Passive bistatic radar weak target detection system and method

Abstract

The invention relates to a passive bistatic radar weak target detection system and method, and belongs to the technical field of target detection. In order to solve the problem of low detection probability of a passive bistatic radar system in the prior art, the target detection system comprises an acquisition module used for acquiring a passive bistatic radar weak target PPI image data set; the preprocessing module is used for carrying out data loading and preprocessing on the data in the data set obtained by the obtaining module; and the improved YOLOv8 network model is used for detecting the data processed by the preprocessing module. By using the detection method of the system, weak signals and background noise can be distinguished more effectively, and the detection accuracy is improved.

Description

Technical Field

[0001] The present invention relates to a passive bistatic radar weak target detection system and method, in particular to a passive bistatic radar weak target detection system and method based on an improved YOLOv8 model, belonging to the technical field of target detection. Background Art

[0002] In the traditional single-base radar transceiver system, there is a problem of being easily discovered and attacked, which limits the ability to detect targets. In order to meet this challenge, the passive dual-base radar system came into being, which has the advantage of resisting the "four major threats". This system does not emit electromagnetic waves, but receives information carried by direct waves from the target radiation source or reflected waves or scattered waves formed after the external radiation source irradiates the target, extracts useful information after information processing, and completes the positioning of the target. Due to the characteristics of the external radar transceiver and its own silence, this system has the advantages of strong concealment, strong anti-interference ability, no pollution and low cost, and has received widespread attention in the past decade.

[0003] However, passive bistatic radar systems also face some challenges, the most prominent of which is the low probability of detection. Passive radars rely on existing external radiation sources for detection, such as broadcast, television signals, communication signals, etc., but the power and signal characteristics of these radiation sources are uncontrollable and are usually much weaker than the transmission signals of dedicated radars, resulting in weak received target echo signals and low signal-to-noise ratio. In addition, passive radars lack their own transmission sources, so they cannot accurately measure target distance and speed by synchronizing the transmission and reception signals like active radars, which increases the complexity of signal processing and reduces the signal-to-noise ratio. At the same time, the radiation sources used by passive radars may change dynamically in both space and time, increasing the uncertainty and complexity of detection, thereby affecting the signal quality and reducing the signal-to-noise ratio.

[0004] In passive radar, weak target detection is one of the key technologies and has important research significance. Deep learning has outstanding performance in this field. Its advantage is that it can automatically extract useful features from complex radar signals. Compared with traditional methods, deep learning can more effectively distinguish weak signals from background noise and improve the accuracy of detection. It is of great significance to improve the accuracy of passive radar weak target detection by combining deep learning algorithms and improving neural networks. Summary of the invention

[0005] The purpose of the present invention is to solve the problem of low detection probability of passive bistatic radar system in the prior art, and to provide a passive bistatic radar weak target detection system and method, which can more effectively distinguish weak signals from background noise and improve the detection accuracy.

[0006] In order to solve the above problems, the present application is implemented through the following technical solutions: A passive bistatic radar weak target detection system, which is special in that it includes: an acquisition module for acquiring a passive bistatic radar weak target PPI image data set; a preprocessing module for loading and preprocessing data in the data set obtained by the acquisition module; The improved YOLOv8 network model is used to detect the data processed by the preprocessing module.

[0007] Furthermore, the improved YOLOv8 network model includes Backbone, Head and DeteAux detection head, and the loss function adopts WIoU loss function; the Backbone is used to extract the features of the target detection map; the Head is used to perform weighted fusion of multi-scale features at different stages; The WIoU loss function: compares the predicted bounding box with the real bounding box to calculate the loss to guide the training of the entire network model; the DetectAux detection head: is used for target detection and integrates features at different levels for the final target detection output.

[0008] Furthermore, the Backbone includes a main module one, the output end of the main module one is provided with three main modules two, the output end of the last main module two is a BiFormer attention mechanism module, and the output end of the BiFormer attention mechanism module is an SPPF module; The main module 1 is composed of two convolutional layers and a C2f module, and the main module 2 is composed of a convolutional layer and a C2f module. The convolutional layer performs convolution operations by sliding the convolution kernel on the image to extract basic features of the image, such as edges, textures, etc.; The C2f module performs feature extraction and fusion operations on the input feature map, refines the information in the feature map, and enhances the expressiveness of the features; The SPPF module performs a spatial pyramid pooling operation, and obtains a feature map containing multi-scale information by performing pooling operations of different scales on the input feature map and fusing the results; The BiFormer attention mechanism module focuses on the key information in the feature map through its attention mechanism, can adaptively weight different areas in the feature map, enhance the network's ability to capture target features, and enable the network to effectively focus on important target areas and features.

