Complex occlusion environment frequency spectrum situation construction method based on semantic segmentation interpolation

Abstract

The invention discloses a complex occlusion environment frequency spectrum situation construction method based on semantic segmentation interpolation, and belongs to the technical field of radio monitoring. Aiming at the problems of signal propagation mutation and difficult spectrum situation recovery caused by unknown information such as a radiation source and a propagation model in a scene and obstacle shielding, the method comprises the following steps of: firstly, carrying out discretization region division on a target scene; secondly, target point detection signal intensity is selected from the divided areas, an observation matrix and a sparse matrix are constructed according to detection positions and detection values, and a rough spectrum situation complementation graph is obtained through matrix filling; semantic segmentation processing is carried out on the basis to obtain position information of the obstacle, and the boundary of the obstacle is determined according to the position information; and finally, a regional spectrum situation result is solved by adopting a block thin plate spline interpolation method, and the overall spectrum situation of the target region is obtained through splicing. The method is suitable for a complex cognitive scene with obstacles in the environment, and has a good spectrum situation reconstruction effect.

Description

Technical Field

[0001] The present invention belongs to the technical field of radio monitoring and relates to a method for constructing spectrum situation in a complex occlusion environment based on semantic segmentation and interpolation. Background Art

[0002] The electromagnetic spectrum refers to the spectrum formed by the arrangement of electromagnetic waves according to their frequencies or wavelengths. It is a limited, non-renewable resource. The electromagnetic spectrum situation refers to the current state, overall situation, and development trends of the electromagnetic environment. Spectrum situation visualization involves processing and analyzing spectrum data and displaying it graphically to provide a more intuitive understanding of spectrum usage and changing trends.

[0003] Spectral state reconstruction methods can be categorized into parametric and non-parametric models. Parametric models primarily rely on the principle of spectrum state superposition. Their application often requires prior information, such as channel propagation characteristics. Typical parametric models include compressed sensing and Bayesian models. Non-parametric models generally reconstruct spectral state through interpolation algorithms and do not rely on prior information about the radiating source. They are therefore well-suited for scenarios where no prior information is available. Typical non-parametric methods include kriging, kernel methods, matrix completion, and tensor completion. In practical applications, prior information about the radiating source and propagation model is often difficult to obtain due to factors such as non-cooperative radiating sources and unknown propagation environment characteristics. This limits the applicability of parametric models that rely on prior information. Non-parametric models assume that spectral data is continuous and self-similar in the unknown region. However, when obstacles are present in the region, the RSS value exhibits abnormal oscillations. Non-parametric methods cannot accurately reconstruct the spectral state in such scenarios.

[0004] Therefore, under the premise of unknown prior information and the existence of obstacles in the construction scene, how to construct an accurate spectrum situation is of great significance. Summary of the Invention

[0005] To address the above technical issues, the present invention proposes a semantic segmentation-assisted TPS block interpolation spectrum situation completion method. This invention primarily addresses two problems: first, quickly and accurately determining the location of obstacles in the scene; second, completing spectrum situation without prior knowledge of the spatial propagation model.

[0006] A spectrum situation construction method for semantic segmentation and block interpolation considering obstacles, characterized by comprising the following steps:

[0007] S1: discretize the target area;

[0008] S2: Randomly select several target points in the coordinate system, deploy sensors, obtain the signal status of the target position, and construct a sparse matrix and an observation matrix;

[0009] S3: Perform matrix filling based on the observation matrix and the sparse matrix to obtain a rough spectrum situation completion map;

[0010] S4: Perform semantic segmentation based on the rough spectral situation completion map to obtain the location information of obstacles;

[0011] S5: Confirm the obstacle boundary based on the location information and perform block-by-block TPS interpolation to solve the spectrum map.

[0012] S6: Splice the spectrum situation reconstruction results of each area to obtain the overall spectrum situation of the target area

[0013] 2. The method according to claim 1, wherein the specific method of step S1 comprises:

[0014] In the target area, a fixed point is selected as the coordinate origin to establish a two-dimensional coordinate system (X, Y) of the target area. The target area is divided into cube blocks of equal size by selecting a certain interval on the two-dimensional coordinate system.

