithm,the alignment deviation is reduced from 17.9 to 14.3,and the model size is also reduced by 13.3%,which effectively improves the alignment accuracy.(3)Using the two methods discussed aboved to optimize the face detection and alignment algorithm,combining the hardware characteristics of the embedded platform TX1 and the Qt framework,multi-threaded software design is used to implement the embedded application of face detection and registration algorithm.At last the algorithm can achieve real-time detection on TX1 at the speed of about 25 fps.