A saliency-based wideband spectrum detection method

Abstract

The application discloses a wideband spectrum detection method based on saliency, which comprises the following steps: performing saliency map conversion on a one-dimensional wideband spectrum data set to convert original data into a one-dimensional saliency vector; inputting the converted saliency vector into a deep saliency detection network for supervised training, extracting low-level detail features, high-level semantic features and global context features of the signal, and fusing the extracted features to obtain fused features; performing linear interpolation upsampling on the fused features to reconstruct the input saliency vector, and calculating cross-entropy loss of the reconstructed vector and the input saliency vector; and finally calculating parameter information such as signal number, frequency, bandwidth and signal-to-noise ratio based on the output result to complete wideband spectrum detection in a complex electromagnetic environment. The application realizes wideband spectrum detection, does not depend on experts, has strong environmental adaptability, is easy to train, has a relatively fast convergence speed, can effectively improve detection probability, and is a wideband spectrum detection method based on saliency.

Description

Technical Field

[0001] This invention relates to the field of signal detection technology, and in particular to a broadband spectrum detection method based on saliency. Background Technology

[0002] With the rapid development of various radio communication systems and the development and application of various electromagnetic communication devices, a large number of electromagnetic signals with complex signal systems and modulation patterns, dense overlap, and dynamic interleaving permeate the electromagnetic environment. Civilian and military signals with different power, bandwidth, modulation methods, coding methods, and time-varying signals are intertwined in the receiving frequency band, making the detected signals exhibit complex characteristics such as massive quantity, high dimension, diversity, and dynamics. These complex electromagnetic environments are manifested in the following aspects: (1) The number of electromagnetic signals in space is increasing, and the background noise is significantly increasing; (2) Various interferences are becoming increasingly serious, and the background noise often exhibits non-Gaussian characteristics; (3) The noise is highly dynamic due to various sudden signals, interferences, and background influences.

[0003] Signal detection and extraction technology is a crucial component of signal processing and modern information theory, and a pillar of information technology. Specifically, signal detection refers to constructing detection features from broadband received signal data to determine the existence of communication signals and to preliminarily estimate and identify communication parameters such as carrier frequency, bandwidth, and system architecture of various narrowband signals. This information effectively guides subsequent modulation identification and individual signal identification. Since the performance of signal detection directly impacts the performance and application of the entire communication system, indiscriminate reconnaissance of signals within the reconnaissance frequency band results in a large number of low-value signals occupying limited signal reconnaissance resources. This not only wastes manpower and time but also leads to subsequent information acquisition failures, putting the system at a disadvantage in communication warfare. Therefore, in complex electromagnetic environments, the ability to quickly and accurately perform signal detection on massive amounts of data is of great significance. This will fully utilize limited signal reconnaissance resources, reduce manpower and time costs, and improve the analysis efficiency of high-value signals such as signals from important and high-threat targets, unknown / unidentified signals, and signals of mission interest.

