Spectral map construction method based on spatially enhanced channel attention

Abstract

The present invention discloses a method for constructing a spectrum map based on spatially enhanced channel attention. The method comprises the following steps: inputting input data of three channels including a building map, a radiation source map and a power spectrum density measurement map into a spectrum mapping model based on spatially enhanced channel attention to obtain a predicted spectrum map. The spectrum mapping model comprises a gated embedding layer, an encoder and a decoder; the decoder comprises a plurality of sequentially connected decoding processing layers and a convolutional layer, and the decoding processing layer comprises sequentially connected transposed convolutional layers and spatially enhanced channel attention layers. The present invention enhances the extraction of detail information of shadow areas with weak PSD by jointly focusing on channel features and spatial information, while also reducing the impact of defective areas, effectively learning the shadow effect caused by building occlusion, and still having good construction performance at a lower sampling rate.

Description

Technical Field

[0001] The present invention belongs to the field of electromagnetic spectrum map completion, and in particular relates to a spectrum map construction method. Background Art

[0002] The electromagnetic spectrum map (Spectrum Map) can provide intuitive and detailed electromagnetic energy distribution and spectrum resource occupancy status from multiple dimensions such as space, frequency, and energy, and present these information features in a visual form on the map. It is also called the Radio Environment Map (REM) or the Electromagnetic Environment Map (EEM). Electromagnetic spectrum data is usually scattered and isolated, making it difficult to integrate into an effective large data set. In addition, large-scale data collection is extremely expensive, and it is difficult and impractical to directly obtain electromagnetic spectrum data at all spatial locations. Therefore, how to use limited electromagnetic environment collection data to construct a complete and accurate radio environment map has become the focus of more and more researchers.

[0003] At present, the construction methods of spectrum maps are mainly divided into two categories: model-driven and data-driven. When constructing spectrum maps, the model-driven method not only requires monitoring data, but also relies on prior knowledge such as transmitter information and radio wave propagation parameters. Although this method has high accuracy, it has poor adaptability and it is difficult to obtain propagation data such as radiation sources in actual scenarios. Therefore, the data-driven method has attracted more attention from scholars. The data-driven method only needs to construct the radio environment map based on the collected data, which can be further subdivided into spatial interpolation construction method and spectrum map construction method based on deep learning. The spline function method estimates the value of unknown points by constructing a "spline" defined by a polynomial, while ensuring the global smoothness of the spline and coinciding with the monitoring data as much as possible. The Kriging interpolation algorithm proposed by T. Imai et al. improves the construction accuracy under complex terrain by interpolating in local areas of different granularities. Deep learning has developed rapidly in the field of artificial intelligence due to its excellent learning and representation capabilities, providing a powerful research tool for spectrum map construction. Researchers at home and abroad have begun to try to use deep learning methods to reconstruct spectrum maps. For example, the neural network architecture designed by H. Zhang et al. achieved the reasoning of the complete radio map by learning the radio shadow fading characteristics in the data set, improving the construction accuracy and generalization. Y. Tu et al. took advantage of the advantages of convolutional neural networks in extracting image features and realized the reconstruction of single-radiator spectrum maps based on CNN.

[0004] Although existing solutions have achieved remarkable results, they still have limitations: on the one hand, most studies are aimed at open or simple scenes. For complex urban scenes, existing methods have difficulty simulating factors such as shadow effects and propagation barriers caused by buildings. On the other hand, the proportion of sampled data has an important impact on the estimation performance of spectrum maps. The sampling rate in existing technologies is generally high, and the performance is poor when the acquisition cost and resources are limited. Summary of the invention

[0005] The present invention proposes a spectrum map construction method based on spatial enhanced channel attention, the purposes of which are: 1. To solve the problem that spectrum maps are difficult to simulate shadow effects and propagation barriers of buildings; 2. To solve the problem that the existing model construction performance is poor when the sampling rate is low.

