Discrete wavelet transform and multilayer convolution-based galaxy image spectrum prediction network

Abstract

The invention discloses a novel galaxy image spectrum prediction network which combines the advantages of discrete wavelet transform (DWT) and a multilayer convolutional neural network so as to realize high-precision prediction from galaxy images to spectrum information. By training and testing galaxy images and spectral data in an SDSS DR18 data set, the method shows performance comparable with or even better than that of an existing advanced method on a plurality of performance indexes. In addition, the invention further comprises a visualization method, and the consistency of the real spectrum and the predicted spectrum is visually compared. Generally speaking, the network structure not only improves the accuracy of spectrum prediction, but also provides a new analysis tool for astronomy research, and has important scientific significance and application value.

Description

Technical Field

[0001] The present invention belongs to the field of SDSS (Sloan Digital Sky Survey) galaxy spectrum prediction, and particularly relates to a galaxy image spectrum prediction network based on discrete wavelet transform and multi-layer convolution. Background Art

[0002] In astronomical research, image and spectral data are the core tools for revealing the mysteries of the universe. The morphology, structure, and evolution process of celestial bodies are usually presented through image data, while spectral data can provide key information about the chemical composition, physical properties, and dynamic evolution of celestial bodies. Traditional astronomical observation methods usually rely on large telescopes to collect galaxy images, and obtaining detailed spectral information requires additional configuration of spectrometers. However, the acquisition of spectral data not only has a high cost but is also limited by observation time and equipment performance. With the rapid development of deep learning technology, the idea of generating spectra based on images has gradually become a possible solution, but this field still faces technical challenges such as relying on physical models, high computational complexity, and poor generalization ability. Therefore, exploring an efficient and accurate deep learning method for directly generating spectral data from galaxy images has become an important issue in current astronomical research. Summary of the Invention

[0003] The present invention aims to develop a galaxy image spectrum prediction network based on discrete wavelet transform and multi-layer convolution, which can accurately predict the corresponding spectral data from the photometric images of galaxies. By constructing and training this network, it is expected to reduce the dependence on traditional spectrometers and improve the acquisition speed and quality of spectral data.

[0004] The technical solution of the present invention is a deep learning network architecture based on discrete wavelet transform (DWT) and multi-layer convolution for predicting the corresponding spectral information from galaxy images. This network architecture mainly includes an input module, a feature extraction module, and an output module, where the feature extraction module is the core innovative part of the present invention and is composed of a discrete wavelet transform module and a multi-layer convolution module.

[0005] The input module receives galaxy images with a size of 64×64. These images serve as the input data for the network, providing basic information for subsequent feature extraction and spectral prediction. The feature extraction module first processes the input images through the discrete wavelet transform module. This module uses two-dimensional discrete wavelet transform technology to decompose the input image into a low-frequency sub-band (LL), a horizontal high-frequency sub-band (HL), a vertical high-frequency sub-band (LH), and a diagonal high-frequency sub-band (HH). This decomposition method can effectively capture the feature information of the image at different scales, providing a multi-scale perspective for subsequent feature extraction. The four decomposed sub-bands are concatenated and output by channel, further enhancing the feature expression ability and providing rich input data for subsequent convolution operations.

[0006] Based on the discrete wavelet transform module, the multi-layer convolution module further processes the feature map to extract richer local feature information. This module consists of multiple convolutional layers, and each convolutional operation is followed by a batch normalization layer and a ReLU activation function. The batch normalization layer is used to accelerate the training process of the network, while improving the stability and generalization ability of the model; the ReLU activation function introduces non-linearity to the network, enabling the network to learn more complex feature representations. In addition, the number of channels in the convolutional layer increases layer by layer. This design enables the network to gradually extract higher-level feature representations, thereby better capturing the local details and complex structures in the image. Through multi-layer convolution operations, the network can effectively fuse the multi-scale features and local detail features of the input image, providing strong feature support for the final spectral prediction.

[0007] Finally, the output module maps the feature vector processed by the feature extraction module to the spectral feature space and outputs the spectral prediction result corresponding to the input image. This module realizes the mapping from the feature vector to the spectral data through a fully connected layer, and its output dimension matches the dimension of the spectral data, thus ensuring the accuracy and reliability of the prediction result.