[0015] 3: According to the method of claim 1, the specific method of step S2 includes:

[0016] Based on the coordinate system constructed by S1, grid points are randomly selected as monitoring nodes, sensors are deployed at these points, and the location information of each monitoring node and the RSS of the monitoring node location are recorded.

[0017] The observation matrix M is constructed based on the location information of the monitoring node. When the element of M is "1", it means that there is a monitoring node at the location of the area. When the element of M is "0", it means that there is no monitoring node at the location of the area.

[0018] According to the location information of the monitoring node and the RSS of the monitoring node location, a sparse matrix Z is constructed. M , when there is a monitoring node in the area, Z M The element takes the signal strength corresponding to the monitoring node. When there is no monitoring node in the area, Z M The element is "0".

[0019] 4. The method according to claim 1, wherein the specific method of step S3 comprises:

[0020] Using the observation matrix M and sparse matrix Z M Perform matrix filling to obtain a rough spectrum situation map.

[0021] Specifically, the present invention uses a truncated nuclear norm (the nuclear norm minus the sum of the largest singular values) to better approximate the rank of the matrix, performs a matrix filling algorithm, and obtains a rough spectrum situation, which can roughly reflect the location information of the obstacle.

[0022] 5. The method according to claim 1, wherein the specific method of step S4 comprises:

[0023] S4.1 Network Framework

[0024] The main framework of the semantic segmentation network is Unet. The model is based on the autoencoder neural network. The encoder is a classification network used to extract features, and the decoder gradually restores the previously lost spatial information of the encoder.

[0025] The characteristics of the Unet network structure are: it can organize the information output by the coding block structure into an FPN feature image pyramid, and perform feature fusion with the results after upsampling in the next stage, which can make up for the features lost by the coding block structure and downsampling structure.

[0026] The encoding and decoding blocks both use a 3×3 convolution kernel with a stride of 1 and a normalization layer. To avoid excessive information loss, the convolutional layer padding mode is set to reflect. To mitigate overfitting, dropout is set to 0.3. The activation function is set to LeakyReLU. Two convolution operations are performed.

[0027] Downsampling architecture: To preserve more image information, the downsampling architecture abandons traditional pooling and instead uses a 3×3 convolution kernel with a stride of 2 and a normalization layer. To minimize information loss, the convolutional layer uses reflect as the padding mode, sets the activation function to LeakyReLU, and performs a single convolution operation to reduce image size.

[0028] Upsampling structure: First, to match the data from the feature fusion stage, a 2×2 convolution kernel is used to halve the number of channels. Second, to preserve as much information as possible, "nearest neighbor interpolation" is used for upsampling, doubling the image size.

[0029] S4.2 Training Process

[0030] Data preparation: The sampled data is matrix-filled to obtain a rough spectrum situation map, which is used as the training sample of Unet. The obstacle information is made into a label. The label is made into an eight-bit image format, that is, each pixel uses an 8-bit binary number to represent its color information, thereby compressing the image to a smaller size. The positions with obstacles are filled with 1. The purpose of labeling is to preserve the obstacle information of the original image so that the predicted image generated during the training process has a real value as the loss, thereby further reducing the loss and adjusting the parameters. It is convenient for model optimization.

[0031] Data preprocessing: The obtained training samples and labels are preprocessed to make them of the same size. The sample data and their corresponding labels are aligned.

[0032] Training process: The Unet network framework is used for training. The input is training samples and labels, and the output is the training effect image and the trained network model parameters. The optimizer uses Adam and the loss function uses CrossEntropyLoss.

[0033] S4.3 Test procedure

[0034] First, the image to be tested is preprocessed to obtain the same size as the training sample. The trained Unet model is used to perform forward propagation calculation on the input tensor to obtain the output prediction tensor out. After dimensional processing of the data, it represents the pixel-level segmentation result of the model on the input image.

[0035] S4.3 Judgment indicators

[0036] The prediction results and the actual values ​​are used to create a confusion matrix, and the intersection-over-union ratio of the predicted target box and the actual target box is analyzed to evaluate the performance of target detection.

[0037] 6. The method according to claim 1, wherein the specific method of step S5 comprises:

[0038] S5.1 Determine obstacle boundaries

[0039] Obstacle information can be obtained from semantic segmentation information, and the target area is divided into free space propagation area and obstacle area according to the semantic segmentation information.

