Research On The Detection Method Of Coronal Mass Ejection Based On Foreground Segmentation Algorithm

A Coronal Mass Ejection(CME)is a massive and violent explosion.Due to this kind of explosion phenomenon will cause serious interference to the earth's environment,the detection of CMEs is of great significance for the prediction of disastrous space weather.Although the predecessors have developed many methods of CMEs detection,most of these methods artificially select features such as grayscale and texture,and use simple threshold segmentation technology to detect,so the detection effect of weak CMEs is not good.And the obtained CMEs result catalog has a big gap compared with the CDAW catalog.Therefore,according to the occurrence state of CMEs,the thesis applies more mature moving target detection methods to CMEs detection.Unlike general natural images,the corona image is a special solar observation image.The CMEs phenomenon is complex and changeable,and the foreground and background are similar in many ways,and there is a background effect of dynamic corona flow.The traditional moving target detection method needs to be improved in detection effect and detection speed.The foreground segmentation method based on deep learning has achieved good results in the detection of video image sequences.Among them,the Fg Seg Net?V2 network has the advantages of fast detection speed and good detection effect in the public related data set.However,there is currently no CME-related image data set,this method needs to manually mark CMEs images,which requires a lot of manpower and time to produce the image data set.At the same time,due to the inaccuracy of the definition of CMEs,some errors and ambiguities may occur in the process of marking some weak and small CMEs.Therefore,the thesis proposes an improved method of manual labeling,which extracts a series of feature parameters in the CDAW catalog and the SEEDS catalog,including occurrence date,occurrence time,central angle,angular width and height,and make them into data files in chronological order.Manually select the CMEs images corresponding to the occurrence time in the data file in the complete CMEs image sequence,and automatically mark the selected CMEs images through program development based on the data file to obtain the CMEs target frame,and then through the binarization process to obtain a dataset containing 7000 CMEs foreground images(ie groundtruth).Finally,the pre-processed CMEs images and the produced CMEs foregroundimage datasets are sent to the Fg Seg Net?V2 network for training respectively,so as to obtain a training model suitable for detecting CMEs images.In order to verify the accuracy and effectiveness of the thesis method,based on the manually marked CDAW catalog,this thesis compares the test result catalog obtained by the method of this thesis to two catalogs obtained by two classic automatic detection methods:CACTus catalog and SEEDS catalog.The results show that,within a fixed detection time,the results catalog obtained by the thesis method has three major advantages in CME detection compared with the CACTus and SEEDS catalogs: first,the number of CMEs detected is the largest,and the detection accuracy is high.On the two test sets,the accuracy rate obtained by the thesis method is improved by 67.7% and 52.9% compared to the CACTus catalog,respectively,and increased by 61.2% and 40.4% compared to the SEEDS catalog.Second,it can detect weak CMEs that are not detected by the CACTus and SEEDS catalogs and the fast CMEs with no mark in the CDAW catalog;third,the two features of the CMEs include center angle and angle width detected by this thesis have high accuracy,That is,the feature value is closest to the CDAW catalog.