A method for eliminating beam coupling effect in main beam synthesis of radio interferometer arrays

Abstract

This invention discloses a method for eliminating the coupling effect of the main beam and synthesized beam in radio interferometry arrays, belonging to the field of radio astronomy imaging. The implementation steps include: S1: constructing the residual processing module RpDH-Deeplab; S2: constructing the encoder structure in the segmentation network; S3: constructing the decoder structure in the segmentation network; S4: combining the results obtained in S1 with the decoder results obtained in S3, and using the activation function Tanh to obtain the output result of the RpDH-Deeplab network. Compared with other methods that separately correct the main beam effect and the synthesized beam, this invention achieves unified correction, avoiding the scenario where the two effects interfere with each other during the correction process. After the deep learning neural network model is trained, the trained neural network model can accurately model the data and extract the features of the main beam effect. The neural network model has excellent generalization ability, and even radio celestial structures not in the dataset are equally applicable to the trained neural network model.

Description

Technical Field

[0001] This invention relates to the field of radio astronomy imaging, and more particularly to a method for eliminating beam coupling effects in the main beam synthesis of radio interferometer arrays. Background Technology

[0002] As indicated by the Radio Interferometer Measurement Equation (RIME), radio observations exhibit both direction-dependent effects (DDEs) and direction-independent effects (DIEs). Accurate correction of these DDEs is a prerequisite for achieving radio astronomy imaging. For DDEs, the self-calculation method included in the traditional Second Generation Calibration (2GC) can achieve calibration. However, for more complex DDEs, the 2GC calibration method is no longer applicable. Therefore, the Third Generation Calibration (3GC) method is needed to correct DDEs, which is a current research trend. The main beam effect is the most significant component of DDEs. Furthermore, the synthetic beam effect also exists during observation, and the effectiveness of its coupling effect correction directly impacts the imaging quality of radio astronomy observations.

[0003] Despite the numerous correction methods applied to radio interferometer arrays (RIAs), their imaging is still affected by the coupling effect of the main beam synthesis beam. This is particularly true for observations of mixed and extended sources at high frequencies, where the coupling effect severely impacts the imaging quality at the edges of the field of view. Therefore, coupling remains a critical problem that needs to be addressed in radio astronomy imaging using RIAs. Furthermore, the massive amounts of data generated with each observation make data processing a major obstacle to rapid radio observations by RIAs. In the era of big data, even more data needs to be processed in parallel, but traditional data processing techniques in radio astronomy are no longer sufficient. Given the ability of deep learning to rapidly extract key information from large datasets, deep learning-based astronomical data processing will play a crucial role and represent a significant trend in the future development of radio astronomy.

[0004] Among the commonly used algorithms for eliminating the coupling effect between the main beam and the synthetic beam, the traditional image domain main beam correction method is the most widely applied. However, this method is only suitable for eliminating image effects at high signal-to-noise ratios and its effect on eliminating coupling effects at high frequencies is very poor. Therefore, it is difficult to achieve good performance in eliminating the coupling effect between the main beam and the synthetic beam of radio interferometer arrays. This patent addresses the problem of processing the coupling effect between the main beam and the synthetic beam of radio interferometer arrays by proposing a method for eliminating the coupling effect between the main beam and the synthetic beam of radio interferometer arrays. Summary of the Invention

[0005] The present invention aims to provide a method for eliminating beam coupling effects in the main beam synthesis of radio interferometer arrays, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for eliminating beam coupling effects in main beam synthesis of radio interferometer arrays, the specific steps of which include:

[0008] S1: Construct the residual processing module RpDH-Deeplab to process the input image containing effects, providing the most comprehensive and effective information F from the input image for subsequent networks. out This helps to more effectively correct the main beam effect;

[0009] S2: Construct the encoder structure in the segmentation network. This structure uses the MobileNet v2 backbone network to extract features from the input image at different levels. For the highest-level features, the DH-ASPP module is used for further feature extraction, ultimately yielding the DH-ASPP module output. out ;

