Method for identifying galaxy image through semi-supervised learning based on uncertainty estimation

Abstract

The invention provides a semi-supervised learning galaxy image recognition method based on uncertainty estimation, which comprises the following steps of: acquiring a plurality of existing galaxy images, and dividing label data and label-free data; constructing a deep neural network model for an image multi-classification task; constructing a training set and a test set, and training and testing the deep neural network model to obtain a trained deep neural network model after a loss function of the deep neural network model is corrected based on a cross entropy loss function of labeled data and an unsupervised loss function of uncertainty estimation based on unlabeled data; and obtaining a to-be-detected galaxy image and importing the to-be-detected galaxy image into the trained deep neural network model to obtain the category of the to-be-detected galaxy image. By implementing the method, the model cost can be reduced, the model efficiency can be improved, and the model generalization ability can be improved by reducing the influence of noise samples, so that the image recognition efficiency and accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a method for recognizing galaxy images through semi-supervised learning based on uncertainty estimation. Background Art

[0002] With the demand for automation and intelligence in galaxy morphology classification tasks, neural network models are used to improve the efficiency and utilization of galaxy images, reducing the cost of studying galaxy formation and evolution.

[0003] In existing technology, improvements in astronomical observation capabilities have led to the collection of a large number of galaxy images. These images are often manually annotated for machine learning model training, which is not only costly and inefficient, but also lacks effective methods for reducing the impact of noise samples to improve model generalization. Therefore, it is necessary to improve the neural network models used in the galaxy image recognition process. This would not only reduce model costs and improve model efficiency, but also reduce the impact of noise samples to improve model generalization, thereby improving image recognition efficiency and accuracy. Summary of the Invention

[0004] The technical problem to be solved by the embodiments of the present invention is to provide a method for semi-supervised learning based on uncertainty estimation to identify galaxy images, which can not only reduce model cost and improve model efficiency, but also reduce the impact of noise samples to improve the model generalization ability, thereby improving image recognition efficiency and accuracy.

[0005] To solve the above technical problems, an embodiment of the present invention provides a method for identifying galaxy images using semi-supervised learning based on uncertainty estimation, the method comprising the following steps:

[0006] Acquire several existing galaxy images and divide them into labeled data and unlabeled data; wherein the unlabeled data consists of galaxy images that have not been manually labeled, and the total amount of such data is N times the total amount of images in the labeled data; the labeled data consists of galaxy images that have been manually labeled, and N is a positive integer greater than or equal to 10;

[0007] Build a deep neural network model for image multi-classification tasks;

[0008] Extract part of the labeled data and the unlabeled data to form a training set, and extract all the remaining labeled data to form a test set, and after the loss function of the deep neural network model is corrected by the cross entropy loss function obtained based on the labeled data and the unsupervised loss function obtained by uncertainty estimation based on the unlabeled data, the deep neural network model is trained and tested by the training set and the test set to obtain a trained deep neural network model; wherein the loss function of the deep neural network model is L total=αL s +ηL e +vL a ;L s is the cross entropy loss function based on the labeled data; L e and L a is the unsupervised loss function obtained by uncertainty estimation based on unlabeled data; α, η, and ν are all preset constants, and α+η+ν=1;

[0009] The image of the galaxy to be tested is obtained and imported into the trained deep neural network model to obtain the category of the image of the galaxy to be tested.

[0010] The specific steps of performing uncertainty estimation based on unlabeled data to obtain an unsupervised loss function include:

[0011] Perform unsupervised deterministic calculations on unlabeled data to obtain normalized uncertainty scores;

[0012] According to the normalized uncertainty score, a simple data indicator function and a fuzzy data indicator function are defined, and simple data and fuzzy data are selected from the unlabeled data according to the simple data indicator function and the fuzzy data indicator function;

[0013] Determine the unsupervised loss function L based on simple data e , and the unsupervised loss function L based on fuzzy data a ; Among them, the unsupervised loss function L e is the cross entropy loss function after consistency regularization; L a is the GCE loss function with noise learning ability.

