Sun dark strip automatic detection method

Through improved U-Net model and APP components, combined with K-means and DBSCAN clustering, the problem of low automation of sun dark bar detection is solved, high accuracy and scalability detection results are achieved, and data set generation is simplified The process improves detection efficiency and accuracy.

CN119091175AActive Publication Date: 2024-12-06NANJING UNIV
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411228839.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-03
Publication Date
2024-12-06
Estimated Expiration
2044-09-03

AI Technical Summary

Technical Problem

The existing solar dark bar detection methods have low degree of automation, average detection accuracy and accuracy, poor method scalability, small number of data sets, low quality, and time-consuming production.

Method used

Using the improved U-Net model combined with APP components, the image training data set is constructed by acquiring full-day surface Hα images for preprocessing, and using K-means clustering and DBSCAN clustering to generate training data to perform automated sun dark bar detection. Supports image processing at different resolutions and time series.

Benefits of technology

It realizes automation of model data generation, improves the accuracy and migration of detection results, supports long-term dark bar tracking, and improves the degree of automation and scalability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119091175A_ABST
    Figure CN119091175A_ABST
Patent Text Reader

Abstract

The invention discloses an automatic detection method for a dark strip of the sun, and the method comprises the steps: S1, obtaining a full-solar-surface H alpha image, carrying out the preprocessing of the full-solar-surface H alpha image, and constructing an image training data set; s2, constructing a U-Net model based on APP improvement as a sun dark stripe detection model, and training the sun dark stripe detection model by using the image data set; s3, processing the preprocessed input image with the time sequence by using the trained sun dark-stripe detection model to obtain a sun dark-stripe detection result; and S4, carrying out SORT tracking on the dark stripe detection result to obtain a dark stripe tracking result. According to the method, semi-automatic and even full-automatic machine learning data generation, target tracking model training, reasoning and target tracking of the dark strip of the sun are realized, and the method has relatively low training cost and better indexes; and an FCN method with any resolution is supported, and good mobility is achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to an automatic detection method for solar dark stripes. Background Art

[0002] Solar activity is a general term for all active phenomena in the solar atmosphere, such as sunspots, flares, and solar flares. Among them, solar filaments are dark stripes produced by the projection of solar prominences on the solar surface, which can reflect the local magnetic field structure of the solar prominence region to a certain extent. Current methods for automatic detection of solar filaments include connected domain processing methods based on Canny edge detection and semantic segmentation methods based on naive U-Net.

[0003] Among them, for the connected domain processing method based on Canny edge detection, this technology requires multiple manually set thresholds, which are sensitive to the characteristics of the image itself. Therefore, for images taken by different stations, it has poor versatility and transferability, low degree of automation, and requires repeated manual adjustments; this technical method has poor dark bar recognition ability and can only recognize relatively obvious dark bars. For dark bars with smaller active areas, this method is likely to regard them as noise due to the problem of connected domain threshold.

[0004] For semantic segmentation methods, the data sets involved need to be completely manually labeled, which is time-consuming, has a low degree of automation, and the data set size is small, and the model indicators are not ideal. At the same time, this method can only process relatively low-resolution data and has no reliable functions for high-resolution images. Summary of the invention

[0005] In response to the above-mentioned deficiencies in the prior art, the solar dark stripe detection method provided by the present invention solves the problems of low degree of automation, general detection precision and accuracy, and poor method scalability of the existing related methods, as well as the problems of small number, low quality, and time-consuming production of existing dark stripe detection data sets.

[0006] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a method for automatically detecting dark stripes of the sun, comprising the following steps:

[0007] S1, obtain the full disk Hα image, preprocess it, and construct the image training data set;

[0008] S2. Build an improved U-Net model based on APP as a solar dark stripe detection model and train it using image datasets;

[0009] S3, using the trained solar streak detection model to process the pre-processed input image with time series to obtain the solar streak detection result;

[0010] S4. Perform SORT tracking on the dark stripe detection result to obtain a dark stripe tracking result.

[0011] Furthermore, the step S1 is specifically as follows:

[0012] S11, acquiring a full-disk Hα image, converting it into a single-channel image, and correcting its limb darkening;

[0013] S12, on the basis of correcting the limb darkening, clipping the circumscribed rectangle of the solar disk according to the solar disk pixel center and radius data to obtain a solar disk image;

[0014] S13. K-means clustering is performed on the solar image according to the color depth of the pixels, and an image training data set is constructed based on the clustering results in which the dark stripe classification visibility is greater than a set threshold.