[0010] S3: Construct the decoder structure in the segmentation network. The decoder's role is to combine the low-level features output from the encoder with the high-level features DH-ASPP. out Feature fusion is performed, and the fusion of multiple features ensures that more contextual information is input into the neural network, thereby ensuring better correction of coupling effects and ultimately obtaining the decoder output. out ;

[0011] S4: The result F obtained from the residual processing module in S1 out Decoder result obtained from S3 outBy combining these parameters and using the activation function Tanh, the output of the RpDH-Deeplab network is obtained. Through continuous training, the parameters in the network structure are updated, enabling the final output of RpDH-Deeplab to accurately correct the main beam synthesis beam effect.

[0012] Preferably, the residual processing module in step S1 consists of two 3×3 convolutional layers, two feature extraction attention blocks (EABs), and residual connections; the EAB module is composed of two parts: feature extraction and attention modules.

[0013] The feature extraction module uses 3×3 convolution to perform preliminary feature extraction on the information of the original input image, and the residual connection preserves the information of the original input image;

[0014] The feature attention module consists of spatial attention, channel attention, and global average pooling. The output features of the feature extraction module are the input features F of the attention part. in In the attention module, spatial attention, channel attention, and hybrid attention mechanisms are first applied to obtain F respectively. s F c and F cs Then for F cs F is obtained by global average pooling operation p ,F p respectively with F c and F s Performing a multiplication operation yields F cp With F sp F sp With F in Multiply, then multiply with F cp Multiply by each product to obtain the final output F. out ;

[0015] As shown in the formula:

[0016] F out =F cs *F p *F in *(F s *F p )

[0017] Preferably, the segmentation network in step S1 is an improvement on Deeplab v3+, which enables it to better extract celestial structures from the input main beam effect image; the segmentation network is an encoding and decoding structure, which is beneficial for better feature extraction.

[0018] The RpDH-Deeplab network uses MobileNet v2 as the backbone of the segmentation network;

[0019] MobileNet v2 employs inverted residuals. Compared to regular residual structures, inverted residuals first perform 1×1 convolutions for dimensionality increase, then 3×3 depthwise separable convolutions, and finally 1×1 convolutions for dimensionality reduction. MobileNet v2 is composed of multiple inverted residual structures stacked together. The expansion factor is the number of repeated structures in the inverted residual structure, the channels refer to the change in the number of feature channels before and after each operation, and the stride applies only to the first layer of each repeated structure.

[0020] MobileNet v2 includes four convolutional operations with a stride of 2, resulting in four changes in feature size: 184×184, 92×92, 46×46, and 23×23, corresponding to F2, F4, F8, and F1 scores. 16 There are 4 different scales of features, of which F 16 For the highest level features; in RpDH-Deeplab, F 16 The input will be fed into the DH-ASPP module for further feature extraction, while F2, F4, and F8 will be fed into the decoder part for feature fusion.

[0021] SPP (Sequential Propagation Point) effectively extracts contextual information from input images. Deeplab v3+ adds dilated convolution to SPP. Dilated convolution increases the receptive field and reduces computation without sacrificing spatial resolution or input information, while capturing more contextual information. It performs well on large structures in detection and segmentation tasks. However, because dilated convolution operates in a chessboard-like manner, the results of convolution operations at a certain layer all come from independent sets in the previous layer, and the results are unrelated. This leads to a lack of correlation between the convolution results, resulting in the loss of local information. Considering that radio astronomy observations contain many weak, small-scale celestial structures, dilated convolution operations can prevent the effective extraction of features from these small-scale structures.

[0022] Therefore, hybrid dilated convolution was added to the RpDH-Deeplab network to extract features at different scales using different dilation rates. This resulted in more connections between the convolution results and covered multiple different sizes, achieving the goal of accurate feature extraction for small-scale astronomical structures. Considering the complexity of the structure after adding hybrid dilated convolution, dense connection operations were added. Based on the ASPP module and combined with the characteristics of radio object structures, dense connections and hybrid dilated convolution (HDC) were added, and finally, the DH-ASPP network structure in RpDH-Deeplab was constructed.