[0004] Among traditional spectrum detection algorithms, matched filtering, energy detection, cyclostationary methods, and eigenvalue detection are the most widely used and classic methods. In addition, wavelet analysis, which extracts the target signal by performing hierarchical analysis of the signal at different scales, has become a research hotspot in signal detection in recent years. When prior knowledge of the signal to be detected is available, matched filtering is the optimal detection method under Gaussian background noise. However, in real-world scenarios, prior knowledge of the signal to be detected is often difficult to obtain. Therefore, the requirement for prior knowledge in matched filtering methods often limits their practical application. Energy detection methods do not require prior information about the signal to be detected and are the most commonly used detection method in practical systems due to their low computational complexity. This method measures the total energy of the received signal over a certain period of time within a defined frequency band and then compares it with a decision threshold to determine whether a signal exists. The energy detection threshold varies with the sampling rate, sampling time, and the variance of the ambient noise, so energy detection methods require accurate estimation of the variance of the surrounding ambient noise. However, under low signal-to-noise ratio conditions, signal variance is often difficult to estimate accurately, leading to a sharp decline in the performance of energy detection methods. Eigenvalue detection, a method that has gradually developed over the past decade, uses the ratio of the largest to smallest eigenvalue of the covariance matrix as a test statistic based on the principle of random matrices. If the ratio is greater than a threshold, the signal is considered present; otherwise, it is considered absent. This method effectively overcomes the impact of noise uncertainty on detection performance and exhibits good performance. However, limited by random matrix theory, the precise threshold of this method is difficult to obtain theoretically and often requires extensive practical experiments for correction, significantly impacting its practical application. The difficulty and bottleneck of wavelet analysis lies in the selection of wavelet basis functions. To date, there is no standard or universal method to obtain wavelet basis functions. Furthermore, the resolution of wavelet transform is limited by sampling frequency and length. All of the above methods are model-driven and require prior knowledge of signal or noise information. However, in real-world environments, signals are often unknown, and noise conditions are dynamically changing, hindering the performance of traditional spectrum sensing methods. Summary of the Invention

[0005] The main objective of this invention is to provide a saliency-based broadband spectrum detection method, which aims to solve the technical problems of traditional broadband signal detection algorithms in practical complex electromagnetic environments, such as high dependence on expert experience, complex threshold settings, poor adaptability, and poor detection performance under large differences in signal bandwidth in complex electromagnetic environments.

[0006] To achieve the above objectives, the present invention provides a broadband spectrum detection method based on saliency, the method comprising the following steps:

[0007] S1: Based on the signal start and end information, perform saliency map transformation on the broadband spectrum data to obtain a binary one-dimensional saliency vector as the label of the broadband spectrum data;

[0008] S2: Input broadband spectrum data and corresponding labels into a deep learning model to extract low-level detail features, high-level semantic features and global context features, and construct fused features;

[0009] S3: Linear interpolation upsampling is performed on the fused features to make the feature dimension the same as the input vector. The input vector is reconstructed using the SoftMax function. The cross-entropy loss between the reconstructed vector and the input salient vector is calculated. After inference by the deep saliency detection network model, the saliency result of the input data is obtained.

[0010] S4: Calculate the start and end points of the broadband spectrum based on the significance results, and calculate the model parameters in the broadband spectrum based on the start frequency and frequency resolution of the broadband spectrum data to complete the detection of the broadband spectrum.

[0011] Optionally, step S2 further includes performing maximum and minimum value normalization on the broadband spectrum data, the expression of which is as follows:

[0012]

[0013] Among them, X norm Here, X represents the normalized sample data, and X represents the original sample data. min X is the minimum value of the sample data. max This represents the maximum value of the sample data.

[0014] Optionally, in step S2, the extraction of low-level detail features and high-level semantic features specifically includes: using a 50-layer deep residual network to extract four layers of low-level features and one layer of high-level features.

[0015] Optionally, step S2 further includes inputting the extracted high-level semantic features into the channel selection module for selection, utilizing spatial and channel attention mechanisms to enhance spatial regions and features that have a high response to salient objects; the expression used by the channel selection module to select high-level semantic features is specifically as follows:

[0016] y = sigmod(fc2(RELU(fc1(F))))

[0017] F out =F·y

[0018] Where fc1 and fc2 represent fully connected layer 1 and fully connected layer 2, F represents the input high-level semantic features, f is the channel weight feature vector obtained by downsampling the high-level features, ReLU and sigmoid represent activation functions, · represents dot product, y represents the final weight vector, and F out This represents the selected feature.

[0019] Optionally, in step S2, the expression for the global context feature extraction is specifically:

[0020]

[0021] y t = sigmod(fc4(RELU(fc3(f gap ))))

[0022]

[0023] Where represents the high-level features of the input, conv represents the convolution operation, and f gap This represents the feature obtained by global average pooling of high-level features, which includes global contextual information. The dot (·) represents the dot product. This represents the final output of the global feature extraction module.

