A deep learning-based wide-area astronomical image dim and weak target detection method

Abstract

The present application relates to a kind of dark and weak target detection methods of wide-area astronomical image based on deep learning, belong to astronomical image processing technical field, including field of view division to original astronomical image, and pretreatment is carried out;Then the threshold segmentation label generation strategy based on star catalogue prior is established to generate label set;Deep learning semantic segmentation model is constructed and trained to obtain optimization model;With the astronomical image after pretreatment as input, output target segmentation map, and eliminate the connected domain target that does not conform to physical characteristics, based on the roundness feature of target area determination and separate fusion target, obtain the final mask of independent target;Based on final mask, the subpixel level center coordinates of each independent target are calculated using weighted centroid method, effectively improve the detection ability and measurement precision of low signal-to-noise ratio dark and weak target;While being able to adapt to the characteristics of astronomical data of high dynamic range, significantly improve target detection performance, especially in the extremely dark and weak target in signal-to-noise ratio 2 to 5 interval.

Description

Technical Field

[0001] This invention relates to the field of astronomical image processing technology, specifically to a method for detecting faint targets in wide-area astronomical images based on deep learning. Background Technology

[0002] Small-aperture wide-area observation equipment has limited detection capabilities, resulting in significantly low signal-to-noise ratios (SNR) for the acquired target signals. Current mainstream methods for detecting faint targets mainly rely on traditional image processing techniques (such as SExtractor and DAOPHOT). However, these traditional methods face the challenge of balancing detection rate and false alarm rate when dealing with targets with extremely low SNR (e.g., SNR < 5). Lowering the threshold leads to a large number of noises being falsely detected as signals, while raising the threshold results in the missed detection of a large number of faint targets. Therefore, there is an urgent need for a deep learning-based method for detecting faint targets in wide-area astronomical images to address these issues. Summary of the Invention

[0003] The purpose of this invention is to provide a method for detecting faint targets in wide-area astronomical images based on deep learning, in order to solve the problems mentioned in the background art.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a method for detecting faint targets in wide-area astronomical images based on deep learning, comprising the following steps:

[0005] The original astronomical image is divided into fields of view and preprocessed.

[0006] For the preprocessed astronomical images, a threshold segmentation label generation strategy based on star catalog priors is established to generate a label set while preserving real faint target information;

[0007] A deep learning semantic segmentation model is constructed, and the model is trained based on preprocessed astronomical images and a label set to obtain an optimized model.

[0008] Using an optimization model, with the preprocessed astronomical image as input, the target segmentation map is output, and connected targets that do not conform to physical characteristics are removed to obtain the initial target mask;

[0009] The initial target mask is subjected to roundness feature determination and the target is separated and fused to obtain the final mask of the independent target;

[0010] The subpixel-level center coordinates of each independent target are calculated using the weighted centroid method based on the final mask.

[0011] Preferably, after performing pixel-by-pixel dark field image subtraction on the original astronomical image, the image field of view is divided into multiple non-overlapping small field-of-view images of size 512×512, centered on the image center pixel and extending outwards.

[0012] Preferably, the preprocessing includes dark field subtraction and non-uniform background suppression, and employs an adaptive contrast stretching strategy.

[0013] Preferably, a threshold segmentation label generation strategy based on star catalog priors is used to generate a deep learning model label set, including:

[0014] Calculate pixel thresholds and extract candidate target connected components from regions that exceed the thresholds;

[0015] The Gaia star catalog coordinates are mapped to image pixel coordinates. For each candidate target connected component and Gaia mapped coordinate position, the signal-to-noise ratio is calculated using the aperture photometry framework.

[0016] Targets with a signal-to-noise ratio not lower than a preset threshold of 2 are considered valid labels.

[0017] Preferably, the pixel threshold for:

[0018] ;

[0019] in, The background standard deviation, As an empirical factor, This represents the background mean.

[0020] Preferably, the radius of the wrapping circle of each source spot in the connected domain of the candidate target is extracted, and one pixel is added to the radius of the wrapping circle. The adjusted radius of the wrapping circle is used as the aperture of the Gaia mapping coordinate. The net signal after background subtraction is integrated to obtain the total number of electrons in the aperture. Then, the target signal-to-noise ratio is calculated according to the aperture photometry method.