[0008] Through the above technical solution, the present invention can, based on the input galaxy images, utilize the synergistic effect of discrete wavelet transform and multi-layer convolution to efficiently extract the multi-scale features and local detail information of the images and map them to the spectral feature space, thereby achieving high-precision prediction of galaxy spectra. This technical solution not only improves the accuracy of spectral prediction but also provides an efficient and reliable spectral prediction method for astronomical research, having important scientific significance and application value.

[0009] The present invention also discloses a galaxy image spectral prediction network based on discrete wavelet transform and multi-layer convolution, including the following steps:

[0010] S1: Data Acquisition and Preprocessing: The experimental data is sourced from the galaxy images in SDSS DR18 and their corresponding spectral information. First, the galaxy images are matched one by one with the spectral data, and a series of screening operations are carried out to remove noise and invalid data. The galaxy images adopt the minimum size of 64×64 provided by SDSS, and at the same time, the spectral data is cropped to the wavelength range of [3.5901, 3.9499], focusing on the key spectral features. After the above processing, a high-quality dataset for training and testing is obtained.

[0011] S2: Model Training: Use the processed dataset above to train the galaxy spectral prediction network, and learn the network parameters through the optimization algorithm, and finally obtain the trained model weights.

[0012] S3: Load the trained model weight parameters, input the test set, obtain the accuracy index of the test set spectral prediction, and sequentially judge the classification performance of the network.

[0013] S4: Use the model parameters and prediction indicators determined in S3 to determine the classification effect and generalization performance of the network.

[0014] Beneficial Effects: Compared with the prior art, the present invention has the following remarkable advantages: Multi-scale feature extraction: Through the discrete wavelet transform module, it can effectively extract the multi-scale features of the image and capture the important information at different scales; Local detail enhancement: The multi-layer convolution module can further extract the local detail features of the image, improving the richness and expression ability of the features; High prediction accuracy: Through the above technical means, the present invention can significantly improve the accuracy of galaxy image spectral prediction and provide more accurate data support for astronomical research. Brief Description of the Drawings

[0015] Figure 1 is the network structure diagram of the present invention;

[0016] Figure 2 is the network diagram of the DWT module in the present invention;

[0017] Figure 3 is the basic flowchart of the present invention;

[0018] Figure 4 is the prediction result diagram of the present invention. Detailed Embodiments

[0019] The following further elaborates on the content of the present invention in conjunction with the drawings.

[0020] Such as Figure 1The figure shows a schematic diagram of the network structure for spectral prediction of galaxy images using discrete wavelet transform and multi-layer convolution. It integrates discrete wavelet transform (DWT) and multi-layer convolution operations to extract multi-scale features of galaxy images. First, the input 64×64 galaxy image is decomposed by the DWT module to capture feature information at different resolutions. Subsequently, these feature maps pass through a series of repeated convolution modules (ConvBlock), each consisting of a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function, gradually extracting deeper features and increasing the number of channels, while downsampling the spatial dimensions through a convolution with a stride of 2. After multiple feature extractions and abstractions, the feature maps are flattened into one-dimensional feature vectors by the Flatten layer, and finally mapped to a 3598-dimensional spectral feature space through a fully connected layer to output the predicted spectral data.

[0021] As Figure 2 The figure shows a schematic diagram of the structure of the DWT module in the network. In this galaxy image spectral prediction network, the DWT (discrete wavelet transform) module is located at the front end of the network and is responsible for extracting multi-scale features from the input 64×64 galaxy image, decomposing it into four sub-bands: LL (low frequency), HL (horizontal high frequency), LH (vertical high frequency), and HH (diagonal high frequency), with the spatial dimension of each sub-band being 32×32. These sub-bands are then concatenated to form a new feature map, whose number of channels is four times that of the input channels, i.e., 12 channels, while the spatial dimension remains unchanged. This process not only enhances the network's ability to capture features at different scales in the image but also provides richer information for subsequent convolution operations through feature fusion, thus improving the accuracy and efficiency of the entire network for spectral prediction of galaxy images.

