A Shortwave Frequency Hopping Signal Sorting Method Based on Improved YOLOv5

Abstract

The present invention provides a shortwave frequency-hopping signal sorting method based on improved YOLOv5. The method includes: First, a gray-scale time-frequency map of the frequency-hopping signal is generated through a time-frequency analysis method as the input of the YOLOv5 object detection network. Second, a CA mechanism is added to the backbone network to capture cross-channel information and position-sensitive information on the premise of ensuring the flexibility and light weight of the model, so as to realize the real-time detection and accurate positioning of the frequency-hopping signal. Third, Soft-SIoU_NMS is used to replace NMS to ensure that the frequency-hopping signal will not be ignored due to low confidence when frequency collisions occur. Finally, NWD is used to replace the intersection over union (IoU) metric in NMS and the regression loss function during small target detection, improving the detection accuracy for fast frequency-hopping signals. Implementing a shortwave frequency-hopping signal sorting method based on improved YOLOv5 according to the present invention has the advantages of fast model convergence, strong robustness, and an mAP reaching 99.5%. It can accurately sort out various frequency-hopping signals under low signal-to-noise ratios, and the sorting rate exceeds 95% at a signal-to-noise ratio of 0 dB.

Description

Technical Field

[0001] The present invention belongs to the field of communication technologies, and particularly relates to a short-wave frequency-hopping signal sorting method based on improved YOLOv5. Background Art

[0002] Frequency hopping (FH) refers to a communication method in which the carrier frequency of a communication signal is continuously changed under the control of a pseudo-random code. This characteristic greatly enhances the anti-interference ability and security reliability of communication. Its applications are becoming more and more widespread both in the military and civilian fields. In military electronic warfare, frequency-hopping communication technology has become a countermeasure for intercepting and cracking the real-time strategic information of the enemy due to its superior anti-interference, anti-interception, anti-fading, and strong multi-address networking capabilities, and also plays a crucial role in ensuring the secure and stable transmission of our information. In the civilian field, mobile communication GSM services, Bluetooth transmissions, and home radio frequencies all adopt frequency-hopping communication technology, which greatly improves the communication quality, can also flexibly allocate frequencies, and can well avoid various interferences in an environment where spectrum resources are scarce. Therefore, frequency-hopping communication technology is the main communication means in the field of short-wave communication signal countermeasures.

[0003] Frequency-hopping signal reconnaissance is a short-wave radio reconnaissance technology that mainly explores and captures frequency-hopping signals in short-wave channels and tests electromagnetic parameters such as their signal types, modulation methods, and carrier frequencies. Frequency-hopping signal reconnaissance is mainly divided into three parts: detection and extraction of frequency-hopping signals, parameter estimation, and sorting. Among them, the purpose of frequency-hopping signal sorting is to separate the intercepted multi-network frequency-hopping signals from each other, and then perform subsequent processing such as parameter estimation on the target frequency-hopping signals, which is a key step in the research of frequency-hopping signal reconnaissance. In practical applications, the complex electromagnetic environment in the short-wave frequency band seriously affects the reconnaissance research of frequency-hopping signals. If the signals of each frequency-hopping network cannot be sorted out from the frequency-domain overlapping multi-network frequency-hopping signals, it is very difficult to convert parameters such as the frequency-hopping period, hopping moment, and frequency-hopping frequency into corresponding communication intelligence, and it is even impossible to obtain communication information. Therefore, it is of great significance to study the frequency-hopping network sorting method. Summary of the Invention

[0004] The purpose of the present invention is to sort out various frequency-hopping signals from a large number of frequency-hopping networks and achieve the sorting of frequency-hopping signals under low signal-to-noise ratios.

[0005] In a first aspect, a short-wave frequency-hopping signal sorting method based on improved YOLOv5 includes:

[0006] S1: Construct a multi-frequency-hopping network mathematical model, and generate a gray-scale time-frequency diagram of the frequency-hopping signal as the input of the YOLOv5 target detection network through a time-frequency analysis method.

[0007] S2: Incorporate the CA mechanism into the backbone network of YOLOv5 to capture cross-channel information and position-sensitive information while ensuring the flexibility and light weight of the model, and achieve real-time detection and accurate positioning of the frequency-hopping signal.

[0008] S3: Using Soft-SIoU_NMS to replace NMS ensures that the frequency-hopping signal will not be ignored due to low confidence when frequency collisions occur.

[0009] S4: Using NWD to replace the Intersection of Union (IoU) metric in NMS and the regression loss function during small target detection improves the detection accuracy for fast frequency-hopping signals.