[0014] The category of the galaxy image to be measured is one of elliptical galaxies, early spiral galaxies, mid-stage spiral galaxies, late-stage spiral galaxies, barred spiral galaxies, irregular galaxies and merged galaxies.

[0015] Before the training set and the test set are formed, the method further comprises:

[0016] Perform image augmentation on the labeled data, including resizing, random cropping, and random horizontal flipping; and,

[0017] Image enhancement processing is performed on unlabeled data, including random rotation, shear transformation, translation transformation, color jittering, tone separation, exposure adjustment, sharpening and automatic contrast enhancement.

[0018] The implementation of the embodiments of the present invention has the following beneficial effects:

[0019] 1. This invention uses a semi-supervised learning framework, which only requires a small amount of data labeled by professionals to train a high-performance model, significantly reducing the reliance on large-scale accurately labeled data in traditional supervised learning, thereby reducing manpower and time costs;

[0020] 2. This invention introduces a sample prediction uncertainty assessment mechanism, which enables the model to identify and carefully handle low-confidence samples, avoiding the propagation of incorrect labels to subsequent training processes. This significantly improves the robustness and generalization performance of the model in the face of noisy data, blurred images, and changing observation conditions.

[0021] 3. The present invention combines uncertainty weight adjustment strategies during model training to dynamically select more representative unlabeled samples for inclusion in the training set, which helps enhance data diversity and alleviates the sample bias problem caused by limited manually labeled samples.

[0022] 4. The present invention assists model training based on uncertainty estimation of unlabeled data, and can also be output as a credibility indicator of the classification results, thereby improving the reliability of the classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, without paying creative labor, other drawings obtained based on these drawings still fall within the scope of the present invention.

[0024] Figure 1 A flowchart of a method for identifying galaxy images using semi-supervised learning based on uncertainty estimation provided by an embodiment of the present invention;

[0025] Figure 2 An algorithm flow chart for introducing an uncertainty assessment mechanism into unlabeled data to distinguish simple data from fuzzy data in a method for semi-supervised learning based on uncertainty estimation to identify galaxy images provided by an embodiment of the present invention;

[0026] Figure 3 for Figure 2 The logical deduction diagram of the loss functions for simple data and fuzzy data. DETAILED DESCRIPTION

[0027] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be described in further detail below with reference to the accompanying drawings.

[0028] like Figure 1FIG. 1 is a method for identifying galaxy images using semi-supervised learning based on uncertainty estimation, according to an embodiment of the present invention. The method includes the following steps:

[0029] Step S1: Acquire several existing galaxy images and divide them into labeled data and unlabeled data; wherein the unlabeled data consists of galaxy images that have not been manually labeled, and the total amount of the unlabeled data is N times the total amount of the labeled data; the labeled data consists of manually labeled galaxy images; N is a positive integer greater than or equal to 10;

[0030] The specific process involves obtaining several existing galaxy images and labeling a small portion of them with galaxy categories to form labeled data. The remaining galaxy images remain unlabeled to form unlabeled data. The total number of unlabeled data images is at least 10 times the total number of labeled images, fully utilizing the unlabeled galaxy images to enhance the generalization capabilities of the subsequent model. The galaxy image categories include, but are not limited to, elliptical galaxies, early spiral galaxies, mid-stage spiral galaxies, late spiral galaxies, barred spiral galaxies, irregular galaxies, and merging galaxies.

[0031] In one example, the seven categories of galaxy morphology in the Galaxy Zoo 2 dataset were selected, such as elliptical galaxies (E), early spiral galaxies (Sa), intermediate spiral galaxies (Sb), late spiral galaxies (Sc), barred spiral galaxies (SB), irregular galaxies (Irr), and merged galaxies (Merger).