[0009] Furthermore, the BiFormer attention mechanism module includes a BRA module and a 2-layer MLP module with an expansion ratio of e, wherein the BRA module divides the input feature map into non-overlapping regions so that each region contains a feature vector, which is specifically accomplished by reshaping the feature map; The MLP module performs further nonlinear transformation on the features output by the BRA module. The expansion ratio e of the MLP determines the proportional relationship between the number of neurons in its internal hidden layer and the number of neurons in the input layer. It can perform more complex mapping of features and enhance the expressiveness of features. The features processed by the MLP module will be fed back or fused with the output of the BRA module to obtain a feature map processed by the BiFormer attention mechanism module.

[0010] Furthermore, the feature graph reshaping is specifically as follows: query, key, and value tensors are derived using linear projection, involving projection weights of queries, keys, and values; region-to-region routing with directed graphs, that is, constructing a directed graph to find participating relationships, including deriving region-level queries and keys, deriving the adjacency matrix of the region-to-region affinity graph through matrix multiplication, pruning the association graph to obtain a routing index matrix, and applying fine-grained token-to-token attention, collecting key-value tensors and focusing attention on the collected key-value pairs.

[0011] Furthermore, the Head is improved based on the existing technology by replacing the neck network with a BiFPN network. The BiFPN network performs bidirectional cross-scale connection and weighted feature fusion, and introduces learnable weights to learn the importance of different input features. The BiFPN_Add2 and BiFPN_Add3 modules are used to repeatedly apply top-down and bottom-up fusion of multi-scale features from different stages of the Backbone. The BiFPN network enables the neural network to better understand and interpret multi-scale information, thereby improving the accuracy and efficiency of faint target detection.

[0012] Furthermore, the BiFPN network combines bidirectional cross-scale connections and fast normalized fusion: first, nodes with only one input edge are removed on the basis of FPN; second, when the original input and output nodes are at the same level, an extra edge is added between the original input and output nodes to fuse more features without increasing too much cost; third, each bidirectional path is regarded as a feature network layer, and the same layer is repeated multiple times to achieve more advanced feature fusion; fast normalized fusion solves the problem that different input features contribute unequally to the output features at different resolutions; Adding an extra weight to each input allows the network to learn the importance of each input feature: the formula is: , Where: is through each Then apply Relu to ensure that, is a small value to avoid numerical instability, is the input feature Furthermore, the two fusion features of the BiFPN level 6 are: , Where: The intermediate result obtained after specific processing is an intermediate feature map; : Indicates the convolution operation, which performs convolution on the result in the brackets to obtain the final ; : Weight value, used to weight different input feature maps; : The input feature map comes from a feature map of a specific level in the previous stage. “in” means “input” input; :right Adjust the size so that it fits Perform appropriate operations; : Input feature map; , used to avoid numerical instability; , Where: : The final output feature map, the result obtained after a series of operations, "out" means "output" output; : New weight value, used to weight different input feature maps; :right Resize it so that it can be operated with other feature maps; is another output feature map that may come from a different stage; Furthermore, the calculation method of the WIoU loss function is as follows: 1) Calculate the loU score of the predicted box and the true box: Let the predicted box , real frame , calculate the intersection area: , Compute the area of ​​the union: , Calculate the IoU score: ; 2) Calculate the area between the two boxes: Use the center points of the predicted box and the real box to calculate the distance between them, and use this distance as the maximum distance between the two boxes to calculate the area between the two boxes: Prediction box The coordinates of the center point are: , the coordinates of the center point of the real frame are , the distance between the two center points ,by Construct a circular region for the radius, the area between the two boxes is: ; 3) According to the area between the two boxes, the weight coefficient is calculated, which measures the relationship between the two boxes and is used to weight the loU score: The area between the two boxes is , the weight coefficient is calculated using the following formula: , Where: is a hyperparameter used to control the shape of the function; 4) By introducing the area and weight coefficient between boxes, WIOU can more accurately evaluate the target detection results and avoid the deviation problem of traditional IoU; calculate , Where: Indicates the number of object boxes, Indicates The coordinates of the object box, Indicates The coordinates of the ground-truth box of the object, Indicates The loU value between the object box and the real annotation box, Indicates the weight value.

