Construction method of astronomical image point source target detection model with depth feature matching

Abstract

The invention belongs to the field of target detection, and particularly relates to a construction method of an astronomical image point source target detection model with depth feature matching. The method comprises the following steps: obtaining a training set of astronomical images containing a point source target; the training set of the astronomical image containing the point source target is input into a PSDetNet network, and the PSDetNet network is composed of a feature extraction module and a matching positioning module; utilizing a feature extraction module to obtain a feature map with point source template features; searching and positioning point source features in the feature map by using a matching and positioning module to obtain a target detection model; the model is trained using an improved cross entropy loss function. According to the invention, high integrity and purity can still be maintained at a low signal-to-noise ratio; an astronomical image with a point source target can be converted into a feature map with a point source similar to a point source template; the network efficiency can be effectively improved, and the point source in the astronomical image can be accurately detected with high purity and high integrity.

Description

Technical Field

[0001] The present invention belongs to the field of target detection, and in particular relates to a method for constructing an astronomical image point source target detection model with deep feature matching. Background Art

[0002] Existing methods for detecting point source targets in astronomical images usually use traditional methods and deep learning methods. In the traditional method, the point target detection process disclosed by "SExtractor (Source Extractor), a classic astronomical image processing tool" is carried out in steps, including background estimation, threshold detection, anti-aliasing and source measurement. The problems are: 1. SExtractor performs poorly when detecting point sources with low signal-to-noise ratio (SNR), especially under weak luminosity point sources or low-contrast targets, and is prone to missed detection. This is because its threshold-based method is insensitive to weak signals. 2. When dealing with overlapping or close point sources, although SExtractor has anti-aliasing function, its anti-aliasing effect is limited in very dense scenes, and it is easy to misjudge multiple point sources as one source or mis-segment. 3. SExtractor relies on a large number of manual parameter adjustments (such as threshold setting, background model, etc.), which need to be tuned for each data set, which may perform inconsistently under different observation conditions, increasing the complexity of use. In the deep learning method, a "deep learning-based astronomical image detection and segmentation model (DeepDISC method)" is provided, which relies on the Detectron2 framework, combines the capabilities of instance segmentation and classification, and can simultaneously detect and classify astronomical targets such as galaxies and stars. DeepDISC uses a convolutional neural network architecture based on ResNet and Transformer, and uses pre-trained models and transfer learning to accelerate model training and improve model performance. The problems are: 1. Poor performance when processing weak signals and high-noise data: DeepDISC's detection ability will be significantly reduced when processing very weak point sources, especially when the signal-to-noise ratio is low. Insufficient dependence on template matching: 2. DeepDISC's instance segmentation based on Mask-RCNN mainly relies on the features learned by the network for target detection, but when processing point sources, it lacks matching for specific templates, resulting in inaccuracy when processing very small or blurred targets. 3. Limited detection ability for large-scale targets: Although DeepDISC can handle more common galaxies and stars, it shows certain limitations when facing very small point sources or large-scale overlapping stars. “The DCAT (Detection and Classification of Astronomical Targets) method is a deep learning framework designed for small aperture wide field telescopes (WFSATs) to automatically detect and classify astronomical targets. It is based on the concept of Faster R-CNN and extracts features from images through a modified ResNet-50 backbone network and a Feature Pyramid Network (FPN) to improve the detection capability of low signal-to-noise ratio (S / N) and complex astronomical scenes.The existing problems are: 1. The detection capability of weak point sources is limited. In complex background or weak signal environment, the accuracy of DCAT decreases, especially when the point sources are sparse and the signals are weak, it is easy to miss the detection. 2. The limitation of fixed image size: DCAT will force the candidate regions of different sizes to be scaled to a fixed size during the ROI pooling process, which may cause the position information of small targets to be lost or the error to increase. This limitation reduces the accuracy of DCAT in detecting small targets. 3. Insufficient adaptability to complex scenes: DCAT performs poorly when dealing with complex astronomical scenes (such as high-density or overlapping targets), especially the lack of anti-aliasing capability, which can easily confuse multiple sources as one target. Summary of the invention

[0003] The present invention provides a method for constructing an astronomical image point source target detection model with deep feature matching, so as to overcome the problems existing in the prior art of limited detection capability of weak point sources and insufficient adaptability to complex scenes.

