Rearch On Multi-scale Face Mask Detection Algorithm Based On Improved Faster-RCNN

Since the mass outbreak of COVID-19,people’s health and lives have been seriously threatened.COVID-19 and its variants are highly infectious,and virus easily spreads in crowded public places.In order to prevent the epidemic from further spreading,it is crucial to wear a mask in public places.Now,the performance of detection model in crowded scenes remains to be improved,which is mainly because there are usually multi-scale and fuzzy targets with poor feature information in the test sample of mask-wearing detection samples in complex situations.Besides,if the image is dimly lit,existing models will not be able to properly identify objects in the sample.The accuracy rate,detection efficiency and other evaluation indexes of the original detection algorithm cannot meet the demand of actual task.Therefore,the use of a high-performance mask-wearing detection technology in public places is necessary for checking and reminding people to wear a mask correctly.On this basis,this thesis put forward a multi-scale mask-wearing detection algorithm MEFaster-RCNN based on improved Faster-RCNN.The main research contents are as follows:(1)First,on the basis of the normal illumination open source data sets,the open source data in various real scenarios are manually annotated.Secondly,for the detection of masks in the dark environment,this thesis recruited volunteers to collect and produce an infrared face mask detection data set,a total of 1050 pieces were collected.And through the image data expansion method,the sample set is expanded to 2460 images.(2)After analyzing the difficulties in mask-wearing detection task,this thesis improved traditional Faster-RCNN according to the size characteristics of training dataset in this task.The improvement included four aspects of using Res Net-50 as feature extraction network,putting forward the new multi-task enhanced RPN model,improved Softer-NMS algorithm,and using CIo U loss function.To be specific,in comprehensive consideration of network parameter quantity and model learning capability,the feature extraction network of this task model was replaced with Res Net-50 network with residual structure.The traditional single RPN network model was improved into multi-task enhanced RPN model to improve detection identification accuracy.RPN network topologies of different sizes were used to identify mask targets of different sizes which were then integrated into the network.Because the anchor box of current size could not adapt to the dataset sample well,this thesis redesigned a new anchor box.With Softer-NMS algorithm,it inhibited the redundant prediction box generated by RPN model,allowing the position of prediction box to be closer to ground truth box.This thesis also introduced CIo U loss function and optimized the positioning loss function and prediction category loss function of prediction box in mask detection task.(3)Aiming at the problem that the sample cannot be detected normally due to insufficient light in the dark environment,a method combining infrared technology and transfer learning is proposed.On the basis of analyzing the characteristics of infrared mask images,the infrared mask dataset was firstly expanded to achieve the purpose of enhancing data diversity.Then,a Res Net-50 network for transfer learning is constructed,and the trained new model is used to identify infrared face mask samples in a dark environment to be detected.(4)The experimental models in the two environments are evaluated separately.This thesis firstly compares the improved algorithm with Faster-RCNN algorithm and other algorithms.Then,in order to verify the feasibility of the infrared mask wearing image detection method based on transfer learning,the experimental results were analyzed and compared in three scenarios: the original image in the dark environment,the infrared shooting image and the transfer learning detection image.In addition,the ablation experiment test of the improved algorithm was completed.Finally,with Tensorflow development tool,it designed the interactive interface of mask-wearing detection experiment platform and realized visualized operation of mask target detection.