Galaxy morphology recognition method and device based on physical priori and deep visual features

Abstract

This application relates to the fields of astronomical data processing and deep learning technology, and provides a method and apparatus for galaxy morphology recognition based on physical priors and deep visual features. The method obtains orientation-robust galaxy representations through multi-scale rotational equivariant feature extraction, and establishes visual and physical branches to jointly model morphological features and structural parameters. The physical branch incorporates astronomical structural parameters as auxiliary supervision to enhance physical interpretability. A gated fusion mechanism is used to achieve adaptive fusion of features from both branches, and end-to-end training is performed using classification loss and structural parameter regression loss to achieve automated galaxy morphology recognition. This method effectively improves the accuracy of fine-grained galaxy morphology classification and the physical interpretability of the model, outperforming existing state-of-the-art methods on some existing datasets, and has significant application value.

Description

Technical Field

[0001] This application relates to the fields of astronomical data processing and deep learning technology, and in particular to a method and apparatus for galaxy morphology recognition based on physical priors and deep visual features. Background Technology

[0002] Galaxietypology classification is a fundamental issue in astrophysical research and a crucial observational tool for understanding galaxy formation mechanisms, evolutionary paths, and the large-scale structure of the universe. In statistical astronomy, galaxy morphology directly impacts sample construction and the reliability of subsequent scientific inferences. Constructing a high-precision and consistent galaxy morphology classification system is of great significance for conducting large-scale sky survey data analysis.

[0003] Traditional galaxy classification methods rely on expert visual interpretation, which faces significant challenges in terms of processing efficiency, consistency, and scalability, making it difficult to meet the needs of large-scale sky survey missions. Some related technologies construct feature representations based on the structural parameters of galaxy images and combine them with traditional machine learning methods for classification; others introduce deep learning methods for galaxy classification.

[0004] However, most existing deep learning methods focus on modeling image texture and appearance features, and still have significant shortcomings when facing fine-grained galaxy morphology classification tasks. For example, existing models often struggle to reliably distinguish subtle structural features such as differences in spiral arm structures, the presence or absence of bar structures, and changes in the disk-nucleus ratio. Furthermore, most existing deep learning models rely solely on visual image features for discrimination, failing to explicitly incorporate structural parameters with clear physical significance in astronomy. This results in difficult-to-interpret model results and makes it difficult to effectively integrate with existing physical theories. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method and apparatus for galaxy morphology recognition based on physical priors and depth visual features, in order to solve the problems of insufficient accuracy and weak interpretability of existing galaxy morphology classification and recognition algorithms.

[0006] A first aspect of this application provides a method for galaxy morphology recognition based on physical priors and depth visual features, including:

[0007] Acquire galaxy image set data, perform preprocessing and enhancement processing on the galaxy image set data to obtain galaxy image training set and galaxy image test set;

[0008] A galaxy morphology classification network model based on a dual-branch structure is constructed. The galaxy morphology classification network model includes a feature extraction network, a first branch network, and a second branch network. The feature extraction network includes at least a multi-scale rotational equivariant feature extraction module and a feature pyramid network. These two modules are used to extract multi-scale and multi-scale rotational equivariant features from galaxy images. The first branch network is a visual branch network used to extract the morphological visual features of galaxy images. The second branch network is a physical branch network used to predict physical structure parameters.

[0009] A galaxy morphology classification network model was trained using a galaxy image training set and a galaxy image test set; the objective function of the model training was a weighted combination of the classification loss function and the physical structure parameter regression loss function.

[0010] Acquire images of the galaxies to be identified, and use a trained galaxy morphology classification network model to determine the morphology of the galaxies based on the images.

[0011] A second aspect of this application provides a galaxy morphology recognition device based on physical priors and depth visual features, comprising:

[0012] The dataset building module is configured to acquire galaxy image set data, preprocess and enhance the galaxy image set data to obtain galaxy image training set and galaxy image test set;

[0013] The model building module is configured to construct a galaxy morphology classification network model based on a dual-branch structure. The galaxy morphology classification network model includes a feature extraction network, a first branch network, and a second branch network. The feature extraction network includes at least a multi-scale rotational equivariant feature extraction module and a feature pyramid network. These two modules are used to extract multi-scale and multi-scale rotational equivariant features from galaxy images. The first branch network is a visual branch network used to extract the morphological visual features of galaxy images. The second branch network is a physical branch network used to predict physical structure parameters.

[0014] The training module is configured to train a galaxy morphology classification network model using a galaxy image training set and a galaxy image test set; wherein the objective function for model training is a weighted combination of the classification loss function and the physical structure parameter regression loss function.

[0015] The identification module is configured to acquire images of galaxies to be identified and use a trained galaxy morphology classification network model to determine the morphology of the galaxies to be identified based on the images.

