Research On Small Face Detection Algorithm Based On Improved CenterNet

In recent years,face detection algorithms based on deep learning have developed rapidly and their performance has been constantly breakthrough.But how to perform multi-scale face detection faster and better is still worth exploring.To solve the problem,this paper proposes a small face detection model Center Face based on improved Center Net.The main research works and achievements of this paper are as follows:(1)This paper reviews the research progress of the face detection algorithms based on deep learning in recent years,summarizes the innovations and shortcomings of the classic anchor based and anchor free algorithm,and focuses on exploring the schemes and progress of various algorithms used in multi-scale object detection.(2)we introduce the basic structure of object detection model and its related components,the convolutional neural network,backbone network,feature enhancement module and other object detection components.(3)based on the object detection algorithm named Center Net,a single stage face detection framework named Center Face is designed.The specific works are as follows:1)In order to solve the multi-scale face detection,the feature pyramid network and receptive field enhancement module are introduced.The feature pyramid fuses the features with semantic information from different sizes of object.Before feature fusion,the receptive field enhancement module carries out more detailed feature enhancement on the output features of each layer.The fusion of these two modules brings a multi-scale receptive field range for the feature,which greatly enhances average precision of the multi-scale face detection,especially small faces.2)To solve the independent optimization of the regression offset branch and the width/height branch of the prediction box,the DIo U loss function is used to jointly monitor the two branches and optimize the regression task performance.(4)On the basis of the above framework,a compromise solution between model complexity and performance is explored.Based on the lightweight network of Mobile Net,we design a detection neck based on the reconstructed receptive field is designed to obtain enhanced output features while ensuring low calculation of the model.(5)The model and solution proposed in this paper are proved to be effective for the multi-scale face detection by ablation experiment.The performance of Model 1 in the three subsets of Wider Face reaches 94.84%,94.17% and 89.43%.Model 2 achieves the detection speed of 50 frames per second,and at the same time,the average precision in the three subsets of Widerface reaches 93.79%,92.67% and 87.64%.The performances of two models are better than that of most classic algorithms.We can say that we have achieved a better and faster multi-scale face detection model.