[0004] In order to achieve the purpose of the present invention, the technical solution provided by the present invention is: a method for constructing an astronomical image point source target detection model with deep feature matching, the steps are as follows:

[0005] Step 1: Get a training set of astronomical images containing point source targets: Use the public dataset AutoSourceID-Light (ASID-L) to mark the position of the point source in each ASID-L image, set the corresponding element in the similarity label to 1, and set other elements to 0, and randomly crop the image. During the cropping process, the similarity label of the cropped image will be transformed according to the original image label. At the same time, the image is enhanced by performing operations such as horizontal and vertical flipping, and the label is adjusted accordingly, thereby obtaining a training set of astronomical images containing point source targets;

[0006] Step 2: A training set of astronomical images containing point source targets is input into a PSDetNet network, wherein the PSDetNet network is composed of a feature extraction module and a matching and positioning module;

[0007] Step 3: Use the feature extraction module to obtain a feature map with point source template features;

[0008] Step 4: Use the matching and positioning module to search and locate the point source features in the feature map to obtain the target detection model;

[0009] Step 5: Use the improved cross entropy loss function to train the model. The formula of the loss function is:

[0010]

[0011] Furthermore, the feature extraction module is an encoder-decoder network structure composed of full convolutional layers, wherein the encoder is composed of 8 residual blocks, each residual block contains 2 convolutional layers; each pair of residual blocks constitutes a single residual block group, in each convolutional layer within the group, the number of neurons remains unchanged, while the number of neurons between groups is doubled successively; the decoder adopts the mirror structure of the encoder, and is also composed of 4 residual block groups.

[0012] Furthermore, the design steps of the above matching and positioning module are as follows:

[0013] (i) Constructing feature dictionary: Using a sliding window of size m×n, starting from the upper left corner of the feature map M×N, and moving to the lower right corner in steps of k. During the sliding process, the feature blocks covered by the window are flattened into feature vectors;

[0014] (ii) Calculate similarity: Flatten the point source template to obtain the template vector V. Perform correlation calculation on each feature vector in the feature dictionary and the template vector to generate a similarity vector S, which consists of M / k×N / k, where each element represents the similarity between the corresponding image block and the point source template;

[0015] (III) Determine the source location: Use statistical methods to select the threshold t, perform threshold processing on the similarity vector S, and obtain S 01 ; After threshold processing, determine S 01 The image block corresponding to the "1" element in contains a point source, and the center of the point source is located at the center of the image block.

[0016] Furthermore, in the above step 4, the point source template is obtained by collecting representative astronomical point source image samples, manually pre-processing the samples to align the source center, aligning and superimposing these samples to obtain the point source template.

[0017] Compared with the prior art, the advantages of the present invention are:

[0018] 1. PSDetNet of the present invention consists of two modules: feature extraction and matching positioning. It adopts a more complex feature matching and template matching positioning method, and can still maintain high integrity and purity under low signal-to-noise ratio:

[0019] In the feature extraction module, the encoder consists of 8 residual blocks, each of which contains 2 convolutional layers. Each pair of residual blocks constitutes a single residual block group. In each convolutional layer within the group, the number of neurons remains unchanged, while the number of neurons between groups doubles successively. The decoder adopts the mirror structure of the encoder and is also composed of 4 residual block groups. The number of neurons in each group gradually decreases until it reaches 1 / 8 of the input layer size, while the size of the feature map in each group increases to 8 times the input layer size; the design steps of the matching and positioning module are added: (1) Feature dictionary construction: Scan the feature map through a sliding window, convert each covered image block into a feature vector, and store it in the feature dictionary. Each feature vector represents the feature of an image block centered on a certain pixel in the original image. (2) Similarity calculation: Flatten the point source template into a template vector, perform correlation calculation with the feature vector in the feature dictionary, and generate a similarity vector. Each element represents the similarity between the corresponding image block and the point source template. (3) Determine the source position: Select a threshold based on a statistical method to filter out image blocks with a similarity higher than the threshold. The area containing the point source is determined by these image blocks, and the center of the source is located by 2D coordinates; the point source can be effectively identified, and through the design of the sliding window, the sliding window matching positioning and feature extraction method, the point source can be well identified at various scales, and the theoretical error is controlled within 2 pixels with high accuracy. Therefore, the network of the present invention can convert an astronomical image with a point source target into a feature map similar to the point source and the point source template.