[0032] The labeled data now includes approximately 60,000 labeled images, each with a resolution of 424×424 pixels and corresponding seven morphological labels. Furthermore, approximately 600,000 unlabeled galaxy images acquired from the Sloan Digital Sky Survey (SDSS) serve as unlabeled data, making the unlabeled data 10 times larger than the labeled data. This data is used for semi-supervised learning.

[0033] Step S2: constructing a deep neural network model for image multi-classification tasks;

[0034] The specific process involves first selecting a deep neural network model suitable for multi-class image classification tasks, such as the Residual Network (ResNet) or the Vision Transformer (ViT). The ResNet model is chosen for its excellent performance and stability in image classification tasks. Its residual connections mitigate the vanishing gradient problem in deep networks, making it suitable for processing complex galaxy image features.

[0035] In one embodiment, PyTorch is used to load the pre-trained ResNet-50 as a deep neural network model, and the last fully connected layer in the deep neural network model is modified, and the output dimension is set to 7 (corresponding to seven types of galaxy morphologies).

[0036] Secondly, determine the initial hyperparameters of the deep neural network model, including the number of training rounds: set to 500 rounds; the learning rate: the initial learning rate is 10 -5 , using the StepLR learning rate adjustment strategy, with a step size of 10 rounds and an adjustment factor of 0.1; batch size: set to 8; optimizer: using stochastic gradient descent (SGD), momentum of 0.9, and weight decay of 10 -4 ; Loss function: According to the characteristics of the data, select Cross Entropy Loss or Generalized Cross Entropy Loss.

[0037] At the end of each training epoch, the model performance is evaluated using the validation set, and the three models with the best evaluation results (based on classification accuracy or loss values ​​on the validation set) are saved with replacement.

[0038] Step S3: extract part of the labeled data and the unlabeled data to form a training set, and extract all the remaining labeled data to form a test set. After the loss function of the deep neural network model is corrected by the cross entropy loss function obtained based on the labeled data and the unsupervised loss function obtained by uncertainty estimation based on the unlabeled data, the deep neural network model is trained and tested by the training set and the test set to obtain a trained deep neural network model; wherein the loss function of the deep neural network model is L total =αL s +ηL e +vL a ;L s is the cross entropy loss function based on the labeled data; L e and L a is the unsupervised loss function obtained by uncertainty estimation based on unlabeled data; α, η, and v are all preset constants, and α + η + v = 1;

[0039] The specific process is as follows: first, 80% of the labeled data and unlabeled data are extracted to form a training set, and the remaining 20% ​​of the labeled data are extracted to form a test set.

[0040] In one embodiment, the training set consists of 48,000 labeled images and 600,000 unlabeled images; the test set consists of 12,000 labeled images, which are only used for performance evaluation.

[0041] Secondly, in order to improve the robustness and generalization ability of the model, the data is enhanced to varying degrees.

[0042] Weak enhancement operations are applied to the labeled data, including: resizing the image to a uniform resolution (e.g., 224×224 pixels); randomly cropping the image to introduce local variations; randomly flipping horizontally to enhance data diversity; and normalizing the image pixel values ​​(based on mean and standard deviation).

[0043] Strong and weak enhancement operations are applied to unlabeled data, and consistency constraints are imposed between different enhancements to simulate more complex image changes, including: random rotation (angle range, for example, -30° to 30°); shear transformation (Shear X / Shear Y, amplitude range, for example, 0.1 to 0.3); translation transformation (Translate X / Translate Y, amplitude range, for example, 0.1 to 0.2); color jitter (adjusting brightness, contrast, saturation, amplitude range, for example, 0.2 to 0.4); tone separation, exposure adjustment, sharpening, and automatic contrast enhancement.

[0044] In one embodiment, the annotation data is augmented (weakly): resized from 424×424 to 224×224. Randomly cropped: cropped to 200×200 and then resized back to 224×224. Randomly flipped horizontally with a probability of 0.5. Normalized: mean [0.485, 0.456, 0.406], standard deviation [0.229, 0.224, 0.225].

