Main beam forming beam coupling effect elimination method for radio interference array

Abstract

The invention discloses a main beam forming beam coupling effect elimination method for a radio interference array, and belongs to the field of radio astronomical imaging, and the method comprises the implementation steps: S1, constructing a residual error processing module RpDH-Deeplab; s2, constructing an encoder structure in the segmentation network; s3, constructing a decoder structure in the segmentation network; s4, combining a result obtained in the step S1 with a decoder result obtained in the step S3, and obtaining an output result of the RpDH-Deeplab network by using an activation function Tanh; compared with other methods for separately correcting the main beam effect and the synthetic beam, unified correction is realized, and the scene that the two effects influence each other in the correction process is avoided; after the deep learning neural network model is trained, the trained neural network model can realize accurate modeling of data and realize extraction of main beam effect features, the neural network model has excellent generalization ability, and the trained neural network model is also suitable for radio celestial body structures not in a data set.

Description

Technical Field

[0001] The invention relates to the field of radio astronomy imaging, and in particular to a method for eliminating main beam synthesis beam coupling effects for a radio interference array. Background Art

[0002] From the Radio Interferometer Measurement Equation (RIME), we know that there are direction-dependent effects (DDEs) and direction-independent effects (DIEs) in radio observations. The premise for achieving radio astronomical imaging is to accurately correct the direction-dependent effects and direction-independent effects. For the direction-independent effect, the self-cal method included in the traditional second-generation calibration method (2GC) can achieve calibration, but for the more complex direction-dependent effect, the 2GC calibration method is no longer applicable. Therefore, the third-generation calibration method (3GC) is needed to correct the direction-dependent effect, which is also a current research trend. The main beam effect is the most important part of the direction-dependent effect. In addition, there is also a synthetic beam effect in the observation process, and its coupling effect correction effect directly affects the imaging quality of radio astronomical observations.

[0003] Although there are many correction methods applied to radio interferometer arrays, their imaging is still affected by the coupling effect of the main beam and the synthetic beam. Especially for the observation of mixed sources and extended sources at high frequencies, the coupling effect seriously affects the imaging quality at the edge of the field of view. Therefore, coupling is still an urgent problem to be solved in radio interferometer arrays for radio astronomical observation imaging. In addition, since each observation generates a large amount of observation data, data processing has become the biggest obstacle to the rapid radio observation of radio interferometer arrays. Then in the era of big data, more data needs to be processed in parallel, but traditional data processing technology in radio astronomy can no longer meet the needs. Since deep learning has the characteristics of quickly extracting key information in big data, astronomical data processing based on deep learning will play an important role and will also be an important 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 most widely used is the traditional image domain main beam correction method. However, this method is only applicable to the elimination of image effects with a high signal-to-noise ratio, and has a poor effect on eliminating coupling effects at high frequencies. Therefore, it is difficult to achieve good performance in eliminating the coupling effect between the main beam and the synthetic beam of a radio interferometer array. This patent aims at solving the problem of dealing with the coupling effect between the main beam and the synthetic beam of a radio interferometer array, and proposes a method for eliminating the coupling effect between the main beam and the synthetic beam of a radio interferometer array. Summary of the invention

[0005] The present invention is intended to provide a method for eliminating the coupling effect of main beam synthesis beams for a radio interference array, so as to solve the problems raised in the above background technology.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A method for eliminating the coupling effect of main beam synthesis beam for a radio interferometer array, the method implementation steps specifically comprising:

[0008] S1: Construct the residual processing module RpDH-Deeplab to process the input image with effect, which can provide the most comprehensive and effective information of the input image to the subsequent network F out , which helps to achieve the correction of main beam effect more effectively;

[0009] S2: Construct the encoder structure in the segmentation network. In this structure, the backbone network MobileNet v2 is used to extract different levels of features of the input image. For the highest level features, the DH-ASPP module is used for further feature extraction, and finally the DH-ASPP module output part DH-ASPP is obtained. out ;