[0016] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0017] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0018] The beneficial effects of this application's embodiments compared to existing technologies are as follows: This application's embodiments obtain orientation-robust galaxy representations through multi-scale rotational equivariant feature extraction, and establish visual and physical branches to jointly model morphological features and structural parameters. The physical branch incorporates astronomical structural parameters as auxiliary supervision to enhance physical interpretability. A gated fusion mechanism is used to achieve adaptive fusion of features from both branches, and end-to-end training is performed using classification loss and structural parameter regression loss to achieve automated galaxy morphology recognition. This method effectively improves the accuracy of fine-grained galaxy morphology classification and the physical interpretability of the model, outperforming existing state-of-the-art methods on some existing datasets, and has significant application value. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a galaxy morphology recognition method based on physical priors and depth visual features provided in an embodiment of this application.

[0021] Figure 2 This is a schematic diagram of the feature extraction network provided in the embodiments of this application.

[0022] Figure 3 This is a schematic diagram of the visual-physical dual-branch feature encoding and fusion structure provided in the embodiments of this application.

[0023] Figure 4 This is a flowchart illustrating another galaxy morphology recognition method based on physical priors and depth visual features provided in this application embodiment.

[0024] Figure 5 This is a schematic diagram of the t-SNE visualization results based on the fusion features of the final classifier provided in the embodiments of this application.

[0025] Figure 6This is a schematic diagram of the t-SNE visualization results based on visual branch features provided in the embodiments of this application.

[0026] Figure 7 This is a schematic diagram of the t-SNE visualization results based on physical branch features provided in the embodiments of this application.

[0027] Figure 8 This is a schematic diagram of a galaxy morphology recognition device based on physical priors and depth visual features provided in an embodiment of this application.

[0028] Figure 9 This is a schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0029] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0030] The following will describe in detail, with reference to the accompanying drawings, a method and apparatus for galaxy morphology recognition based on physical priors and depth visual features according to embodiments of this application.

[0031] As mentioned above, some solutions have incorporated deep learning methods for galaxy classification. During the paradigm-forming phase, Dieleman et al. were the first to use a Convolutional Neural Network (CNN) to directly learn the problem tree from the citizen science project GZ (Galaxy Zoo), validating the feasibility of end-to-end deep learning in galaxy morphology classification. Subsequently, Domínguez Sánchez et al. systematized and engineered this paradigm, constructing a large-scale morphology catalog covering approximately 670,000 galaxies from the Sloan Digital Sky Survey (SDSS). Its predictive performance on the Hubble T-type regression task significantly outperformed traditional methods. To further improve classification accuracy, research focused on the continuous optimization of network structures: from residual network variants and customized CNNs, to network structures incorporating neural frequent differential equations to characterize continuous dynamic properties, and finally to the Convolutional Vision Transformer (CvT) architecture that integrates the local receptive field of CNNs with the global modeling capabilities of the Transformer. These methods have gradually improved accuracy to over 98% in coarse classification tasks.

[0032] However, most deep learning methods focus on modeling image texture and appearance features, and still have significant shortcomings when facing fine-grained galaxy morphology classification tasks.

[0033] In view of this, embodiments of this application provide a galaxy morphology recognition method based on physical priors and deep visual features. This method can effectively integrate physical structure priors in a deep learning framework and improve the model's ability to recognize fine-grained galaxy morphology, so as to achieve high-precision, interpretable and scalable automated processing of large-scale survey data.

[0034] Figure 1 This is a flowchart illustrating a galaxy morphology recognition method based on physical priors and depth visual features provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps:

[0035] In step S101, galaxy image set data is acquired, and the galaxy image set data is preprocessed and enhanced to obtain galaxy image training set and galaxy image test set.

[0036] In step S102, a galaxy morphology classification network model based on a dual-branch structure is constructed.

[0037] The galaxy morphology classification network model includes a feature extraction network, a first branch network, and a second branch network. The feature extraction network includes at least a multi-scale rotational variability feature extraction module and a feature pyramid network. The multi-scale rotational variability feature extraction module and the feature pyramid network are used to extract multi-scale features and multi-scale rotational variability features from galaxy images. The first branch network is a visual branch network used to extract the morphological visual features of galaxy images. The second branch network is a physical branch network used to predict physical structure parameters.

[0038] In step S103, a galaxy morphology classification network model is trained using a galaxy image training set and a galaxy image test set.

[0039] The objective function for model training is a weighted combination of the classification loss function and the physical structure parameter regression loss function.

[0040] In step S104, an image of the galaxy to be identified is acquired, and the morphology of the galaxy to be identified is determined based on the image of the galaxy to be identified using the trained galaxy morphology classification network model.

[0041] In some embodiments of this application, the method may be executed by a server or by a terminal device with certain processing capabilities.

[0042] In some embodiments of this application, galaxy image set data can be acquired first, and then preprocessed and enhanced to obtain a galaxy image training set and a galaxy image test set. In some examples, anomaly screening can be performed on the galaxy image set data to construct the training and test sets; simultaneously, the corresponding physical structure parameters can be calculated based on the galaxy image set data as auxiliary supervision information; finally, data augmentation processing can be performed on the input galaxy image set data to improve the diversity of training samples.