[0006] The technical solution of the present invention is as follows:

[0007] A spectrum map construction method based on spatially enhanced channel attention, which includes building maps, radiation source maps and power spectral density measurements Figure 3 The input data of each channel is input into the spectrum mapping model based on spatial enhanced channel attention to obtain the predicted spectrum map;

[0008] The building map, radiation source map and power spectrum density measurement map correspond to the same target rectangular space area; in the power spectrum density measurement map, the target rectangular space area is discretized into a grid and only includes power spectrum density measurement values ​​at a number of positions;

[0009] The spectrum mapping model includes a gated embedding layer, an encoder and a decoder;

[0010] The gated embedding layer is used for feature extraction;

[0011] The input of the encoder is the output feature of the gated embedding layer, which includes several layers of convolutional layers connected in sequence, which are used to learn effective information and compress it to obtain low-dimensional feature encoding;

[0012] The decoder comprises a plurality of decoding processing layers connected in sequence and a convolutional layer, which is used to reconstruct a complete spectrum map according to the information contained in the low-dimensional feature code learned by the encoder;

[0013] The decoding processing layer includes a transposed convolution layer and a spatial enhancement channel attention layer connected in sequence;

[0014] The number of layers of the decoder is consistent with that of the encoder, and jump connections are made between corresponding layers;

[0015] The last convolutional layer in the decoder is used to map the feature map output by the last decoding processing layer into a complete spectrum map.

[0016] As a further improvement of the spectrum map construction method based on spatially enhanced channel attention: the gated embedding layer includes a first path and a second path in parallel, the first path includes a convolutional layer, and the second path sequentially arranges a convolutional layer and a Gaussian error linear unit, and the features obtained by the input data passing through the first path and the second path respectively are merged through element-wise multiplication to obtain the output features of the gated embedding layer.

[0017] As a further improvement of the spectrum map construction method based on spatial enhanced channel attention, the jump connection means that the number of layers of the decoder and the number of layers of the encoder are , the encoder The output of the layer is the input of the first decoding processing layer in the decoder. The output of the layer is the same as the output of the decoder The output of the layer decoding processing layer is concatenated and used as the first The layer decoding process layer input, here .

[0018] As a further improvement of the spectrum map construction method based on spatial enhanced channel attention: the convolution layer and the transposed convolution layer in the decoder both use the Leaky ReLU function as the activation function.

[0019] As a further improvement of the spectrum map construction method based on spatially enhanced channel attention: the spatially enhanced channel attention layer includes a channel attention part and a spatial attention part, the input features pass through the channel attention part and the spatial attention part respectively, and then the features output by the channel attention part and the spatial attention part are implicitly fused through convolution.

[0020] As a further improvement of the spectral map construction method based on spatial enhanced channel attention: for a given input feature , the output feature of the channel attention part is expressed as:

[0021] ;

[0022] in, is the probability distribution conversion function, is a learnable parameter for scaling, , and For input features The query, key, and value obtained after the convolution layer and the depth-separable convolution layer are expressed as:

[0023] , , ;

[0024] , and is the learnable weight matrix.

[0025] As a further improvement of the spectral map construction method based on spatially enhanced channel attention, for a given input feature , the output feature of the spatial attention part is expressed as:

[0026] ;

[0027] ;

[0028] in, , and is the convolution operation, and are average pooling and maximum pooling operations respectively, For splicing operation, For normalization operation, is a S-type function.

[0029] As a further improvement of the spectrum map construction method based on spatially enhanced channel attention, the output characteristics of the spatially enhanced channel attention layer are:

[0030] ;

[0031] The convolution kernel is a 1×1 convolution operation. and They are the output features of the channel attention part and the spatial attention part respectively.

[0032] As a further improvement of the spectrum map construction method based on spatially enhanced channel attention, the training process of the spectrum mapping model is: input samples into the spectrum mapping model to obtain an output spectrum map, then calculate the loss function based on the output spectrum map and the corresponding true spectrum map, and then update the model parameters using reverse gradient operation based on the loss function until the loss function converges to stability.

