A Coronal Mass Ejection Detection Method Based on Deep Learning

Through deep learning-based methods, coronal matter ejection detection is carried out, and polarization and convolutional neural networks are used to solve the problems of low detection accuracy and noise impact in the existing technology, achieving high-precision detection and robustness of weak CMEs promote.

CN112101444BActive Publication Date: 2024-11-22KUNMING UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010941895.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-09-09
Publication Date
2024-11-22
Estimated Expiration
2040-09-09

AI Technical Summary

Technical Problem

The existing automatic detection method of coronal material ejection is susceptible to noise and non-CME highlighting features, resulting in missed or multi-checking. Manually defined features cannot effectively characterize CME, and thresholding processing has poor effect on weak CME detection.

Method used

Using a deep learning-based method, the polarization and median filtering of CME images are used to denoise, and the improved feature extraction network and regional suggestion network are used, combined with convolutional neural networks, object detection and bounding box regression are performed, and non-maximum Value suppression eliminates redundancy boxes for more accurate CME detection.

Benefits of technology

Improve the accuracy of CME detection, detect more undetected CMEs, and are more accurate in the center angle and angle width of weak CMEs, are robust, and are not affected by comets and noise, overcome the feature selection of traditional methods. question.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112101444B_ABST
    Figure CN112101444B_ABST
Patent Text Reader

Abstract

The present invention discloses a coronal mass ejection detection method based on deep learning, which belongs to the technical field of automated coronal mass ejection detection. The method comprises the following steps: Step 1, CME image preprocessing and data set preparation; Step 2, feature extraction; Step 3, generating candidate region suggestions; Step 4, target region pooling; Step 5, performing CME subclassification; Step 6, bounding box regression; Step 7, removing duplicate redundant frames. The present invention introduces a deep learning method for CME target detection, which overcomes the problems that artificially designed features cannot well characterize CME semantics when threshold segmentation or traditional machine learning classifiers are used for CME detection and identification, and the classification effect of thresholds and traditional machine learning classifiers is poor, resulting in poor detection effect for weak CMEs. Compared with other methods, the present invention has better robustness.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of automatic detection of coronal mass ejections, and in particular to a coronal mass ejection detection method based on deep learning. Background Art

[0002] Coronal mass ejections (CMEs) are a strong space weather phenomenon, a phenomenon in which coronal material is ejected from the solar corona into interplanetary space. Because CMEs contain huge speed and energy and differ from the background solar wind in magnetic field, speed, and temperature, CMEs can cause solar wind disturbances when propagating in interplanetary space, and in severe cases can cause extreme space weather such as magnetic storms. These can cause satellite failures and data loss in space, collapse of ground power systems, and interruption of shortwave communications. In order to prevent the impact and harm of CMEs on human life, it is very valuable to detect CMEs and provide early warning of CMEs.

[0003] Satellites such as SOHO, Wind and STEREO provide humans with a large amount of valuable CME image data. Based on these image data, people have proposed various methods to detect CMEs. The CACTus method proposed by Berghmans et al. first preprocesses the LASCO C2 and C3 images, and then identifies CMEs through Hough line detection. The CACTus catalog generated by CACTus is the first automatically detected CME catalog. SEEDS proposed by Olmedo et al. uses the grayscale information of CME images to detect the leading and trailing edges of CMEs. The SEEDS catalog, CACTus catalog, and CDAW catalog generated by them are the three major reference catalogs for CMEs.

[0004] The CME target appears as a bright, complex texture enhanced structure on the CME image, followed by a dark area with insufficient brightness. A weak CME is faint in brightness, and the bright features are not obvious. Traditional automatic detection methods have achieved good results in detecting CMEs, but they also have the following shortcomings: 1) They are easily affected by various noises and non-CME highlight features of planets and comets, resulting in missed or over-detected CMEs; 2) Artificially defined CME features cannot characterize CMEs well; 3) Most methods are based on threshold processing, and have poor detection effects on weak CMEs.