[0010] Preferably, the step S1 specifically includes:

[0011] Within a certain observation time, after the frequency-hopping signal detection and extraction, the channel environment only contains frequency-hopping signals and strong noise. Construct a mathematical model of multi-station frequency-hopping signals and Gaussian white noise. According to the non-stationary characteristics of the frequency-hopping signal, use STFT to generate a grayscale time-frequency diagram. According to the networking rules of frequency-hopping signals and the parameter characteristics of fast and slow frequency-hopping signals in the complex short-wave channel environment, perform pairwise mixing to construct a time-frequency diagram of multi-station frequency-hopping signals as the sorting data set.

[0012] Preferably, the step S2 specifically includes:

[0013] The backbone network is the feature extraction module of the YOLOv5 model. The Spatial Pyramid Pooling Fast (SPPF) module under the backbone network improves the operation efficiency by serially connecting multiple Maxpool layers while ensuring the same calculation results. Add the CA mechanism after the SPPF module, decompose the Squeeze-and-Excitation Network (SENet) into two parallel one-dimensional feature encoding processes, aggregate features along two spatial directions, reduce the loss of position information caused by two-dimensional global pooling, and more effectively integrate spatial coordinate information into the generated attention map. Through the above operations, not only can cross-channel information be captured, but also direction and position-aware features can be captured, enabling the model to more accurately locate and identify objects.

[0014] Preferably, the step S3 specifically includes:

[0015] During the execution of Non-maximum Suppression (NMS), the confidence of adjacent anchors will be forced to zero. If frequency collisions occur in a certain hop of two frequency-hopping signals, that is, in the overlapping area, it will lead to the failure of hop detection for the frequency-hopping signal, reducing the Mean Average Precision (mAP) of the algorithm. At the same time, using IoU as an indicator to filter out duplicate anchors in NMS cannot accurately reflect the degree of overlap between two anchors.

[0016] YOLOv5 uses CIoU_Loss to calculate the bounding box loss. The aspect ratio of its bounding box describes a relative value, which is somewhat ambiguous and does not consider the balance problem of easy and difficult samples.

[0017] To address the deficiencies in the original YOLOv5 algorithm above, the SIoU indicator is used to replace the IoU indicator to improve the calculation accuracy of NMS, and SIoU_Loss is used to replace CIoU_Loss to improve the model training speed and inference accuracy.

[0018] Preferably, the step S4 specifically includes:

[0019] Existing object detection network models commonly use IoU as a measurement method for the loss function, but it is quite sensitive to the position deviation of small targets, which will lead to a decline in detection performance in small target detection based on anchors. During the observation time, the time-frequency characteristics of fast frequency-hopping signals are manifested as small targets with dense short cycles. If an algorithm based on IoU measurement is used for sorting such frequency-hopping signals, it will lead to a serious decline in mAP and sorting accuracy. Therefore, by judging the size of the bounding box, if the bounding box is larger than 16×16, the IoU measurement is used; if the bounding box is smaller than 16×16, the Normalized Wasserstein Distance (NWD) is introduced to replace the IoU measurement for small target detection as an evaluation indicator for bounding box detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0021] Figure 1 It is a schematic flowchart of a shortwave frequency-hopping sorting method based on improved YOLOv5 provided by an embodiment of the present invention;

[0022] Figure 2 is the network structure diagram of implementing the YOLOv5 model in the present invention;

[0023] Figure 3 is the structure diagram of implementing the CA mechanism in the present invention;

[0024] Figure 4 is the backbone structure diagram of YOLOv5 after implementing the CA mechanism in the present invention;

[0025] Figure 5 is the calculation diagram of implementing Angle cost in the present invention

[0026] Figure 6 is the 8 types of short-wave frequency-hopping signals implemented in the present invention;

[0027] Figure 7 is the comparison of bounding box regression loss implemented in the present invention;

[0028] Figure 8 is the comparison of mAP@0.5 implemented in the present invention;

[0029] Figure 9 is the comparison of classification loss implemented in the present invention;

[0030] Figure 10 is the comparison of the frequency-hopping signal sorting performance of different algorithms implemented in the present invention. Detailed implementation manners

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0032] S1: Construct a multi-frequency-hopping network station mathematical model, and generate a gray time-frequency diagram of the frequency-hopping signal as the input of the YOLOv5 object detection network through time-frequency analysis methods.

[0033] S2: Add the CA mechanism to the backbone network of YOLOv5 to capture cross-channel information and position-sensitive information while ensuring the flexibility and light weight of the model, and realize the real-time detection and accurate positioning of frequency-hopping signals.