[0033] As a further improvement of the spectrum map construction method based on spatial enhanced channel attention, the loss function used is a combined loss function:

[0034] ;

[0035] in, is the pixel-level loss function, is the adversarial loss function, is the structural loss function, , , is the weight value;

[0036] (1) Pixel-level loss function The calculation process is:

[0037] ;

[0038] in, and They are the real spectrum map and the spectrum map output by the model. is the width and height of the map in pixels, is the Frobenius norm;

[0039] (2) Adversarial Loss Function The calculation method is:

[0040] ;

[0041] in, and They are the real spectrum map and the spectrum map output by the model. is the discriminator of the adversarial generative network, is the generator of the adversarial generative network, It means to find the expected value;

[0042] (3) Structural loss function The calculation method is:

[0043] ;

[0044] in, and They are the real spectrum map and the spectrum map output by the model. is the multi-scale structural similarity index;

[0045] The calculation process is:

[0046] Map the real spectrum And the spectrum map of the model output Each one is carried out in turn Downsampling times, a total of For images of different scales; for each pair of images, calculate the contrast comparison function and the structure comparison function respectively;

[0047] The contrast comparison function is calculated as:

[0048] ;

[0049] in, and It is For images, and They are and The respective standard deviations, is a constant term;

[0050] The calculation method of the structure comparison function is:

[0051] ;

[0052] in, and It is For images, yes and The covariance of and They are and The respective standard deviations, is a constant term;

[0053] Then, according to Compute a brightness comparison function for an image:

[0054] ;

[0055] in, and It is For images, and They are and The respective average values, and They are and The respective standard deviations, is a constant term;

[0056] The multi-scale structural similarity index is calculated as:

[0057] ;

[0058] in, is a parameter that controls the relative importance of brightness similarity, For the The relative importance parameter that controls the contrast similarity at the corresponding scale of the image, For the A parameter that controls the relative importance of structural similarity at corresponding scales of images.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] 1. The present invention adopts a U-shaped autoencoder network and a spatial enhancement channel attention layer in the decoder part, which jointly focuses on channel features and spatial information, enhances the extraction of detail information in shadow areas with weak PSD, and reduces the impact of defective areas, effectively learning the shadow effect caused by building occlusion. Experimental results show that this method has better construction effect than other algorithms that have been used in open or simple scenes, and solves the problem of low performance of spectrum map construction in complex urban scenes.

[0061] 2. The present invention adopts a combined loss function, which considers three aspects: pixel-level loss, adversarial loss and structural loss respectively. It improves the construction performance from multiple angles and scales, and can enhance the authenticity of the generated spectrum map while ensuring the rapid and stable convergence of the model.

[0062] 3. By conducting experiments at extremely low data sampling rates, using random sampling rates of 0.01 to 0.05 during training, and extending the sampling rate range from an extremely low 0.001 to 0.1 during testing, it can be proved that under sampling conditions far lower than the prior art, the present invention still has the lowest construction error compared to other models. It can be seen that this method can greatly reduce the acquisition cost, effectively solve the problem of poor performance of existing model construction under low sampling rates, and is suitable for resource-constrained acquisition scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a structural diagram of the spectrum mapping model in the present invention;

[0064] Figure 2 This is the structural diagram of the spatial enhanced channel attention layer. , and are the number of channels, height and width of the input features respectively;

[0065] Figure 3 is a comparison diagram of the root mean square error between the method of the present invention and other methods, in which ours refers to the method of the present invention;

[0066] Figure 4 is a comparison diagram of normalized mean square error between the method of the present invention and other methods, in which ours refers to the method of the present invention;

[0067] Figure 5 This is the real spectrum map in the comparison experiment;

[0068] Figure 6 This is the real spectrum map obtained by this method in the comparison experiment;

[0069] Figure 7 The spectrum map obtained by using RadioUNet (Wireless U-Network) in the comparison experiment;

[0070] Figure 8 The spectrum map obtained by using Deep AE (deep spectrum mapping method) in the comparison experiment;

[0071] Fig. 9 The spectrum map obtained using cDCGAN (conditional deep convolutional generative adversarial network) in the comparative experiment;

[0072] Fig.10 This is the spectrum map obtained by using Kriging algorithm in the comparison experiment;

[0073] Fig.11 This is the spectrum map obtained using KNN (k nearest neighbor method) in the comparative experiment. DETAILED DESCRIPTION

[0074] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0075] A spectrum map construction method based on spatially enhanced channel attention, which includes building maps, radiation source maps and power spectral density measurements Figure 3 The input data of each channel is input into the spectrum mapping model based on spatially enhanced channel attention to obtain the predicted spectrum map.

[0076] The building map, radiation source map and power spectrum density measurement map correspond to the same target rectangular space area. In the power spectrum density measurement map, the target rectangular space area is discretized as When measuring, randomly deploy A wireless receiving device, , the received The power spectrum density measurement graph can be obtained by marking the power spectrum density measurement values ​​at the locations in the grid.