[0045] In another embodiment, unlabeled data enhancement (strong enhancement): random rotation: -30° to 30°. Shear transformation: Shear X / Y, amplitude 0.1 to 0.3. Translation transformation: Translate X / Y, amplitude 0.1 to 0.2. Color jitter: brightness, contrast, saturation, amplitude 0.2 to 0.4. Other enhancements: tone separation, exposure adjustment, sharpening, automatic contrast enhancement. Combination strategy: randomly select three enhancement combinations. Among them, unlabeled data enhancement (weak enhancement) is the same as labeled data enhancement (weak enhancement).

[0046] Then, conventional supervised learning methods are used to train the labeled data; for unlabeled data, an uncertainty assessment mechanism is introduced to distinguish simple data from fuzzy data, in order to optimize the training process of unlabeled data.

[0047] At this time, during the training process, the cross entropy loss function obtained based on the labeled data and the unsupervised loss function obtained by uncertainty estimation based on the unlabeled data are used to correct the loss function of the deep neural network model, so that the loss function of the modified deep neural network model is L total =αL s +ηL e+vL a , and through the loss function L total =αL s +ηL e +vL a , calculate the cross entropy loss of the labeled data and the consistent cross entropy loss of the simple data and the fuzzy data to optimize the minimum total loss.

[0048] In one embodiment, the specific steps of performing uncertainty estimation based on unlabeled data to obtain an unsupervised loss function include: performing unsupervised deterministic calculation on the unlabeled data to obtain a normalized uncertainty score; defining a simple data indicator function and a fuzzy data indicator function based on the normalized uncertainty score, and selecting simple data and fuzzy data from the unlabeled data based on the simple data indicator function and the fuzzy data indicator function; determining an unsupervised loss function L based on the simple data e , and the unsupervised loss function L based on fuzzy data a ; Among them, the unsupervised loss function L e is the cross entropy loss function after consistency regularization; L a is the GCE loss function with noise learning ability.

[0049] In another embodiment, Figure 2 and Figure 3 As shown in the figure, for unlabeled data, an uncertainty assessment mechanism is introduced to distinguish simple data from fuzzy data. The specific implementation is as follows:

[0050] (1) Multiple forward propagation. For each unlabeled galaxy image, we perform 10 gradient-free forward propagations to obtain the softmax probability distribution of each propagation. Multiple propagation simulates the uncertainty of model predictions by introducing randomness (such as dropout or random transformations of data augmentation) to enhance the robustness of the evaluation.

[0051] (2) Calculation of uncertainty index. Calculate the average probability distribution of 10 forward propagations, and based on this, calculate the entropy of the average probability distribution (Mean Entropy) to reflect the overall stability of the model prediction. Calculate the entropy of the Softmax probability distribution of each forward propagation and take its average value (Average Entropy) to reflect the degree of dispersion of a single prediction. Add the entropy of the average probability distribution to the average value of the entropy of each propagation to obtain a comprehensive uncertainty index. The higher the uncertainty index, the more unstable the model's prediction of the sample is, and it may be fuzzy data; conversely, a lower uncertainty indicates that the sample may be simple data.

[0052] (3) Data differentiation. Based on uncertainty indicators, unlabeled data is divided into simple data and fuzzy data. Simple data usually has high confidence and low uncertainty, and is suitable for training using standard loss functions. Fuzzy data has medium confidence and high uncertainty, and requires a more robust loss function to deal with potential label noise.

[0053] At this time, simple data: confidence ≥ 0.8, uncertainty ≥ 0.4 (such as Figure 2 As shown in ), using the cross entropy loss function after consistency regularization (as Figure 3 Fuzzy data: confidence level between 0.6 and 0.8, uncertainty ≤ 0.4 (e.g. Figure 2 As shown in ), using the GCE loss function with noise learning ability (as Figure 3 ), to better handle label noise and uncertainty.

