| Nowadays,there has been a lot of work related to facial technology,and there are many related implementations.The implementation methods of the implementation include relying on the rich computing resources of the Internet cloud platform,using traditional machine learning methods to reduce the amount of calculation,and Rely on high-end embedded platforms with GPUs,but these implementation methods are difficult to promote as a solution that takes into account real-time,security,complex environment adaptability,low cost and low power consumption.In view of the above problems,this project intends to design a lightweight face detection,face key point detection and face verification algorithm,self-made data set,and randomly use a variety of techniques to enhance the data set to improve the model training process,According to the characteristics of the embedded platform,comprehensively consider the amount of model calculation,parameter amount,model performance and model operator types,and only use pooling,relu activation function,1*1 convolution,3*3convolution or fully connected layer and control The number of convolution kernel channels and model layers balances performance and low computing power requirements.The final trained face detection model is only 109 KB before optimization,the face key point detection model is only 471 KB before optimization,and the face verification model is about 3344 KB before optimization.Face detection can detect multiple faces of different sizes and poses in unconstrained scenes at the same time and give their specific positions.At the same time,it can label the left pupil,right pupil,nose tip,left mouth corner and right mouth corner of the face,and realize simple Closed face verification function.Later,the face detection,face key point detection and face verification models are deployed on low-cost,low-memory and low-computing embedded platforms,and the test results are good,suitable for promotion as a low-cost and low-power solution. |
PDF Full Text Request