[0040] S5.2 Block TPS Interpolation

[0041] Semantic segmentation divides the scene into blocks. Within each block, signals are considered to be propagating through the same medium, and signal strength does not change suddenly. Therefore, TPS interpolation is used to perform block segmentation and recover the spectrum status. The main idea of ​​TPS interpolation is to fit the sampled data using the TPS function to obtain the interpolation function, and then use the interpolation function to estimate the data at unknown locations to complete the interpolation of the target area. The spectrum status generation algorithm based on TPS interpolation uses the TPS function to fit the RSS data collected by the sensor, estimate the RSS value of the unknown point, and combine the RSS distribution of the entire target area to generate a spectrum map.

[0042] Beneficial effects

[0043] Compared to traditional parametric methods, including compressed sensing, dictionary learning, and Bayesian models, this method requires no prior knowledge of the electromagnetic environment and is more universal. Compared to direct interpolation algorithms for spectral situation reconstruction based on non-parametric models, this method overcomes the difficulty of large interpolation errors caused by obstacles in the scene.

[0044] This method performs matrix filling on sparse sampling points to obtain a rough spectrum situation. It then uses semantic segmentation to divide the region into free space propagation and obstacle areas, and performs block interpolation restoration. By splicing different areas together, the spectrum situation within the entire area can be restored. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to make the purpose, technical solutions and beneficial effects of the present invention more clear, the present invention provides the following drawings for illustration:

[0046] Figure 1 This is a complete flow chart of a semantic segmentation-assisted TPS block interpolation spectrum situation completion method of the present invention;

[0047] Figure 2 This is a schematic diagram of the Unet network framework of the present invention;

[0048] Figure 3 This is a flowchart of a semantic segmentation-assisted TPS block interpolation spectrum situation completion method of the present invention;

[0049] Figure 4 This is the NAE error curve of the present invention; DETAILED DESCRIPTION

[0050] In order to make the steps of the present invention more detailed and clear, the present invention is further described in detail below with reference to the accompanying drawings and implementation examples.

[0051] like Figure 1 As shown, this is a flow chart of Example 1 of an electromagnetic spectrum situation mapping method disclosed in this application.

[0052] In step S1, the target area is discretized. In the two-dimensional target area with a spatial size of (X, Y), the target area is divided into I×J equally spaced grids, with a total of I×J grid points.

[0053] In step S2, considering that there are N monitoring nodes in the target area, sensors are deployed at this point, (x n ,y n ) represents the position of the nth sensor, {R1, R2, …, R N} represents the RSS data collected by N sensors.

[0054] The observation matrix M is constructed based on the location information of the monitoring node. When the element of M is "1", it means that there is a monitoring node at the location of the area. When the element of M is "0", it means that there is no monitoring node at the location of the area.

[0055]

[0056] According to the location information of the monitoring node and the RSS corresponding to the monitoring node location, a sparse matrix Z is constructed. M , when there is a monitoring node in the area, Z M The element takes the signal strength corresponding to the monitoring node. When there is no monitoring node in the area, Z M The element is "0".

[0057]

[0058] In step S3, the observation matrix M and the sparse matrix Z are used M Perform matrix filling to obtain a rough spectrum situation map.

[0059] Specifically, the formula can be used:

[0060]

[0061] stZ=W,P M (Z)=P M (Z M )

[0062] To solve the result of matrix filling, where Z is the result after matrix filling to be solved, A l ,B l The first l columns of U and V corresponding to the decomposition of matrix Wsvd are truncated matrices, where l is determined by the maximum position of the first-order difference and second-order difference of the singular value. Define P M For the orthogonal projection algorithm:

[0063]

[0064] Then the augmented Lagrangian function of the above formula is:

[0065]

[0066] Where β>0 is the penalty term coefficient. Initialize Z1=Z M , W1=Z1, Y1=Z1. The TNNR-ADMM optimization method updates the variables alternately by minimizing the augmented Lagrangian function. The convex optimization problem can be solved by the following steps:

[0067] (1) Fixed W k and Y k , update Z k+1 :

[0068]

[0069] Ignoring the constant term, the equation can be written as:

[0070]

[0071] (2) Fixed X k+1 and Y k Calculate W k+1 :

[0072]

[0073] First, let To ensure that the observed part of the restored matrix values ​​remains unchanged, perform the following operations:

[0074]

[0075] (3) Fixed X k+1 and W k+1 Calculate Y k+1 :

[0076] Y k+1 =Y k +β(Z k+1 -W k+1 )

[0077] The matrix filling algorithm is performed to obtain a rough spectrum situation, which can roughly reflect the location information of the obstacle.