[0054] In one example, for each unlabeled image, 10 gradient-free forward passes are performed, using dropout (with probability 0.5) and strong augmentation to introduce randomness, and the softmax probability distribution (a 7-dimensional vector corresponding to seven types of galaxies) of each pass is recorded.

[0055] The methods for determining the mean entropy of the average probability distribution include: the mean Softmax probability P of 10 propagations mean and entropy And, determine the average value of the single transmission entropy including the entropy of each transmission and the average entropy Finally, the comprehensive uncertainty calculation formula is determined as U=H mean +H avg At the same time, the data distinction rule is set: simple data corresponds to U ≥ 0.4, confidence ≥ 0.8; fuzzy data corresponds to U ≤ 0.4, confidence ∈ [0.6, 0.8).

[0056] Assume that an unlabeled image undergoes 10 forward propagations and the average Softmax probability is P mean =[0.85,0.05,0.03,0.02,0.02,0.01,0.02], and the confidence level is 0.85 (>0.8).

[0057] calculate

[0058]

[0059] calculate

[0060] Calculation: U = 0.674 + 0.668 = 1.342.

[0061] Given that U>0.4 and confidence ≥0.8 meet the simple data conditions, calculate the loss: L CE =―log(0.85)≈0.162.

[0062] Assume that an unlabeled image undergoes 10 forward propagations and the average Softmax probability is P mean =[0.65,0.15,0.08,0.05,0.03,0.02,0.02], confidence level = 0.65 (0.6≤0.65<0.8)

[0063] Calculate H mean =―[0.65log(0.65)+0.15log(0.15)+0.08log(0.08)+0.05log(0.05)+0.03log(0.03)+2.035,log(0.02)]≈―[0.65×(―0.431)+0.15 ×(-1.897)+0.08×(-2.526)+0.05×(-2.996)+0.03×(-3.507)+2×0.02×(-3.912)]≈0.280+0.285+0.202+0.150+0.105+0.156=0.178

[0064] calculate

[0065] Calculation U = 0.178 + 0.186 = 0.364 < 0.4

[0066] Given that U<0.4 and confidence ∈[0.6,0.8) meets the fuzzy data condition, calculate the loss:

[0067] Then, the deep neural network model is trained and tested using the training set and the test set to obtain a trained deep neural network model. At this point, the training stop condition is: the loss function value converges to the preset threshold (for example, the loss change of the validation set is less than 10 -4 ); the maximum number of training rounds (500 rounds) is reached. During the training process, the three models with the best performance on the validation set are regularly saved.

[0068] Finally, the trained deep neural network model is evaluated and adjusted as follows:

[0069] To comprehensively evaluate the performance of the galaxy morphology classification model, we systematically analyzed its classification performance on different galaxy morphologies (e.g., elliptical, spiral, and irregular galaxies) by comparing the model's predictions on the validation set with manually annotated true labels. We also used evaluation metrics suitable for multi-classification tasks. The following is a detailed description of the main evaluation metrics:

[0070] (1) Accuracy. Accuracy reflects the proportion of samples correctly classified by the model on the validation set and is a core indicator for evaluating overall classification performance. It measures the model's comprehensive performance on all galaxy morphologies by counting the ratio of the number of correctly predicted samples to the total number of samples. For multi-classification tasks, accuracy comprehensively considers the proportion of correct predictions for all categories and is suitable for evaluating the overall reliability of the model.

[0071] (2) Precision. Precision measures the proportion of samples predicted by the model to have a specific galaxy morphology that actually have that morphology. This metric is particularly suitable for assessing the reliability of the model's predictions for a certain type of galaxy (e.g., rare morphologies). For multi-classification tasks, the precision of each category is calculated separately, and the model performance is comprehensively evaluated through macro-averaging (taking the average of the precisions of each category) or weighted averaging (weighting by the number of samples in each category).