[0021] Preferably, the deep learning semantic segmentation model introduces a multi-branch feature enhancement module. The multi-branch feature enhancement module adopts a parallel structure, simultaneously deploying a 3×3 standard convolutional branch and a 5×5 depthwise separable convolutional branch to capture the fine local features and contextual information of the target, respectively, and performs feature fusion through 1×1 convolution to achieve effective extraction of multi-scale features.

[0022] Preferably, the loss function used in training the deep learning semantic segmentation model includes cross-entropy loss and Dice loss, and a phased dynamic learning rate adjustment strategy is introduced to smoothly reduce the learning rate from the initial value to the lower limit of the learning rate. The phased dynamic learning rate adjustment strategy is as follows:

[0023] ;

[0024] in This represents the current learning round number. The total number of training rounds. .

[0025] Preferably, the target segmentation map is thresholded to obtain a binary mask, and target fields with fewer than a preset threshold of 2 pixels in the target mask are removed.

[0026] Preferably, the roundness feature determination and separation of the fusion target includes: calculating the roundness of each connected component and performing the following determination:

[0027] If the circularity is lower than the preset threshold of 0.7, the region is determined to be the fusion target. The internal pixel with the largest Euclidean distance from the background is found in the fusion target region as the seed point, and a step-by-step expansion strategy is adopted with this point as the center. The circularity of the current expansion region is calculated each time. If the circularity is lower than the preset threshold of 0.7, the expansion continues; otherwise, it stops.

[0028] Beneficial Effects: This invention utilizes a threshold segmentation label generation strategy based on star catalog priors, combined with an automated label generation method integrating threshold segmentation and Gaia star catalog localization, to obtain high-quality supervised learning labels. This strategy fully preserves real, faint target information. Furthermore, it introduces a staged dynamic learning rate adjustment strategy and a loss function strategy into the deep learning semantic segmentation model, enabling the model to learn the features of targets with extremely low signal-to-noise ratios (SNR), significantly improving the detection rate of faint targets in the SNR range of 2 to 5. Additionally, the invention's method of determining the roundness features of the initial target mask and separating and fusing targets effectively solves the target merging problem in dense star fields and achieves sub-pixel-level localization accuracy. Compared to traditional threshold segmentation methods, this invention effectively improves the recall capability for extremely faint targets while maintaining high accuracy, resulting in a significant improvement in overall faint target detection performance. Attached Figure Description

[0029] Figure 1 This is a flowchart of the deep learning-based wide-area astronomical image faint target detection method of the present invention;

[0030] Figure 2 This is a schematic diagram of the overall image preprocessing process of the present invention;

[0031] Figure 3 This is a partial schematic diagram illustrating the noise suppression effect brought about by the dark field subtraction in this invention;

[0032] Figure 4 This is a schematic diagram of the wide-area image field of view division according to the present invention;

[0033] Figure 5 This is a visual illustration of the image tags of the present invention;

[0034] Figure 6 This is a diagram of the deep learning model architecture used in this invention;

[0035] Figure 7This is a schematic diagram showing the changes of the training set and validation set on the overall loss function during the training process of the model of this invention;

[0036] Figure 8 This is an example diagram of the target separation algorithm based on circularity of the present invention.

[0037] Figure 9 The graph shows the statistical results of the ATD-DL method of the present invention and the traditional star detection methods SExtractor and DAOFind on various indicators.

[0038] Figure 10 This is a comparison chart of the mean values ​​of the three sets of data with the highest F1 scores between the ATD-DL method of this invention and the SExtractor and DAOFind methods;

[0039] Figure 11 This paper compares the recall rates of the ATD-DL method of the present invention with those of the traditional star detection methods SExtractor and DAOFind in the three sets of data with the highest F1 scores across various signal-to-noise ratio ranges.

[0040] Figure 12 This is a visual comparison of the target segmentation results of the ATD-DL of the present invention, the target segmentation results of SExtractor, and the original image.

[0041] Figure 13 This is a table showing five sets of parameter combinations with highly matched accuracy in an embodiment of the present invention. Detailed Implementation

[0042] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0043] Example: Figure 1 As shown, a method for detecting faint targets in wide-area astronomical images based on deep learning includes the following steps:

[0044] S1. Divide the field of view of the original astronomical image and perform preprocessing, referring to... Figure 2 As shown, it includes:

[0045] S11. Subtract the dark field image pixel by pixel from the acquired single-frame 1600×1600 pixel image, as follows: Figure 3 As shown, the discrete bright spots caused by the spatial difference of dark current are removed;

[0046] S12. The image obtained in step S11 is divided into multiple non-overlapping small field-of-view images of size 512×512, centered on the image's center pixel and extending outwards. Figure 4 As shown;

[0047] The resulting single-field-of-view image is then enhanced, including:

[0048] S11. By dividing the image into 50×50 pixel tile units, and then using the median statistic as an estimator within each tile, a spatially varying background model is constructed. The model is then subtracted pixel by pixel from the original image to obtain a target signal image in which the background is effectively suppressed.