[0043] Simultaneously, a galaxy morphology classification network model based on a dual-branch structure can be constructed. This model includes a feature extraction network, a first branch network, and a second branch network. The feature extraction network comprises at least a multi-scale rotational variability feature extraction module and a feature pyramid network. These modules are used to extract multi-scale and multi-scale rotational variability features from galaxy images. The first branch network is a visual branch network used to extract the morphological visual features of galaxy images. The second branch network is a physical branch network used to predict physical structure parameters.

[0044] In some embodiments of this application, a galaxy morphology classification network model can be trained using a galaxy image training set and a galaxy image test set. The objective function for model training is a weighted combination of a classification loss function and a physical structure parameter regression loss function.

[0045] Finally, images of the galaxy to be identified can be obtained, and the morphology of the galaxy to be identified can be determined based on the trained galaxy morphology classification network model.

[0046] According to the technical solution provided in this application, a direction-robust galaxy representation is obtained through multi-scale rotational equivariant feature extraction. Visual and physical branches are established to jointly model morphological features and structural parameters. The physical branch incorporates astronomical structural parameters as auxiliary supervision to enhance physical interpretability. A gated fusion mechanism is used to achieve adaptive fusion of features from both branches, and end-to-end training is performed using classification loss and structural parameter regression loss to achieve automated galaxy morphology recognition. This method effectively improves the accuracy of fine-grained galaxy morphology classification and the physical interpretability of the model. It outperforms existing state-of-the-art methods on some existing datasets and has significant application value.

[0047] In some embodiments of this application, preprocessing and enhancement of the galaxy image set data to obtain a galaxy image training set and a galaxy image test set may include:

[0048] Screening for outliers in galaxy image datasets;

[0049] The galaxy image set data after removing abnormal samples is divided into an initial galaxy image training set and a galaxy image test set according to a preset ratio;

[0050] Calculate the physical structure parameters of each galaxy image in the initial galaxy image training set; the physical structure parameters are the standardized parameters.

[0051] The physical structure parameters corresponding to each galaxy image are added to the initial galaxy image training set to obtain the initial galaxy image training set and the initial galaxy image test set with auxiliary supervision information; the auxiliary supervision information is the physical structure parameters.

[0052] The initial galaxy image training set with added auxiliary supervision information is augmented using a preset augmentation strategy to obtain the galaxy image training set.

[0053] The preset enhancement strategies include at least one of the following: random rotation, brightness and contrast perturbation, noise injection, blurring, or appearance enhancement based on imaging style transfer.

[0054] In other words, we can first acquire a set of galaxy images, then perform anomaly screening on this raw set of images, including removing images with severe noise pollution, imaging defects, or missing targets. Subsequently, we construct an initial training set and a test set of galaxy images according to a preset ratio. This preset ratio can be set according to actual needs, for example, 9:1.

[0055] Next, the physical structure parameters corresponding to each galaxy image in the initial galaxy image training set can be calculated. These physical structure parameters may include CAS parameters (concentration C, asymmetry A, and smoothness S), Gini coefficient G, second-order moment exponent M20, shape asymmetry, MID statistics (multimodal M, intensity I, and bias D), and ellipticity based on elliptic contour fitting. Index, half-luminosity radius, bulge-to-total luminosity ratio (B / T, where B represents the luminosity of the bulge and T represents the total luminosity of the entire galaxy), color gradient, etc.

[0056] All calculated physical structure parameters can be standardized to eliminate the influence of dimensions and accelerate model training convergence. In one example, Z-score standardization can be chosen as the standardization method.

[0057] The standardized physical structure parameters can be added to the initial galaxy image training set. In this set, each training sample contains a triple: a galaxy image, a morphological category label, and a vector of physical structure parameters. These parameters can be used as supplementary supervisory information during model training to enhance the model's physical interpretability.

[0058] Furthermore, to enhance the robustness of the galaxy morphology classification network model under real-world observation conditions, various data augmentation operations can be applied to each input galaxy image in the initial galaxy image training set. These operations include random rotation, brightness and contrast perturbation, noise injection, blurring, and appearance enhancement based on imaging style transfer. These enhancement strategies effectively simulate different observation conditions, imaging styles, and orientation variations, increasing the diversity of training samples, reducing the model's dependence on non-morphological factors, and thus enhancing the model's generalization ability.

[0059] Finally, the enhanced initial galaxy image training set can be used as the galaxy image training set.

[0060] In some embodiments of this application, the feature extraction network may further include a backbone network for performing conventional multi-scale feature extraction operations. In some examples, this backbone network may be implemented using the ConvNeXt network architecture.

[0061] The galaxy morphology classification network model provided in this application employs the ConvNeXt network to extract multi-scale features and introduces a multi-scale rotation-equivariant feature extraction module (Rotation-Equivariant Adapter, REA) to improve the model's robustness to changes in galaxy orientation. Based on this, visual and physical branch networks are further constructed to extract morphological visual features from galaxy images and predict physical structure parameters, respectively. Finally, a feature fusion module fuses the features output from the two branch networks and completes forward propagation calculations, outputting classification prediction results and structure parameter prediction results.