[0077] The spectrum mapping model includes a gated embedding layer, an encoder and a decoder.

[0078] 1. Gated embedding layer.

[0079] The gated embedding layer is used for feature extraction.

[0080] Specifically, the gated embedding layer includes a first path and a second path in parallel, the first path includes a convolutional layer, and the second path sequentially sets a convolutional layer and a Gaussian error linear unit (GELU), and the features obtained by the input data passing through the first path and the second path are combined by element-level multiplication to obtain the output features of the gated embedding layer. This parallel path structure can extract more fine-grained features and improve the expressiveness of features.

[0081] 2. Encoder.

[0082] The input of the encoder is the output feature of the gated embedding layer, which includes several layers of convolutional layers connected in sequence, which are used to learn effective information and compress it to obtain low-dimensional feature coding.

[0083] In this embodiment, the encoder has a total of 5 convolution layers, each of which has a convolution kernel of 5×5.

[0084] 3. Decoder.

[0085] The decoder comprises a plurality of decoding processing layers connected in sequence and a convolutional layer. The decoder is used to reconstruct a complete spectrum map according to the information contained in the low-dimensional feature code learned by the encoder.

[0086] The decoding processing layer includes a transposed convolution layer and a spatially enhanced channel attention layer (SECAL) connected in sequence, which improves the construction accuracy during feature upsampling.

[0087] The number of layers of the decoder is the same as that of the encoder, and jump connections are made between corresponding layers: Let the number of layers be , the encoder The output of the layer is the input of the first decoding processing layer in the decoder. The output of the layer is the same as the output of the decoder The output of the layer decoding processing layer is concatenated and used as the first The layer decoding process layer input, here This structure can retain the features learned by the encoder and speed up the convergence.

[0088] The last convolutional layer in the decoder is used to map the feature map output by the last decoding processing layer into a complete spectrum map.

[0089] Furthermore, the convolutional layer and the transposed convolutional layer in the decoder both use the Leaky ReLU function as the activation function to prevent the gradient vanishing problem and enhance the expressiveness of the model.

[0090] Furthermore, the spatially enhanced channel attention layer includes a channel attention part and a spatial attention part. The input features pass through the channel attention part and the spatial attention part respectively, and then the features output by the channel attention part and the spatial attention part are implicitly fused through convolution.

[0091] Channel self-attention focuses on the channel dimension of the feature map, and adjusts the feature response between channels by assigning different importance to the features of different channels, making the model more selective in deciding which features are more important. This helps the model focus on the important features of the input incomplete spectrum map, such as radiation source features, building obstruction effects, etc. However, the channel self-attention mechanism does not consider the specific location of the features in the entire space, and ignores the relationship between feature blocks. At the same time, the defective areas and effective sampling areas of the defective radio map (power spectral density measurement map) and the building map also contain key information, which is important for reconstructing an accurate spectrum map. Figure 10 Therefore, the present invention combines the spatial attention mechanism with the channel attention mechanism, further pays attention to the different positions of the input feature map, extracts global context information, and learns the attention score matrix of the spatial position, so that the model pays attention to the location information of the radiation source and the detail information of the shadow area with weaker PSD intensity. After introducing the spatial attention mechanism, the model can better capture the detailed changes of the edges of the buildings and learn the shadow effects caused by the occlusion of complex buildings, thereby improving the accuracy of reconstructing the spectrum map. Moreover, unlike the traditional method of directly combining the output features of the two parts, this method is implicitly fused through convolution, thereby achieving double enhancement of the input features.

[0092] In the channel attention part, deep convolution is first used to extract features, and then matrix multiplication and softmax operations are performed to calculate the attention weights between channels. The channel features are redistributed through the weights to highlight the responses of key channels.

[0093] For a given input feature , the output feature of the channel attention part is expressed as:

[0094] ;

[0095] in, is the probability distribution conversion function, is a learnable parameter for scaling, , and For input features The query, key, and value obtained through the convolution layer and the depth-separable convolution layer are expressed as follows:

[0096] , , ;

[0097] , and is the learnable weight matrix.