[0049] S14. The ZScale algorithm is applied to the image obtained in step S13. This algorithm first performs random sampling of the input data and calculates the median and median absolute difference (MAD) to reduce the influence of outliers (such as bright stars). Then, iterative linear fitting and outlier removal processes are performed. Finally, the optimal display intensity range is generated based on the fitting results, and the image pixels are cropped so that weak signals can be linearly mapped to a wider dynamic range. Finally, a normalization formula is applied to generate a grayscale PNG image.

[0050] Based on the above, by combining dark field subtraction and non-uniform background suppression, and by adopting an adaptive contrast stretching strategy, the problem of contrast compression of weak signals caused by the high dynamic range of astronomical images is effectively solved, laying the foundation for image data.

[0051] S2. For the preprocessed astronomical images, a threshold segmentation label generation strategy based on star catalog priors is established to generate a label set while preserving real, faint target information; the threshold segmentation label generation strategy based on star catalog priors includes:

[0052] S21. Extract the connected components of candidate targets according to the threshold segmentation formula, and calculate the pixel threshold. Its expression is as follows:

[0053] ;

[0054] in The background standard deviation, As an empirical factor, As the background mean, in one embodiment, Set the threshold to 0.5 to extract more targets; extract candidate target connected components from regions larger than the threshold.

[0055] S22. Map the Gaia star catalog coordinates to image pixel coordinates. For each candidate target connected region and its Gaia mapped coordinate position, calculate its signal-to-noise ratio using the aperture photometric framework. First, extract the radius of the enclosing circle of each source spot in the candidate target connected region, and add 1 pixel to the enclosing circle radius, represented as:

[0056] ;

[0057] in For the first Spatial coordinates of pixels, Using the centroid as the coordinate system and the aperture as the reference point, integrate the net signal after background subtraction to obtain the total number of electrons within the aperture. Then the target signal-to-noise ratio Calculated using the classic aperture photometry method:

[0058] ;

[0059] in Accumulation within the aperture value, The aperture pixel area, The average value of the sky in concentric rings. Camera gain (e⁻ / ADU) The standard deviation of the background ring is given; this takes into account both source photon noise and background Poisson noise.

[0060] In one specific embodiment, the Gaia DR3 star catalog is used as a reference, and its sub-milliarsecond positions are mapped to image pixel positions. Since the detection limit of Gaia DR3 (G≈20.7 magnitude) is deeper than the observation image used in this embodiment, we use a fixed 3×3 pixel aperture centered on the pixel position mapped to each source for aperture metering and calculate its signal-to-noise ratio (SNR). The 3×3 pixel aperture is approximately twice the full width at half maximum (FWHM) of the point spread function (PSF) measured in the observation image of this embodiment, which can cover point sources. Finally, targets with an SNR not lower than a preset threshold of 2 are used as valid labels to generate the mask image required for model training, such as... Figure 5 As shown in the figure; green: labels obtained based on connected components; blue: target labels without connected components and with a signal-to-noise ratio that meets the requirements; red: target coordinates obtained by Gaia DR3 mapped to pixel-level center coordinates.

[0061] S3. Construct a deep learning semantic segmentation model, and train the deep learning semantic segmentation model based on the preprocessed astronomical images and label set to obtain an optimized model; specifically:

[0062] S1. A deep learning semantic segmentation model using an improved U-Net architecture, such as... Figure 6 As shown, the encoder part of the deep learning semantic segmentation model inherits the progressive downsampling idea, max pooling and double convolution structure of the classic U-Net model. It uses two levels of 3×3 convolution as feature upscaling units, and each layer sequentially completes spatial downsampling and channel expansion. However, the traditional UNet model uses double convolution operation with 3×3 convolution kernels in each layer. The effective receptive field after stacking is only 5×5. For dark targets, only a few pixels have effective information. Although the receptive field can cover the target itself, it lacks sufficient contextual information to distinguish the real target from image noise. To address this, a multi-branch Feature Enhancement Module was designed and introduced into the deep learning semantic segmentation model. This module adopts a parallel structure, simultaneously deploying a 3×3 standard convolutional branch and a 5×5 depthwise separable convolutional branch to capture the fine local features of the target and a wider range of contextual information, respectively. Finally, feature fusion is performed through 1×1 convolution to achieve effective extraction of multi-scale features, enhancing the network's ability to represent dim targets. In addition, the introduction of this depthwise separable convolution not only achieves dual receptive fields but also significantly reduces the number of parameters of the 5×5 large convolutional kernel, improving the model's training efficiency and generalization ability.