[0062] Feature extraction networks can perform feature extraction in the following ways:

[0063] The features of galaxy images are extracted using a multi-scale feature module to obtain an intermediate feature map;

[0064] The multi-scale rotational isovariant feature extraction module takes the intermediate feature map as input and maps the intermediate features of the two-dimensional plane to the discrete rotational group space through boosting convolution.

[0065] Group convolution is used to perform convolution operations on a preset set of discrete rotation angles so that the intermediate features remain equivariant under different rotation angles.

[0066] The group feature maps obtained by convolution are subjected to max pooling and average pooling respectively to aggregate responses at different rotation angles, thereby generating a feature representation with rotation invariance.

[0067] The rotation-invariant features are sequentially subjected to lateral convolution projection, top-down upsampling, and convolution smoothing to obtain the feature pyramid features.

[0068] In other words, to improve the model's robustness to arbitrary directional changes in galaxies, this embodiment further constructs an REA module in the feature extraction network. This REA module is implemented based on group equivariant convolution theory, ensuring the network maintains consistent feature responses under planar rotation transformations while preserving the inherent morphological and structural information of the galaxy.

[0069] refer to Figure 2 When the input galaxy image is received, the backbone network can first extract the multi-scale features of the galaxy image to obtain the intermediate feature map {C2, C3, C4}; then, by boosting convolution, the intermediate features of the two-dimensional plane are mapped to the discrete rotation group space, so that the features can be represented in the rotation group domain.

[0070] Subsequently, group convolution can be performed on a preset set of discrete rotation angles to ensure that features remain equivariant under different rotation angles. After completing the group convolution operation, max pooling and average pooling can be performed on the resulting group feature maps to aggregate responses from different rotation angles, thereby generating rotation-invariant features.

[0071] Finally, the rotation-invariant features described above can be input into a Feature Pyramid Network (FPN) for fusion. Specifically, the FPN achieves feature fusion through 1×1 horizontal convolutional projection, top-down upsampling, and 3×3 convolutional smoothing, resulting in multi-scale pyramid features {P2, P3, P4}. These pyramid features can simultaneously represent the detailed structure and overall morphology of galaxies, providing a unified multi-scale input for the subsequent dual-branch representation learning module.

[0072] In some embodiments of this application, the first branch network is encoded in the following manner:

[0073] Using the feature pyramid features as input, average pooling and max pooling operations are performed on the feature map at each scale to obtain pooled features of feature maps at different scales.

[0074] Pooled features from feature maps of different scales are concatenated, normalized, and then input into a multilayer perceptron for nonlinear mapping to generate low-dimensional visual embedding vectors.

[0075] The second branch network is encoded in the following manner:

[0076] Using the feature pyramid features as input, average pooling and max pooling operations are performed on the feature map at each scale to obtain pooled features of feature maps at different scales.

[0077] Pooled features from feature maps of different scales are concatenated, normalized, and then input into a multilayer perceptron for feature mapping to generate a low-dimensional physical embedding vector.

[0078] The regression prediction module is used to determine the predicted values ​​of the physical structure parameters of galaxies based on low-dimensional physical embedding vectors.

[0079] To simultaneously model the visual morphological features and physical structural properties of galaxy images, this application embodiment uses the aforementioned pyramid features {P2, P3, P4} as shared inputs to construct visual branch networks and physical branch networks respectively, thereby achieving parallel encoding of different information.

[0080] refer to Figure 3 In the visual branch of the network, average pooling and max pooling operations can be performed on feature maps at each scale to extract compact statistical feature descriptions. Then, the pooled features at different scales are concatenated, normalized, and input into a multilayer perceptron (MLP) for nonlinear mapping to generate low-dimensional visual embedding vectors. These visual embeddings can effectively capture visual information such as the brightness distribution, texture structure, spiral arm morphology, and core features of galaxies, providing key appearance feature support for subsequent classification.

[0081] The physics branch network can employ the same multi-scale feature input and pooling strategies as the visual branch network, performing average pooling and max pooling on features at each scale, followed by cross-scale feature concatenation and normalization. Subsequently, feature mapping is performed using an MLP to generate a low-dimensional physical embedding vector. Based on this, a regression prediction module can further process the latent physical features in this low-dimensional physical embedding vector to predict the physical structure parameters of the galaxy. These predicted physical structure parameters have the same format as the physical structure parameters calculated for each galaxy image in the initial galaxy image training set. Furthermore, these predicted physical structure parameters can be supervised during training using an auxiliary regression loss, thereby guiding the model to learn physically meaningful feature representations.

[0082] In some embodiments of this application, the galaxy morphology classification network model further includes a gated fusion module and a classification module; the gated fusion module is used to receive and fuse low-dimensional visual embedding vectors and low-dimensional physical embedding vectors to obtain fused embedding vectors; the classification module is used to obtain galaxy morphology prediction results based on the fused embedding vectors.