[0010] S3: Construct the decoder structure in the segmentation network. The decoder part is used to combine the low-level features output from the encoder with the high-level features DH-ASPP out Perform feature fusion. The fusion of multiple features ensures that more context information is input into the neural network to ensure better correction of the coupling effect, and finally obtain the output of the decoder. out ;

[0011] S4: The result F obtained by the residual processing module in S1 out Decoder result obtained with S3 outCombined with the activation function Tanh, the output of the RpDH-Deeplab network is obtained. Through continuous training, the parameters in the network structure are updated, so that the final output of RpDH-Deeplab can accurately correct the main beam synthesis beam effect.

[0012] Preferably, the residual processing module in step S1 is composed of two 3×3 convolutional layers, two feature extraction attention modules Extension feature Attention Block, i.e., EAB, and residual connection; the EAB module is composed of two parts: feature extraction and attention module;

[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 retains the information of the original input image;

[0014] The feature attention module consists of spatial attention, channel attention and global average pooling modules. For the output feature of the feature extraction module, that is, the input feature F of the attention part in In the attention module, spatial attention mechanism, channel attention mechanism and hybrid attention mechanism are used to obtain F s 、F c and F cs , then for F cs The global average pooling operation is used to obtain F p ,F p Respectively with F c and F s Perform multiplication to get F cp With F sp , F sp With F in Multiply and then add F cp Multiply them together to get 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 part described in step S1 is improved on the basis of Deeplab v3+, so that it can better extract the celestial body structure in the input main beam effect image; the segmentation network belongs to the encoding and decoding structure, which is conducive to better extraction of features;

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

[0019] MobileNet v2 uses an inverted residual structure (InvertedResiduals). Compared with the general residual structure, the inverted residual structure first performs a 1×1 convolution operation to increase the dimension, then performs a 3×3 depth-separable convolution, and finally performs a 1×1 convolution to reduce the dimension; MobileNet v2 is composed of multiple inverted residual structures. The expansion factor is the number of repeated structures in the inverted residual structure, the channel refers to the change in the number of feature channels before and after each operation, and the step size is only for the first layer of each repeated structure;

[0020] MobileNet v2 has a total of 4 convolution operations with a stride of 2, so the feature size changes 4 times, namely 184×184, 92×92, 46×46 and 23×23, corresponding to F2, F4, F8 and F 16 There are 4 features of different scales, among which F 16 is the highest level feature; in RpDH-Deeplab, F 16 It will be input into the DH-ASPP module for further feature extraction, while F2, F4 and F8 will be input into the decoder for feature fusion;

[0021] SPP can extract the context information of the input image very well. Deeplab v3+ adds the dilated convolution operation based on SPP. While increasing the receptive field in the network and reducing the amount of calculation, the dilated convolution does not sacrifice the spatial resolution and input information. At the same time, it captures more context information and can show good effects on large-scale structures in detection and segmentation tasks. However, since the dilated convolution is calculated in a chessboard-like form, the results obtained in a certain layer of convolution operation all come from an independent set of the previous layer. There is no relationship between the results, resulting in no correlation between the results of the convolution operation, so local information loss will occur. Considering that radio astronomy observations contain many faint small-scale celestial structures, the dilated convolution operation will cause the features of small-scale celestial structures to be unable to be effectively extracted;

[0022] Therefore, a hybrid dilated convolution is added to the RpDH-Deeplab network, and different dilation rates are used to extract features of different scales, so that there are more connections between the convolution results. At the same time, multiple different sizes are covered to achieve the purpose of accurate feature extraction of small-scale astronomical structures. Considering the complexity of the structure after adding the hybrid dilated convolution, a dense connection operation is added. On the basis of the ASPP module, and combined with the characteristics of the radio astronomical structure, dense connection and hybrid dilated convolution (HDC) are added, and finally the network structure of DH-ASPP in the RpDH-Deeplab is constructed.