[0098] In the spatial attention part, convolution and normalization operations are performed after the average pooling and maximum pooling operations to obtain the soft attention feature matrix. Compared with the feature map extracted by the traditional single-channel spatial attention, it can capture more fine-grained local spatial information, thereby performing spatial feature enhancement on important areas in the image.

[0099] For a given input feature , the output feature of the spatial attention part is expressed as:

[0100] ;

[0101] ;

[0102] in, , and is the convolution operation, and are average pooling and maximum pooling operations respectively, For splicing operation, For normalization operation, is a S-type function.

[0103] In this embodiment, and The convolution kernel is a 3×3 convolution operation. The convolution kernel is a 7×7 convolution operation.

[0104] The output characteristics of the spatial enhancement channel attention layer are:

[0105] ;

[0106] The convolution kernel is a 1×1 convolution operation.

[0107] The spatially enhanced channel attention layer introduces a dual-branch design in the spatial attention part. This multi-scale receptive field fusion enables the model to obtain spatial information at both fine-grained and coarse-grained levels, rather than being limited to spatial features at a single scale. Compared with traditional methods, the spatially enhanced channel attention layer of the present invention ensures a wider receptive field in this way, and can capture more global and local key information, thereby improving sensitivity to spatial information. In general, the channel attention part focuses on the importance between different channels, and can capture features such as radiation source information and building distribution. Spatial attention uses average pooling and maximum pooling to retain overall information, further strengthen the distribution of features in the spatial dimension, and also extracts the most important features, effectively capturing the detail information caused by the building, and ultimately making the model more accurate when extracting global and local features.

[0108] The training process of the spectrum mapping model is as follows: input samples into the spectrum mapping model to obtain an output spectrum map, then calculate the loss function based on the output spectrum map and the corresponding real spectrum map, and then update the model parameters using reverse gradient operation based on the loss function until the loss function converges to stability, so that the reconstructed spectrum map is as close to the real spectrum map as possible.

[0109] The loss function used in the present invention is a combined loss function:

[0110] ;

[0111] in, is the pixel-level loss function, is the adversarial loss function, is the structural loss function, , , is the weight value.

[0112] (1) Pixel-level loss function The calculation process is:

[0113] ;

[0114] in, and They are the real spectrum map and the spectrum map output by the model. is the width and height of the map in pixels, is the Frobenius norm.

[0115] (2) Adversarial Loss Function The calculation method is:

[0116] ;

[0117] in, and They are the real spectrum map and the spectrum map output by the model. is the discriminator of the adversarial generative network, is the generator of the adversarial generative network, Indicates the expected value.

[0118] (3) Structural loss function The calculation method is:

[0119] ;

[0120] in, and They are the real spectrum map and the spectrum map output by the model. is a multi-scale structural similarity index.

[0121] The calculation process is:

[0122] Map the real spectrum And the spectrum map of the model output Each one is carried out in turn Downsampling times, a total of For images of different scales, for each pair of images, the contrast comparison function and the structure comparison function are calculated respectively.

[0123] The contrast comparison function is calculated as:

[0124] ;

[0125] in, and It is For images, and They are and The respective standard deviations, is a constant term.

[0126] The calculation method of the structure comparison function is:

[0127] ;

[0128] in, and It is For images, yes and The covariance of and They are and The respective standard deviations, is a constant term.

[0129] Then, according to Compute a brightness comparison function for an image:

[0130] ;

[0131] in, and It is For images, and They are and The respective average values, and They are and The respective standard deviations, is a constant term.

[0132] The multi-scale structural similarity index is calculated as:

[0133] ;

[0134] in, is a parameter that controls the relative importance of brightness similarity, For the The relative importance parameter that controls the contrast similarity at the corresponding scale of the image, For the A parameter that controls the relative importance of structural similarity at corresponding scales of images.

[0135] In this embodiment, the training uses the public data set RadioMapSeer, which contains 700 city maps of large cities such as London, Berkin, Ljublijana, Ankara, etc. Each map corresponds to the location of 80 radiation sources, with a total of 56,000 spectrum maps. The heights of the radiation source, receiver and building are set to 1.5m, 1.5m and 25m respectively. The power of the radiation source is 23dBm, the frequency is 5.9GHz, the map size is 256m×256m, and each pixel represents 1 square meter. The defective spectrum map is sampled from the complete spectrum map for training and testing. 40,000 maps of the data set are used for training models, 8,000 for verification, and 8,000 for testing. During training, 8 pictures are selected as a batch input network for training each time, and only a small batch of gradients is calculated for each iteration, so that the loss function converges quickly to stability. After each round of training, the network is verified using the verification set, and the results of this round are compared with the results of the previous round, and the network model with the best effect is saved.