[0083] Continue to refer to Figure 3The low-dimensional visual embedding vector output by the visual branch network and the low-dimensional physical embedding vector output by the physical branch network can be concatenated and input into a learnable gating fusion module. This gating module dynamically adjusts the contributions of the two types of features through an adaptive weight allocation mechanism to achieve effective fusion of visual and physical information. Then, the fused feature representation can be input into a subsequent MLP for further feature transformation, and a linear classifier can be used to output the galaxy morphology category prediction result.

[0084] Through the above-described dual-branch coding structure, this application embodiment achieves collaborative modeling of visual features and physical prior information, thereby improving the model's ability to discriminate fine-grained galaxy morphology and enhancing the physical interpretability of the classification results.

[0085] In some embodiments of this application, a classification loss function is used to supervise the galaxy morphology prediction results output by the model. This classification loss function... It can be represented as ;in, Indicates category index, Indicates the total number of categories. Indicates category The true category label, Indicates category The predicted probability distribution. That is, this classification loss function is used to constrain the model to learn discriminative morphological feature representations, thereby improving classification accuracy.

[0086] Physical structure parameter regression loss function It can be represented as ;in, The parameter dimension representing the physical structure parameters. Represents the smoothing L1 loss function. The first galaxy image The true values ​​of each physical structural parameter The first galaxy image The predicted values ​​of several physical structure parameters are used. The regression loss function for these physical structure parameters guides the model to learn physically meaningful structural embeddings, giving it stronger physical consistency and interpretability in classification tasks.

[0087] To simultaneously optimize classification performance and structural parameter prediction performance, the classification loss function and the physical structure parameter regression loss function are weighted and combined to construct the final training objective function. Where λ is the weighting coefficient, used to balance the relationship between classification accuracy and structural consistency.

[0088] During training, the input galaxy image and its corresponding category label, along with the physical structure parameter label, can be input into the model. The prediction result is obtained through forward propagation, and the total loss is calculated based on the joint loss function mentioned above. Subsequently, the backpropagation algorithm is used to update the network parameters, thereby achieving iterative optimization of the model parameters.

[0089] After the model training is completed, the image of the galaxy to be identified can be input into the trained galaxy morphology classification network model to obtain a feature representation that integrates visual and physical structure information, and output the corresponding galaxy morphology category to achieve high-precision automated classification.

[0090] Through the above implementation process, this application provides an automatic galaxy morphology classification method that integrates prior physical structure and deep visual features. By data preprocessing and enhancement, multi-scale rotational feature extraction, and a dual-branch coding structure, it jointly models the visual information and physical properties of galaxies. Based on this, a model is trained through joint optimization of classification and structural parameter regression to achieve high-precision, interpretable automatic identification of galaxy morphology categories.

[0091] Figure 4 This is a flowchart illustrating another galaxy morphology recognition method based on physical priors and depth visual features provided in an embodiment of this application. Figure 4 As shown, the first step is to construct a dataset and preprocess it. Then, a network model based on physical priors and visual features is designed and trained using a joint loss function. During training, image enhancement methods can be used to enhance the training set samples. After training, the trained model parameters can be saved. Finally, images of galaxies to be identified can be obtained from the test set as test inputs, and the trained model can be used to infer the galaxy morphology classification results.

[0092] The technical solution provided in this application, through joint modeling of visual features and physical structure priors, can effectively distinguish subtle structural differences such as spiral arm morphology, bar structure, and disk-core ratio, significantly improving the classification performance of fine-grained tasks. It introduces structural parameters with clear astronomical significance as physical priors and provides explicit supervision during training, enabling the model to learn feature expressions consistent with the physical structure of galaxies, thereby improving the interpretability and scientific credibility of the classification results. It introduces a multi-scale rotational feature extraction mechanism and data augmentation strategies, enabling the model to adapt to galaxy images with different signal-to-noise ratios, resolutions, and imaging directions, improving generalization performance in actual sky survey data. Employing an end-to-end training method, the structure is modular and highly scalable, adaptable to different sky survey datasets. Experimental results on the Galaxy10 dataset show that the technical solution provided in this application significantly outperforms existing mainstream deep learning methods in classification accuracy across 10 fine-grained classification tasks, providing a more reliable technical solution for automated fine classification of large-scale sky survey data.

[0093] The galaxy morphology recognition method based on physical priors and deep visual features provided in this embodiment can be evaluated for accuracy by calculating accuracy, precision, recall, and F1 score. Let the total number of test samples be N, and the total number of categories be K. For any category c, a true positive (TP) is defined. c ), false positive (FP) c ) and false negative (FN) c The number of correct predictions, the number of incorrect predictions, and the number of items misclassified as other categories represent the number of items in that category.

[0094] Overall classification accuracy is defined as: ;

[0095] For each category c, precision and recall are defined as follows: ;

[0096] The F1 score is the harmonic average of precision and recall, defined as: .

[0097] Furthermore, to reduce the impact of imbalanced category distribution on the evaluation results, this application employs a weighted approach to statistically analyze precision, recall, and F1 score. The weights are determined by the proportion of samples in each category, thereby obtaining evaluation results that better reflect the overall classification performance. In addition, this application constructs a confusion matrix to visualize the classification results for each category, revealing the confusion relationships between different morphological categories and assisting in analyzing the model's performance and misclassification patterns in fine-grained classification tasks.

