Atmospheric Particulate Matter SEM Image Recognition Based On YOLOV3 Algorithm

In recent years,atmospheric particulate matter has become one of the important pollutants in China’s atmospheric environment,which not only seriously endangers human health,but also closely relates to the global ecosystem.The sources and harms of atmospheric particulate matter with different morphological characteristics are also various.It is of great significance to study the image recognition of atmospheric particulate matter in order to analyze morphological characteristics and formulate treatment countermeasures.In this paper,scanning electron microscope(SEM)recognition of single atmospheric particle matter is studied.The main contents are as follows:1.In this paper,PQ200 and 2050 sampling equipment are used to collect samples of atmospheric particulate matter from 7 different sampling sites in Qingdao,and the completed samples are analyzed by scanning electron microscopy.Then 672 SEM images containing four types of particle matter with different morphological characteristics are obtained,which are expanded to 5,376 images by image enhancement,and the SEM image data set of atmospheric particle matter is established.2.For the identification of atmospheric particulate matter SEM images,this paper applies the deep learning technology based on regression.Based on the YOLOV3 algorithm,an improved YOLOV3 method for SEM image recognition of atmospheric particulate matter is proposed,which realizes the end-to-end optimization of SEM image recognition of particulate matter.3.Considering the high resolution of the SEM image of particulate matter,the input resolution of the original YOLOV3 algorithm is improved to adapt to the high-resolution particulate matter SEM image.The convolutional layer at the end of the darknet-53 feature extraction network is replaced by the dense connection layer to enhance the feature extraction capability of the original algorithm.The multi-scale prediction of YOLOV3 is combined with the spatial transformation network to enhance the spatial transformation capability of the original algorithm.The formula of average stacking degree is used for re-clustering analysis to obtain the initial candidate box which is more suitable for the SEM image data set of particulate matter.The experimental results show that the improved YOLOV3 SEM image recognition algorithm(YOLOV3-DN-STN)proposed in this paper has better recognition effect than the other five algorithms.The recall rate reaches 91.51%,5.25%higher than the original algorithm,and the accuracy reaches 93.08%,4.68%higher than the original algorithm.YOLOV3-DN-STN algorithm significantly improves the recognition accuracy of the original YOLOV3 algorithm,and realizes the effective recognition of SEM images of particle matter,which lays the foundation for the next research of particle matter images.