Single And Multi-wavelength Solar Event Detection Based On Convolutional Neural Networks

Object detection is a common and challenging task in computer vision.Traditional object detection is divided into three processes: extracting candidate regions through sliding windows,extracting relevant features,and classifying them.The shortcomings of this method include complicated time and redundancy of sliding windows.In addition,the features of manual design do not have good robustness in object diversity.With the development of deep learning technology,object candidate regions and related features can be extracted by convolutional neural networks.Object detection tasks become end-to-end.The detection algorithm of deep learning has greatly improved in speed and accuracy.In the field of astronomy,the traditional detection algorithms have the disadvantages of pre-processing solar images on the task of detecting Coronal Dimming(CD)and Coronal Wave(CW),and at the same time the data size of solar images is increasing.In order to detect CD and CW more accurately and quickly,this paper introduces a object detection method based on convolutional neural networks.The first task of this paper is to build a public dataset to provide an open data source for studying solar events CD and CW.Secondly,CD and CW are applied to single wavelength through deep learning model Faster R-CNN(Faster Region-based Convolutional Neural Networks)on the constructed public dataset,on which we run a compaprative analysis with trained R-FCN(Region-based Fully Convolutional Networks).Experiments show that this method can accurately detect solar events CD and CW.Finally,using multi-wavelength feature information,a shallow feature fusion Faster R-CNN(SF-Faster R-CNN)and a deep feature fusion Faster R-CNN(DF-Faster R-CNN)solar event detection method.The experimental results show that SF-Faster R-CNN is more suitable for multi-wavelength solar event detection than R-FCN,and it is stronger than single-wavelength model in recognition of multiple solar events;DFFaster R-CNN is better than SF-Faster R-CNN and it has improved the accuracy of solar event detection,approaching the single-wavelength model.At the same time,the detection effect of SF-Faster R-CNN on multiple solar events has been enhanced.From the experimental results of single wavelength,SF-Faster R-CNN and DFFaster R-CNN,convolutional neural network has higher accuracy in solar event detection,and can solve the detection tasks of solar event CD and CW.At the same time,it has certain applicability in the field of astronomy research.