[0098] In this embodiment, the performance of the proposed DBMR-GalaxyNet model is evaluated based on the Galaxy10 dataset. The model achieves an overall classification accuracy of 90.3% on this dataset, with weighted precision, weighted recall, and weighted F1 score reaching 90.1%, 90.3%, and 90.1%, respectively. Specific classification results are shown in Table 1. These results demonstrate that the method provided in this embodiment can maintain high classification performance and stability even in complex classification scenarios with a large number of categories and subtle morphological differences.

[0099] Table 1. Classification results of the model proposed in the embodiments of this application.

[0100]

[0101] To verify the effectiveness of the method provided in this application, it was compared with representative methods such as Galaxy-CvT, Galaxy-WRN, GC-SWGAN, and CLIP-GMC. As shown in Table 2, the method provided in this application achieved a highest accuracy of 90.3% on the Galaxy10 dataset, significantly outperforming Galaxy-CvT (85.3%), Galaxy-WRN (76.8%), GC-SWGAN (83.6%), and CLIP-GMC (84.3%). This result demonstrates that the multi-scale rotational equivariant feature and bi-branch structure prior fusion strategy proposed in this application can effectively improve the recognition ability of fine-grained galaxy morphology.

[0102] Table 2. Comparison of classification results between the model proposed in this application and other methods.

[0103]

[0104] To analyze the feature representation capability of the model, the features learned by the model are further visualized and analyzed. Figure 5 This diagram illustrates the visualization results of t-SNE based on the fused features of the final classifier. Figure 6 This diagram illustrates the visualization results of t-SNE based on visual branch features. Figure 7 A schematic diagram of the visualization results of t-SNE based on physical branch features is shown. t-SNE is a dimensionality reduction method for high-dimensional data visualization. Figures 5 to 7 The serial numbers 0 to 9 represent different morphological categories and correspond to the numbers in Table 1.

[0105] like Figure 5 As shown, different morphological categories exhibit a relatively clear clustering structure in the feature space, with compact intra-class distribution and obvious inter-class separation. For example... Figure 6 and Figure 7As shown, further analysis of the feature distributions of the visual and physical branches reveals that the visual branch excels at distinguishing categories with significant appearance differences, while the physical branch plays a supplementary role in distinguishing categories with subtle structural differences. Fusing the two yields a more discriminative feature representation, thereby improving overall classification performance.

[0106] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0107] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0108] Figure 8 This is a schematic diagram of a galaxy morphology recognition device based on physical priors and depth visual features, provided in an embodiment of this application. Figure 8 As shown, the device includes:

[0109] The dataset construction module 801 is configured to acquire galaxy image set data, preprocess and enhance the galaxy image set data to obtain galaxy image training set and galaxy image test set.

[0110] The model building module 802 is configured to build a galaxy morphology classification network model based on a dual-branch structure. The galaxy morphology classification network model includes a feature extraction network, a first branch network, and a second branch network. The feature extraction network includes at least a multi-scale rotational equivariant feature extraction module and a feature pyramid network. The multi-scale rotational equivariant feature extraction module and the feature pyramid network are used to extract multi-scale features and multi-scale rotational equivariant features of galaxy images. The first branch network is a visual branch network used to extract morphological visual features of galaxy images. The second branch network is a physical branch network used to predict physical structure parameters.

[0111] Training module 803 is configured to train a galaxy morphology classification network model using a galaxy image training set and a galaxy image test set; wherein, the objective function for model training is a weighted combination of the classification loss function and the physical structure parameter regression loss function.

[0112] The identification module 804 is configured to acquire an image of the galaxy to be identified and use a trained galaxy morphology classification network model to determine the morphology of the galaxy to be identified based on the image.

[0113] According to the technical solution provided in this application, a direction-robust galaxy representation is obtained through multi-scale rotational equivariant feature extraction. Visual and physical branches are established to jointly model morphological features and structural parameters. The physical branch incorporates astronomical structural parameters as auxiliary supervision to enhance physical interpretability. A gated fusion mechanism is used to achieve adaptive fusion of features from both branches, and end-to-end training is performed using classification loss and structural parameter regression loss to achieve automated galaxy morphology recognition. This method effectively improves the accuracy of fine-grained galaxy morphology classification and the physical interpretability of the model. It outperforms existing state-of-the-art methods on some existing datasets and has significant application value.