[0023] Preferably, the DH-ASPP structure includes 4 mixed dilated convolutions, namely Part-1, Part-2, Part-3 and Part-4, each of which is composed of 3 dilated convolution modules, each of which contains a convolution layer, a batch normalization layer and an activation function, and different dilated rates are configured to achieve the purpose of interconnecting the dilated convolution results;

[0024] Part-5 is a pooling operation, convolution operation and upsampling operation. A small encoder-decoder module, Encoder-Decoder, ED, is constructed to achieve the final extraction of features, as shown in the formula:

[0025]

[0026] Among them, Cat represents the splicing operation, Represents a mixed dilated convolution combination, where i1, i2, and i3 are three different convolution rates of each mixed dilated convolution part.

[0027] Preferably, the decoder part of the RpDH-Deeplab model in step S3 is used to output the captured context information in the form of the same resolution as the input image, thereby achieving fine pixel-level segmentation;

[0028] The function of the decoder is to fuse the low-level features output from the encoder with the high-level features. The fusion process is shown in the formula:

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

[0030] Among them, Decorder out is the final output of the decoder, U NIndicates N-fold upsampling operation, Cat is concatenation operation, Encoder out is the output of the encoder.

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

[0032] 1. A 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 synthetic beam, the method in this paper achieves unified correction and avoids the scenario where the two effects affect each other during the correction process.

[0034] 2. Good transferability of the method. After the deep learning neural network model is trained, it can be used once and for all. The trained neural network model can accurately model the data and then extract the main beam effect characteristics. Since the neural network model has excellent generalization ability, the trained neural network model is also applicable even for radio astronomical structures that are not in the data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A flow chart of a method for eliminating the coupling effect of main beam synthesis beams used in a radio interferometer array;

[0036] Figure 2 It 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 is the structural block diagram of the encoder;

[0039] Figure 5 It is a schematic diagram of the DH-ASPP network structure;

[0040] Figure 6 A schematic diagram of the decoder structure. DETAILED DESCRIPTION

[0041] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments:

[0042] The specific implementation process is as follows:

[0043] like Figure 1 As shown, a method for eliminating the coupling effect of main beam synthesis beam for a radio interferometer array, the method implementation steps specifically include:

[0044] S1: Construct the residual processing module RpDH-Deeplab to process the input image with effect, which can provide the most comprehensive and effective information of the input image to the subsequent network F out , which helps to achieve the correction of main beam effect more effectively;

[0045] S2: Construct the encoder structure in the segmentation network. In this structure, the backbone network MobileNet v2 is used to extract different levels of features of the input image. For the highest level features, the DH-ASPP module is used for further feature extraction. The dense connection operation and the mixed hole convolution reduce the calculation parameters of the model, thereby greatly reducing the amount of calculation. In addition, it also prevents the occurrence of overfitting, making the overall model more generalizable. Finally, the output part of the DH-ASPP module DH-ASPP out ;

[0046] S3: Construct the decoder structure in the segmentation network. The decoder part is used to combine the low-level features output from the encoder with the high-level features DH-ASPP out Perform feature fusion. The fusion of multiple features ensures that more context information is input into the neural network to ensure better correction of the coupling effect, and finally obtain the output of the decoder. out ;

[0047] S4: The result F obtained by the residual processing module in S1 out Decoder result obtained with S3 out Combined with the activation function Tanh, the output of the RpDH-Deeplab network is obtained. Through continuous training, the parameters in the network structure are updated, so that the final output of RpDH-Deeplab can accurately correct the main beam synthesis beam effect.

[0048] Considering that the deep learning model needs to extract contours and structural features, a neural network model of image segmentation type that is more sensitive to contours should be selected; and the input of celestial body brightness information in images with coupling effects can be realized using the residual connection in ResNet, so it can be determined that the main architecture of the deep learning model is an image segmentation network with residual connections.