[0072] (3) Recall. Recall measures the model's ability to correctly identify samples of a certain category and reflects the model's sensitivity to detecting specific galaxy morphologies. This metric is particularly important in scenarios with class imbalance, such as when there are fewer samples of certain galaxy morphologies (such as irregular galaxies). Similar to precision, recall integrates multi-classification performance through macro-averaging or weighted averaging.

[0073] (4) F1-Score. The F1-Score combines precision and recall, providing a balanced assessment of the model's classification performance by calculating the harmonic mean of the two. The F1-Score is particularly suitable for galaxy morphology classification tasks with imbalanced classes, effectively measuring the trade-off between high precision and high recall. For multi-classification tasks, the F1-Score is calculated for each class separately, and the macro-average or weighted average is taken as the overall performance indicator.

[0074] (5) Confusion Matrix. The confusion matrix records the correspondence between the predicted results and the true labels for each galaxy morphology category in matrix form, visually demonstrating the model's classification performance on different categories. For a task with K categories, the confusion matrix is ​​a K×K matrix, where the matrix elements represent the number of samples whose true category is one but predicted to be another. By analyzing the confusion matrix, we can identify the model's tendency to misclassify specific galaxy morphologies and guide subsequent optimization.

[0075] During the evaluation process, the above indicators are calculated regularly on the validation set, and the three models with the best performance are recorded based on key indicators such as F1 score or accuracy for subsequent hyperparameter adjustment and final model selection.

[0076] During the evaluation process, the above metrics are regularly calculated on the validation set, and the three models with the best performance, based on key metrics such as F1 score or accuracy, are recorded for subsequent hyperparameter adjustment and final model selection. Therefore, based on the evaluation results, a limited-range grid search method is used to optimize the model's hyperparameters and training strategy to improve classification accuracy and generalization ability. Specific adjustments include:

[0077] (1) Training rounds. Test different training rounds (e.g., 300, 500, 1000) to strike a balance between training time and model convergence, ensuring that the model fully learns the data features while avoiding overfitting.

[0078] (2) Learning rate adjustment strategy. Try various learning rate decay schemes, such as StepLR (step size 10, decay factor 0.1) or CosineAnnealingLR, and optimize the dynamic adjustment of the learning rate to accelerate convergence and improve performance.

[0079] (3) Batch size. Test different batch sizes (e.g., 4, 8, and 16) to optimize video memory utilization and gradient estimation stability to adapt to hardware resources and improve training efficiency.

[0080] (4) Optimizer. Comparison with stochastic gradient descent (SGD, momentum 0.9, weight decay 10 -4 ) and Adam optimizer (default parameters β1 = 0.9, β2 = 0.999, weight decay 10 -4 ) performance and select an optimization strategy that is more suitable for the galaxy morphology classification task.

[0081] Through multiple rounds of iterative optimization, the optimal hyperparameter combination is determined to ensure that the model achieves high accuracy and generalization ability in the galaxy morphology classification task. For the ResNet-based model, the example configuration is as follows: training rounds: 500 rounds; initial learning rate: 10 -5 , using the StepLR strategy (step size 10, decay rate 0.1); batch size: 8; optimizer: SGD, momentum 0.9, weight decay 10 -4 ; Enable mixed precision training to speed up calculations and reduce video memory usage.

[0082] Step S4: Obtain the image of the galaxy to be tested, and import it into the trained deep neural network model to obtain the category of the image of the galaxy to be tested.

[0083] The specific process is to obtain the image of the galaxy to be tested and import it into the deep neural network model trained in step S3, so as to obtain the category of the image of the galaxy to be tested as one of elliptical galaxies, early spiral galaxies, mid-stage spiral galaxies, late spiral galaxies, barred spiral galaxies, irregular galaxies and merged galaxies.