[0024] Optionally, in step S2, the fusion strategy for the fusion features is specifically as follows:

[0025]

[0026]

[0027]

[0028]

[0029]

[0030]

[0031]

[0032]

[0033] in, f represents high-level semantic features l t Represents low-level detailed features. Representing global context features, upsample represents bilinear interpolation upsampling, t represents stage number, W represents the generated mask, and concat represents concatenation by channel; the feature fusion module first combines the low-level features f l t The input is fed into the convolutional layer conv1 to obtain the same number of channels as the high-level features. Then, high-level characteristics The mask is obtained by bilinear interpolation upsampling after inputting it into the convolutional layer conv2. mask and After performing a dot product operation and then applying the ReLU activation function, the first type of fused feature from the high and low layers is obtained. The same method can be used to obtain another fused feature of high and low layer features. and global low-level fusion features Finally, the three preliminary fusion features are concatenated by channel and then passed through a convolutional layer to obtain the final fusion feature.

[0034] Optionally, after step S2, the method further includes: refining and enhancing the fused features using multiplication and addition modules, the specific expression of which is:

[0035]

[0036]

[0037] Among them, f in The input features are represented by f, and conv represents the convolution operation. First, f is... in The input is fed into a 3x3 convolutional layer to obtain... Then The inputs are fed into two convolutional layers to obtain the mask W and the bias b, respectively. The obtained W and b are then compared... Perform the operation to obtain the final output f out .

[0038] Optionally, in step S3, when calculating the cross-entropy loss between the reconstructed vector and the input salient vector, the loss function of the sampling results of the three intermediate upper layers is also calculated, and its expression is as follows:

[0039]

[0040]

[0041] Among them, G ij S represents the true labels of points i and j. ij Represents the predicted label, l dom The loss representing the final predicted result. The auxiliary loss representing the intermediate result, γ i This represents the weighting coefficient.

[0042] Optionally, in step S4, the start and end points of the broadband spectrum are calculated based on the significance results. Specifically, the part with a length exceeding the threshold and continuously equal to 1 is regarded as a signal, and the start and end points of this part are considered as the start and end points of the signal.

[0043] Optionally, in step S4, the model parameters in the broadband spectrum are calculated based on the starting frequency and frequency resolution of the broadband spectrum data to complete the broadband spectrum detection step, specifically including:

[0044] f start =f startin +Start index *f step

[0045] f end =f startin +End index *f step

[0046] f c =f start +(f end -f start ) / 2

[0047] bw = (End) index -Start index )*f step

[0048] Among them, f startin f represents the starting frequency of the original broadband spectrum data. step Indicates frequency resolution, Start index Indicates the start point of the signal, End index Indicates the termination point of the signal, f start f represents the starting frequency of the signal. end f represents the signal termination frequency. c bw represents the center frequency of the signal, and bw represents the bandwidth of the signal.

[0049] The beneficial technical effects of this invention are as follows:

[0050] (1) Reduced reliance on experts. This invention uses data-driven deep learning technology, which can mine potential patterns and information within data through complex network structures and weight connections. By continuously adapting the network weights to the distribution of data, it automatically learns the data features. No thresholds or parameters need to be set during the analysis process, which greatly reduces reliance on expert knowledge and reduces the manpower and time costs of analysis.

[0051] (2) Strong environmental adaptability. This invention is trained based on real-world data. During the training process, it deeply mines the low-level features, high-level features, and global context features of the real-world data, and fully integrates these three types of features through a feature fusion module. This allows the fused features to possess the advantages of these three types of features, enabling more effective extraction of salient information from the input data, suppression of noise effects, and recovery of more structural and detailed information about the target. It can ensure the integrity of salient target detection in complex environments and reduce noise interference. Compared with traditional methods that set thresholds based on background noise, it has stronger environmental adaptability and can guarantee signal detection performance even when background noise changes.