[0034] S3: Use Soft-SIoU_NMS to replace NMS to ensure that frequency-hopping signals will not be ignored due to low confidence when frequency collisions occur.

[0035] S4: Use NWD to replace the Intersection of Union (IoU) metric in NMS and the regression loss function for small target detection, improving the detection accuracy for fast frequency-hopping signals.

[0036] Specifically, step S1 includes:

[0037] Construct a multi-hop frequency-hopping network station mathematical model, and generate a gray time-frequency diagram of the frequency-hopping signal through time-frequency analysis as the input of the YOLOv5 object detection network.

[0038] Multi-hop frequency-hopping network station mathematical model: Within a certain observation time, assume that the short-wave multi-network station frequency-hopping signal mathematical model after frequency-hopping signal detection and extraction is as follows:

[0039]

[0040] In the formula, r(t) is the short-wave multi-network station frequency-hopping signal, s k (t) represents the signal of the k-th frequency-hopping network station within the observation time T, and n(t) is the additive white Gaussian noise with a mean of 0 and a variance of σ 2 .

[0041] Assume that the frequency-hopping period of the k-th frequency-hopping signal is The rectangular window is And it satisfies

[0042]

[0043] The received model expression of the k-th frequency-hopping signal is

[0044]

[0045] Among them, A k is the signal amplitude, θ k represents the initial phase, f n k represents the hopping frequency of the n-th time slot in the k-th frequency-hopping signal.

[0046] Since the frequency of the frequency-hopping signal changes irregularly over time and belongs to a typical non-stationary signal, it is difficult to comprehensively analyze such signals in both the time domain and the frequency domain. Therefore, in order to more effectively detect the frequency-hopping signal, time-frequency analysis technology is often used for analysis and processing. Therefore, the present invention performs a short-time Fourier transform (STFT) on the signal collected within the observation time, with a window length of 2048 and a window function type of Hamming window.

[0047] The STFT of the short-wave multi-network station frequency-hopping signal is defined as

[0048]

[0049] where k is the number of discrete points, is the phase transformation, h(kΔt - mΔt) is the window function, Δt is the sampling interval, and STFT r (t, f) is the discrete two-dimensional distribution of the short-wave multi-network station frequency-hopping signal r(t) in time and frequency. Its basic idea is to divide the time-domain signal into multiple sub-time-domain signal segments and perform windowing processing, and calculate the Fourier transform of each sub-time-domain signal segment respectively. Therefore, STFT is a linear transformation and will not generate cross terms during the time-frequency processing, and the computational complexity is relatively low. To meet the actual engineering requirements, the present invention uses STFT as the time-frequency analysis tool and performs grayscale processing to generate the signal grayscale time-frequency diagram.

[0050] Figure 2 is the network structure diagram of the YOLOv5 model implemented in the present invention.

[0051] As the network model with the strongest engineering applicability in the one-stage object detection algorithm, YOLOv5 modifies the backbone and Neck networks on the basis of YOLOv3 and adds some improvement techniques on the basis of YOLOv4, so that its speed and accuracy have been greatly improved in performance. The network structure of the YOLOv5 model is as Figure 2As shown in the figure, it mainly consists of an input end, a backbone, a Neck, and a prediction. During the model training phase, the input end uses the Mosaic data augmentation of YOLOv4 and presents two improvement techniques: adaptive anchor calculation can adaptively calculate the best anchor according to the dataset name during training, improving the detection accuracy. To address the problem that a large amount of information redundancy caused by different sizes of black edges at both ends in the original scaling method affects the inference speed of the algorithm, adaptive image scaling can adaptively add the least amount of black edges to the scaled image, improving the model inference speed. The backbone uses the latest CSP-Darknet53. In version v6.0, the Focus module is replaced with a 6×6 Conv layer to make GPU device training more efficient. The CSP_X module is replaced with the C3_X module to reduce duplicate gradient information while reducing the computational amount and shortening the model inference time. SPPF (Spatial PyramidPooling Fast) improves the operation efficiency by serially connecting multiple Maxpool layers while ensuring the same calculation result. To further extract and fuse the image feature information output by the backbone, YOLOv5 adopts the FPN+PAN structure in the Neck network to achieve multi-scale feature fusion of high-level semantic information and low-level detail features of the image, and introduces the C3_X_F structure in the PAN structure to strengthen the network feature fusion ability. Prediction, as the output end of the YOLOv5 model, mainly performs loss function and NMS calculations, and then outputs feature maps of three scales of large, medium, and small, detecting targets of different scales in the original image. For different detection algorithms, the number of branches at the output end is also different. CIoU_Loss is used as the regression loss function and IoU_NMS is used as non-maximum suppression in YOLOv5.