[0005] Through the search of existing patents and related technologies, it is found that the existing automated detection methods for coronal mass ejections are:

[0006] A coronal mass ejection identification method based on Kalman filtering (CN106056125A) preprocesses the coronal sequence images: alignment, filtering and noise reduction, and conversion to polar coordinates; then, the preprocessed images are updated with a background using Kalman filtering, and then binarized to obtain a binary sequence image; and the binarized image is used to identify and screen the material ejection in the coronal sequence images.

[0007] An intelligent coronal mass ejection event observation method (CN103487844A) is characterized in that the observation method includes: A. Turning on the coronal mass ejection observation system, acquiring image data in real time according to the parameters of the observation system, and storing the data in a memory; B. Acquiring a differential image, and judging whether a coronal mass ejection event occurs, if a coronal mass ejection event occurs, executing step C; if no coronal mass ejection event occurs, executing step E; C. Alarming and adjusting the image observation mode; D. Judging whether a halo-shaped coronal mass ejection event occurs, if a halo-shaped coronal mass ejection event occurs, alarming again; and executing step E; E. Acquiring image data and storing the data in a hard disk.

[0008] A coronal mass ejection detection method based on multi-feature fusion (CN105046259A), characterized in that it includes the following steps: step 1, converting the corona observation image after differential processing into polar coordinate display; step 2, using different scales to segment the image processed by step 1 to obtain different sub-blocks, and obtaining the brightest block; step 3, respectively extracting the grayscale feature GB, texture feature TB and HOG feature HB of the brightest block in step 2; step 4, based on the extracted grayscale feature GB, texture feature TB and HOG feature HB, using a decision tree as a base classifier, and using the AdaBoost algorithm to enhance a strong classifier, finally obtaining a classification result, and completing the detection.

[0009] The above patented methods all use threshold methods or design artificial features to extract traditional grayscale features, texture features, and hog features, etc., and use threshold segmentation or traditional machine learning classifiers to distinguish CME from non-CME. However, artificially designed features cannot well represent CME, and the classification effect of thresholds and traditional machine learning classifiers is poor, resulting in poor detection of weak CME. Summary of the invention

[0010] In order to solve the above technical deficiencies, the present invention provides a coronal mass ejection detection method with high detection accuracy. The present invention greatly improves the accuracy of CME detection; not only more CMEs that are not detected in the SEEDS and CDAW catalogs are detected, but also the central angle and angular width of the CME are detected more accurately.

[0011] In order to solve the above technical problems, the technical solution of the present invention is as follows: A coronal mass ejection detection method based on deep learning comprises the following steps:

[0012] Step 1, CME image preprocessing and data set preparation: The CME d2 image is converted into polar coordinates and denoised by median filtering to obtain the CME image, and the target frame is semi-automatically annotated to prepare the data set;

[0013] Step 2, feature extraction: input the CME image in Step 1 into the improved feature extraction network to obtain the feature map;

[0014] Step 3, generate candidate region proposals: input the feature map obtained in step 2 into the region proposal network RPN to obtain the candidate region proposals of interest;

[0015] Step 4, target area pooling: take the feature map output by Step 2 and the candidate area proposal output by Step 3 as input, map the candidate area proposal to the corresponding position of the feature map, and perform pooling operation on the mapped feature map to obtain a target area feature map of uniform size;

[0016] Step 5: Perform CME subclassification: Calculate the target type based on the target area feature map obtained in Step 4, distinguish which type of CME the target belongs to, and obtain the classification score;

[0017] Step 6, bounding box regression: Based on the target area feature map obtained in Step 4, perform bounding box regression to refine the bounding box and finally obtain the precise position of the detection box;

[0018] Step 7, eliminate redundant frames: According to step 5 and step 6, the target feature map is obtained, and the non-maximum suppression method is used to remove the redundant target frames of the same CME target to obtain the final CME target area.

[0019] As a further description of the above technical solution: the specific implementation method of step Step2 is: the sample image is input into the vgg16 network, passes through five Conv Layer layers, and the Layer4 features and Layer5 features are merged to form a feature map as the output.