[0015] Furthermore, in step S1, the preprocessing further includes:

[0016] The different clustering results corresponding to the same solar disk image are integrated, and majority election and groundtruth annotation are performed;

[0017] The groundtruth annotation results are re-clustered, noise is removed, and manual correction is performed.

[0018] Further, in step S2, the solar dark stripe detection model includes an improved U-Net model architecture and an APP component;

[0019] The improved U-Net model architecture includes a DoubleConv component, first to nth DownSample components, first to nth UpSample components and a 2D convolution layer connected in sequence; the output end of the 2D convolution layer is connected to the APP component;

[0020] The DoubleConv component is also connected to the nth UpSample component through a concatenate function, and the first to nth DownSample components are connected to the n-1th to the first UpSample components in a one-to-one correspondence through a concatenate function;

[0021] Wherein, n is the adjustable number of UpSample components and DownSample components.

[0022] Furthermore, when the number of output channels of the DoubleConv component is base_channels, the number of input and output channels of the i-th DownSample component is 2 respectively. i-1 * base_channels and 2 i*base_channels, the number of input and output channels of the i-th UpSample component is 2 respectively n-i+1-f * base_channels and 2 n-i-f *base_channels, the number of input and output channels of the 2D convolutional layer is base_channels respectively;

[0023] Where i is the index of the DownSample component or the UpSample component, where f is 1 if bi_linear is true, otherwise f is 0, and bi_linear is the parameter passed to the UpSample component.

[0024] Furthermore, the DoubleConv component includes a first 3×3 convolutional layer, a first batch normalization layer BN, a first LeakyReLU activation function, a second 3×3 convolutional layer, a second batch normalization layer BN, a second LeakyReLU activation function and a random dropout layer connected in sequence.

[0025] Furthermore, the APP component includes depth number of parallel adaptive pooling layers with the same structure, and the output end of each adaptive pooling layer is sequentially connected to a 1×1 convolutional layer, a Dropout layer, and a Sigmoid activation layer through a concatenate function;

[0026] Each of the adaptive pooling layers includes an adaptive pooling strategy, a 1×1 convolution layer, and an upsampling layer connected in sequence; the adaptive pooling strategy is maximum pooling or average pooling;

[0027] The target size of the output of the mth adaptive pooling layer is out_image_size is the output size of the APP component, and depth is the depth of the APP component, that is, the number of adaptive pooling layers.

[0028] Further, the DownSample component includes an APP component and a DoubleConv component connected in sequence;

[0029] For the UpSample component, when bi_linear is true, the UpSample component includes an UpSample layer with a mode of bi_linear and a DoubleConv component; the processing result of the UpSample layer is merged with the processing result of the corresponding DownSample component after center cropping, and connected to the DoubleConv component through the concatenate function;

[0030] When bi_linear is false, the UpSample component includes a 2×2 transposed convolution layer and a DoubleConv component; the processing result of the 2×2 transposed convolution layer is center-cropped and merged with the processing result of the corresponding DownSample component, and connected to the DoubleConv component through the concatenate function;

[0031] When the input size of the UpSample component is H in ×W in ×C in When the output size is 2H in ×2W in ×C out .

[0032] Furthermore, according to the complexity of the detected image, the improved U-Net model architecture and APP components are freely connected to expand the structure of the solar dark stripe detection model.

[0033] Furthermore, the step S3 specifically comprises:

[0034] S31, using the trained solar dark stripe detection model to infer the input image, and outputting the detection dark stripe labeling result;

[0035] S32, performing DBSCAN clustering on the output results to obtain classification results of independent dark stripes;

[0036] S33, recording the pixel coordinates and image number of each independent dark stripe, and obtaining the solar dark stripe detection result.

[0037] The beneficial effects of the present invention are:

[0038] 1. The training data generation of the model in the present invention can be basically automated, and the Kmeans clustering speed is fast. No human participation is required except for manual screening, so a large amount of training data can be generated quickly;

[0039] 2. The improved U-Ne model structure is used as the solar dark stripe detection model in the present invention, and its depth supports manual adjustment, so as to support training data of different resolutions and hardware requirements;

[0040] 3. The target detection result indicators obtained based on the solar dark stripe detection model in the present invention are good.

[0041] 4. The present invention provides a solution for tracking long-period dark stripes, which can process image groups with time series instead of single images;

[0042] 5. The method of the present invention has a high overall degree of automation, good portability and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of the method for automatically detecting solar dark stripes provided by the present invention.

