The Design And Implementation Of An 8-bit CNN Face Detection And Tracking System With Online Learning

Face detection technology is currently widely used and is the basis of current face recognition and verification applications.Face detection technology based on convolutional neural networks has far surpassed traditional methods in accuracy,but the huge amount of calculations and parameters are a major disadvantage of convolutional neural networks,and it has also become an important obstacle to its deployment in actual scenes.In addition,most of the current convolutional neural networks use offline training,and their parameters will not change after deployment,so it is difficult to adjust adaptively to the actual deployment scenario.Therefore,designing a lightweight face detection algorithm that can achieve a good balance between detection accuracy and computational complexity,and can perform online learning for deployment scenarios,is of great significance for the deployment of face detection algorithms for terminal devices..Firstly,the development process and typical methods of face detection and target detection algorithm,network quantization algorithm,target tracking algorithm and online learning algorithm have been introduced.In the the aspect of face detection algorithm,the current mainstream lightweight convolution operator is analyzed,and in order to ensure the accuracy of the algorithm,the feature pyramid of the network is improved,and the pixel shuffling method is adopted without increasing the amount of calculation.The feature map is enlarged and the channel is reduced to ensure that the features extracted by the network can be fully utilized,and a face detection network that can achieve a good balance between accuracy and calculation is designed.In terms of network quantization,in order to reduce the complexity of the algorithm,the parameter range is compressed to a certain extent during training.In terms of face tracking,the principle of current multi-target tracking algorithm has been analyzed and the face tracker according to the multi-target detection algorithm has been designed.In addition,an online learning strategy has been proposed,which replicates the prediction branches of the network,fixes the parameters of one branch and updates the parameters of the other branch online,so that the network has strong generalization ability and adaptability to deployment scenarios.Finally,the online learning face detection and tracking algorithm based on quantization network is verified on embedded system.The proposed face detection algorithm,achieves the detection accuracy of 92.3%,90.7% and78.2% respectively on the three test sets of wide face,and its model size is only 4.6MB.After quantization to 8bits,the accuracy is 91.6%,90.3%,and 77.1%,respectively,and the accuracy drops are all about 1%.After quantization,the model size is only 1.3MB.The proposed online learning method can improve the accuracy by about 0.1% after a single traversal on wide face.The designed algorithm can reach a running speed of 12 FPS on the Raspberry Pi 4B development board.The research has certain reference significance for the application of lightweight face detection algorithms on terminal devices,especially for online learning after the network is deployed.