[0078] In step S4, the rough spectrum situation in step S3 is fed into the neural network for semantic segmentation to determine the location information of the obstacle.

[0079] S4.1 Network Framework

[0080] The main framework of the semantic segmentation network is Unet. The model is based on the autoencoder neural network. The encoder is a classification network used to extract features, and the decoder gradually restores the previously lost spatial information of the encoder.

[0081] The characteristics of the Unet network structure are: it can organize the information output by the coding block structure into an FPN feature image pyramid, and perform feature fusion with the results after upsampling in the next stage, which can make up for the features lost by the coding block structure and downsampling structure.

[0082] The encoding and decoding blocks both use a 3×3 convolution kernel with a stride of 1 and a normalization layer. To avoid excessive information loss, the convolutional layer padding mode is set to reflect. To mitigate overfitting, dropout is set to 0.3. The activation function is set to LeakyReLU. Two convolution operations are performed.

[0083] Downsampling architecture: To preserve more image information, the downsampling architecture abandons traditional pooling and instead uses a 3×3 convolution kernel with a stride of 2 and a normalization layer. To minimize information loss, the convolutional layer uses reflect as the padding mode, sets the activation function to LeakyReLU, and performs a single convolution operation to reduce image size.

[0084] Upsampling structure: First, to match the data from the feature fusion stage, a 2×2 convolution kernel is used to halve the number of channels. Second, to preserve as much information as possible, "nearest neighbor interpolation" is used for upsampling, doubling the image size.

[0085] S4.2 Training Process

[0086] Data preparation: The sampled data is matrix-filled to obtain a rough spectrum situation map, which is used as the training sample of Unet. The obstacle information is made into a label. The label is made into an eight-bit image format, that is, each pixel uses an 8-bit binary number to represent its color information, thereby compressing the image to a smaller size. The positions with obstacles are filled with 1. The purpose of labeling is to preserve the obstacle information of the original image so that the predicted image generated during the training process has a real value as the loss, thereby further reducing the loss and adjusting the parameters. It is convenient for model optimization.

[0087] Data preprocessing: The obtained training samples and labels are preprocessed to make them of the same size. The sample data and their corresponding labels are aligned.

[0088] Training process: The Unet network framework is used for training. The input is training samples and labels, and the output is the training effect image and the trained network model parameters. The optimizer uses Adam and the loss function uses CrossEntropyLoss.

[0089] S4.3 Test procedure

[0090] First, preprocess the image to be tested to the same size as the training sample. Then, use the trained UNet model to perform a forward pass on the input tensor, generating the output prediction tensor out. For each pixel in the output prediction tensor out, the class number with the highest predicted probability is calculated along the channel dimension. This is done by using the torch.argmax() function to take the maximum value along the second dimension (i.e., the number of channels), resulting in a two-dimensional integer tensor of shape (H, W). The output prediction tensor out is then dimensionalized to remove unnecessary batch size and channel dimensions. Note that during training, since input data is typically loaded in batches (batchsize), the shape of the training images is typically (B, C, H, W), where B represents the batch size, C represents the number of channels (i.e., the number of classes), and H and W represent the image height and width, respectively. However, during testing, since only a single image is processed, a new batch size dimension is added using the torch.unsqueeze() method, reducing the input tensor shape to (1, C, H, W). The torch.squeeze() method can be used to remove the redundant batch size dimension and channel number dimension, resulting in a two-dimensional integer tensor with a shape of (H, W). However, in order to meet the requirements of subsequent visualization functions, a batch size dimension needs to be added again (that is, using the torch.unsqueeze() method).

[0091] The final output prediction result out is a three-dimensional integer tensor with a shape of (1, H, W), which represents the pixel-level segmentation result of the model on the input image.

[0092] S4.3 Judgment indicators

[0093] The prediction results and the actual values ​​are used to create a confusion matrix, and the intersection-over-union ratio of the predicted target box and the actual target box is analyzed to evaluate the performance of target detection.