[0044] Figure 2 This is a schematic diagram of the structure of the solar dark stripe detection model provided by the present invention.

[0045] Figure 3 This is a schematic diagram of the Double Conv component structure provided by the present invention.

[0046] Figure 4 This is a schematic diagram of the APP component structure provided by the present invention.

[0047] Figure 5 This is a schematic diagram of the Down Sample component structure provided by the present invention.

[0048] Figure 6 This is a schematic diagram of the Up Sample component structure provided by the present invention.

[0049] Figure 7 A schematic diagram of the process of dark stripe detection using the method provided by the present invention. DETAILED DESCRIPTION

[0050] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0051] The embodiment of the present invention provides a method for automatically detecting dark stripes of the sun. Figure 1 As shown, the following steps are included:

[0052] S1, obtain the full disk Hα image, preprocess it, and construct the image training data set;

[0053] S2. Build an improved U-Net model based on APP as a solar dark stripe detection model and train it using image datasets;

[0054] S3, using the trained solar streak detection model to process the pre-processed input image with time series to obtain the solar streak detection result;

[0055] S4. Perform SORT tracking on the dark stripe detection result to obtain a dark stripe tracking result.

[0056] Step S1 of the embodiment of the present invention is specifically as follows:

[0057] S11, acquiring a full-disk Hα image, converting it into a single-channel image, and correcting its limb darkening;

[0058] S12, on the basis of correcting the limb darkening, clipping the circumscribed rectangle of the solar disk according to the solar disk pixel center and radius data to obtain a solar disk image;

[0059] S13. K-means clustering is performed on the solar image according to the color depth of the pixels, and an image training data set is constructed based on the clustering results in which the dark stripe classification visibility is greater than a set threshold.

[0060] Exemplarily, in the embodiment, the preprocessing process of the full-disk Hα image is performed based on the FTTS file of the full-disk Hα image.

[0061] Optionally, use the sldtk tool to correct for limb darkening.

[0062] Optionally, when performing K-means clustering on solar disk images, multiple K values ​​are set, such as 10, 20, and 30, and multiple clusterings are performed respectively.

[0063] In step S1 of this embodiment, the preprocessing further includes:

[0064] Different clustering results corresponding to the same solar disk image are integrated, and majority election and groundtruth annotation are performed; the groundtruth annotation results are re-clustered, noise is removed, and manual correction is performed.

[0065] Optionally, when a majority election is performed, the annotations that exceed half are marked as groundtruth.

[0066] Optionally, a DBSCAN method is used for clustering, and when removing noise points, for example, if the number of pixels of a certain class is less than 10, it is removed.

[0067] Optionally, when manual correction is used, Photoshop or LabelImg tools can be used to connect discontinuous or incomplete results in the clustering.

[0068] In step S2 of the embodiment of the present invention, the solar dark stripe detection model includes an improved U-Net model architecture and an APP component;

[0069] like Figure 2 As shown, the improved U-Net model architecture includes a DoubleConv component, the first to nth DownSample components, the first to nth UpSample components and a 2D convolution layer connected in sequence; the output end of the 2D convolution layer is connected to the APP component;

[0070] The DoubleConv component is also connected to the nth UpSample component through a concatenate function, and the first to nth DownSample components are connected to the n-1th to the first UpSample components in a one-to-one correspondence through a concatenate function;

[0071] Wherein, n is the adjustable number of UpSample components and DownSample components.

[0072] In this embodiment, in the improved U-Net model architecture, when the number of output channels of the DoubleConv component is base_channels, the number of input and output channels of the i-th DownSample component is 2 respectively. i-1 *base_channels and 2i*base_channels, the number of input and output channels of the i-th UpSample component is 2 respectively n-i+1-f * base_channels and 2 n-i-f *base_channels, the number of input and output channels of the 2D convolutional layer is base_channels respectively;

[0073] Where i is the index of the DownSample component or the UpSample component, where f is 1 if bi_linear is true, otherwise f is 0, and bi_linear is the parameter passed to the UpSample component.

[0074] Specifically, compared with the traditional structure, in the improved U-Net model architecture, the parameters passed include base_channels, num_pooling, and requires_activation compared to UpSample. The U-Net depth can be artificially adjusted by adjusting the first two parameters accordingly.