[0052] Further, step S2 includes:

[0053] With the continuous research and verification of lightweight networks, the channel attention mechanism (Squeeze-and-Excitation Network, SENet) can bring significant performance improvement to the model. However, SENet only focuses on constructing the interdependence between channels and ignores the spatial location information. The Convolutional Block Attention Module (CBAM) introduces a spatial attention module (Spartial Attention Module, SAM) on the basis of SENet, which pays attention to both channel features and spatial features. However, large-scale convolutional kernels can only capture local information and ignore the long-range dependence problem. The Coordinate Attention (CA) mechanism decomposes SENet into two parallel one-dimensional feature encoding processes, aggregates features along two spatial directions, reduces the loss of location information caused by two-dimensional global pooling, and more effectively integrates spatial coordinate information into the generated attention map. Specifically, the CA mechanism uses two one-dimensional global pooling operations to aggregate the input features along the vertical and horizontal directions into two separate directional feature maps, and then simplifies and encodes these two feature maps with embedded direction-specific information into two perception maps, each of which captures the long-range correlation of the input feature map along one spatial direction. Therefore, the location information can be preserved in the generated attention map, and then the two attention maps are applied to the input feature map through the Hadamard product to enhance the representation of useful features and reduce useless information.

[0054] The CA mechanism has the following advantages:

[0055] (1) Flexibility and light weight, which can be easily inserted into the backbone and core modules of lightweight networks to strengthen the representation of location information and enhance feature expression.

[0056] (2) It can not only capture cross-channel information, but also capture direction-aware and location-aware information, helping the model to more accurately locate and identify the target of interest.

[0057] (3) It significantly improves the performance of tasks such as dense prediction (fast hopping frequency signal detection).

[0058] (4) It can avoid a large amount of computational overhead when obtaining large-area attention information in lightweight networks.

[0059] Figure 3 This is the structural diagram of the CA mechanism implemented in the present invention. The algorithm process is as follows:

[0060] (1) To avoid compressing all spatial information into the channels, global average pooling (GAP) is not used. To capture long-range spatial interactions with precise location information, GAP is decomposed.

[0061]

[0062] Among them, represents the output feature map of the c-th channel with height h, represents the output feature map of the c-th channel with width w, where W and H are the width and height of the input feature map respectively, and x c is the input feature map of the c-th channel. The input feature map Input with size C×H×W is pooled along the X direction and Y direction respectively to generate feature maps with sizes C×H×1 and C×1×W.

[0063] (2) The feature maps z h and z w generated in the X and Y directions in process (1) are subjected to a Concat operation to achieve feature aggregation, and the concatenated feature map is subjected to a convolution operation and an activation operation to output a feature map with a reduced number of channels

[0064] f = δ(F1([z h , z w )) (7)

[0065] In the formula, δ is a non-linear activation function, and F1 is a 1×1 convolution function to achieve dimensionality reduction.

[0066] (3) Along the spatial dimension, f is then split into and Then, 1×1 convolutions are respectively used for dimensionality increase operations, and combined with the sigmoid activation function to obtain the final attention vectors and

[0067] g h = σ(F h (f h )) (8)

[0068] g w = σ(F w (f w )) (9)

[0069] The output of the CA mechanism can be written as:

[0070]

[0071] Figure 4This is the backbone structure diagram of YOLOv5 after the CA mechanism is added in the implementation of the present invention. Secondly, the specific steps of step S3 include:

[0072] In the target detection network, there are usually many high-confidence anchors around the real target. At this time, in order to remove duplicate anchors, non-maximum suppression (NMS) filters out low-confidence anchors with an overlap ratio greater than the threshold through an iterative-traversal-elimination process, achieving the purpose of having one and only one detection result for each object. The specific steps of the NMS algorithm:

[0073] (1) Sort the anchors from high to low according to the confidence level;

[0074] (2) Take out the first anchor in the anchor list. At this time, this anchor is the anchor with the highest confidence level. Calculate the IoU between this anchor and all the remaining anchors. After all the calculations are completed, add this anchor to another list;

[0075] (3) Discard the anchors corresponding to the IoU greater than the IoU threshold (it is considered that the same target is detected);

[0076] (4) Repeat steps (2) and (3) until the anchor list is empty.

[0077] Since NMS will force the confidence levels of adjacent anchors to zero during execution, if a frequency collision occurs in a certain hop of two frequency-hopping signals, that is, in the overlapping area, it will lead to the failure of detecting the hop of the frequency-hopping signal, reducing the mAP of the algorithm. At the same time, using IoU as the index to filter out duplicate anchors in NMS cannot accurately reflect the degree of overlap between two anchors.

