Research On Detection Method Of CME Based On Deep Learning

Coronal Mass Ejection(CME)is one of the most violent explosions in the solar atmosphere.It usually carries a large number of high-energy particles,which has a great influence on the space environment and human activities.Studying the signs of CME activities will help us have a clearer understanding of their movement process,make space weather warnings,and reduce disasters.The current catalogue of CME is mainly produced by hand and traditional image processing methods.Manual catalogue is obviously subjective and time-consuming.The recognition results of traditional methods always have some errors.The excellent performance of deep learning in object detection,image segmentation,object tracking and other tasks prompts us to apply it to CME's Detection.We use deep learning technology to study CME,and proposes two methods to detect CME.Based on these two methods,we design and implement a CME catalogue system.This paper has three main contributions:(1)CME detection based on deep descriptor transformation.We proposes a method for detecting CME based on deep descriptor transformation.This method includes three modules.First,we design an image classification network which includes a feature extraction module and a classification module.We train the image classification network by obtaining the training tags from the existing CME catalogues and extract the CME's semantic features from the pretrained network.Secondly,we use depth descriptor transformation to locate the common objects in the image set to identify the CME area,then applying Graph Cut algorithm to fine-tune the detected CME area.Finally,we convert the binary image with the detected CME pixels from Cartesian coordinates to polar coordinate.We define rules to track CME events:If a CME event can move at least two frames and reach the edge of the coronagraph's field of view,it is marked as an event.We calculate and count CME's basic motion parameters and compare them with the existing CME catalogues.The experimental results show that the method in this paper can capture CME as soon as possible and identify weak CME.(2)CME detection based on context perception.From the perspective of obtaining the edge details of the CME,we propose a detection method for the CME based on context perception.We hand-made a one-month CME trimap dataset,and design an end-to-end alpha prediction network which takes the white-light coronagraph differential image and the corresponding trimap as input.In the encoding process,in order to obtain the better spatial context information of the marginal area of the CME,we adopted the attention mechanism to extract broader and deeper semantic features and retain more spatial information.At the same time,we propose a measurement method that combines alpha loss and composition loss.The loss function constrains the high-level features of the unknown area to make the prediction results of the CME's marginal area more accurate.Finally,We define the tracking rules to clean the detection results and track the CME event.We count the basic motion parameters and compare the tracking results with the existing methods.Experiments show that this method can obtain a more perfect edge structure of the CME and get a more realistic trajectory.(3)Implement a catalogue system for CME.We integrated the method of detecting coronal mass ejection in this article,calculated a series of CME events,and designed and implemented a CME cataloging system.The system catalogs and demonstrates the movement parameters of the daily CME.We publish them on the website of the Purple Mountain Observatory,which is shared with ordinary users for reference and learning.