[0020] As a further description of the above technical solution: the specific implementation method of step Step3 is: use the k-means algorithm to obtain the nine length and width clusters of the data set, the RPN network generates nine anchor boxes according to each point of these nine clusters on the feature map of Step 2, and then use Softmax binary classification to determine whether each anchor box is the area of ​​interest, and the regression network obtains the position of the area of ​​interest.

[0021] As a further description of the above technical solution: the specific implementation method of step Step 4 is: divide the mapped feature area into blocks of the same size, the number of blocks is the same as the dimension of the output feature map, and then perform a maximum pooling operation on each block to convert a set of candidate area feature maps of different sizes into a feature map of a fixed uniform size.

[0022] As a further description of the above technical solution: the specific implementation method of step Step7 is: non-maximum suppression, for a set of detection frames obtained in step Step6, for each window, select the detection frame with the highest score in the neighborhood, suppress the detection frame with low score, and remove redundant detection frames.

[0023] Compared with the prior art, the present invention has the following beneficial effects: the present invention introduces a deep learning method for CME target detection in polar coordinate grayscale images, which overcomes the problem of missed detection of weak CMEs caused by threshold segmentation based on grayscale feature characteristics and texture features in traditional methods; the present invention has good robustness and is basically not affected by comets, auroras and noise.

[0024] This paper combines the idea of ​​deep learning convolutional neural network and uses convolutional features to replace traditional manual features, avoiding the problem of feature selection in traditional detection problems. Deep convolutional features have better expression ability and provide a more generalized and concise idea for solving CME detection problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart of the present invention;

[0026] Figure 2 This is the improved Faster R-CNN structure diagram;

[0027] Figure 3 This is the structure diagram of the improved feature extraction network VGG. DETAILED DESCRIPTION

[0028] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments, but the present invention is not limited to the following technical solution.

[0029] Example 1

[0030] like Figure 1-Figure 3 As shown, the present invention proposes a coronal mass ejection detection method based on an improved Faster R-CNN, comprising the following steps:

[0031] Step 1. CME image preprocessing and dataset preparation: Download CME d2 images from the SOHO website (https: / / sohowww.nascom.nasa.gov), convert d2 images into polar coordinates and use median filtering to denoise to obtain CME polar coordinate difference images. According to the pascal_voc dataset framework, a dataset was prepared based on the information in the SEEDS directory and the CDAW directory, including 2678 training sets and 246 validation sets. The dataset was input into the improved Faster R-CNN for finetuning training, and the trained network model was obtained after 350,000 training cycles.

[0032] Step 2, feature extraction: Input the CME image processed in Step 1 into the trained feature extraction network VGG16 to extract the feature map of the target CME image; merge the Layer4 features and Layer5 features of VGG16 to form a feature map to obtain the feature map of CME, which is shared for the subsequent region proposal network (RPN) and Fast R-CNN classification layer.

[0033] Step 3, generate candidate region proposals: input the feature map obtained in step 2 into the region proposal network (RPN) to obtain a set of region proposals, which includes using the k-means algorithm to obtain nine types of clusters of length and width and sliding a 3×3 window point by point on the feature map. Each point in the feature map, that is, the center point of each sliding window position, is mapped to the original image, and 9 candidate target regions of interest with nine types of length and width are generated based on k-means. Then, the Softmax network is used to determine whether it belongs to the foreground or the background, that is, whether it is a CME target or a background of no interest, and then the bounding box regression is used to correct the proposal box.

[0034] Step 4, target area pooling: take the feature map output by Step 2 and the candidate area proposal output by Step 3 as input, map the candidate area proposal to the corresponding position of the feature map, and perform pooling operation on the mapped feature map; the specific implementation method is: divide the mapped feature area into blocks of the same size (the number of blocks is the same as the dimension of the output feature map), and then perform the maximum pooling operation on each block, and convert a set of candidate area feature maps of different sizes into feature maps of fixed uniform size for subsequent work.