[0094] In step S5, the obstacle boundary is confirmed and the spectrum situation is solved by block TPS interpolation.

[0095] S5.1 First, the control point matrix is ​​determined based on the location information of the monitoring nodes:

[0096]

[0097] Where (xi ,y i ) represents the coordinates of a monitoring node.

[0098] S5.2 Determine the height matrix

[0099]

[0100] where R i Represents the monitoring point coordinates and the corresponding RSS. The three zeros at the end are for uniform filling.

[0101] Set the radial basis function U(x) = r 2 logr means that the deformation of a point on a surface will be affected by the deformation of all control points

[0102]

[0103] where r ij =||p i -p j || represents the distance between two control points, let the matrix L be:

[0104]

[0105] have:

[0106] Y=L(Ω|m0,m1,m2)

[0107] Where Ω=(ω1,…ω N ) The last three lines introduce a set of constraints on the parameters:

[0108]

[0109]

[0110]

[0111] We have:

[0112] (Ω|m0,m1,m2) T =L -1 Y

[0113] Of course, the parameter set (Ω|m0,m1,m2) can also be obtained by solving the linear equations T Once this parameter set is calculated, our interpolation function is also known. Given any point on the plane, we can interpolate it to the target plane through the interpolation function, thereby restoring the complete spectrum from the sampled data.

[0114] In step S6, the spectrum situation reconstruction results of each area are spliced ​​together to obtain the overall spectrum situation of the target area point.

[0115] The application effect of the present invention is described in detail below in conjunction with simulation.

[0116] 1) Simulation conditions

[0117] The experiment is based on spectrum map data generated by simulation. The test scenario is: two unknown sources are located in the plane area of ​​interest, which is 100×100m. 2 , divided into an equally spaced I × J grid, where I = J = 101. Sensors with a sampling rate of ρ (0.07-0.2) receive signals from various radiators and periodically transmit the signal power within the frequency band of interest to the fusion center. The free-space path loss exponent is defined as 2, the obstacle path loss exponent is defined as 5, the shadow fading variance in the scene is defined as 2 dB, and the spatial radiators are assumed to be omnidirectional antennas.

[0118] 2) Simulation results

[0119] In this embodiment, Figure 4 The graph shows the change in the normalized absolute error (NAE) of the algorithm in this embodiment as the sampling rate changes from 0.07 to 0.2. The figure shows that the NAE does not decrease significantly with increasing sampling rate. This is due to the sampling rate restrictions imposed by the matrix filling algorithm. When the sampling rate restrictions are met, the block interpolation algorithm proposed in this embodiment is effective in recovering the spectral state of scenes containing obstacles.

Claims

1. A method for constructing spectrum situation in complex occlusion environment based on semantic segmentation and interpolation, characterized by: The following steps are involved: S1: Discrete region division of the target scene; S2: Randomly select several target points in the divided area, deploy sensors, obtain the received signal strength (RSS) of the target location, and construct a sparse matrix and an observation matrix; S3: Perform matrix filling based on the observation matrix and the sparse matrix to obtain a rough spectrum situation completion map; S4: Perform semantic segmentation based on the rough spectral situation completion map to obtain the location information of obstacles; S5: The obstacle boundary is confirmed based on the position information, and the regional spectrum situation result is obtained by using the block thin plate splines (TPS) interpolation method; S6: Splice the spectrum situation reconstruction results of each area to obtain the overall spectrum situation of the target area.

2. The method according to claim 1, characterized in that The specific method of step S1 includes: In the target area, a certain point is selected as the coordinate origin, and a two-dimensional coordinate system (X, Y) of the target area is established. A certain interval is selected on the two-dimensional coordinate system to divide it into equal intervals and divide it into areas of the same size.

3. The method according to claim 1, characterized in that The specific method of step S2 includes: Based on the coordinate system constructed by S1, grid points are randomly selected as monitoring nodes, sensors are deployed at these points, and the location information of each monitoring node and the RSS of the monitoring node location are recorded. The observation matrix M is constructed based on the location information of the monitoring node. When the element of M is "1", it means that there is a monitoring node at the location of the area. When the element of M is "0", it means that there is no monitoring node at the location of the area. According to the location information of the monitoring node and the RSS of the monitoring node location, a sparse matrix Z is constructed. M , when there is a monitoring node in the area, Z M The element takes the signal strength corresponding to the monitoring node. When there is no monitoring node in the area, Z M The element is "0".

