Research On Face Detection Technology Based On Light-weight Deep Neural Network

With the rapid development of deep learning and its wide application in the field of computer vision,the performance of face detection algorithms has also been greatly improved,but what comes with it is the huge amount of parameters and calculations brought by deep convolutional neural networks,making face detection algorithms must be attached to high-performance computing clusters for inference and application.Therefore,it is important to propose a light-weight face detection algorithm that can perform real-time inference on mobile devices.Which can significantly expand the application scenario of face detection and has important research and application value.This article firstly expounds the background and significance of light-weight face detection algorithms,and conducts a detailed investigation of the current state of face detection technology research at home and abroad.For the face detection algorithm based on light-weight deep neural network that can perform real-time inference on mobile devices,this paper does the following:1)Combine a light-weight deep convolutional neural network with a single-stage face detection algorithm to design a light-weight face detection algorithm based on a personalized Anchor.Due to the criticality of the prior box for face detection tasks,this paper designs Anchor based on K-means clustering and mathematical statistics of actual face detection application scenarios to make the predicted face boundingbox more realistic and improve the accuracy of face detection models.In addition,this paper uses transfer learning theory to use CelebA dataset as pre-training data before face detection algorithm training,to provide a more suitable pre-trained model for light-weight face detection algorithms,making it better and faster convergence during training.2)The light-weight face detection algorithm based on personalized Anchor is further optimized in terms of efficiency and accuracy.Firstly,two model compression methods are used to compress the light-weight face detection network.One is to propose a channel compression hyperparameter? to control the proportion of the entire convolutional neural network channel compression to balance the speed and accuracy of the algorithm.The second is to INT8 quantize the model,which make the model take less memory and speed up while inference.Secondly,the feature extraction convolutional neural network of the light-weight face detection algorithm is improved by using a multiscale feature fusion method,which combines the shallow features with the deep features,so that the output features have both high-level semantics and higher resolution.Which enriches the extracted features,expands the scale of detectable faces,and improves the accuracy of light-weight face detection algorithms.The light-weight face detection algorithm proposed in this paper achieved good results on the FDDB face detection public data set,with an F-measure of 88.2%and a face detection speed of 11.1 FPS on the CPU while the model size is only 5M,Proved the advancedness and effectiveness of the algorithm.