[0084] The implementation of the present invention has the following beneficial effects:

[0085] 1. This invention uses a semi-supervised learning framework, which only requires a small amount of data labeled by professionals to train a high-performance model, significantly reducing the reliance on large-scale accurately labeled data in traditional supervised learning, thereby reducing manpower and time costs;

[0086] 2. This invention introduces a sample prediction uncertainty assessment mechanism, which enables the model to identify and carefully handle low-confidence samples, avoiding the propagation of incorrect labels to subsequent training processes. This significantly improves the robustness and generalization performance of the model in the face of noisy data, blurred images, and changing observation conditions.

[0087] 3. The present invention combines uncertainty weight adjustment strategies during model training to dynamically select more representative unlabeled samples for inclusion in the training set, which helps enhance data diversity and alleviates the sample bias problem caused by limited manually labeled samples.

[0088] 4. The present invention assists model training based on uncertainty estimation of unlabeled data, and can also be output as a credibility indicator of the classification results, thereby improving the reliability of the classification results.

[0089] A person skilled in the art will understand that all or part of the steps in the above-mentioned embodiment method can be completed by hardware related to program instructions, and the program can be stored in a computer-readable storage medium, such as ROM / RAM, disk, CD, etc.

[0090] The above disclosure is only a preferred embodiment of the present invention, which certainly cannot limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for identifying galaxy images using semi-supervised learning based on uncertainty estimation, characterized in that: The method comprises the following steps: Acquire several existing galaxy images and divide them into labeled data and unlabeled data; wherein the unlabeled data consists of galaxy images that have not been manually labeled, and the total amount of such data is N times the total amount of images in the labeled data; the labeled data consists of galaxy images that have been manually labeled, and N is a positive integer greater than or equal to 10; Build a deep neural network model for image multi-classification tasks; Extract part of the labeled data and the unlabeled data to form a training set, and extract all the remaining labeled data to form a test set, and after the loss function of the deep neural network model is corrected by the cross entropy loss function obtained based on the labeled data and the unsupervised loss function obtained by uncertainty estimation based on the unlabeled data, the deep neural network model is trained and tested by the training set and the test set to obtain a trained deep neural network model; wherein the loss function of the deep neural network model is L total =αL s +ηL e +vL a ;L s is the cross entropy loss function based on the labeled data; L e and L a is the unsupervised loss function obtained by uncertainty estimation based on unlabeled data; α, η, and v are all preset constants, and α + η + v = 1; The image of the galaxy to be tested is obtained and imported into the trained deep neural network model to obtain the category of the image of the galaxy to be tested.

2. The method for identifying galaxy images using semi-supervised learning based on uncertainty estimation according to claim 1, wherein: The specific steps of performing uncertainty estimation based on unlabeled data to obtain an unsupervised loss function include: Perform unsupervised deterministic calculations on unlabeled data to obtain normalized uncertainty scores; According to the normalized uncertainty score, a simple data indicator function and a fuzzy data indicator function are defined, and simple data and fuzzy data are selected from the unlabeled data according to the simple data indicator function and the fuzzy data indicator function; Determine the unsupervised loss function L based on simple data e , and the unsupervised loss function L based on fuzzy data a ; Among them, the unsupervised loss function L e is the cross entropy loss function after consistency regularization; L a is the GCE loss function with noise learning ability.

3. The method for identifying galaxy images using semi-supervised learning based on uncertainty estimation according to claim 2, wherein: The category of the galaxy image to be measured is one of an elliptical galaxy, an early spiral galaxy, a mid-stage spiral galaxy, a late spiral galaxy, a barred spiral galaxy, an irregular galaxy and a merging galaxy.

4. The method for identifying galaxy images using semi-supervised learning based on uncertainty estimation according to claim 1, wherein: Before the training set and the test set are formed, the method further comprises: Perform image augmentation on the labeled data, including resizing, random cropping, and random horizontal flipping; and, Image enhancement processing is performed on unlabeled data, including random rotation, shear transformation, translation transformation, color jittering, tone separation, exposure adjustment, sharpening and automatic contrast enhancement.

Images