Research On The Detection Method Of Solar Dimples Based On Improved VNet Algorithm

Solar filaments exist in the corona,they make up of cold dense plasma.Strong filaments eruption activities affect the earth's magnetosphere,in serious cases,the communication equipment will be damaged,because communication interruption,air transport navigation failure and so on.In addition,the filaments act as tracers of the solar atmosphere's magnetic field,It is helpful to study the structure and evolution of solar magnetic field.Therefore,the accurate detection of dark strips is the basis and premise of related research,it has important scientific significance.There are inaccurate data sets in the existing filament detection methods,the accuracy of detection results is not high,small filaments can not be detection,In this paper,for the first time,we make a filaments database by combining the solar magnetogram,at the same time,we improve the network on the basis of the encoder-decoder segmentation network,we put forward M-DensenUNet segmentation network and improved VNet segmentation network,training network model by our filaments database,and detection the filaments by this model.In view of the above,this paper mainly carries on the following three aspects of the research work.First,we make a filaments database by combining the solar magnetogram for the first time.due to the existing filaments database,they marking the filaments database by the image threshold segmentation,as a result,sunspots and small noise points in the image are also highlighted and missed small filaments,cause the network to extract the wrong characteristic information,causes the filaments strip to miss the detection and the background to misidentify.This paper makes use of the relationship between the distribution of solar magnetic field and the position of filament,according to the polarity reversal lines in the magnetogram at the same time,the position of the filaments strips is further determined,and according to the brightness and texture characteristics of the filaments,we made the filaments database by hand annotation,it improves the accuracy of the database and provide data samples for subsequent experiments.Second,we proposed a solar filaments detection method based on multiscale dense connection M-DensenUNet.DenseUNet network with dense connection structure solves the redundancy problem of ResnetUNet network by increasing the connection between shallow and deep feature graphs,which increases the reuse of feature information,reduces the number of parameters and improves the detection accuracy.In order to prevent the loss of detail information in the process of subsampling,we proposed a solar filaments detection M-DensenUNet method based on multi-scale dense connection,we introduced multi-scale convolution structure on the basis of DenseUNet network and combined feature maps of different scales to reduce the loss of filaments texture information and further improve the result of filaments segmentation.Third,we propose a method of solar stripe detection based on improved VNet.we consider that the multi-scale dense connection network occupies very high memory and is subject to large background interference,so it is easy to misidentify part of the background as filaments.Therefore,we use VNet network with strong anti-jamming ability to train and filaments detection in this paper.At the same time,we improved the VNet network to remove the influence of background interference and the problem of the faint strip breaking.Firstly,the Inception module is introduced to obtain the different scale information of the dark bar,so as to better utilize the internal computing resources of the network and extract more image features.Secondly,attention mechanism is introduced to enhance the feature of dark bars and suppress useless background information.Finally,the depth monitoring module is introduced in the upper sampling part,so that the most accurate features can be learned from the feature images of different scales,which can not only suppress the interference background but also improve the detection accuracy of the filaments.The results of the improved VNet network and other segmentation networks were compared and analyzed,and the average accuracy of the improved VNet network on the test database reached 0.9883,and the F1 value reached 0.8385.The experimental results show that the method can effectively identify H? solor filaments in the full disk.