Research And Implementation Of Face Recognition Algorithm Based On Davinci Processor

Convolutional Neural Network(CNN)is an artificial neural network that is widely used in the field of face recognition.Compared with other recognition algorithms,it only requires less image preprocessing operations,and the performance has reached the best level.This dissertation focuses on the face recognition algorithm based on the lightweight CNN model,in order to achieve the human's recognition level.In addition,the paper transplants the face recognition algorithm to the embedded system.Specifically speaking,the paper includes the following three aspects:(1)In the face recognition algorithm,the paper uses the residual network ResNet to design an 11-layer CNN network called ResNet11.The computational complexity of this network is far less than that of the current face recognition model.The dissertation trains the ResNet11 model with different numbers and resolutions of pictures,and evaluates their impacts on the algorithm's performance.In the selection of loss function,the paper tries softmax loss,center-loss and their combination,and finds the optimal combination ratio.With a 96*96 input resolution,the amount of multiply-accumulate operations of the trained ResNet11 model is approximately 210 M and this model achieves a recognition rate of 97.55% on the LFW database.(2)In the face detection algorithm,the paper trains a face detection model based on normalized pixel difference(NPD)feature and AdaBoost.We collect a large number of face and non-face samples,and conduct further training on the original model called NPD-frontal.Under the premise of ensuring the detection rate and detection time,the false detection rate of our new model is reduced to 12% of the original.In the same way,we test it on the FDDB database,the results also show that our new model has a lower false detection rate than NPD-frontal.(3)In order to implement the face recognition algorithm on the embedded platform,this paper uses the TMS320DM8168 multi-core processor as the implementation platform to build a video processing and display system.The system has good scalability and provides a flexible algorithm interface.Using this algorithm interface,the system integrates face recognition algorithm and face detection algorithm.At the same time,we transplant the face recognition algorithm to the AM5728 multi-core processor.Finally we evaluate the performance of our algorithms on the DM8168 and AM5728 multi-core processors.