[0136] Figure 3 and Figure 4 The model obtained by this method (ours in the figure) is compared with other models. The results show that the scheme adopted by the present invention has the best performance at all sampling rates compared with other models and algorithms, especially in the low sampling rate range. It is worth noting that although a random sampling rate of 0.01 to 0.05 was used in training, the sampling rate range was extended from an extremely low 0.001 to 0.1 during testing. At a sampling rate as low as 0.001, this method still maintains the lowest test RMSE (root mean square error) and NMSE (normalized mean square error), highlighting its outstanding learning ability. At a sampling rate of 0.01, the RMSE of this scheme is about 12.5% ​​lower than that of RadioUNet and about 28.8% lower than that of Deep AE. In addition, the construction accuracy increases with the increase of sampling rate because when the sampling data increases, the model can learn more effective propagation information to better learn the distribution characteristics.

[0137] Figures 5 to 11 The results of constructing a spectrum map in the test sample set using different models and algorithms at a sampling rate of 0.05 are shown. The results show that the proposed method achieves the highest visual effect and can highly simulate the texture details produced by the channel propagation effect in complex urban scenes. RadioUnet, DeepAE and cDCGAN contribute to capturing the overall propagation features, but there are different degrees of distortion in restoring the propagation details. The Kriging method has the worst effect and can hardly restore the propagation features.

[0138] In this embodiment, an ablation experiment is also performed to illustrate whether channel attention and spatial attention are helpful for the task. The test results at four sampling rates are shown in Table 1. Among them, UNet represents the model after removing SECAL from the method proposed in the present invention, UNet+C Attn is the model of the method proposed in the present invention ablated to only include channel attention, and Ours is the model of this method. The results show that this method can effectively capture channel features and identify texture information in important spatial areas, proving the effectiveness of SECAL.

[0139] Table 1 - Ablation experiment results at four sampling rates

[0140]

[0141] The algorithm program in this embodiment is implemented based on Python and is performed on 4 Nvidia 3090 GPUs. The libraries used include torch (deep learning library), numpy (scientific mathematical operation library) and matplotlib (for drawing).

[0142] It should be noted that it is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. The scope of the present invention is defined by the claims rather than the above description.

Claims

1. A spectrum map construction method based on spatially enhanced channel attention, characterized in that: Input data including three channels of building map, radiation source map and power spectral density measurement map are input into the spectrum mapping model based on spatial enhanced channel attention to obtain the predicted spectrum map; The building map, radiation source map and power spectrum density measurement map correspond to the same target rectangular space area; In the power spectrum density measurement graph, the target rectangular space region is discretized into a grid and only includes power spectrum density measurement values ​​at a number of positions; The spectrum mapping model includes a gated embedding layer, an encoder and a decoder; The gated embedding layer is used for feature extraction; The input of the encoder is the output feature of the gated embedding layer, which includes several layers of convolutional layers connected in sequence, which are used to learn effective information and compress it to obtain low-dimensional feature encoding; The decoder comprises a plurality of decoding processing layers connected in sequence and a convolutional layer, which is used to reconstruct a complete spectrum map according to the information contained in the low-dimensional feature code learned by the encoder; The decoding processing layer includes a transposed convolution layer and a spatial enhancement channel attention layer connected in sequence; The number of layers of the decoder is consistent with that of the encoder, and jump connections are made between corresponding layers; The last convolutional layer in the decoder is used to map the feature map output by the last decoding processing layer into a complete spectrum map; The spatial enhanced channel attention layer includes a channel attention part and a spatial attention part. The input features pass through the channel attention part and the spatial attention part respectively, and then the features output by the channel attention part and the spatial attention part are implicitly fused through convolution. For a given input feature X de , the output feature of the channel attention part is expressed as: Among them, Softmax(·) is the probability distribution conversion function, are learnable parameters for scaling, Q, K, and V are input features X de The query, key, and value obtained after the convolution layer and the depth-separable convolution layer are expressed as: Q=φ Q X de ,K=φ K X de ,V=φ V X de ; φ Q ,φ K and φ V is a learnable weight matrix; For a given input feature X de , the output feature of the spatial attention part is expressed as: X=Concat(avgp(Conv1(X de )),maxp(Conv2(X de ))); Among them, Conv1(·), Conv2(·) and Conv3(·) are convolution operations, avgp(·) and maxp(·) are average pooling and maximum pooling operations respectively, Concat(·) is a concatenation operation, Norm(·) is a normalization operation, and Sigmod(·) is a S-type function; The output characteristics of the spatial enhancement channel attention layer are: Conv4(·) is a 1×1 convolution operation with a convolution kernel. and They are the output features of the channel attention part and the spatial attention part respectively.