[0023] Preferably, the DH-ASPP structure includes four hybrid dilated convolutions: Part-1, Part-2, Part-3 to Part-4. Each hybrid dilated convolution consists of three dilated convolution modules. Each convolution module contains a convolutional layer, a batch normalization layer, and an activation function. By configuring different dilation rates, the dilated convolution results are interconnected.

[0024] Part 5 involves pooling, convolution, and upsampling operations. A small encoder-decoder module, ED, is constructed to perform the final feature extraction, as shown in the formula:

[0025]

[0026] Here, Cat represents the concatenation operation. This represents a hybrid dilated convolution combination, where i1, i2, and i3 are three different convolution rates for each hybrid dilated convolution part.

[0027] Preferably, the decoder part of the RpDH-Deeplab model described in step S3 captures contextual information and outputs it in the same resolution as the input image, thereby achieving fine pixel-level segmentation.

[0028] The decoder's role is to fuse the low-level features output from the encoder with the high-level features. The fusion process is shown in the equation:

[0029] Decorder out =U N (Cat(F2,U2(Cat(F4,U4(Cat(F8,U8(Encoder out )))))))

[0030] Decorder out For the final output of the decoder, U NThis indicates an N-fold upsampling operation, Cat represents the concatenation operation, and Encoder represents the encoding. out This is the output of the encoder.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. Deep learning method for correcting coupling effects between main beam and synthetic beam in radio interferometer arrays

[0033] Compared with other methods that separately correct the main beam effect and the synthesized beam, the method in this paper achieves unified correction, avoiding the scenario where the two effects interfere with each other during the correction process.

[0034] 2. Excellent transferability of the method. Once trained, deep learning neural network models can be used "once and for all". A well-trained neural network model can accurately model data, thereby extracting the characteristics of the main beam effect. Due to the excellent generalization ability of neural network models, even radio structures not in the dataset are still applicable to well-trained neural network models. Attached Figure Description

[0035] Figure 1 A flowchart of a method for eliminating beam coupling effects in the main beam synthesis of radio interferometer arrays;

[0036] Figure 2 This is a schematic diagram of the residual processing module;

[0037] Figure 3 This is a schematic diagram of the RpDH-Deeplab network framework structure;

[0038] Figure 4 Here is a block diagram of the encoder structure;

[0039] Figure 5 A schematic diagram of the DH-ASPP network structure;

[0040] Figure 6 This is a schematic diagram of the decoder structure. Detailed Implementation

[0041] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments:

[0042] The specific implementation process is as follows:

[0043] like Figure 1 The method shown is for eliminating beam coupling effects in the main beam synthesis of radio interferometer arrays. The specific steps of the method implementation include:

[0044] S1: Construct the residual processing module RpDH-Deeplab to process the input image containing effects, providing the most comprehensive and effective information F from the input image for subsequent networks. out This helps to more effectively correct the main beam effect;

[0045] S2: Construct the encoder structure in the segmentation network. This structure uses the MobileNet v2 backbone network to extract features at different levels from the input image. For the highest-level features, the DH-ASPP module is used for further feature extraction. Dense connections and hybrid dilated convolutions reduce the model's computational parameters, significantly lowering the computational cost. Furthermore, it prevents overfitting, resulting in a stronger overall model generalization ability. The final output of the DH-ASPP module is obtained. out ;

[0046] S3: Construct the decoder structure in the segmentation network. The decoder's role is to combine the low-level features output from the encoder with the high-level features DH-ASPP. out Feature fusion is performed, and the fusion of multiple features ensures that more contextual information is input into the neural network, thereby ensuring better correction of coupling effects and ultimately obtaining the decoder output. out ;

[0047] S4: The result F obtained from the residual processing module in S1 out Decoder result obtained from S3 out By combining these parameters and using the activation function Tanh, the output of the RpDH-Deeplab network is obtained. Through continuous training, the parameters in the network structure are updated, enabling the final output of RpDH-Deeplab to accurately correct the main beam synthesis beam effect.

