Research On Efficient Facial Landmark Detection Algorithm For Mobile Devices

Facial landmark detection is the foundation and the important part of face recognition,which is widely used in intelligent unlocking,face payment,3D reconstruction and other fields.With the arrival of the 4G era,the mobile Internet has been greatly developed,so the facial landmark detection technology for mobile devices has a broad development space.Compared with the traditional facial landmark detection algorithm that relies on handcraft features,the facial landmark detection algorithm based on deep learning can extract stable and deep features with scale invariance by itself,so the detection accuracy is higher and the performance is more superior.However,the facial landmark detection algorithm based on deep learning usually has the characteristics of large amount of calculation and large memory space,which is difficult to apply to mobile devices with relatively scarce storage space and computing resources.In this background,this paper proposes a light weight and efficient facial landmark detection algorithm for mobile devices.The detection algorithm has a three-layer cascaded convolutional neural network structure,which can detect facial landmark on the face in the image from the whole to the part,and improve the detection speed while maintaining detection accuracy.Based on the proposed algorithm,singular value decomposition is used for further optimization to achieve real-time detection.The main contents of this article are as follows:(1)In this paper,the research status of facial landmark detection algorithm at home and abroad is deeply studied,and several typical methods of facial landmark detection are analyzed and compared.It is found that facial landmark detection algorithms based on deep learning methods have more superior performance.Therefore,it is difficult to apply such algorithm on mobile devices,and the research direction of this article is determined.(2)The theoretical basis of convolutional neural networks is deeply studied.Firstly,the convolutional layer,pooling layer,activation layer,then fully connected layer and other structures of the convolutional neural network are studied in depth,and the principles and characteristics of the related structures are analyzed.Finally,several typical convolutional neural network structures are briefly described.(3)Through the previous analysis,two different types of cascaded convolutional neural networks and lightweight convolutional neural networks for mobile devices are studied.Based on the above research,an efficient facial landmark detection algorithm for mobile devices is proposed.and the algorithm is further optimized by using singular value decomposition.This algorithm can detect the facial landmark from rough to fine by multilayer convolution neural network,and be compared with other algorithms.The experimental results show that the proposed algorithm can trade-off the accuracy and speed of detection,and reflects certain performance advantages.After further optimization,it achieves the effect of real-time detection and can be applied to mobile devices.