[0049] Based on the influence of coupling effect on the image formation, this paper deeply studies the relevant theories of deep learning and proposes the RpDH-Deeplab model to correct the coupling effect. The overall framework of the RpDH-Deeplab model is the residual processing module and the 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, so that RpDH-Deeplab can achieve more efficient and accurate coupling effect correction. Figure 3 shown.

[0051] The residual processing module described in step S1 is composed of two 3×3 convolutional layers, two feature extraction attention modules, namely, EAB, 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 retains the information of the original input image;

[0053] The feature attention module consists of spatial attention, channel attention and global average pooling modules. For the output feature of the feature extraction module, that is, the input feature F of the attention part in In the attention module, spatial attention mechanism, channel attention mechanism and hybrid attention mechanism are used to obtain F s 、F c and F cs , then for F cs The global average pooling operation is used to obtain F p ,F p Respectively with F c and F s Perform multiplication to get F cp With F sp , F sp With F in Multiply and then add F cp Multiply them together to get 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 realizes the repeated extraction of important information about the space and channels of the input image. Therefore, the entire residual processing module can provide the most comprehensive and effective information about the input image to the subsequent network, which helps to more effectively correct the main beam effect.

[0057] The segmentation network part described in step S1 is improved on the basis of Deeplab v3+, so that it can better extract the celestial body structure in the input main beam effect image; the segmentation network belongs to the encoding and decoding structure, which is conducive to better feature extraction. The encoder part is as follows Figure 4 shown.

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

[0059] MobileNet v2 uses an inverted residual structure (InvertedResiduals). Compared with the general residual structure, the inverted residual structure first performs a 1×1 convolution operation to increase the dimension, then performs a 3×3 depth-separable convolution, and finally performs a 1×1 convolution to reduce the dimension; MobileNet v2 is composed of multiple inverted residual structures. The expansion factor is the number of repeated structures in the inverted residual structure, the channel refers to the change in the number of feature channels before and after each operation, and the step size is only for the first layer of each repeated structure;

[0060] MobileNet v2 has a total of 4 convolution operations with a stride of 2, so the feature size changes 4 times, namely 184×184, 92×92, 46×46 and 23×23, corresponding to F2, F4, F8 and F 16 There are 4 features of different scales, among which F 16 is the highest level feature; in RpDH-Deeplab, F 16 It will be input into the DH-ASPP module for further feature extraction, while F2, F4 and F8 will be input into the decoder for feature fusion;

[0061] SPP can extract the context information of the input image very well. Deeplab v3+ adds the dilated convolution operation based on SPP. While increasing the receptive field in the network and reducing the amount of calculation, the dilated convolution does not sacrifice the spatial resolution and input information. At the same time, it captures more context information and can show good effects on large-scale structures in detection and segmentation tasks. However, since the dilated convolution is calculated in a chessboard-like form, the results obtained in a certain layer of convolution operation all come from an independent set of the previous layer. There is no relationship between the results, resulting in no correlation between the results of the convolution operation, so local information loss will occur. Considering that radio astronomy observations contain many faint small-scale celestial structures, the dilated convolution operation will cause the features of small-scale celestial structures to be unable to be effectively extracted;

[0062] Therefore, a hybrid dilated convolution is added to the RpDH-Deeplab network, and different dilation rates are used to extract features of different scales, so that there are more connections between the convolution results, while covering multiple different sizes, to achieve the purpose of accurate feature extraction of small-scale astronomical structures; considering the complexity of the structure after adding the hybrid dilated convolution, a dense connection operation is added; thanks to the advantages of the residual connection, the module after adding the dense connection reduces the calculation parameters of the model, thereby greatly reducing the amount of calculation. In addition, it also prevents the occurrence of overfitting, making the overall model more generalizable; on the basis of the ASPP module, and combined with the characteristics of the radio astronomical structure, dense connection and hybrid dilated convolution (HDC) are added, and finally the network structure of DH-ASPP in the RpDH-Deeplab is constructed, as shown in Figure 4 shown.