[0052] (3) Easy to train and fast convergence. This invention establishes a saliency-based deep learning model for broadband signal detection tasks; the deep learning model is trained under supervision using the actual sampled signal training set, thus transforming the broadband detection problem into pixel-level detection, where one pixel corresponds to one label, salient regions are considered positive samples, and non-salient regions are considered negative samples. Since one pixel corresponds to one label, the saliency-based broadband detection method does not require a large amount of data for training, has a smaller computational load compared to other deep learning models, is easier to train, and has a faster convergence speed. Furthermore, the deep neural network model can be deployed on high-performance GPUs or dedicated AI processing chips for rapid parallel inference, enabling fast computation of the deep neural network model.

[0053] (4) Small detection error and high detection probability. This invention targets real-world broadband spectrum data and uses high-precision floating-point operators to construct a deep neural network model. The deep neural network model is trained with high-precision labeled data and trained using a supervised training method. Multiple features of the data are extracted and fused using the backbone network. Linear interpolation is performed at different feature fusion stages, and cross-entropy is calculated with the input saliency map for multiple reconstructions, ensuring that the features obtained in the multi-layer feature extraction process are conducive to saliency map generation. The trained deep neural network model is used to process the broadband spectrum data, automatically generating a saliency map of the signal, and calculating parameters such as the signal's center frequency and bandwidth based on the saliency map. This achieves highly stable and high-precision detection of real-world broadband spectrum signals. It not only has high detection stability and accuracy but also high parameter estimation accuracy. Furthermore, simulations using computational methods can be used to verify the results from aspects such as frequency offset acquisition range, signal-to-noise ratio working threshold, and level reception range. The verification results show that this invention has a smaller detection error and a higher detection probability compared to traditional methods. Attached Figure Description

[0054] Figure 1 This is a schematic flowchart of an embodiment of the broadband spectrum detection method based on saliency of the present invention.

[0055] Figure 2 This is a schematic diagram illustrating the principle and flow of the broadband spectrum detection method based on saliency of the present invention.

[0056] Figure 3 This is a schematic diagram of the saliency-based signal detection model of the present invention.

[0057] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0058] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0059] Currently, in related technical fields, traditional broadband signal detection algorithms suffer from several drawbacks when applied to complex electromagnetic environments: high reliance on expert experience, complex threshold settings, poor adaptability, and unsatisfactory detection performance under conditions of large differences in signal bandwidth in complex electromagnetic environments.

[0060] To address this issue, various embodiments of the saliency-based broadband spectrum detection method of this invention are proposed. The saliency-based broadband spectrum detection method provided by this invention transforms a one-dimensional broadband spectrum dataset into a one-dimensional salient vector by performing saliency map transformation. The transformed salient vector is then input into a deep saliency detection network for supervised training to extract low-level detail features, high-level semantic features, and global context features of the signal. These extracted features are then fused to obtain fused features. The fused features are then linearly interpolated and upsampled to reconstruct the input salient vector, and the cross-entropy loss between the reconstructed vector and the input salient vector is calculated. Finally, based on the output results, parameters such as the number of signals, frequency, bandwidth, and signal-to-noise ratio are calculated to complete broadband spectrum detection in complex electromagnetic environments. This method achieves broadband spectrum detection without relying on experts, exhibits strong environmental adaptability, is easy to train, has a fast convergence speed, and can effectively improve the detection probability.

[0061] This invention provides a saliency-based broadband spectrum detection method, referring to... Figure 1 , Figure 1 This is a schematic flowchart of an embodiment of the broadband spectrum detection method based on saliency of the present invention.