[0020] 2. The PSDetNet provided by the present invention is composed of two modules, constructed using a full convolutional layer, and can be trained using small-size samples, so it can process images of any size and can perform end-to-end point source target recognition in astronomical images. The following three are combined to effectively improve network efficiency. The model can accurately detect point sources in astronomical images with high purity and integrity: (1) The feature map generated by feature extraction is matched with the template through a sliding window to obtain a similarity vector. This similarity vector is used as the output of the network. (2) The loss function calculates the error by comparing the predicted similarity vector (model output) with the true label. Due to the sparsity of the point source, the loss function strengthens the detection of the point source through a higher weight w. Its introduction allows the error weight between the point source and the background to be adjusted in the loss function, thereby reducing the impact on the background area and helping the model to learn the features of the point source more effectively. (3) Back propagation: Through the back propagation algorithm, the gradient information of the loss function is fed back to each layer of the network, thereby adjusting the parameters in the feature extraction module so that it can better extract features that help to accurately locate the point source. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is the network structure diagram of PSDetNet;

[0022] Figure 2 Create process maps for datasets;

[0023] Figure 3 This is the acquisition diagram of the point source template. DETAILED DESCRIPTION

[0024] The technical solution of the present invention is further described below by specific implementation methods. It should be understood by those skilled in the art that the embodiments are only used to help understand the present invention and should not be regarded as specific limitations of the present invention.

[0025] A method for constructing a detection model for astronomical image point source target detection with deep feature matching, the steps are as follows:

[0026] Step 1: Get a training set of astronomical images containing point source targets:

[0027] The present invention utilizes the public dataset AutoSourceID-Light (ASID-L).

[0028] ASID-L consists of images taken by the MeerLICHT telescope, which has a 65 cm optical telescope with a field of view of 2.7 square degrees and is equipped with a 10.5k × 10.5k pixel CCD. The images in ASID-L were acquired from fields with different source densities, including fields centered on the Omega Cen globular cluster, the Fornax galaxy cluster, and an "empty" field centered on the Chandra Deep Field-South (CDF-S). The positions and magnitudes of point sources in these images were determined using the Gaia Early Data Version 3 catalog as a reference.

[0029] To construct the dataset, the present invention annotated the location of the point source in each ASID-L image, set the corresponding element in the similarity label to 1, and set the other elements to 0. Subsequently, these images were randomly cropped to obtain approximately 50,000 small images with a size of 256×256 pixels. During the cropping process, the similarity labels of the cropped images were transformed according to the original image labels, and these images were also enhanced by performing operations such as horizontal and vertical flipping, and the labels were adjusted accordingly, resulting in more than 200,000 samples, resulting in a training set of astronomical images containing point source targets. See Figure 2 .

[0030] Step 2: The training set of astronomical images containing point source targets is input into the PSDetNet network. The PSDetNet network structure is shown in the figure below: Figure 1As shown in FIG. 1 , the PSDetNet network consists of a feature extraction module and a matching and positioning module. The feature extraction module takes the astronomical image as input and generates an output feature map. It adopts a mirror encoder-decoder structure and consists of residual blocks. The matching and positioning module includes matching the point source template with the feature block to detect whether the feature block contains a point source target. If so, its position is accurately determined. The specific description is as follows:

[0031] The feature extraction module is designed as an encoder-decoder network structure consisting of fully convolutional layers, where the encoder consists of 8 residual blocks, each of which contains 2 convolutional layers. Each pair of residual blocks constitutes a single residual block group. In each convolutional layer within a group, the number of neurons remains unchanged, while the number of neurons between groups is doubled successively. The decoder adopts the mirror structure of the encoder and is also composed of 4 residual block groups. The number of neurons in each group gradually decreases until it reaches 1 / 8 of the input layer size, while the feature map size in each group increases to 8 times the input layer size.

[0032] The matching and positioning module is designed to construct a feature dictionary, calculate similarity, and determine the location of the point source target in three steps:

[0033] 1. Construct a feature dictionary: Use a sliding window of size m×n, starting from the upper left corner of the feature map M*N, and moving to the lower right corner in steps of k. During the sliding process, the feature blocks covered by the window are flattened into feature vectors. Therefore, a feature dictionary D of size (m×n)×(M / k×N / k) can be constructed. Each vector in the feature dictionary represents the feature of an image block centered on a pixel in the original image.

[0034] 2. Calculate similarity: Flatten the point source template to obtain the template vector V. Each feature vector in the feature dictionary is correlated with the template vector to generate a similarity vector S. S consists of m / k×n / k, where each element represents the similarity between the corresponding image block and the point source template, indicating the possibility that the image block contains a point source.