[0078] In addition, the loss function in YOLOv5 during training includes bounding box loss, confidence loss, and classification loss. At present, CIoU_Loss is mainly used to calculate the bounding box loss. The aspect ratio of its bounding box describes a relative value, which is somewhat ambiguous and does not consider the balance problem of easy and difficult samples. For the above IoU index and bounding box loss problems in NMS, the present invention uses the SIoU

[24] index to replace the IoU index to improve the calculation accuracy of NMS, and uses SIoU_Loss to replace CIoU_Loss to improve the model training speed and inference accuracy.

[0079] Figure 5 This is the calculation diagram of Angle cost in the implementation of the present invention. SIoU_Loss is mainly composed of 4 cost functions:

[0080] (1) Angular cost

[0081] By adding an angular perception LF component, make predictions on the X or Y axis and then continuously approach along the relevant axis. The convergence process will first attempt to minimize α. If then minimize The calculation process is as Figure 5 shown.

[0082] The angular cost function is defined as:

[0083]

[0084] (2) Distance cost

[0085] The definition of the distance cost incorporates the angular cost defined above:

[0086]

[0087] γ = 2 - Λ(12d). Combining the definitions of the angular cost and the distance cost, it can be obtained that when α → 0, Δ is smaller. Conversely, when Δ is larger.

[0088] (3) Shape cost

[0089] The shape cost is defined as:

[0090]

[0091] Equations (13b) and (13c) determine the magnitude of the shape cost. The θ value controls how much attention the shape cost requires, and each dataset corresponds to a unique θ value. If θ = 1, the shape will be immediately optimized, thus affecting the free movement of the shape.

[0092] (4) IoU cost

[0093] Define the IoU cost:

[0094]

[0095] Combining the above cost functions, the regression loss function of SIoU is:

[0096]

[0097] To solve the problem of missed detection caused by frequency collisions in the hop of frequency hopping signals in NMS, the present invention uses Soft-NMS to replace NMS and improves it by replacing the IoU metric with SIoU during the process, constructing the Soft-SIoU_NMS algorithm. The algorithm execution process of Soft-SIoU_NMS does not simply delete the anchor with SIoU greater than the threshold, but reduces the confidence of its anchor. By introducing the confidence reset function f(SIoU(M, b i ))), this function will attenuate the confidence of adjacent boxes overlapping with the bounding box M. The more an anchor overlaps with M, the more severe the attenuation of its confidence. SIoU and its threshold are used as the basis for reducing the confidence. After traversing the anchor list, those with a confidence less than the threshold are filtered out. Therefore, its goal is to reduce the anchor confidence score.

[0098] IoU_NMS sets the confidence of all anchors with IoU greater than the threshold to 0

[0099]

[0100] The confidence reset function of Soft-SIoU_NMS is defined as:

[0101]

[0102] In equations (16), (17a) and (17b), b i is the anchor to be processed and D is the storage for the final anchor, s i is the updated confidence of b i and N t represents the threshold. Therefore, Soft-SIoU_NMS does not require retraining the original model, has a small computational load, is easy to implement, and can more efficiently select anchors with the same computational complexity as IoU_NMS, thereby improving the mAP and the sorting accuracy of frequency hopping signals.

[0103] Finally, the specific steps of step S4 include:

[0104] Existing target detection network models all use the Intersection of Union (IoU) as the metric of the loss function. However, it is quite sensitive to the position deviation of small targets, which will lead to unsatisfactory detection performance in anchor-based small target detection algorithms. In short-wave channels, multi-network station hopping signals with different hopping rates are mixed with each other. During the observation time, the time-frequency characteristics of fast-hopping signals are manifested as small targets with dense short cycles. If an algorithm based on IoU metric is used for such hopping signal sorting, the mAP and sorting accuracy rate will decrease. Therefore, in the present invention, by judging the size of the bounding box, if the bounding box is larger than 16×16, the IoU metric is used. If the bounding box is smaller than 16×16, the Normalized Wasserstein Distance (NWD) is introduced to replace the IoU metric for small target detection as the evaluation index for bounding box detection.

[0105] The sensitivity of the IoU metric to the position of the bounding box stems from its scale transformation of discrete position deviation objects, which will cause the small position deviation of small targets to have an anchor flip, resulting in similar positive and negative sample features and difficult network convergence. In addition, it is difficult to find a good IoU threshold for the sensitivity of IoU to provide positive and negative samples with high confidence. The NWD models the bounding box as a two-dimensional Gaussian distribution. Even if there is no intersection, the similarity between bounding boxes can still be calculated through the corresponding Gaussian distribution. This NWD metric can be easily embedded into any anchor-based label assignment, NMS, and loss function instead of the IoU metric.