[0062] In this embodiment, the saliency-based broadband spectrum detection method includes the following steps:

[0063] S1: Based on the signal start and end information, perform saliency map transformation on the broadband spectrum data to obtain a binary one-dimensional saliency vector as the label of the broadband spectrum data;

[0064] S2: Input broadband spectrum data and corresponding labels into a deep learning model to extract low-level detail features, high-level semantic features and global context features, and construct fused features;

[0065] S3: Linear interpolation upsampling is performed on the fused features to make the feature dimension the same as the input vector. The input vector is reconstructed using the SoftMax function. The cross-entropy loss between the reconstructed vector and the input salient vector is calculated. After inference by the deep saliency detection network model, the saliency result of the input data is obtained.

[0066] S4: Calculate the start and end points of the broadband spectrum based on the significance results, and calculate the model parameters in the broadband spectrum based on the start frequency and frequency resolution of the broadband spectrum data to complete the detection of the broadband spectrum.

[0067] It's important to note that target saliency is a fundamental concept in psychology and neuroscience, serving as a crucial tool for studying the perceptual mechanisms of the human visual system. In computer science, the study of visual saliency in images was first initiated by scholars in computer vision and continues to this day. Subsequently, saliency has been widely applied in shape analysis and processing within computer graphics. After more than a decade of development, shape saliency has become a vital bridge connecting human visual perception and shape analysis. In natural scenes, the human visual system's detection of object saliency generally involves two stages: first, a parallel, rapid, and simple pre-attention process; then, a sequential, slow, and complex attention process. In the first stage, low-level features (e.g., edges, orientation, light intensity) are quickly perceived; in the second stage, these features act as candidate elements, influencing the intelligent behavior of target detection. Based on this idea, the signal detection problem in complex electromagnetic environments can be considered a one-dimensional saliency detection problem.

[0068] Deep learning is an emerging data-driven intelligent technology that can uncover potential patterns and information within data through complex network structures and weighted connections. By allowing network weights to continuously adapt to the data distribution, it automatically learns data features, significantly reducing reliance on expert knowledge in the analysis process. It is a powerful tool for reducing the manpower and time costs of analysis. The essence of deep learning is its distributed and parallel computing architecture, which ensures high computational speed and processing power. With the continuous upgrading of parallel processing architectures such as GPUs, the development of deep learning technology has accelerated, shortening model training time and enabling practical applications. However, most current deep learning-based signal detection methods focus on the features of the target signal while neglecting background environmental features, resulting in insufficient representation of electromagnetic spatial information. Furthermore, complex electromagnetic environments contain a large number of meaningless, low-value signals, while the number of high-value signals is significantly disparate from low-value signals. This necessitates analysis and annotation by personnel with specialized knowledge, and the amount of labeled samples is insufficient to meet the data requirements of deep learning models, while also suffering from imbalanced positive and negative samples.

[0069] Attention mechanisms are highly efficient because they selectively focus on relevant information within the overall data set, while ignoring irrelevant information to some extent. Attention mechanisms represent an improvement over encoder-decoder-based neural machine translation systems in natural language processing. They can give more attention to local focal points and reduce computational complexity by filtering out useless information. Attention mechanisms and their variants have wide applications in discourse, speech, and vision processing. Deep learning models, with their unique deep feature extraction and learning capabilities, possess a distinct advantage in detecting anomalies in dynamic and complex environments. Therefore, deep learning with attention mechanisms can effectively achieve automatic feature learning and reduce computational complexity.

[0070] Based on this, in order to explain this application more clearly, a specific example of a broadband spectrum detection method based on saliency is provided below.