[0048] Considering that deep learning models need to extract contour and structural features, a neural network model that is more sensitive to contours should be chosen. The input of celestial brightness information in images containing coupling effects can be achieved using residual connections in ResNet. Therefore, the main architecture of the deep learning model can be determined to be an image segmentation network with residual connections.

[0049] Based on the impact of coupling effects on image formation, this paper, after in-depth research into relevant deep learning theories, proposes the RpDH-Deeplab model to correct for these coupling effects. The overall framework of the RpDH-Deeplab model consists of a residual processing module and a segmentation network module, such as... Figure 2 As shown;

[0050] The residual processing module inputs more effective information from the input image into the neural network, providing more effective original image information for coupling effect correction. This enables RpDH-Deeplab to achieve more efficient and accurate coupling effect correction. The residual module, for example... Figure 3 As shown.

[0051] The residual processing module described in step S1 consists of two 3×3 convolutional layers, two feature extraction attention blocks (EABs), and residual connections; the EAB module is composed of two parts: feature extraction and attention modules.

[0052] The feature extraction module uses 3×3 convolution to perform preliminary feature extraction on the information of the original input image, and the residual connection preserves the information of the original input image;

[0053] The feature attention module consists of spatial attention, channel attention, and global average pooling. The output features of the feature extraction module are the input features F of the attention part. in In the attention module, spatial attention, channel attention, and hybrid attention mechanisms are first applied to obtain F respectively. s F c and F cs Then for F cs F is obtained by global average pooling operation p ,F p respectively with F c and F s Performing a multiplication operation yields F cp With F sp F sp With F in Multiply, then multiply with F cp Multiply by each product to obtain the final output F. out ;

[0054] As shown in the formula:

[0055] F out =F cs *F p *F in *(F s *F p )

[0056] The attention module iteratively extracts important spatial and channel information from the input image. Therefore, the entire residual processing module provides the most comprehensive and effective information from the input image to subsequent networks, helping to more effectively correct the main beam effect.

[0057] The segmentation network described in step S1 is an improvement on Deeplab v3+, enabling it to better extract celestial structures from the input main beam effect image. The segmentation network is an encoder-decoder structure, which facilitates better feature extraction. The encoder part is as follows... Figure 4 As shown.

[0058] Compared to X-ception in the original Deeplab v3+, the RpDH-Deeplab network uses MobileNetv2 as the backbone of the segmentation network. While ensuring performance, it greatly reduces the number of parameters, improves the convergence speed of the network, and is more suitable for situations with a large amount of data in radio astronomy.

[0059] MobileNet v2 employs inverted residuals. Compared to regular residual structures, inverted residuals first perform 1×1 convolutions for dimensionality increase, then 3×3 depthwise separable convolutions, and finally 1×1 convolutions for dimensionality reduction. MobileNet v2 is composed of multiple inverted residual structures stacked together. The expansion factor is the number of repeated structures in the inverted residual structure, the channels refer to the change in the number of feature channels before and after each operation, and the stride applies only to the first layer of each repeated structure.

[0060] MobileNet v2 includes four convolutional operations with a stride of 2, resulting in four changes in feature size: 184×184, 92×92, 46×46, and 23×23, corresponding to F2, F4, F8, and F1 scores. 16 There are 4 different scales of features, of which F 16 For the highest level features; in RpDH-Deeplab, F 16 The input will be fed into the DH-ASPP module for further feature extraction, while F2, F4, and F8 will be fed into the decoder part for feature fusion.