[0063] Specifically, the input image first extracts primary features through a multi-branch feature enhancement module; then, at each downsampling stage, the feature map resolution is halved by max pooling with a stride of 2, and then the double convolution operation is repeated, while the number of output channels is multiplied (64→128→256→512), forming a four-level cascaded encoding path; this design significantly reduces the computational cost while preserving edge details, and provides high-resolution structural priors for the upsampling path by directly feeding the features of each layer into the decoder through skip connections;

[0064] S32. Based on the model architecture constructed in step S31, the loss function used during training is a combination of cross-entropy loss and Dice loss, where the cross-entropy loss... The calculation expression is:

[0065] ;

[0066] in, ∈{0,1} represents the true label of pixel (i,j). ∈(0,1) represents the model's predicted probability. and These represent the height and width of the image, respectively; cross-entropy loss calculates the difference between the predicted probability and the true label pixel by pixel, mainly to ensure the accuracy of pixel-level classification, and is used to handle subtle differences between targets and backgrounds in astronomical images;

[0067] Dice loss The calculation expression is:

[0068] ;

[0069] Where N represents the batch size of the training process, i.e., the number of samples (images) processed in one training iteration, and M represents the total number of pixels in a single image. It is a real label (0 or 1). It is the probability value predicted by the model (between 0 and 1); Dice loss evaluates the spatial overlap between the predicted region and the real region, i.e., the overall recognition effect of the target region.

[0070] S33. Based on the model architecture of step S31, the training learning rate is set using a phased dynamic learning rate adjustment strategy through a cosine annealing scheduler, as follows:

[0071] During training, the learning rate is dynamically adjusted according to the following cosine annealing formula. :

[0072] ;

[0073] in This represents the current learning round number. The total number of training rounds. The cosine annealing scheduler uses a cosine function to smoothly and non-linearly decay the learning rate from its initial value to its minimum value within a training cycle. This strategy allows the learning rate to smoothly decrease from its initial value to the lower limit of the learning rate, enabling the model to converge quickly in the early stages of training and finely adjust the parameters in the later stages.

[0074] S34. Based on the overall model architecture and training strategy formed in steps S31, S32, and S33, the model loss obtained in this embodiment is as follows: Figure 7 As shown, the training loss rapidly decreased from its initial value of 1.287, dropping to approximately 0.72 within the first 20 epochs, and then entered a slow decay phase, eventually stabilizing at around 0.565 after 100 epochs, demonstrating good convergence and stability overall. The validation loss fluctuated somewhat during the middle of training, but still showed a downward trend, eventually converging to a level similar to the training loss. The difference between the two gradually narrowed as training progressed, ultimately maintaining a small difference of approximately 0.01. This evolutionary characteristic indicates that the model did not exhibit overfitting during training.

[0075] S4. Using the optimization model, the preprocessed astronomical image is used as input to output the target segmentation map, and connected targets that do not conform to physical characteristics are removed to obtain the initial target mask.

[0076] In one embodiment, the target segmentation map is thresholded to obtain a binary mask, and target fields with fewer than a preset threshold of 2 pixels in the target mask are removed.

[0077] S5. Perform roundness feature determination on the initial target mask and separate and fuse the targets to obtain the final mask for each independent target; specifically:

[0078] S51. For each connected component, first calculate its circularity. As a criterion for adhesion, roundness is defined as:

[0079] ;

[0080] in For the area, The perimeter of the region;

[0081] S52. Roundness determination: If the roundness is below a preset threshold of 0.7, the region is determined to be a fusion target. The internal pixel with the largest Euclidean distance from the background is selected as the seed point within the fusion target region. Using this point as the center, a step-by-step expansion strategy is adopted. At each expansion, the roundness of the current expansion region is calculated. If the roundness is below the preset threshold of 0.7, expansion continues; otherwise, it stops. This process separates the first sub-target. The remaining region, after morphological processing, forms the second sub-target. Finally, an independent sub-target mask separated from the fusion target is obtained. Figure 8 As shown.