[0071] See Figure 2Based on the signal start and end information of broadband one-dimensional spectrum data, the original data is transformed into a saliency map. Regions with signal are considered salient and represented by 1, while regions without signal are considered non-salient and represented by 0, resulting in a binary one-dimensional 0-1 saliency vector as the label. Simultaneously, the original data is normalized by minimax and maxmax, and the vector, along with the corresponding label, is input into the deep learning model. A deep residual network is used as the backbone network to extract various features from the input data, including low-level detail features and high-level semantic features. The extracted high-level feature map is input into the channel selection module, which utilizes spatial and channel attention mechanisms to enhance spatial regions and features with high salient object response, resulting in more representative high-level semantic features. These high-level semantic features are then input into the global feature extraction module to capture the relationships between multiple salient objects or different parts of salient objects from a global perspective, obtaining global context features. The enhanced high-level semantic features, low-level detail features, and global feature extraction are then combined. Contextual features are input to the feature fusion module for fusion, integrating the functions of three types of features to suppress noise while recovering more structural and detailed information. The fused feature map is then input to the feature enhancement module, which refines the feature mapping by utilizing the internal characteristics of the feature map to obtain the final fused feature map. Linear interpolation upsampling is performed on the fused features to make the feature dimension the same as the input vector. The SoftMax function is used to map the deconvolutioned data to real numbers between 0 and 1 to reconstruct the input vector. The cross-entropy loss between the reconstructed vector and the input saliency vector is calculated. After inference by the deep saliency detection network model, the saliency result of the input data is obtained. Based on the saliency result, the starting and ending points of the signal are calculated, and the number of signals, the center frequency of the signal, the signal bandwidth, and other parameter information contained in the broadband spectrum are calculated based on the starting frequency and frequency resolution of the original broadband spectrum data, completing the signal detection and parameter extraction in complex electromagnetic environments.

[0072] The input raw data is normalized by maximum and minimum values, and the cost function is calculated as follows:

[0073]

[0074] Among them, X norm Here, X represents the normalized sample data, and X represents the original sample data. min X is the minimum value of the sample data. max This represents the maximum value of the sample data.

[0075] See Figure 3 The backbone network of the model uses a 50-layer deep residual network to extract four low-level features and one high-level feature.

[0076] See Figure 3 The feature selection module selects high-level features from the output using the following formula:

[0077] y=sigmod(fc2(RELU(fc1(F)))) (2)

[0078] F out =F·y (3)

[0079] Where fc1 and fc2 represent fully connected layer 1 and fully connected layer 2, F represents the input high-level features, f is the channel weight feature vector obtained by downsampling the high-level features, ReLU and sigmoid represent activation functions, · represents dot product, y represents the final weight vector, and F out This represents the selected feature.

[0080] See Figure 3 The global feature extraction module extracts contextual information from high-level data, using the following formula:

[0081]

[0082] y t = sigmod(fc4(RELU(fc3(f gap (5)

[0083]

[0084] Where F represents the high-level features of the input, conv represents the convolution operation, and f gap This represents the feature obtained by global average pooling of high-level features, which includes global contextual information. The dot (·) represents the dot product. This represents the final output of the global feature extraction module.

[0085] See Figure 3 The feature fusion module has three inputs: low-level features, high-level features, and global features. High-level features come from the output of the previous level, low-level features come from the output of the backbone network, and global features come from the global information extraction module. The feature fusion module fuses the two types of features through upsampling and dot multiplication, which enhances the target's noise suppression. The features fused in the first step are then fused again through convolution to obtain the final result. The specific fusion strategy is as follows:

[0086]

[0087]

[0088]

[0089]

[0090]

[0091]

[0092]

[0093]

[0094] in, Represents characteristics of high-level personnel. Represents low-level characteristics. Represents global features, upsample represents bilinear interpolation upsampling, t represents stage number, W represents the generated mask, and concat represents concatenation by channel.

[0095] The feature fusion module first combines low-level features The input is fed into the convolutional layer conv1 to obtain the same number of channels as the high-level features. Then, high-level characteristics The mask is obtained by bilinear interpolation upsampling after inputting it into the convolutional layer conv2. mask and After performing a dot product operation and then applying the ReLU activation function, the first type of fused feature from the high and low layers is obtained. The same method can be used to obtain another fused feature of high and low layer features. and global low-level fusion features Finally, the three preliminary fusion features are concatenated by channel and then passed through a convolutional layer to obtain the final fusion feature.