[0075] In this embodiment, if Figure 3 As shown, the DoubleConv component includes a first 3×3 convolutional layer, a first batch of normalization layers BN, a first LeakyReLU activation function, a second 3×3 convolutional layer, a second batch of normalization layers BN, a second LeakyReLU activation function, and a random dropout layer, which are connected in sequence.

[0076] For example, in Figure 3In the DoubleConv component structure shown, the kernel size of the convolution layer is 3, the padding size is 1, and the parameters of the LeakyReLU activation function (leaky_negative_slope), the number of input channels (in_channels), the number of output channels (out_channels), the number of channels in the middle layer (the output of the first 3×3 convolution layer) (mid_channels), and the dropout probability of the random dropout layer (dropout_p) can all be set by passing parameters.

[0077] In this embodiment, if Figure 4 As shown, the APP component includes depth adaptive pooling layers with the same structure in parallel, and the output end of each adaptive pooling layer is sequentially connected to a 1×1 convolutional layer, a Dropout layer, and a Sigmoid activation layer through a concatenate function;

[0078] Each of the adaptive pooling layers includes an adaptive pooling strategy, a 1×1 convolution layer, and an upsampling layer connected in sequence; the adaptive pooling strategy is maximum pooling or average pooling;

[0079] The target size of the output of the mth adaptive pooling layer is out_image_size is the output size of the APP component, and depth is the depth of the APP component, that is, the number of adaptive pooling layers.

[0080] Specifically, the above-mentioned APP component is also called the pyramid pooling component, and its parameters include the number of input channels (in_channels), the number of output channels (out_channels), the output size (out_image_size), the depth (depth), whether the activation layer is required (requires_activation) and the pooling strategy (pooling_strategy).

[0081] For example, for any input x, it is upsampled after passing through each adaptive pooling layer so that the image size finally becomes out_image_size and then concatenated on the channel, and then input into a 1×1 convolutional layer to change the number of channels to out_channels, and finally whether to pass the activation function is selected according to whether the layer needs to be activated or not.

[0082] Exemplarily, the upsampling layer adopts a bilinear interpolation upsampling operation.

[0083] In this embodiment, if Figure 5 As shown, the DownSample component includes an APP component and a DoubleConv component connected in sequence;

[0084] When the input size of the DownSample component is H in ×W in ×C in When , the output size is

[0085]

[0086] Specifically, for the DownSample component of the above structure, the parameter passed is one more out_image_size than the DoubleConv component, which is used to specify the output image size of the APP component. The out_channels of the APP component is specified to be the same as that of DownSample itself, and the output image size can be freely set to support images of different resolutions. Generally, the length and width are set to half of the previous layer. Compared with the original U-Net model that uses maximum pooling, it is replaced by the APP component to improve the information extraction capability during the downsampling process and encode and integrate image information under different receptive fields.

[0087] In this embodiment, if Figure 6 As shown, for the UpSample component, the UpSample component in this embodiment has one more bi_linear parameter than DoubleConv, and there are two architecture forms according to the value of the bi_linear parameter.

[0088] Specifically, when bi_linear is true, the UpSample component includes an UpSample layer with mode bi_linear and a DoubleConv component; the processing result of the UpSample layer is center-cropped and merged with the processing result of the corresponding DownSample component, and connected to the DoubleConv component through the concatenate function; wherein the parameter of the number of channels in the middle layer of the DoubleConv component is set to half the number of input channels;

[0089] When bi_linear is false, the UpSample component includes a 2×2 transposed convolution layer and a DoubleConv component; the processing result of the 2×2 transposed convolution layer is center-cropped and merged with the processing result of the corresponding DownSample component, and connected to the DoubleConv component through the concatenate function; wherein the number of output channels of the 2×2 transposed convolution layer is half of the number of input channels, and the number of intermediate channels of the DoubleConv component is the same as the number of output channels.

[0090] Based on the above UpSample component structure, when the input size of the UpSample component is H in ×W in ×Cin When the output size is 2H in ×2W in ×C out .

[0091] Exemplarily, in step S2, when the solar streak detection model is trained, the data is processed as follows in each training iteration:

[0092] The original image and the annotated image are uniformly scaled to the specified image size; the original image and the annotated image are synchronously flipped up and down with a certain probability, such as a 50% probability; the original image and the annotated image are synchronously randomly rotated within a certain range with a certain probability, such as an 80% probability, and the rotation angle is ±45°; the original image is randomly gamma transformed within a certain range with a certain probability, such as an 80% probability, and the gamma value is between 0.5 and 1.5; the original image is Min-Max normalized.