[0061] SPP (Sequential Propagation Point) effectively extracts contextual information from input images. Deeplab v3+ adds dilated convolution to SPP. Dilated convolution increases the receptive field and reduces computation without sacrificing spatial resolution or input information, while capturing more contextual information. It performs well on large structures in detection and segmentation tasks. However, because dilated convolution operates in a chessboard-like manner, the results of convolution operations at a certain layer all come from independent sets in the previous layer, and the results are unrelated. This leads to a lack of correlation between the convolution results, resulting in the loss of local information. Considering that radio astronomy observations contain many weak, small-scale celestial structures, dilated convolution operations can prevent the effective extraction of features from these small-scale structures.

[0062] Therefore, hybrid dilated convolution (HDC) was added to the RpDH-Deeplab network to extract features at different scales using different dilation rates. This resulted in more connections between the convolution results and covered multiple different sizes, achieving accurate feature extraction for small-scale astronomical structures. Considering the complexity of the structure after adding HDC, dense connection operations were added. Benefiting from the advantages of residual connections, the addition of dense connections reduced the computational parameters of the model, thus significantly reducing the computational load. Furthermore, it prevented overfitting and improved the overall model's generalization ability. Based on the ASPP module and considering the characteristics of radio celestial structures, dense connections and hybrid dilated convolution (HDC) were added, ultimately constructing the DH-ASPP network structure in RpDH-Deeplab, as shown below. Figure 4 As shown.

[0063] The DH-ASPP structure includes four hybrid dilated convolutions: Part-1, Part-2, Part-3 to Part-4. Each hybrid dilated convolution consists of three dilated convolution modules. Each convolution module contains a convolutional layer, a batch normalization layer, and an activation function. By configuring different dilation rates, the dilated convolution results are interconnected.

[0064] Part 5 involves pooling, convolution, and upsampling operations. A small encoder-decoder module, ED, is constructed to perform the final feature extraction, as shown in the formula:

[0065]

[0066] Here, Cat represents the concatenation operation. This represents a hybrid dilated convolution combination, where i1, i2, and i3 are three different convolution rates for each hybrid dilated convolution part.

[0067] The decoder part of the RpDH-Deeplab model described in step S3 captures contextual information and outputs it at the same resolution as the input image, thereby achieving fine pixel-level segmentation. The decoder network diagram is as follows: Figure 6 As shown;

[0068] The decoder's role is to fuse the low-level features output from the encoder with the high-level features. The fusion process is shown in the equation:

[0069] Decorder out =U N (Cat(F2,U2(Cat(F4,U4(Cat(F8,U8(Encoder out )))))))

[0070] Decorder out For the final output of the decoder, U N This indicates an N-fold upsampling operation, Cat represents the concatenation operation, and Encoder represents the encoding. out This is the output of the encoder. The fusion of multiple features ensures that more contextual information is input into the neural network, thereby ensuring better correction of coupling effects.