[0106] The definition of NWD is as follows:

[0107] (1) Gaussian distribution modeling of the bounding box

[0108] In a low signal-to-noise ratio environment, time-frequency noise will be densely distributed around the hopping signal. Therefore, the time-frequency characteristics of the hopping signal often have background pixels that are not strictly rectangular strips. In the bounding box of the hopping signal, foreground pixels and background pixels are concentrated in the center and boundary of the bounding box respectively. To better describe the weights of different pixels in the bounding box, the bounding box can be modeled as a two-dimensional Gaussian distribution, where the weights of the bounding box pixels decrease from the center position to the boundary position. For a horizontal bounding box R = (c x , c y , w, h), where (c x , c y ) represents the center coordinates, and w and h represent the width and height respectively. Its inscribed ellipse expression is:

[0109]

[0110] In the formula, (μ x , μy ) is the central coordinate of the inscribed ellipse, σ x and σ y are the semi - axis lengths along the x - and y - axes respectively. Thus, μ x = c x , μ y = c y , The probability density function of the two - dimensional Gaussian distribution is as follows:

[0111]

[0112] In the formula, c represents the coordinates (x, y) of the two - dimensional Gaussian distribution, μ is the mean vector, and Σ represents the covariance matrix. If it satisfies

[0113] (c - μ) T Σ -1 (c - μ)= 1 (20)

[0114] Then equation (18) will be the density contour line of the two - dimensional Gaussian distribution. Therefore, R=(c x , c y , w, h) can be modeled as the two - dimensional Gaussian distribution N(μ, Σ)

[0115]

[0116] Therefore, the similarity between bounding box A and bounding box B can be converted into the distribution distance between two Gaussian distributions.

[0117] (2) Normalized Gaussian Wasserstein distance

[0118] Use the Wasserstein distance to calculate the distribution distance. For two two - dimensional Gaussian distributions m1 = N(μ1, Σ1) and m2 = N(μ2, Σ2), the second - order Wasserstein distance between m1 and m2 is defined as:

[0119]

[0120] Simplified to:

[0121]

[0122] In equation (22), Tr(·) represents taking the trace of the matrix, and ||·|| F is the Frobenius norm.

[0123] Therefore, from equation (23), it can be obtained that bounding box A=(cx a , cy a , w a , h a) of the Gaussian distribution N a and the bounding box B=(cx b ,cy b ,w b ,h b ) of the Gaussian distribution N b The Wasserstein distance between them is:

[0124]

[0125] Since W2 2 (N a ,N b ) as a distance metric is not in the range of [0,1], it cannot be directly used as a similarity index. Therefore, exponential normalization is used to obtain the NWD metric:

[0126]

[0127] where ε is a constant closely related to the dataset. Generally, setting it to the average size of the dataset target can achieve the best performance. Therefore, in the present invention, ε = 13.5.

[0128] (3) NMS and regression loss based on the NWD metric

[0129] Regarding the sensitivity of the IoU metric to small targets (fast frequency hopping signals), the IoU value of the prediction box will be lower than the threshold N t , resulting in the problems of misjudgment and missed detection of frequency hopping signals. Using the NWD metric to replace the IoU metric as a new index for NMS can overcome the scale sensitivity problem. Additionally, by designing a loss function based on the NWD metric:

[0130] L NWD = 1 - NWD(N p ,N gt ) (26)

[0131] In the formula, N p is the Gaussian distribution model of the prediction box P, and N gt is the Gaussian distribution model of the ground truth box GT. Based on the NWD metric, gradients can be provided even when |P∩G| = 0 or |P∩G| = P or G, overcoming the sensitivity problem of the IoU metric to the position deviation of small targets.

[0132] After introducing the NWD metric, the content described in S3 is updated, and the loss function SIoU is updated to SNMD, and Equation (15) is updated to:

[0133]

[0134] Meanwhile, Soft-SIoU_NMS is updated to Soft-SNWD_NMS, and equations (17a) and (17b) are updated to:

[0135]

[0136] The experimental environment configuration of the present invention is shown in the following table:

[0137]