[0096] See Figure 3 The feature enhancement module uses multiplication and addition modules to further refine and enhance the feature maps that have passed through the feature selection and feature fusion modules. The specific operations are as follows:

[0097]

[0098]

[0099] Among them, f in The input features are represented by f, and conv represents the convolution operation. First, f is... in The input is fed into a 3x3 convolutional layer to obtain... Then The inputs are fed into two convolutional layers to obtain the mask W and the bias b, respectively. The obtained W and b are then compared... The final output f is obtained by performing the operation according to formula (16). out .

[0100] The loss function used when calculating the reconstruction error of the model is the cross-entropy loss. In addition to the final result, the loss function is also calculated for the intermediate layer upsampling results.

[0101]

[0102]

[0103] Among them, G ij S represents the true labels of points i and j. ij Represents the predicted label, l dom The loss representing the final predicted result. The auxiliary loss representing the intermediate result, γ i This represents the weighting coefficient.

[0104] See Figure 2 After obtaining the final significance result, the part with a length exceeding the threshold and continuously equal to 1 is regarded as the signal, and the start and end points of this part are considered as the start and end points of the signal.

[0105] See Figure 2 After obtaining the start and end points of the signal, the signal center frequency and signal bandwidth parameters are calculated based on the start frequency and frequency resolution of the original broadband spectrum data. The specific calculation method is as follows:

[0106] f start =f startin +Start index *f step (19)

[0107] f end =f startin +End indwx *f step (20)

[0108] f c =f start +(f end -f start ) / 2 (21)

[0109] bw = (End) index -Start index )*f step (twenty two)

[0110] Among them, f startin f represents the starting frequency of the original broadband spectrum data. step Indicates frequency resolution, Start index Indicates the start point of the signal, End index Indicates the termination point of the signal, f startf represents the starting frequency of the signal. end f represents the signal termination frequency. c bw represents the center frequency of the signal, and bw represents the bandwidth of the signal.

[0111] This embodiment provides a saliency-based broadband spectrum detection method, including saliency map transformation of a one-dimensional broadband spectrum dataset, converting the original data into a one-dimensional saliency vector; inputting the transformed saliency vector into a deep saliency detection network for supervised training to extract low-level detail features, high-level semantic features, and global context features of the signal; fusing the extracted features to obtain fused features; reconstructing the input saliency vector by linear interpolation upsampling of the fused features; calculating the cross-entropy loss between the reconstructed vector and the input saliency vector; and finally calculating parameters such as the number of signals, frequency, bandwidth, and signal-to-noise ratio based on the output results to complete broadband spectrum detection in complex electromagnetic environments. This invention achieves broadband spectrum detection, is independent of experts, has strong environmental adaptability, is easy to train, has a fast convergence speed, and can effectively improve the detection probability, thus providing a saliency-based broadband spectrum detection method.

[0112] The above are merely preferred embodiments of the invention and do not limit the patent scope of the invention. Any equivalent structural or procedural changes made using the contents of the invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the invention.

Claims

1. A saliency-based wideband spectrum detection method, characterized in that, The method comprises the following steps: S1: significant map conversion is performed on the wideband spectrum data based on signal start-stop information to obtain a binary one-dimensional significant vector as a label of the wideband spectrum data; S2: the wideband spectrum data and the corresponding label are input into a deep learning model to extract low-level detail features, high-level semantic features and global context features, and a fusion feature is constructed, in the step S2, the low-level detail features and the high-level semantic features are extracted, specifically including: four low-level features and one high-level feature are extracted by using a 50-layer deep residual network; in the step S2, an expression of the global context feature extraction is specifically: wherein, represents an input high-level feature, represents a convolution operation, represents a feature including global context information obtained by global average pooling of the high-level feature, represents a dot product, represents a final output of the global feature extraction module; S3: linear interpolation up-sampling is performed on the fusion feature to make the feature dimension same as the input vector, a SoftMax function is used to reconstruct the input vector, cross-entropy loss of the reconstructed vector and the input significant vector is calculated, and a significant result of the input data is obtained through inference of a deep saliency detection network model; S4: start and end points of the wideband spectrum are calculated according to the significant result, and a model parameter in the wideband spectrum is calculated according to a start frequency and a frequency resolution of the wideband spectrum data, and detection of the wideband spectrum is completed, in the step S4, the start and end points of the wideband spectrum are calculated according to the significant result, specifically: a part with a length exceeding a threshold and being continuous to 1 is regarded as a signal, and start and end points of the part are regarded as the start and end points of the signal.