[0093] Optionally, in the above process, the original image and the annotated image are synchronously translated within a certain range of proportion with a certain probability, such as 80% probability, and the translation degree is ±30%.

[0094] For example, during the model training process, Focal Loss is used as the loss function, AdamW optimizer is used, and a certain number of iterations, such as 250 rounds, are trained.

[0095] In the above solar streak detection model in this embodiment, illustratively, the most basic model is based on the U-Net architecture with the requires_activation parameter being False, and an additional APP component with out_image_size being the original image size and out_channels being 1 is connected, and depth can be freely set, and 4 is taken as an example in this embodiment. Compared with a single U-Net model, an APP component is added after it, which improves the decoding ability of the model without increasing too many trainable parameters, making up for the shortcomings of the simple decoding process and single receptive field of U-Net, and having stronger learning ability and faster model convergence speed.

[0096] In an embodiment of the present invention, according to the complexity of the detected image, an improved U-Net model architecture and APP components are freely connected to expand the structure of the solar dark stripe detection model.

[0097] Exemplarily, the model can be arbitrarily extended using the structure of U-Net-APP-U-Net-APP-...

[0098] Optionally, use the above model as an auxiliary head for training, for example, use a U-Net-APP as the main model and a U-Net as the auxiliary head to improve the model indicators.

[0099] In step S3 of the embodiment of the present invention, the method for processing the preprocessed input image with time series is specifically as follows:

[0100] S31, using the trained solar dark stripe detection model to infer the input image, and outputting the detection dark stripe labeling result;

[0101] S32, performing DBSCAN clustering on the output results to obtain classification results of independent dark stripes;

[0102] S33, recording the pixel coordinates and image number of each independent dark stripe, and obtaining the solar dark stripe detection result.

[0103] Exemplarily, in step S3, the operations of preprocessing the image include: reading the FITS file of the full disk Hα and converting it into a single-channel image; correcting the limb darkening using the sldtk tool; cropping the circumscribed rectangle of the disk according to the disk pixel center and radius data in the FITS file; rotating according to the p angle size provided in the FITS file to correct the ecliptic obliquity; and performing Min-Max normalization on the image.

[0104] Optionally, in the process of obtaining the above output results, multiple U-Net trained models of different depths can be used for integration, such as mask merging.

[0105] In step S4 of this embodiment, the solar dark stripe detection result obtained by the above model inference is input into the SORT model tool and combined with the time series to perform dark stripe tracking.

[0106] In one example of this embodiment, Figure 7 FIG. 4 is a schematic diagram of a process of dark stripe detection using the method of the present invention, wherein: Figure 7 (a) is the original image. Figure 7 (b) is the annotated image after the “training data generation” step. Figure 7 (c) is the inference result after model training. Figure 7 (d) is the result of inference result after denoising by DBSCAN clustering method. Figure 7 (e) Figure 7 (d) Based on the above, the dark stripes are colored and the bounding rectangles are drawn. Figure 7 (a) Figure 7 (b) with Figure 7(c) It can be found that the reasoning results after training basically identify all the dark bars in the labeled image, and the outlines of some dark bars are more complete; comparison Figure 7 (c) with Figure 7 (d) Figure 7 (e) It can be found that after DBSCAN clustering, some noise is removed and the dark bars are distinguished, and each dark bar is identified independently.

[0107] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0108] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for automatically detecting dark stripes of the sun, characterized in that: The following steps are involved: S1, obtain the full disk Hα image, preprocess it, and construct the image training data set; S2. Build an improved U-Net model based on APP as a solar dark stripe detection model and train it using image datasets; S3, using the trained solar streak detection model to process the pre-processed input image with time series to obtain the solar streak detection result; S4. Perform SORT tracking on the dark stripe detection result to obtain a dark stripe tracking result.

2. The method for automatically detecting dark solar stripes according to claim 1, characterized in that: The step S1 is specifically as follows: S11, acquiring a full-disk Hα image, converting it into a single-channel image, and correcting its limb darkening; S12, on the basis of correcting the limb darkening, clipping the circumscribed rectangle of the solar disk according to the solar disk pixel center and radius data to obtain a solar disk image; S13. K-means clustering is performed on the solar image according to the color depth of the pixels, and an image training data set is constructed based on the clustering results in which the dark stripe classification visibility is greater than a set threshold.