[0138] Figure 6 They are 8 types of shortwave frequency-hopping signals implemented in the present invention. Based on the networking rules of frequency-hopping signals and the parameter characteristics of fast and slow frequency-hopping signals in the complex shortwave channel environment, taking 8 types of frequency-hopping signals as the basis, FH1, FH2, FH3, FH4, FH6, FH7, FH8, and FH9 are used as class names respectively. As Figure 6 shown, when constructing the time-frequency diagrams of multi-network station frequency-hopping signals by pairwise mixing, they are used as the sorting data set. The observation duration is 10 ms, the sampling rate is 10 MHz, and the signal-to-noise ratio is in the range of 0 - 15 dB. As shown in the experimental environment configuration table. The input image resolution is 400×500, the horizontal axis is time, and the vertical axis is frequency. The number of samples of each mixed frequency-hopping signal time-frequency diagram at each SNR is 400, and a total of 12,400 time-frequency diagrams are randomly divided into 9,919 training sets, 1,241 validation sets, and 1,240 test sets according to the ratio of 8:1:1. The frequency-hopping frequency set of each time-frequency diagram is constantly changing to meet the time-frequency diversity of multi-network station frequency-hopping signals. During training, the image resolution is 640×640, the optimizer is SGD, the learning rate is 0.01, the attenuation is 0.0005, the batch_size is 32, and the epoch is 300.

[0139] To verify the performance of the improved model, the present invention uses precision (P), recall (R), average precision (AP), mAP, and the frequency-hopping signal sorting rate (SR) to evaluate the detection performance and signal sorting performance of the model.

[0140] The calculation formulas of the above indicators are as follows:

[0141]

[0142]

[0143] Among them, TP represents the number of correctly detected frequency-hopping signals, FP represents the number of noises identified as frequency-hopping signals, and FN represents the number of undetected frequency-hopping signals.

[0144]

[0145] Among them, N is the number of label categories, AP is the integral of P with respect to R in the range [0, 1], that is, the area under the P-R curve. The larger the AP value, the higher the model accuracy. And mAP represents the mean of AP for each target category and is the main evaluation index of detection performance.

[0146]

[0147] Among them, K i is the number of labels of the i-th frequency-hopping signal, represents the number of correctly recognized i-th frequency-hopping signals on the corresponding frequency set, represents the number of recognized i-th frequency-hopping signals, represents the sorting rate of the i-th frequency-hopping signal, and SR represents the sorting rate of the model for frequency-hopping signals.

[0148] Figure 7 This is the implementation of the bounding box regression loss in the present invention. One of the improved YOLOv5 algorithms proposed in the present invention: After replacing the IoU metric with the NWD metric and using SNWD as the bounding box regression loss, all parameters tend to be stable at 250 epochs, and the bounding box loss value drops to about 0.005. It not only solves the problems of easy-hard sample imbalance and aspect ratio ambiguity of the original network model using the CIoU loss function, but also improves mAP. Figure 8 This is the implementation of the mAP@0.5 comparison in the present invention. The comparison curve of mAP@0.5 between the improved YOLOv5 and the original YOLOv5, where mAP@0.5 represents the mAP value when the threshold is 0.5. The mAP of the improved YOLOv5 has been stable at 15 epochs, and the mAP reaches 99.5% under the premise that the threshold is 0.5, which is 4.7% higher than the mAP of the original YOLOv5.

[0149] Figure 9 This is the implementation of the classification loss comparison in the present invention. The comparison of the classification loss between the algorithm proposed in the present invention and the traditional CNN network-based algorithm is introduced. The classification loss based on the CNN algorithm shows a large fluctuation phenomenon during the iteration process and only tends to be stable after 100 epochs, while the classification loss of the improved YOLOv5 is smoother, and the loss value is lower than 0.01 at 10 epochs and has converged stably at 30 epochs.

[0150] Figure 10It is a comparison of the performance of frequency-hopping signal sorting for different algorithms implemented in the present invention. The present invention presents a comparison of the sorting rate performance of the algorithm proposed in the present invention and different traditional frequency-hopping signal sorting algorithms at different SNRs. The sorting rate of the frequency-hopping signal sorting algorithm based on K-Means clustering is only 22.6% at 0 dB, and it is severely affected by noise at low SNRs. The sorting rates of the frequency-hopping signals based on the CNN algorithm and the BP neural network algorithm are 46.5% and 52.7% respectively at 0 dB and both increase with the increase of SNR. The sorting rate approaches 90% at SNR = 5 dB. The improved YOLOv5 algorithm proposed in the present invention has a sorting rate of 96.2% at SNR = 0 dB, and as SNR increases, the sorting rate reaches a maximum of 98.6% at SNR = 4 dB. Therefore, the algorithm proposed in the present invention has better anti-noise performance compared with other existing algorithms and has the highest sorting rate of frequency-hopping signals at low SNRs.

