Circuit Design Of Ficial Landmark Detection Based On Convolution Neural Network

As the basis of face recognition and analysis,facial landmark detection has received more and more attention.Compared with traditional algorithms,convolutional neural networks can often achieve higher precision.However,the networks are often limited to computing platforms such as GPUs because of intensive parameters and large amount of computation.As the increasing demand for real-time embedded facial landmark detection,it is necessary to optimize the existing facial landmark detection method to improve the accuracy and speed.A facial landmark detection algorithm based on multi-task learning is designed in this thesis.24 auxiliary tasks are optimized together with facial landmark detection.The efficiency of model is greatly improved by unified kernel size and stride-2 convolution.The accuracy of the face feature points can be improved by early stopping while may lead to over-fitting.This problem can be solved by reducing the learning rate and only fine-tuning the high-level parameters.In addition,the convolutional calculation circuit with configurable feature map size is designed in this thesis,including address generator,2D convolutionn unit,cache unit and other key modules,which improve the computational efficiency.The synthesis and simulation were completed under SMIC's 40 nm process.Finally,the facial landmark detection system was tested on FPGA at 100 MHz.The test shows that the mean error is 7.4% in AFLW,which is 0.6% lower than TCDCN.Compared with MTCNN,the model scale is 2/3 smaller.The accuracy is hardly reduced after 16-bit fixed point processing of the facial landmark detection system based on Miz702 N FPGA,and the speed reaches 26 fps.It provides a reference scheme for the application of facial landmark detection in various mobile terminals such as intelligent robots and mobile phones.