2. The saliency-based wideband spectrum detection method of claim 1, wherein, In the step S2, the wideband spectrum data is also subjected to maximum and minimum value normalization, and an expression of a cost function is specifically: wherein, is the normalized sample data, is the original sample data, is the minimum value of the sample data, is the maximum value of the sample data.

3. The saliency-based wideband spectrum detection method of claim 1, wherein, In the step S2, the extracted high-level semantic features are also input into a channel selection module for selection, and spatial and channel attention mechanisms are used to enhance spatial regions and feature enhancement of the significant object high response; an expression of the channel selection module for selecting the high-level semantic features is specifically: wherein, and denote the fully connected layer 1 and the fully connected layer 2, denote the input high-level semantic feature, f is a channel weight feature vector obtained by downsampling the high-level feature, RELU and sigmod represent activation functions, denote dot product, y represents the final weight vector, denote the selected features.

4. The saliency-based wideband spectrum detection method of claim 1, wherein, In the step S2, a fusion strategy of the fusion feature is specifically: wherein, represents high-level semantic features, represents low-level detail features, represents global context features, represents bilinear interpolation up-sampling, t represents stage number, and W represents a generated mask, represents channel concatenation; the feature fusion module first inputs low-level features into a convolutional layer conv1 to obtain the same number of channels as high-level features , then inputs high-level features into a convolutional layer conv2 to perform bilinear interpolation up-sampling to obtain a mask , performs a point multiplication operation between the mask and , and then performs a RELU activation function to obtain a first fusion feature of high-level and low-level features ; the same method can be used to obtain another fusion feature of high-level and low-level features and a global low-level fusion feature , and finally, the three preliminary fusion features are concatenated by channels and then passed through a convolutional layer to obtain a final fusion feature .

5. The saliency-based wideband spectrum detection method of claim 1, wherein, After the step S2, the method further comprises: the fusion feature is perfected and enhanced by using a multiplication and addition module, and an expression thereof is specifically: wherein, represents an input feature, represents a convolution operation, first inputting into a 3*3 convolution layer to obtain , then inputting into two convolution layers to obtain a mask W and a bias b respectively, and performing operation on the obtained W and b to obtain the final output . .

6. The saliency-based wideband spectrum detection method of claim 1, wherein, In the step S3, when the cross-entropy loss of the reconstructed vector and the input significant vector is calculated, three intermediate upper sampling results are also subjected to loss function calculation, and an expression thereof is specifically: wherein, a true label representing a point i,j, a predicted label representing a point i,j, a loss representing a final prediction result, an auxiliary loss representing an intermediate result, a weight coefficient.

7. The saliency-based wideband spectrum detection method of claim 1, wherein, In the step S4, the model parameter in the wideband spectrum is calculated according to the start frequency and the frequency resolution of the wideband spectrum data to complete the detection step of the wideband spectrum, specifically including: wherein, represents a start frequency of the original wideband spectrum data, represents a frequency resolution, represents a start point of the signal, represents an end point of the signal, represents a start frequency of the signal, represents an end frequency of the signal, represents a center frequency of the signal, represents a bandwidth of the signal.

Images