[0022] As Figure 3 The figure shows the basic flow chart of the present invention, including the following steps:

[0023] S1: To ensure the accuracy and reliability of the galaxy image spectral prediction network, the right ascension and declination information of galaxies was systematically retrieved from the SDSS DR18 dataset, and the corresponding galaxy images and spectral data were downloaded using these coordinates. In the data preprocessing stage, strict screening and cleaning were performed on the downloaded data, including removing noise and invalid data, and cropping the spectral data to the wavelength range of [3.5901, 3.9499] to focus on key spectral features.

[0024] S2: During model training, we use the preprocessed dataset to train the proposed galaxy image spectral prediction network based on discrete wavelet transform (DWT) and multi-layer convolution. This step involves learning and optimizing the network weights to minimize the difference between the predicted spectrum and the true spectrum. The network utilizes the multi-resolution analysis ability of DWT to extract multi-scale features of galaxy images, enabling the network to capture a wide range of information from low-frequency trends to high-frequency details. Additionally, by using wavelet transform in the convolutional layer and entropy coding, the network can perform feature learning more effectively, thereby enhancing the ability to learn image compression. The multi-layer convolutional structure of the network further abstracts and integrates these features, enhancing the model's ability to identify key spectral features. In this way, the model training process not only improves the accuracy of spectral prediction but also provides a new analysis tool for astronomical research, contributing to a deeper understanding of the physical properties and evolution process of galaxies. After training, the predicted weights can be obtained;

[0025] S3: Model validation and performance evaluation are crucial stages to ensure the effectiveness and reliability of the model in practical applications. In this step, we load the trained model weights and evaluate the model using the test set. We adopt several commonly used performance evaluation metrics to measure the prediction performance of the model, including mean squared error (MSE), mean absolute error (MAE), root mean squared error (RMSE), and Pearson correlation coefficient.

[0026] The present invention realizes the efficient prediction of galaxy images to spectral information through a systematic method from data acquisition to model evaluation. This method not only improves the accuracy of spectral prediction but also provides a new tool for astronomical research, contributing to a deeper understanding of the physical properties and evolution process of galaxies.

[0027] Table 1 shows the influence of different numbers of ConvBlock modules on the model performance in the present invention, where the measurement metrics include mean squared error (MSE), mean absolute error (MAE), and root mean squared error (RMSE):

[0028] Table 1 Setting selection of the repetition times of ConvBlock modules in the present invention

[0029] MSE MAE RMSE DWT+ConvBlock 0.4446 0.3555 0.6129 DWT+ConvBlock×2 0.4051 0.3496 0.6027 DWT+ConvBlock×3 0.3846 0.3412 0.5712 DWT+ConvBlock×4 0.3281 0.3397 0.5428 DWT+ConvBlock×5 0.4428 0.3419 0.5986

[0030] Table 2 shows the experimental results of comparing the spectral prediction network of the present invention with other algorithms based on convolutional neural network (CNN) in terms of performance. The prediction networks involved in the comparison include traditional LSTM, ResNet18, and U-Net++, as well as the versions after combining these networks with discrete wavelet transform (DWT), namely DWT+LSTM, DWT+ResNet18, and DWT+U-Net++. The metrics for comparative evaluation cover mean square error (MSE), mean absolute error (MAE), and root mean square error (RMSE), which can comprehensively reflect the performance of each network model in the spectral prediction task.

[0031] Table 2 Comparison of Experimental Results between the Present Invention and Other Algorithms Based on Convolutional Neural Network

[0032] Prediction network MSE MAE RMSE LSTM 0.5571 0.5843 0.7492 DWT+LSTM 0.5061 0.5472 0.7192 ResNet18 0.8731 0.7492 0.9743 DWT+ResNet18 0.4763 0.5186 0.6785 U-Net++ 0.6544 0.6294 0.8472 DWT+U-Net++ 0.5735 0.6185 0.7492 The spectral prediction network of the present invention 0.3281 0.3397 0.5428

[0033] As Figure 4 shown, the prediction result graphs of the present invention clearly show the galaxy images and their corresponding spectra. In each graph, the galaxy image is shown on the left, while the corresponding spectral graph is on the right. Among them, the purple curve represents the true spectrum obtained through observation, and the red curve represents the spectrum predicted by the model of the present invention. By comparison, it can be found that the predicted spectrum and the true spectrum show a high degree of consistency in terms of overall continuity and change trend. In addition, at multiple key feature points of the spectrum, the two also show good synchronization, which further proves the accuracy and reliability of the model of the present invention in capturing the spectral characteristics of galaxies.