[0035] In this embodiment, 2069 representative astronomical point source image samples were collected. These samples were manually preprocessed to align the point source center. These samples were aligned and superimposed to obtain a point source template. The point source template production process is as follows: Figure 3 shown.

[0036] 3. Determine the source location: Use statistical methods to select the threshold t, perform threshold processing on the similarity vector S, and obtain S01. After threshold processing, it is determined that the image block corresponding to the "1" element in S01 contains a point source, and the center of the point source is located at the center of the image block.

[0037] Step 3: Use the feature extraction module to obtain a feature map with point source template features;

[0038] Step 4: Use the matching and positioning module to search and locate the point source features in the feature map to obtain the target detection model.

[0039] Step 5: Use the improved cross entropy loss function to train the model. The formula of the loss function is:

[0040]

[0041] Where: w is the adjustment parameter, set to 0.99, is a sample image, is the predicted similarity vector.

[0042] The method for using the target detection model constructed by the present invention is as follows: after the constructed target detection model PSDetNet network is input with an astronomical image, a feature extraction module is first used to convert the astronomical image with a point source target into a feature map similar to the point source target and the point source template, and then a matching positioning module is used to compare with the pre-constructed point source template in a block-by-block manner, so that the specific position of the point source target in the astronomical image can be obtained.

[0043] In order to evaluate the generalization ability of the model on heterogeneous data, we collected a total of 1,203 astronomical images containing point source targets captured by telescopes such as Hubble and Webb. These images were randomly cropped to different sizes to test the model's ability to handle multi-resolution images. The results show that PSDetNet performs well with a purity of 0.97 and a completeness of 0.93, demonstrating its excellent performance in detecting true point sources and effectively reducing false positives.

[0044] The above are only specific embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for constructing an astronomical image point source target detection model with deep feature matching, characterized in that: Here are the steps: Step 1: Get a training set of astronomical images containing point source targets: Use the public dataset AutoSourceID-Light to mark the position of the point source in each ASID-L image, set the corresponding element in the similarity label to 1, and set other elements to 0, and randomly crop the image. During the cropping process, the similarity label of the cropped image will be transformed according to the original image label. At the same time, the image is enhanced by performing operations such as horizontal and vertical flipping, and the label is adjusted accordingly, thereby obtaining a training set of astronomical images containing point source targets; Step 2: A training set of astronomical images containing point source targets is input into a PSDetNet network, wherein the PSDetNet network is composed of a feature extraction module and a matching and positioning module; Step 3: Use the feature extraction module to obtain a feature map with point source template features; Step 4: Use the matching and positioning module to search and locate the point source features in the feature map to obtain the target detection model; Step 5: Use the improved cross entropy loss function to train the model. The formula of the loss function is:

2. The method for constructing an astronomical image point source target detection model with deep feature matching according to claim 1, characterized in that: The feature extraction module in step 2 is an encoder-decoder network structure composed of full convolutional layers, wherein the encoder is composed of 8 residual blocks, each residual block contains 2 convolutional layers; each pair of residual blocks constitutes a single residual block group, in each convolutional layer within the group, the number of neurons remains unchanged, while the number of neurons between groups is doubled successively; the decoder adopts a mirror structure of the encoder, consisting of 4 residual block groups.

3. The method for constructing an astronomical image point source target detection model with deep feature matching according to claim 2, characterized in that: The design steps of the matching positioning module in step 2 are as follows: (i) Constructing a feature dictionary: Using a sliding window of size m×n, starting from the upper left corner of the feature map M*N, and moving to the lower right corner in steps of k; during the sliding process, the feature blocks covered by the window are flattened into feature vectors; (ii) Calculating similarity: Flatten the point source template to obtain the template vector V, calculate the correlation between each feature vector in the feature dictionary and the template vector, and generate a similarity vector S, which consists of M / k×N / k, where each element represents the similarity between the corresponding image block and the point source template; (III) Determine the source location: Use statistical methods to select the threshold t, perform threshold processing on the similarity vector S, and obtain S 01 ; After threshold processing, determine S 01 The image block corresponding to the "1" element contains a point source, and the center of the point source is located at the center of the image block.

4. The method for constructing an astronomical image point source target detection model with deep feature matching according to claim 3, characterized in that: In step 4, the point source template is obtained by collecting representative astronomical point source image samples, manually pre-processing the samples to align the source center, aligning and superimposing the samples to obtain the point source template.

Images