[0151] The ablation experiment table implemented in the present invention is shown in the following table:

[0152]

[0153] The present invention trains each improvement idea under the same experimental conditions respectively. The experimental results are shown in the ablation experiment table. Among them, "×" represents the improvement idea not used in the network model, "√" represents the use of this improvement idea, mAP@0.5:0.95 is the mAP value between the thresholds of 0.5 and 0.95, Inference(ms) represents the model inference time, and NMS(ms) represents the non-maximum suppression time. It can be seen from the ablation experiment table that the improved network 1 introduces the CA mechanism into the original YOLOv5 network, the mAP@0.5 is increased by 2.0%, and the mAP@0.5:0.95 is increased by 6.4%, but the inference time rises to 199.7ms and the NMS is still large; the improved network 2 introduces SNWD as the loss function on the basis of the improved network 1, which improves the detection accuracy and convergence speed of small target signals. Compared with the mAP@0.5 of the improved network 1, it is increased by 1.6%, and the mAP@0.5:0.95 is increased by 3.2%. The inference time rises by 4.0ms and the NMS is reduced to 1.3ms; the improved network 3 introduces Soft_SNWD_NMS on the basis of the improved network 2, which solves the problem of reduced accuracy caused by frequency collision of hopping signals. Compared with the mAP@0.5 of the improved network 2, it is increased by 1.1%, and the mAP@0.5:0.95 is increased by 1.3%. The inference time rises by 1.6ms and the NMS is still 1.3ms. To sum up, the improved algorithm proposed by the present invention is the improved network 3. Compared with the original YOLOv5 network, the mAP@0.5 is increased by 4.7%, the mAP@0.5:0.95 is increased by 10.9%, the NMS is decreased from 2.5ms to 1.3ms, which is only 52% of the original network. Although the inference time rises by 19.2ms, it still meets the requirements of model lightweight and engineering real-time performance.

Claims

1. A method for sorting short-wave frequency-hopping signals based on improved YOLOv5, characterized in that, The method includes: S1: Construct a multi-hop frequency hopping network station mathematical model, and generate a gray time-frequency diagram of the frequency hopping signal as the input of the YOLOv5 object detection network through time-frequency analysis methods; S2: Add a CA mechanism to the backbone network of YOLOv5 to capture cross-channel information and position-sensitive information on the premise of ensuring the flexibility and light weight of the model, and achieve real-time detection and accurate positioning of frequency hopping signals; S3: Use Soft-SIoU_NMS to replace NMS to ensure that frequency hopping signals will not be ignored due to low confidence when frequency collisions occur; The confidence reset function of Soft-SIoU_NMS is defined as: In Formula (17a) and Formula (17b) b i is the anchor to be processed and D stores the final anchor, s i is b i the updated confidence N t represents the threshold value; S4: Use NWD to replace the IoU metric in NMS and the regression loss function during small target detection to improve the detection accuracy of fast frequency hopping signals; The specific steps of S4 include: By judging the size of the bounding box, if the bounding box is larger than 16×16, the IoU metric is used; if the bounding box is smaller than 16×16, NWD is introduced to replace the IoU metric for small target detection as the evaluation index for bounding box detection.

2. The method for sorting short-wave frequency-hopping signals based on improved YOLOv5 according to claim 1, characterized in that, The specific steps of S1 include: Within a certain observation time, after the channel environment is detected and extracted by the frequency hopping signal, only the frequency hopping signal and strong noise exist. Construct a mathematical model of the multi-network station frequency hopping signal and Gaussian white noise; According to the non-stationary characteristics of the frequency hopping signal, use STFT to generate a gray time-frequency diagram; According to the networking rules of the frequency hopping signal in the complex short-wave channel environment and the parameter characteristics of the fast and slow frequency hopping signals, perform pairwise mixing to construct a time-frequency diagram of the multi-network station frequency hopping signal as a sorting data set.

3. The method for sorting short-wave frequency-hopping signals based on improved YOLOv5 according to claim 1, characterized in that, The specific steps of S2 include: The backbone network is the feature extraction module of the YOLOv5 model. The SPPF module under the backbone network improves the operation efficiency by serially connecting multiple Maxpool layers on the premise of ensuring the same calculation result; Add a CA mechanism after the SPPF module, decompose the channel attention mechanism SENet into two parallel one-dimensional feature encoding processes, aggregate features along two spatial directions, reduce the loss of position information caused by two-dimensional global pooling, and more effectively integrate spatial coordinate information into the generated attention map; Through the above operations, not only can cross-channel information be captured, but also direction and position-aware features can be captured, so that the model can more accurately locate and identify objects.

Images