[0114] In some implementations, the galaxy image set data is preprocessed and enhanced to obtain a galaxy image training set and a galaxy image test set, including: screening the galaxy image set data for outliers; dividing the galaxy image set data after removing outliers into an initial galaxy image training set and a galaxy image test set according to a preset ratio; calculating the physical structure parameters of each galaxy image in the initial galaxy image training set; the physical structure parameters are the parameters after standardization; adding the physical structure parameters corresponding to each galaxy image to the initial galaxy image training set to obtain an initial galaxy image training set and an initial galaxy image test set with added auxiliary supervision information; the auxiliary supervision information is the physical structure parameters; and using a preset enhancement strategy to perform sample enhancement on the initial galaxy image training set with added auxiliary supervision information to obtain the galaxy image training set; wherein the preset enhancement strategy includes at least one of the following: random rotation, brightness and contrast perturbation, noise injection, blurring, or appearance enhancement based on imaging style transfer.

[0115] In some implementations, the feature extraction network further includes a backbone network. The feature extraction network performs feature extraction as follows: the backbone network extracts features from the galaxy image to obtain an intermediate feature map; the multi-scale rotation equivariant feature extraction module takes the intermediate feature map as input and maps the intermediate features of the two-dimensional plane to a discrete rotation group space through boosting convolution; the group convolution is performed on a preset set of discrete rotation angles to ensure that each intermediate feature remains equivariant under different rotation angles; the group feature map obtained by convolution is subjected to max pooling and average pooling respectively to aggregate the responses of different rotation angles, thereby generating a rotation-invariant feature representation; the rotation-invariant features are sequentially subjected to lateral convolution projection, top-down upsampling, and convolution smoothing to obtain feature pyramid features.

[0116] In some implementations, the first branch network is encoded as follows: using the feature pyramid features as input, average pooling and max pooling operations are performed on the feature map at each scale to obtain pooled features of feature maps at different scales; the pooled features of feature maps at different scales are concatenated, and after normalization, they are input into a multilayer perceptron for nonlinear mapping to generate a low-dimensional visual embedding vector.

[0117] In some implementations, the second branch network is encoded as follows: using the feature pyramid features as input, average pooling and max pooling operations are performed on the feature map at each scale to obtain pooled features of feature maps at different scales; the pooled features of feature maps at different scales are concatenated, and after normalization, they are input into a multilayer perceptron for feature mapping to generate a low-dimensional physical embedding vector; the regression prediction module is used to determine the predicted values ​​of the physical structure parameters of the galaxy based on the low-dimensional physical embedding vector.

[0118] In some implementations, the galaxy morphology classification network model further includes a gated fusion module and a classification module; the gated fusion module is used to receive and fuse low-dimensional visual embedding vectors and low-dimensional physical embedding vectors to obtain fused embedding vectors; the classification module is used to obtain galaxy morphology prediction results based on the fused embedding vectors.

[0119] In some implementations, the classification loss function is used to supervise the model's output predictions of galaxy morphology; the classification loss function for ;in, Indicates category index, Indicates the total number of categories. Indicates category The true category label, Indicates category Predicted probability distribution; physical structure parameter regression loss function for ;in, The parameter dimension representing the physical structure parameters. Represents the smoothing L1 loss function. The first galaxy image The true values ​​of each physical structural parameter The first galaxy image Predicted values ​​for each physical structure parameter.

[0120] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0121] Figure 9This is a schematic diagram of an electronic device provided in an embodiment of this application. Figure 9 As shown, the electronic device 9 of this embodiment includes a processor 901, a memory 902, and a computer program 903 stored in the memory 902 and executable on the processor 901. When the processor 901 executes the computer program 903, it implements the steps in the various method embodiments described above. Alternatively, when the processor 901 executes the computer program 903, it implements the functions of each module / unit in the various device embodiments described above.

[0122] Electronic device 9 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 9 may include, but is not limited to, processor 901 and memory 902. Those skilled in the art will understand that... Figure 9 This is merely an example of electronic device 9 and does not constitute a limitation on electronic device 9. It may include more or fewer components than shown, or different components.

[0123] The processor 901 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0124] The memory 902 can be an internal storage unit of the electronic device 9, such as a hard disk or RAM of the electronic device 9. The memory 902 can also be an external storage device of the electronic device 9, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc., equipped on the electronic device 9. The memory 902 can also include both internal and external storage units of the electronic device 9. The memory 902 is used to store computer programs and other programs and data required by the electronic device.

[0125] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0126] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0127] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for galaxy morphology identification based on physical priors and deep visual features, characterized in that, include: Acquire galaxy image set data, and perform preprocessing and enhancement processing on the galaxy image set data to obtain galaxy image training set and galaxy image test set; A galaxy morphology classification network model based on a dual-branch structure is constructed. This model includes a feature extraction network, a first branch network, and a second branch network. The feature extraction network comprises at least a multi-scale rotational variability feature extraction module and a feature pyramid network, which are used to extract multi-scale and multi-scale rotational variability features from galaxy images. The first branch network is a visual branch network used to extract the morphological visual features of galaxy images. The second branch network is a physical branch network used to predict physical structure parameters. The galaxy morphology classification network model is trained using the galaxy image training set and the galaxy image test set; wherein, the objective function of the model training is a weighted combination of the classification loss function and the physical structure parameter regression loss function; Acquire an image of the galaxy to be identified, and use a trained galaxy morphology classification network model to determine the morphology of the galaxy based on the image. 2.The method of claim 1, wherein, The galaxy image set data is preprocessed and enhanced to obtain a galaxy image training set and a galaxy image test set, including: The galaxy image set data is subjected to anomaly sample screening; The galaxy image set data after removing abnormal samples is divided into an initial galaxy image training set and a galaxy image test set according to a preset ratio; Calculate the physical structure parameters of each galaxy image in the initial galaxy image training set; the physical structure parameters are the parameters after standardization. The physical structure parameters corresponding to each galaxy image are added to the initial galaxy image training set to obtain an initial galaxy image training set and an initial galaxy image test set with auxiliary supervision information; the auxiliary supervision information is the physical structure parameters. The initial galaxy image training set with added auxiliary supervision information is augmented using a preset augmentation strategy to obtain the galaxy image training set. The preset enhancement strategy includes at least one of the following: random rotation, brightness and contrast perturbation, noise injection, blurring, or appearance enhancement based on imaging style transfer.