4. The method according to claim 1, characterized in that The specific method of step S3 includes: Using the observation matrix M and sparse matrix Z M Perform matrix filling to obtain a rough spectrum situation map. Specifically, the present invention uses a truncated nuclear norm (the nuclear norm minus the sum of the largest singular values) to better approximate the rank of the matrix, performs a matrix filling algorithm, and obtains a rough spectrum situation, which can roughly reflect the location information of the obstacle.

5. The method according to claim 1, characterized in that: The specific method of step S4 includes: S4.1 Network Framework The main framework of the semantic segmentation network is Unet. The model is based on the autoencoder neural network. The encoder is a classification network used to extract features, and the decoder gradually restores the previously lost spatial information of the encoder. The characteristics of the Unet network structure are: it can organize the information output by the coding block structure into an FPN feature image pyramid, and perform feature fusion with the results after upsampling in the next stage, which can make up for the features lost by the coding block structure and downsampling structure. The encoding and decoding blocks both use a 3×3 convolution kernel with a stride of 1 and a normalization layer. To avoid excessive information loss, the convolutional layer padding mode is set to reflect. To mitigate overfitting, dropout is set to 0.

3. The activation function is set to LeakyReLU. Two convolution operations are performed. Downsampling structure: To retain more image information, the downsampling structure abandons the traditional pooling operation and instead uses a 3×3 convolution kernel with a stride of 2 plus a normalization layer. In order to avoid excessive information loss, the convolution layer padding mode uses reflect; the activation function is set to LeakyReLU; a convolution operation is performed; and the image size is reduced to achieve downsampling. Upsampling Structure: To match the data from the feature fusion stage, a 2×2 convolution kernel is first used to halve the number of channels. Secondly, to preserve as much information as possible, "nearest neighbor interpolation" is used for upsampling, doubling the image size. S4.2 Training Process Data preparation: The sampled data is matrix-filled to obtain a rough spectrum situation map, which is used as the training sample of Unet. The obstacle information is made into a label. The label is made into an eight-bit image format, that is, each pixel uses an 8-bit binary number to represent its color information, thereby compressing the image to a smaller size. The positions with obstacles are filled with 1. The purpose of labeling is to preserve the obstacle information of the original image so that the predicted image generated during the training process has a real value as the loss, thereby further reducing the loss and adjusting the parameters. It is convenient for model optimization. Data preprocessing: The obtained training samples and labels are preprocessed to make them of the same size. The sample data and their corresponding labels are aligned. Training process: The Unet network framework is used for training. The input is the training samples and labels. The output is the training effect diagram and the trained network model parameters. The optimizer uses Adam and the loss function uses CrossEntropyLoss cross entropy. S4.3 Test procedure First, the image to be tested is preprocessed to obtain the same size as the training sample. The trained Unet model is used to perform forward propagation calculation on the input tensor to obtain the output prediction tensor out. After dimensional processing of the data, it represents the pixel-level segmentation result of the model on the input image. S4.3 Judgment indicators The prediction results and the actual values ​​are used to create a confusion matrix, and the intersection-over-union ratio of the predicted target box and the actual target box is analyzed to evaluate the performance of target detection.

6. The method according to claim 1, characterized in that: The specific method of step S5 includes: S5.1 Determine obstacle boundaries Obstacle information can be obtained from semantic segmentation information, and the target area is divided into free space propagation area and obstacle area according to the semantic segmentation information. S5.2 Block TPS Interpolation Semantic segmentation divides the scene into blocks. Within each block, signals are considered to be propagating through the same medium, and signal strength does not change suddenly. Therefore, TPS interpolation is used to perform block segmentation and recover the spectrum status. The main idea of ​​TPS interpolation is to fit the sampled data using the TPS function to obtain the interpolation function, and then use the interpolation function to estimate the data at unknown locations to complete the interpolation of the target area. The spectrum status generation algorithm based on TPS interpolation uses the TPS function to fit the RSS data collected by the sensor, estimate the RSS value of the unknown point, and combine the RSS distribution of the entire target area to generate a spectrum map.

Images