[0035] Step 5: Perform CME subclassification: Calculate the target type based on the target area feature map obtained in Step 4, identify which type of CME the target belongs to, and form a classification score; specifically, use a fully connected layer to classify each target area proposal.

[0036] Step 6, bounding box regression: According to the target area feature map obtained in Step 4, the bounding box is refined by bounding box regression, and finally the precise position of the detection box is obtained.

[0037] Step 7, eliminate duplicate redundant target frames: Use non-maximum suppression to obtain a set of detection frames for step Step 6. For each window, select the detection frame with the highest score and low suppression score in the neighborhood, remove redundant detection frames, and find the best object detection position. The specific method is: sort the classification scores of each detection frame obtained in step Step 6, and then select the detection frame with the highest classification score, use it as the standard, traverse the remaining detection frames in turn, and discard it when the degree of overlap with the standard detection frame is higher than the threshold. After traversing all the remaining detection frames, reselect the detection frame with the largest classification score in the remaining detection frames as the standard detection frame, and repeat the above operation until all the detection frames are traversed to obtain a target frame corresponding to each CME target.

[0038] The above is only a preferred specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can make equivalent substitutions or changes within the technical scope disclosed in the present application according to the technical solution and utility model concept of the present application, which should be covered by the protection scope of the present application.

Claims

1. A coronal mass ejection detection method based on deep learning, characterized in that: The following steps are involved: Step 1, CME image preprocessing and data set preparation: The CME d2 image is converted into polar coordinates and denoised by median filtering to obtain the CME image, and the target frame is semi-automatically annotated to prepare the data set; Step 2, feature extraction: input the CME image in Step 1 into the improved feature extraction network to obtain the feature map; Step 3, generate candidate region proposals: input the feature map obtained in step 2 into the region proposal network RPN to obtain the candidate region proposals of interest; Step 4, target area pooling: take the feature map output by Step 2 and the candidate area proposal output by Step 3 as input, map the candidate area proposal to the corresponding position of the feature map, and perform pooling operation on the mapped feature map to obtain a target area feature map of uniform size; Step 5: Perform CME subclassification: Calculate the target type based on the target area feature map obtained in Step 4, distinguish which type of CME the target belongs to, and obtain the classification score; Step 6, bounding box regression: Based on the target area feature map obtained in Step 4, perform bounding box regression to refine the bounding box and finally obtain the precise position of the detection box; Step 7, eliminate redundant frames: According to step 5 and step 6, the target feature map is obtained, and the non-maximum suppression method is used to remove the redundant target frames of the same CME target to obtain the final CME target area.

2. The method for detecting coronal mass ejections based on deep learning according to claim 1, characterized in that: The specific implementation method of Step 2 is: the sample image is input into the vgg16 network, passes through five Conv Layer layers, and merges the Layer4 features and the Layer5 features to form a feature map as the output.

3. The method for detecting coronal mass ejections based on deep learning according to claim 1, characterized in that: The specific implementation method of Step 3 is as follows: using the k-means algorithm to obtain nine clusters of length and width in the data set, the RPN network generates nine anchor boxes according to each point of these nine clusters on the feature map of Step 2, and then the Softmax binary classification is used to determine whether each anchor box is a region of interest, and the position of the region of interest is obtained by the regression network.

4. The method for detecting coronal mass ejections based on deep learning according to claim 1, characterized in that: The specific implementation method of step 4 is: divide the mapped feature area into blocks of the same size, the number of blocks is the same as the dimension of the output feature map, and then perform a maximum pooling operation on each block to convert a set of candidate area feature maps of different sizes into a feature map of a fixed uniform size.

5. The method for detecting coronal mass ejections based on deep learning according to claim 1, characterized in that: The specific implementation method of step Step7 is: non-maximum suppression, for a set of detection frames obtained in step Step6, for each window, select the detection frame with the highest score in the neighborhood, suppress the detection frame with low score, and remove redundant detection frames.