[0082] S6. Based on the final mask, the weighted centroid method is used to calculate the sub-pixel-level center coordinates of each independent target, specifically:

[0083] The sub-pixel center of the target is calculated using the weighted centroid method. The definition is as follows:

[0084] ;

[0085] in Represents a set of pixels within the same connected component. The electron count is performed at the corresponding position; finally, a sub-pixel-level center detection result of the target is generated, which is consistent with the coordinate format of the traditional method.

[0086] This application also provides a specific experiment to comprehensively analyze the effectiveness and superiority of the methods provided above; see details below. Figure 9 , Figure 10 , Figure 11 , Figure 12 , Figure 13The diagram shows a comparison of the results of traditional astronomical target detection methods SExtractor and DAOPHOT (which actually uses its star target detection algorithm DAOFind) based on the method of this invention, as well as the beneficial effects of the multi-branch feature enhancement module developed in this paper for the detection of faint targets.

[0087] Depend on Figure 9 It can be seen that the ATD-DL method of this invention maintains stable recall and accuracy, and on average is higher than the two traditional methods. When the F1 score reaches its respective maximum value, that is, 0.792 for the ATD-DL method of this invention, it is 12.18% and 50.00% higher than 0.706 for SExtractor and 0.528 for DAOFind, respectively. Regarding the robustness of the model's detection capability (… Figure 10 For each method, the three sets of data with the highest F1 scores were selected and their averages were calculated. The ATD-DL method of this invention showed more stable performance across various metrics compared to the traditional method, with an average score of 0.791, representing improvements of 15.14% and 50.67% compared to SExtractor's 0.687 and DAOFind's 0.525, respectively. To demonstrate the ability of the method of this invention in detecting extremely faint targets, the recall rate of each model at its maximum F1 score within various small-scale signal-to-noise ratio ranges was statistically analyzed with a step size of 1. Figure 11 Compared to SExtractor, the recall rate of the ATD-DL method of this invention increases significantly as the signal-to-noise ratio (SNR) decreases. The increase gradually rises from 3.02% in the SNR greater than 10 range to 20.476% in the SNR 3-4 range, indicating that the performance improvement of ATD-DL is largely attributed to the detection rate of extremely faint targets. Both ATD-DL and SExtractor extract a unified centroid from the segmentation results. However, SExtractor's detection performance suffers from excessive threshold lowering, causing continuous mask expansion and severe centroid deviation, resulting in correctly segmented targets being assigned incorrect coordinates.

[0088] Specifically, compared to the SExtractor centroid result corresponding to the F1 peak, ATD-DL improves the precision and recall metrics by 10.00% and 24.56%, respectively. A direct comparison is made between the segmentation masks of the two methods and the original image at the optimal F1 score. Figure 12 As can be seen, SExtractor tends to enlarge the area of ​​bright sources and incorporate surrounding low signal-to-noise ratio targets into the same connected domain, which significantly weakens its ability to segment dark and weak targets.

[0089] To rigorously evaluate the effectiveness of the Feature Enhancement Module (FEM) proposed by ATD-DL for weak target detection, ablation experiments were specifically conducted. The baseline model adopts the standard U-Net architecture, with its first encoder layer using a traditional 3×3 dual convolutional module. To avoid the overall performance metrics masking the improvement in weak target detection, a more targeted evaluation strategy was adopted, focusing only on the number of weak targets detected with a signal-to-noise ratio (SNR) between 2 and 5. To eliminate the influence of different accuracy levels, five parameter combinations with highly matched accuracy were selected in the comparison between the baseline U-Net and the enhanced model (UNet+FEM). The corresponding results are shown in [Figure number missing]. Figure 13 The results show that, while maintaining comparable overall accuracy (e.g., 97.49% for U-Net and 97.50% for UNet+FEM in parameter number 1), the enhanced model significantly improved the number of detected dark targets in all comparison parameters. Notably, even under highly stringent operating conditions exceeding 99% accuracy, FEM maintained strong sensitivity to dark sources. Particularly in parameter number 5, the number of detected dark targets nearly doubled. These results demonstrate that FEM effectively enhances the model's ability to learn and identify features of dark targets without sacrificing overall detection accuracy, thereby significantly improving the detection rate of dark sources.