2. The method for constructing a spectrum map based on spatially enhanced channel attention according to claim 1, characterized in that: The gated embedding layer includes a first path and a second path in parallel, the first path includes a convolutional layer, and the second path is sequentially provided with a convolutional layer and a Gaussian error linear unit. The features obtained by respectively passing the input data through the first path and the second path are combined by element-wise multiplication to obtain the output features of the gated embedding layer.

3. The spectrum map construction method based on spatial enhanced channel attention according to claim 1, characterized in that: The jump connection means that the number of layers of the decoder and the number of layers of the encoder are L, the output of the Lth layer in the encoder is the input of the first decoding processing layer in the decoder, and the output of the lth layer of the encoder and the output of the Llth decoding processing layer in the decoder are spliced ​​as the input of the L-l+1th decoding processing layer in the decoder, where l=1,2,…,L-1.

4. The method for constructing a spectrum map based on spatially enhanced channel attention according to claim 1, characterized in that: The convolutional layer and the transposed convolutional layer in the decoder both use the Leaky ReLU function as the activation function.

5. The method for constructing a spectrum map based on spatially enhanced channel attention according to any one of claims 1 to 4, characterized in that: The training process of the spectrum mapping model is as follows: samples are input into the spectrum mapping model to obtain an output spectrum map, then a loss function is calculated based on the output spectrum map and the corresponding true spectrum map, and then the model parameters are updated using a reverse gradient operation based on the loss function until the loss function converges to stability.

6. The method for constructing a spectrum map based on spatially enhanced channel attention according to claim 5, characterized in that: The loss function used is a combined loss function: in, is the pixel-level loss function, is the adversarial loss function, is the structural loss function, γ1, γ2, γ3 are weight values; (1) Pixel-level loss function The calculation process is: Among them, Φ and are the real spectrum map and the spectrum map output by the model, respectively, N is the number of pixels in the width and height of the map, is the Frobenius norm; (2) Adversarial Loss Function The calculation method is: Among them, Φ and are the real spectrum map and the spectrum map output by the model, respectively. D(·) is the discriminator of the adversarial generative network, and G(·) is the generator of the adversarial generative network. It means to find the expected value; (3) Structural loss function The calculation method is: Among them, Φ and They are the real spectrum map and the spectrum map output by the model. is the multi-scale structural similarity index; The calculation process is: The real spectrum map Φ and the spectrum map output by the model Each image is downsampled M-1 times in turn, and a total of M pairs of images of different scales are obtained; for each pair of images, the contrast comparison function and the structure comparison function are calculated respectively; The contrast comparison function is calculated as: Among them, p m and g m is the mth pair of images, and They are p m and g m The respective standard deviations, c1 is a constant term; The calculation method of the structure comparison function is: Among them, p m and g m is the mth pair of images, Yes m and g m The covariance of and They are p m and g m The respective standard deviations, c2 is a constant term; Then, the brightness comparison function is calculated based on the Mth pair of images: Among them, p M and g M is the Mth pair of images, and They are p M and g M The respective average values, and They are p M and g M The respective standard deviations, c3 is a constant term; The multi-scale structural similarity index is calculated as: Among them, α M is a parameter that controls the relative importance of brightness similarity, β m is the relative importance parameter for controlling the contrast similarity at the scale corresponding to the mth pair of images, λ m is the relative importance parameter for controlling the structural similarity at the scale corresponding to the mth pair of images.

Images