[0063] The DH-ASPP structure includes 4 mixed dilated convolutions, Part-1, Part-2, Part-3 to Part-4. Each mixed dilated convolution consists of 3 dilated convolution modules. Each convolution module contains a convolution layer, a batch normalization layer, and an activation function. Different dilated rates are configured to achieve the purpose of interconnecting the dilated convolution results.

[0064] Part-5 is a pooling operation, convolution operation and upsampling operation. A small encoder-decoder module, Encoder-Decoder, ED, is constructed to achieve the final extraction of features, as shown in the formula:

[0065]

[0066] Among them, Cat represents the splicing operation, Represents a mixed dilated convolution combination, where i1, i2, and i3 are three different convolution rates of each mixed dilated convolution part.

[0067] The decoder part of the RpDH-Deeplab model in step S3 is used to output the captured context information in the form of the same resolution as the input image, thereby achieving fine pixel-level segmentation. The decoder network block diagram is shown in Figure 6 As shown;

[0068] The function of the decoder is to fuse the low-level features output from the encoder with the high-level features. The fusion process is shown in the formula:

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

[0070] Among them, Decorder out is the final output of the decoder, U N Indicates N-fold upsampling operation, Cat is concatenation operation, Encoder out is the output of the encoder. The fusion of multiple features ensures that more context information is input into the neural network to ensure better correction of the coupling effect.

[0071] The above is only an embodiment of the present invention, and the common knowledge such as the known specific technical solutions and / or characteristics in the solution is not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several modifications and improvements can be made, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for eliminating the coupling effect of main beam synthesis beam for radio interferometer array, characterized in that: The steps for implementing the method specifically include: S1: Construct the residual processing module RpDH-Deeplab to process the input image with effect, which can provide the most comprehensive and effective information of the input image to the subsequent network F out , which helps to achieve the correction of main beam effect more effectively; S2: Construct the encoder structure in the segmentation network. In this structure, the backbone network MobileNet v2 is used to extract different levels of features of the input image. For the highest level features, the DH-ASPP module is used for further feature extraction, and finally the DH-ASPP module output part DH-ASPP is obtained. out ; S3: Construct the decoder structure in the segmentation network. The decoder part is used to combine the low-level features output from the encoder with the high-level features DH-ASPP out Perform feature fusion. The fusion of multiple features ensures that more context information is input into the neural network to ensure better correction of the coupling effect, and finally obtain the decoder output Decoder out ; S4: The result F obtained by the residual processing module in S1 out Decoder result obtained with S3 out Combined with the activation function Tanh, the output result of the RpDH-Deeplab network is obtained; through continuous training, the parameters in the network structure are updated, so that the final output of RpDH-Deeplab can achieve accurate correction of the main beam synthesis beam effect.

