Lightweight Face Mask Detection Combined With Data Augmentation

Face mask detection is a popular and important research topic nowadays,especially in the field of public health and safety applications.How to ensure the speed and accuracy of face mask detection is its difficulty.The existing deep learning-based face mask detection methods have two problems.(1)In low luminance,due to background information interference,feature information blurring leads to poor image detection,and the existing method does not consider the brightness of the situation using fixed network parameters image enhancement,will cause the enhanced image too dark overexposure problem,loss of feature information,enhance the background information interference,not only waste computer arithmetic power,but also increase the inference time,reducing the detection accuracy.(2)Current deep learning-based target detection methods are difficult to deploy on mobile platforms with small computing power,and their lightweight versions are deployed by directly reducing the number of backbone network channels and using lightweight convolution to reduce the number of parameters of the model.However,this approach tends to result in the key features of the target not being extracted accurately,and even the situation that small target face features are not extracted at all,and there are serious mis-detection and under-detection problems.To address the effect of low illumination in Problem 1,this paper proposes an image brightness adaptive enhancement network,which includes two modules: a lightweight classification network and an image brightness enhancement algorithm.In the first part,the classification network introduces the inverse residual structure,which is implemented by deep separable convolution and attention mechanism.It solves the training overfitting problem caused by conventional classification models that focus too much on the deep semantic information of images and ignore the shallow texture information of luminance features,and ensures a high accuracy of image classification while greatly reducing the parameters required by the network,with the number of network parameters being 1/28 of that of Res Net-101((Residual Network))and the accuracy improved by 6 percentage points.In the second part,automatic image brightness enhancement can be achieved without manual assignment and training,which is not only computationally small and fast but also better than the trained network model,and can effectively solve the problem of overdarkness and overexposure after image enhancement,reduce the interference of background information,and thus improve the target detection accuracy.To address the problem of wrong and missed detection of small target faces in the lightweight network in Problem 2,this paper comprehensively explores the mechanism of inadequate feature extraction and declining detection accuracy of the lightweight network through extensive experiments,and proposes a lightweight face-worn mask detection network combining key features of faces.It contains two modules: face key feature extraction and multi-scale mask wearing detection.In the first part,according to the face key feature extraction module can extract the face key features according to the improved multi-tasking volume and network,and input the face position information to the multi-scale detection module to improve the recognition rate of small target faces;in the second part,for the problem of inadequate feature extraction of small target faces in the lightweight network,this paper makes innovative improvements in several aspects,such as backbone network,feature fusion,bounding box loss function In the second part,to address the problem of inadequate extraction of small target face features in the lightweight network,innovative improvements are made in the aspects of backbone network,feature fusion and bounding box loss function.Specifically,the backbone network proposes a lightweight multi-branch stacking structure and parallel downsampling structure to extract features of different scales,which can reduce the loss of small target feature extraction while parameter lightweighting,and can solve the problem of redundant feature parameters extracted by conventional convolution while lightweight convolution cannot effectively extract key features;the feature enhancement part performs feature fusion at different levels of the network based on the input feature information of different scales in a cross-connected manner.The feature enhancement part is based on the fusion of features at different levels of the network in a cross-connected manner,which can combine multiple sensory fields of the network and enhance the robustness of the network for large and small target feature extraction;the boundary frame loss loss part,in response to the problem of gradient disappearance and unbalanced samples of training instances when using IOU(Intersection over Union),this paper adopts Focal-EIOU(Focal-Effective Intersection).Focal-Effective Intersection over Union)explicitly measures the differences of three geometric factors of overlap area,centroid and edge length in the bounding box regression,so that the EIOU loss is concentrated on high quality samples and the advantages of convergence speed and localization accuracy are obtained.In this paper,we conduct extensive tests and experiments on the above methods and models on datasets in Py Torch deep learning framework around the above problem.The results show that the proposed luminance classification model is able to capture detailed information in the face of daytime shadow images and has a higher classification accuracy compared with existing schemes.At the same time,the multi-branch stacking and parallel downsampling structure proposed in this paper as the backbone network can effectively solve the serious problem of accuracy loss of lightweight small targets,and it can achieve the requirement of real-time detection.The covariance of the network in this paper is 1/6 of YOLOv7,and the difference of m AP(mean average precision)is only 0.3%,which is nearly better than its lightweight version YOLOv7-tiny’s high value of accuracy is improved by nearly 5%.All of the above work outperforms existing schemes,and related papers are currently in submission.