3. The method for automatically detecting dark solar stripes according to claim 2, characterized in that: In the step S1, the preprocessing further includes: The different clustering results corresponding to the same solar disk image are integrated, and majority election and groundtruth annotation are performed; The groundtruth annotation results are re-clustered, noise is removed, and manual correction is performed.

4. The method for automatically detecting dark solar stripes according to claim 1, characterized in that: In step S2, the solar streak detection model includes an improved U-Net model architecture and an APP component; The improved U-Net model architecture includes a DoubleConv component, first to nth DownSample components, first to nth UpSample components and a 2D convolution layer connected in sequence; the output end of the 2D convolution layer is connected to the APP component; The DoubleConv component is also connected to the nth UpSample component through a concatenate function, and the first to nth DownSample components are connected to the n-1th to the first UpSample components in a one-to-one correspondence through a concatenate function; Wherein, n is the adjustable number of UpSample components and DownSample components.

5. The method for automatically detecting dark solar stripes according to claim 4, characterized in that: When the number of output channels of the DoubleConv component is base_channels, the number of input and output channels of the i-th DownSample component is 2 respectively. i-1 * base_channels and 2 i *base_channels, the number of input and output channels of the i-th UpSample component is 2 respectively n-i+1-f * base_channels and 2 n-i-f *base_channels, the number of input and output channels of the 2D convolutional layer is base_channels respectively; Where i is the index of the DownSample component or the UpSample component, where f is 1 if bi_linear is true, otherwise f is 0, and bi_linear is the parameter passed to the UpSample component.

6. The method for automatically detecting dark solar stripes according to claim 4, characterized in that: The DoubleConv component includes a first 3×3 convolutional layer, a first batch normalization layer BN, a first LeakyReLU activation function, a second 3×3 convolutional layer, a second batch normalization layer BN, a second LeakyReLU activation function and a random dropout layer, which are connected in sequence.

7. The method for automatically detecting dark solar stripes according to claim 4, characterized in that: The APP component includes depth number of parallel adaptive pooling layers with the same structure, and the output end of each adaptive pooling layer is sequentially connected to a 1×1 convolutional layer, a Dropout layer, and a Sigmoid activation layer through a concatenate function; Each of the adaptive pooling layers includes an adaptive pooling strategy, a 1×1 convolution layer, and an upsampling layer connected in sequence; the adaptive pooling strategy is maximum pooling or average pooling; The target size of the output of the mth adaptive pooling layer is out_image_size is the output size of the APP component, and depth is the depth of the APP component, that is, the number of adaptive pooling layers.

8. The method for automatically detecting dark solar stripes according to claim 4, characterized in that: The DownSample component includes an APP component and a DoubleConv component connected in sequence; When the input size of the DownSample component is H in ×W in ×C in When , the output size is For the UpSample component, when bi_linear is true, the UpSample component includes an UpSample layer with a mode of bi_linear and a DoubleConv component; The processing result of the UpSample layer is merged with the processing result of the corresponding DownSample component after center cropping, and connected to the DoubleConv component through the concatenate function; When bi_linear is false, the UpSample component includes a 2×2 transposed convolution layer and a DoubleConv component; the processing result of the 2×2 transposed convolution layer is center-cropped and merged with the processing result of the corresponding DownSample component, and connected to the DoubleConv component through the concatenate function; When the input size of the UpSample component is H in ×W in ×C in When the output size is 2H in ×2W in ×C out .

9. The method for automatically detecting dark solar stripes according to claim 4, characterized in that: According to the complexity of the detected image, the improved U-Net model architecture and APP components are freely connected to expand the structure of the solar dark stripe detection model.

10. The method for automatically detecting dark solar stripes according to claim 1, characterized in that: The step S3 specifically includes: S31, using the trained solar dark stripe detection model to infer the input image, and outputting the detection dark stripe labeling result; S32, performing DBSCAN clustering on the output results to obtain classification results of independent dark stripes; S33, recording the pixel coordinates and image number of each independent dark stripe, and obtaining the solar dark stripe detection result.

Citation Information

Patent Citations

  • Coronal mass ejection detection method based on improved U-Net network

    CN112101440A

  • Small target semantic segmentation sun dark bar detection method combined with target detection

    CN112102274A

  • Method for detecting dark strips in full-solar-surface image

    CN115249319A

  • Sun dark stripe image detection method based on TransUnet

    CN118097180A

  • Solar activity area automatic detection method based on density clustering algorithm

    CN119091174A