[0034] In summary, the network structure of the present invention demonstrates comparable or even superior performance to existing advanced methods in specific tasks. It can accurately predict the corresponding spectral information through galaxy images, providing a new and efficient analysis tool for astronomical research.

Claims

1. A galaxy image spectrum prediction network based on discrete wavelet transform and multi-layer convolution, characterized by: Input module, feature extraction module and output module; The feature extraction module includes a discrete wavelet transform module and a multi-layer convolution module, which is used to extract multi-scale feature information from the input galaxy image and improve the expression ability of the model through layer-by-layer abstraction; The discrete wavelet transform module performs a two-dimensional discrete wavelet transform on the input image to extract multi-scale features of the image; The multi-layer convolution module further extracts local feature information of the feature map through layer-by-layer convolution operation; The output module maps the extracted features to a spectral feature space and outputs a spectral prediction result corresponding to the input image.

2. The galaxy image spectrum prediction network based on discrete wavelet transform and multi-layer convolution according to claim 1, characterized in that: The discrete wavelet transform module decomposes the input image into a low frequency sub-band (LL), a horizontal high frequency sub-band (HL), a vertical high frequency sub-band (LH) and a diagonal high frequency sub-band (HH) through a two-dimensional discrete wavelet transform; The sub-bands are spliced ​​and output according to channels to achieve multi-scale feature extraction of the image.

3. The galaxy image spectrum prediction network based on discrete wavelet transform and multi-layer convolution according to claim 2, characterized in that: The implementation of the discrete wavelet transform module includes the following steps: The input image is downsampled by rows and columns to obtain four sub-bands; the sub-bands are linearly combined to generate a low-frequency sub-band (LL), a horizontal high-frequency sub-band (HL), a vertical high-frequency sub-band (LH) and a diagonal high-frequency sub-band (HH); the four sub-bands are concatenated by channel and the transformed feature map is output.

4. The galaxy image spectrum prediction network based on discrete wavelet transform and multi-layer convolution according to claim 1, characterized in that: The multi-layer convolution module includes multiple convolution layers and batch normalization layers, and each convolution layer is connected to a ReLU activation function; the number of channels of the multi-layer convolution module increases layer by layer to extract richer local feature information.

5. According to the galaxy image spectrum prediction network based on discrete wavelet transform and multi-layer convolution according to claim 4, in the multi-layer convolution module, the implementation method of the convolution layer includes the following steps: The input feature map is convolved using a 3×3 convolution kernel; a batch normalization layer is applied after each convolution layer to accelerate training and improve the stability of the model; and a ReLU activation function is applied to introduce nonlinear feature mapping.

6. According to the galaxy image spectrum prediction network based on discrete wavelet transform and multi-layer convolution according to claim 4, in the multi-layer convolution module, the stride and padding method of the convolution layer are dynamically adjusted according to the size of the feature map to maintain the spatial resolution of the feature map.

7. According to the galaxy image spectrum prediction network based on discrete wavelet transform and multi-layer convolution according to claim 1, the feature extraction module also includes a downsampling module for capturing multi-scale feature information to help the model identify key features in images of different sizes and complexities; the downsampling module implements downsampling of the feature map through a convolution operation with a stride of 2 to extract coarser-grained feature information.

8. According to the galaxy image spectrum prediction network based on discrete wavelet transform and multi-layer convolution as described in claim 1, in the whole feature extraction process, discrete wavelet transform modules and multi-layer convolution modules are alternately stacked to realize multi-scale feature extraction and layer-by-layer abstraction of the input image; in the network structure, the number of multi-layer convolution modules is 3, 4, 6, and 3 respectively, and the depth and width of the convolution layer can be modified at each stage to meet the tasks and requirements of different data sets.

Images