Research On Mask Wearing Detection Algorithm Based On Deep Learning

In recent years,the New Coronavirus epidemic has been rampant around the world and has had a significant impact on people’s productive lives.New Coronavirus is highly contagious.Wearing masks in public places can cut off the transmission route of the virus and effectively stop the spread of the epidemic.Therefore,it is important to use computational vision technology to detect mask wearing behavior in public places.The mask-wearing detection algorithm faces complex scenes,dense and different scales of face data,while the practical application needs to complete the fast detection of face targets;secondly,the existing mask detection dataset lacks samples of mask-wearing errors and cannot meet the actual demand.To address the above problems,this thesis proposes two mask wearing detection algorithms based on deep learning target detection algorithms.The main research content of this thesis is as follows.:(1)A new mask wearing detection dataset was created for training and testing of mask wearing detection studies.The dataset includes samples of mask wearing errors and covers a variety of realistic social scenarios.The dataset was well annotated according to the requirements of practical applications.(2)A mask wearing detection algorithm based on feature fusion and coordinated attention mechanism is proposed.The algorithm is based on the SSD target detection algorithm,and the corresponding improvements are made by combining the actual situation of mask wearing detection.First,in the actual detection environment,there are differences in the size of face targets wearing masks,small feature differences between wearing masks and wearing errors,and complex and diverse detection scenarios.To address these problems,the feature extraction network of the algorithm is reconstructed by combining the feature fusion network and the coordinated attention mechanism;secondly,to address the problem of unbalanced positive and negative samples,the Quality Focal Loss loss function is introduced to adjust the positive and negative sample weights,and the classification scores and IOU scores of the algorithm are combined to represent them.The experiments on the created mask wearing detection dataset show that the algorithm improves the average accuracy mean by 5.62% compared to the original SSD,reaching 96.28%.It proves that the algorithm has excellent performance on the mask wearing detection task.(3)A mask wearing detection algorithm based on depth-separable convolution and improved bounding box loss is proposed.The mask wearing detection task requires correct and fast recognition of the target.Traditional full convolutional networks are computationally intensive and affect the detection speed of the model.Therefore,depth-separable convolution is used to reduce the computational effort of the algorithm and to improve the detection speed of the mask-wearing target.Meanwhile,in order to improve the detection accuracy of the algorithm,firstly,the GIOU loss function is used to replace the original bounding box regression loss in the SSD algorithm;secondly,the flexible non-maximum suppression algorithm is introduced to solve the problem that the non-maximum suppression algorithm is easy to suppress the low-scoring positive samples.The experimental results show that the detection speed and detection accuracy of the algorithm are improved,and it has good real-time and accuracy in the mask wearing detection task.In this thesis,there are a total of 42 figures,9 tables,and 74 references.