[0090] The embodiments of the present invention have been described in detail above with reference to the examples. However, the present invention is not limited to the above embodiments. For those skilled in the art, after learning the contents described in the present invention, several equivalent changes and substitutions can be made without departing from the principle of the present invention. These equivalent changes and substitutions should also be considered to fall within the protection scope of the present invention.

Claims

1. A method for detecting faint targets in wide-area astronomical images based on deep learning, characterized in that: Includes the following steps: The original astronomical image is divided into fields of view and preprocessed. For the preprocessed astronomical images, a threshold segmentation label generation strategy based on star catalog priors is established to generate a label set while preserving real faint target information; A deep learning semantic segmentation model is constructed, and the model is trained based on preprocessed astronomical images and a label set to obtain an optimized model. Using an optimization model, with the preprocessed astronomical image as input, the target segmentation map is output, and connected targets that do not conform to physical characteristics are removed to obtain the initial target mask; The initial target mask is subjected to roundness feature determination and the target is separated and fused to obtain the final mask of the independent target; The subpixel-level center coordinates of each independent target are calculated using the weighted centroid method based on the final mask.

2. The method for detecting faint targets in wide-area astronomical images based on deep learning according to claim 1, characterized in that: After performing pixel-by-pixel dark field image subtraction on the original astronomical image, the image field of view is divided into multiple non-overlapping small field-of-view images of size 512×512, centered on the image center pixel and extending outwards.

3. The method for detecting faint targets in wide-area astronomical images based on deep learning according to claim 1, characterized in that: The preprocessing includes dark field subtraction and non-uniform background suppression, and employs an adaptive contrast stretching strategy.

4. The method for detecting faint targets in wide-area astronomical images based on deep learning according to claim 1, characterized in that: A threshold segmentation label generation strategy based on star catalog priors is used to generate a label set for deep learning models, including: Calculate pixel thresholds and extract candidate target connected components from regions that exceed the thresholds; The Gaia star catalog coordinates are mapped to image pixel coordinates. For each candidate target connected component and Gaia mapped coordinate position, the signal-to-noise ratio is calculated using the aperture photometry framework. Targets with a signal-to-noise ratio not lower than a preset threshold of 2 are considered valid labels.

5. The method for detecting faint targets in wide-area astronomical images based on deep learning according to claim 4, characterized in that: The pixel threshold for: ； in, The background standard deviation, As an empirical factor, This represents the background mean.

6. The method for detecting faint targets in wide-area astronomical images based on deep learning according to claim 4, characterized in that: Extract the radius of the wrapping circle of each source spot in the connected domain of the candidate target, and add 1 pixel to the radius of the wrapping circle. Use the adjusted radius of the wrapping circle as the aperture of the Gaia mapping coordinate. Integrate the net signal after background subtraction to obtain the total number of electrons in the aperture. Then calculate the target signal-to-noise ratio in the form of aperture photometry.

7. The method for detecting faint targets in wide-area astronomical images based on deep learning according to claim 1, characterized in that: The deep learning semantic segmentation model introduces a multi-branch feature enhancement module. The multi-branch feature enhancement module adopts a parallel structure, deploying a 3×3 standard convolutional branch and a 5×5 depthwise separable convolutional branch to capture the fine local features and contextual information of the target respectively. It also performs feature fusion through 1×1 convolution to achieve effective extraction of multi-scale features.

8. The method for detecting faint targets in wide-area astronomical images based on deep learning according to claim 7, characterized in that: The deep learning semantic segmentation model uses cross-entropy loss and Dice loss as loss functions during training, and introduces a phased dynamic learning rate adjustment strategy to smoothly decrease the learning rate from its initial value to the lower bound. The phased dynamic learning rate adjustment strategy is as follows: ； in This represents the current learning round number. The total number of training rounds. .

9. The method for detecting faint targets in wide-area astronomical images based on deep learning according to claim 1, characterized in that: The target segmentation map is thresholded to obtain a binary mask, and target fields with fewer than 2 pixels in the target mask are removed.

10. The method for detecting faint targets in wide-area astronomical images based on deep learning according to claim 1, characterized in that: The roundness feature determination and separation of the fusion target includes: calculating the roundness of each connected component and performing the following determination: If the circularity is lower than the preset threshold of 0.7, the region is determined to be the fusion target. The internal pixel with the largest Euclidean distance from the background is found in the fusion target region as the seed point, and a step-by-step expansion strategy is adopted with this point as the center. The circularity of the current expansion region is calculated each time. If the circularity is lower than the preset threshold of 0.7, the expansion continues; otherwise, it stops.

Images