[0071] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for eliminating beam coupling effects in the main beam synthesis of radio interferometer arrays, characterized in that, The specific steps for implementing the method include: S1: Construct a residual processing module in the RpDH-Deeplab network to process the input image containing effects, thereby providing the most comprehensive and effective information F from the input image for subsequent networks. out This helps to more effectively correct the main beam effect; S2: Construct the encoder structure of the segmentation network within the RpDH-Deeplab network. This structure uses the MobileNet v2 backbone network to extract features at different levels from the input image. For the highest-level features, the DH-ASPP module is used for further feature extraction, ultimately yielding the DH-ASPP module output. out ; S3: Construct the decoder structure of the segmentation network in the RpDH-Deeplab network. The role of the decoder is to combine the low-level features output from the encoder with the high-level features DH-ASPP. out Feature fusion is performed, and the fusion of multiple features ensures that more contextual information is input into the neural network, thereby ensuring better correction of coupling effects and ultimately obtaining the decoder output. out ; S4: The result F obtained from the residual processing module in step S1 is... out The decoder output obtained in step S3 out By combining these parameters and using the activation function Tanh, the output of the RpDH-Deeplab network is obtained. Through continuous training, the parameters in the network structure are updated, enabling the final output of the RpDH-Deeplab network to accurately correct the main beam synthesis beam effect. The segmentation network described in step S2 is an improvement on Deeplab v3+, enabling it to better extract celestial structures from the input main beam effect image; the segmentation network is an encoding / decoding structure, which is beneficial for better feature extraction. The RpDH-Deeplab network uses MobileNet v2 as the backbone of the segmentation network; MobileNet v2 employs an inverse residual structure. Compared to a typical residual structure, the inverse residual structure first performs a 1×1 convolution operation to increase dimensionality, then a 3×3 depthwise separable convolution, and finally a 1×1 convolution to reduce dimensionality. MobileNet v2 is composed of multiple inverse residual structures stacked together. The expansion factor is the number of repeated structures in the inverse residual structure, the channels refer to the change in the number of feature channels before and after each operation, and the stride is only applied to the first layer of each repeated structure. MobileNetv2 includes four convolutional operations with a stride of 2, resulting in four changes in feature size: 184×184, 92×92, 46×46, and 23×23, corresponding to F2, F4, F8, and F1 scores. 16 There are 4 different scales of features, of which F 16 The highest level feature; in the RpDH-Deeplab network, F 16 The input will be fed into the DH-ASPP module for further feature extraction, while F2, F4, and F8 will be fed into the decoder part for feature fusion. Based on the ASPP module and taking into account the characteristics of radio celestial bodies, dense connections and hybrid dilated convolutions were added, and the network structure of the DH-ASPP module in the RpDH-Deeplab network was finally constructed.

2. The method for eliminating beam coupling effect in main beam synthesis of radio interferometer arrays according to claim 1, characterized in that, The residual processing module described in step S1 consists of two 3×3 convolutional layers, two feature extraction attention modules, and residual connections; the feature extraction attention module is composed of two parts: a feature extraction module and an attention module. The feature extraction module uses 3×3 convolution to perform preliminary feature extraction on the information of the original input image, and the residual connection preserves the information of the original input image; The attention module consists of a spatial attention module, a channel attention module, and a global average pooling module; the output features of the feature extraction module are the input features F of the attention part. in In the attention module, spatial attention, channel attention, and hybrid attention mechanisms are first applied to obtain F respectively. s F c and F cs Then for F cs F is obtained by global average pooling operation p ,F p respectively with F c and F s Performing a multiplication operation yields F cp With F sp F sp With F in Multiply, then multiply with F cp Multiply by each product to obtain the final output F. out ; As shown in the formula: F out =F cs *F p *F in *(F s *F p )。 3. The method for eliminating beam coupling effect in main beam synthesis of radio interferometer arrays according to claim 1, characterized in that: The DH-ASPP module structure includes four hybrid dilated convolutions: Part-1, Part-2, Part-3 to Part-4. Each hybrid dilated convolution consists of three dilated convolution modules. Each dilated convolution module contains a convolutional layer, a batch normalization layer, and an activation function. By configuring different dilation rates, the dilated convolution results are interconnected. The DH-ASPP module also includes Part-5, which performs pooling, convolution, and upsampling operations to construct a small encoder-decoder module, ED, to extract the final features, as shown in the formula: Here, Cat represents the concatenation operation. This represents a hybrid dilated convolution combination, where i1, i2, and i3 are three different convolution rates for each hybrid dilated convolution part.

4. The method for eliminating beam coupling effect in main beam synthesis of radio interferometer arrays according to claim 1, characterized in that, The decoder part of the RpDH-Deeplab network described in step S3 captures contextual information and outputs it in the same resolution as the input image, thereby achieving fine pixel-level segmentation. The decoder's role is to fuse the low-level features output from the encoder with the high-level features. The fusion process is shown in the equation: Decorder out =U N (Cat(F2,U2(Cat(F4,U4(Cat(F8,U8(Encoder out ))))))) Decorder out For the final output of the decoder, U N This indicates an N-fold upsampling operation, Cat represents the concatenation operation, and Encoder represents the encoding. out This is the output of the encoder.

Images