3. The galaxy morphology recognition method based on physical priors and depth visual features according to claim 1, characterized in that, The feature extraction network also includes a backbone network multi-scale rotational equivariant feature extraction module; The feature extraction network performs feature extraction in the following manner: The backbone network is used to extract features from galaxy images to obtain intermediate feature maps; The multi-scale rotational isovariant feature extraction module takes the intermediate feature map as input and maps the intermediate features of the two-dimensional plane to the discrete rotational group space through boosting convolution. Group convolution is used to perform convolution operations on a preset set of discrete rotation angles so that the intermediate features remain equivariant under different rotation angles. The group feature maps obtained by convolution are subjected to max pooling and average pooling respectively to aggregate responses at different rotation angles, thereby generating a feature representation with rotation invariance. The rotation-invariant features are sequentially subjected to lateral convolution projection, top-down upsampling, and convolution smoothing to obtain the feature pyramid features.

4. The galaxy morphology recognition method based on physical priors and depth visual features according to claim 3, characterized in that, The first branch network is encoded in the following manner: Using the feature pyramid features as input, average pooling and max pooling operations are performed on the feature map at each scale to obtain pooled features of feature maps at different scales. Pooled features from feature maps of different scales are concatenated, normalized, and then input into a multilayer perceptron for nonlinear mapping to generate low-dimensional visual embedding vectors.

5. The galaxy morphology recognition method based on physical priors and depth visual features according to claim 4, characterized in that, The second branch network is encoded in the following manner: Using the feature pyramid features as input, average pooling and max pooling operations are performed on the feature map at each scale to obtain pooled features of feature maps at different scales. Pooled features from feature maps of different scales are concatenated, normalized, and then input into a multilayer perceptron for feature mapping to generate a low-dimensional physical embedding vector. The regression prediction module is used to determine the predicted values ​​of the physical structure parameters of the galaxy based on the low-dimensional physical embedding vector.

6. The galaxy morphology recognition method based on physical priors and depth visual features according to claim 5, characterized in that, The galaxy morphology classification network model also includes a gated fusion module and a classification module; The gated fusion module is used to receive and fuse the low-dimensional visual embedding vector and the low-dimensional physical embedding vector to obtain the fused embedding vector. The classification module is used to obtain galaxy morphology prediction results based on the fused embedding vector.

7. The galaxy morphology recognition method based on physical priors and depth visual features according to claim 1, characterized in that, The classification loss function is used to supervise the galaxy morphology prediction results output by the model; The classification loss function for ;in, Indicates category index, Indicates the total number of categories. Indicates category The true category label, Indicates category The predicted probability distribution; The physical structure parameter regression loss function for ;in, The parameter dimension representing the physical structure parameters. Represents the smoothing L1 loss function. The first galaxy image The true values ​​of each physical structural parameter The first galaxy image Predicted values ​​for each physical structure parameter.

8. A galaxy morphology recognition device based on physical priors and depth visual features, characterized in that, include: The dataset construction module is configured to acquire galaxy image set data, preprocess and enhance the galaxy image set data to obtain galaxy image training set and galaxy image test set; The model building module is configured to construct a galaxy morphology classification network model based on a dual-branch structure. The galaxy morphology classification network model includes a feature extraction network, a first branch network, and a second branch network. The feature extraction network includes at least a multi-scale rotational variability feature extraction module and a feature pyramid network, which are used to extract multi-scale features and multi-scale rotational variability features from galaxy images. The first branch network is a visual branch network used to extract the morphological visual features of galaxy images. The second branch network is a physical branch network used to predict physical structure parameters. The training module is configured to train the galaxy morphology classification network model using the galaxy image training set and the galaxy image test set; wherein the objective function for model training is a weighted combination of the classification loss function and the physical structure parameter regression loss function; The identification module is configured to acquire an image of the galaxy to be identified and use a trained galaxy morphology classification network model to determine the morphology of the galaxy to be identified based on the image.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the galaxy morphology recognition method based on physical priors and depth visual features as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the galaxy morphology recognition method based on physical priors and depth visual features as described in any one of claims 1 to 7.

Images