Research On Face Detection Algorithm Based On Lightweight Convolutional Neural Network

With the development of computer and network technology recent years,how to detect faces accurately and rapidly from massive video or image data has become an important research topic.Traditional face detection algorithms mainly use hand-crafted features for detection.Due to the inadequate ability of hand-crafted features to represent the face in complex scenes,the performance of traditional algorithms under these circumstances will be greatly reduced.Related studies show that the convolutional neural network has the ability to extract high-level visual features from the image,so more and more researchers apply the convolutional neural network to face detection field.In addition,the models based on the convolutional neural network used in the face detection task are large at this stage and are not fully suitable for the face detection task in mobile or embedded devicesTo deal with the problem of face detection in complex scenes,lack of training samples with labels,and the large size of existing models,we changed the composition of the network based on convolutional neural networks,and designed a face detection algorithm based on a lightweight convolution neural network with depthwise separable convolution in this paper And a multi-scale fusion method of facial features for training was proposed to improve the robustness of the algorithm.In addition,the attention module is applied to face detection tasks,and a face detection algorithm based on a lightweight convolutional neural network with the attention module is proposed to help the model locate the face quickly.The main work of this paper is as follows:(1)Change of the composition of the network.The depthwise separable convolutional layer is introduced into the face detection algorithm,and a face detection algorithm based on a lightweight convolutional neural network with depthwise separable convolution is proposed to be suitable for face detection tasks in mobile or embedded devices(2)Optimization of the feature extraction method.This paper proposed to fuse facial features of different scales and different levels to form multi-scale fusion facial features containing facial information of different scales and different levels to improve the robustness of the algorithm.And this paper proposed to construct a feature pyramid network to make full use of multi-scale fusion facial features(3)Solution of the problem of lack of training samples with labels.In the face detection task,the training samples of the face data set are not enough to train a large model.Applying transfer learning to face detection tasks,this paper proposed to use a large-scale data set that is similar to the face data set to pre-train the network(4)Optimization of the network structure.Applying the attention module to the face detection task,this paper proposed a face detection algorithm based on a lightweight convolutional neural network with the attention module.The attention modules are added after each stage of the network to help the model locate the face quickly and improve the accuracy of the face detection result.