2. The method for eliminating the coupling effect of main beam synthesis beam for radio interferometer array according to claim 1, characterized in that: The residual processing module described in step S1 is composed of two 3×3 convolutional layers, two feature extraction attention modules, namely, EAB, and residual connections; the EAB module is composed of two parts: feature extraction and attention modules; 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 retains the information of the original input image; The feature attention module consists of spatial attention, channel attention and global average pooling modules. For the output feature of the feature extraction module, that is, the input feature F of the attention part in In the attention module, spatial attention mechanism, channel attention mechanism and hybrid attention mechanism are used to obtain F s 、F c and F cs , then for F cs The global average pooling operation is used to obtain F p ,F p Respectively with F c and F s Perform multiplication to get F cp With F sp , F sp With F in Multiply and then add F cp Multiply them together to get 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 the coupling effect of main beam synthesis beam for radio interferometer array according to claim 1, characterized in that: The segmentation network described in step S1 is improved on the basis of Deeplab v3+, so that it can better extract the celestial body structure in the input main beam effect image; the segmentation network belongs to the encoding and decoding structure, which is conducive to better feature extraction; The RpDH-Deeplab network uses MobileNetv2 as the backbone network of the segmentation network; MobileNet v2 uses an inverted residual structure. Compared with the general residual structure, the inverted residual structure first performs a 1×1 convolution operation to increase the dimension, then performs a 3×3 depth-separable convolution, and finally performs a 1×1 convolution to reduce the dimension. MobileNetv2 is composed of multiple inverted residual structures superimposed. The expansion factor is the number of repeated structures in the inverted residual structure, the channel refers to the change in the number of feature channels before and after each operation, and the step size is only for the first layer of each repeated structure; MobileNetv2 has a total of 4 convolution operations with a stride of 2, so the feature size changes 4 times, namely 184×184, 92×92, 46×46 and 23×23, corresponding to F2, F4, F8 and F 16 There are 4 features of different scales, among which F 16 is the highest level feature; in RpDH-Deeplab, F 16 It will be input into the DH-ASPP module for further feature extraction, while F2, F4 and F8 will be input into the decoder for feature fusion; SPP can extract the context information of the input image very well. Deeplab v3+ adds the dilated convolution operation based on SPP. While increasing the receptive field in the network and reducing the amount of calculation, the dilated convolution does not sacrifice the spatial resolution and input information. At the same time, it captures more context information and can show good effects on large-scale structures in detection and segmentation tasks. However, since the dilated convolution is calculated in a chessboard-like form, the results obtained in a certain layer of convolution operations all come from an independent set of the previous layer. There is no relationship between the results, resulting in no correlation between the results of the convolution operation, so local information loss will occur. Considering that radio astronomy observations contain many faint small-scale celestial structures, the dilated convolution operation will cause the features of small-scale celestial structures to be unable to be effectively extracted; Therefore, a hybrid dilated convolution is added to the RpDH-Deeplab network, and different dilation rates are used to extract features of different scales, so that there are more connections between the convolution results. At the same time, multiple different sizes are covered to achieve the purpose of accurate feature extraction of small-scale astronomical structures. Considering the complexity of the structure after adding the hybrid dilated convolution, a dense connection operation is added. On the basis of the ASPP module, and combined with the characteristics of the radio astronomical structure, dense connection and hybrid dilated convolution (HDC) are added, and finally the network structure of DH-ASPP in the RpDH-Deeplab is constructed.

4. The method for eliminating the main beam synthesis beam coupling effect for a radio interferometer array according to claim 3 and claim 1, characterized in that: The DH-ASPP structure includes four mixed dilated convolutions, Part-1, Part-2, Part-3 to Part-4. Each mixed dilated convolution consists of three dilated convolution modules. Each convolution module contains a convolution layer, a batch normalization layer and an activation function. Different dilated rates are configured to achieve the purpose of interconnecting the dilated convolution results. Part-5 is a pooling operation, convolution operation and upsampling operation. A small encoder-decoder module, Encoder-Decoder, ED, is constructed to achieve the final extraction of features, as shown in the formula: Among them, Cat represents the splicing operation, Represents a mixed dilated convolution combination, where i1, i2, and i3 are three different convolution rates of each mixed dilated convolution part.

5. The method for eliminating the main beam synthesis beam coupling effect for a radio interferometer array according to claim 1, characterized in that: The decoder part of the RpDH-Deeplab model in step S3 is used to output the captured context information in the form of the same resolution as the input image, thereby achieving fine pixel-level segmentation; The function of the decoder part is to fuse the low-level features output from the encoder with the high-level features. The fusion process is shown in the formula: Decorder out =U N (Cat(F2,U2(Cat(F4,U4(Cat(F8,U8(Encoder out ))))))) Among them, Decorder out is the final output of the decoder, U N Indicates N-fold upsampling operation, Cat is concatenation operation, Encoder out is the output of the encoder.

Images