[0013] In WIOU, the weight value of each object box depends on its degree of overlap with the true annotation box. Object boxes with greater overlap have higher weights, and object boxes with less overlap have lower weights. The dynamic non-monotonic focusing mechanism uses "outliers" to evaluate the quality of anchor boxes and provides a sensible gradient gain distribution strategy. This strategy not only reduces the competitiveness of high-quality anchor boxes, but also reduces the harmful gradients generated by low-quality examples. This allows WIoU to focus more on anchor boxes of ordinary quality, thereby improving the overall performance of the detector. Using WIoU for faint target detection improves the detection performance and training accuracy of target detection, and enhances the network's ability to locate the detected target.

[0014] Furthermore, the DeteAux detection head includes: a feature extraction layer: a series of convolutional layers are used to extract features and extract higher-level feature representations from the input feature map; these convolutional layers can gradually extract feature information at different levels and capture different feature attributes of the target; and convert the original feature map into a feature representation with more semantic information; Supervision module: The supervision module interacts with the extraction results of the feature extraction layer, compares the extracted features, the features fused by the fusion module and the real annotations, calculates the loss function, and calculates the gradient based on the loss function. The gradient is used for back propagation and adjusts the parameters of the detection head to improve the detection performance; Fusion module: Fusion of feature information of different scales or levels, fusion of features from different layers to obtain a more comprehensive target representation, improve detection accuracy and generalization ability, the output of the fusion module is the supervision module; Output module: Receives the fused feature information from the fusion module, performs convolution and classification to generate the passive radar weak target detection results as the final output.

[0015] Auxiliary detection heads are often placed in the middle layer or between multiple layers of the neural network and applied during training as a means of additional supervision. It helps to enhance the robustness and generalization ability of the network, especially in the early stages of training or when dealing with complex scenes. Auxiliary detection heads are usually able to capture feature information at different scales or levels, thereby more comprehensively parsing the targets in the passive radar image. With the auxiliary detection head, additional supervision signals can be introduced, which helps to better propagate gradients, speed up training and prevent gradient vanishing. In addition, it also provides additional regularization, reduces the risk of overfitting, and enhances the generalization ability of the passive bistatic radar weak target detection model.

[0016] Feature maps of different scales are used for target detection through the Detect module. The detection head contains a series of convolutional layers and deconvolutional layers to generate passive radar weak target detection results. In order to detect targets more accurately, the Detect module is improved and the DetectAux auxiliary detection head is used to enhance the generalization ability of the passive radar weak target detection model.

[0017] Another technical solution of the present invention is: adopting the above system to perform a passive bistatic radar weak target detection method, which is special in that it includes the following steps: step 1, obtaining a passive bistatic radar weak target PPI image data set; Step 2, preprocessing and dividing the passive dual-base radar weak target image data obtained in step 1; Step 3, inputting the training set in step 2 into the improved YOLOv8 network model for iterative training, and when the improved YOLOv8 network model iterates to a preset number of training rounds, the passive dual-base radar weak target detection is realized; Step 4: Input the data in the test set of step 2 into the YOLOv8 network model of step 3 for detection.

[0018] Furthermore, the preprocessing and partitioning in step 2 include data normalization, data enhancement and data partitioning, wherein the data normalization: normalizing the pixel values ​​of the image data so that the pixel values ​​are distributed in the range of [0, 1] or [-1, 1] to improve the training efficiency and stability of the model and avoid the adverse effects of too large or too small values ​​on the model training; the data enhancement: randomly rotating, flipping, scaling the normalized data, adding noise, increasing the diversity of the data and improving the generalization ability of the model; The data division is as follows: the enhanced data set is divided into a training set, a validation set and a test set. The training set is used for model training, the validation set is used to evaluate the performance of the model and adjust the hyperparameters during the training process, and the test set is used to finally evaluate the generalization ability of the model.

[0019] Furthermore, the step 3 specifically includes the following steps: step 3.1, forward propagation: the image in the training set of step 2 enters the forward propagation stage, and through multi-layer convolution and pooling operations, finally generates a prediction result, that is, the prediction of the input image by the YOLOv8 network model, that is, the location information of the target; Step 3.2, calculate the loss: Determine the performance of the model by comparing the difference between the model's prediction results and the actual label; the loss value is calculated by the loss function, which represents the gap between the prediction of the YOLOv8 network model and the actual situation. The smaller this value is, the closer the prediction of the YOLOv8 network model is to the actual situation; Step 3.3, back propagation: Calculate the gradient of the model parameters based on the loss value, and back propagate this gradient information back to the model to update the parameters; Step 3.4, parameter update: Use the gradient descent optimization algorithm to update the parameters of the YOLOv8 network model so that the YOLOv8 network model can be continuously optimized to improve the detection accuracy of weak targets; Repeat steps 3.1 to 3.4 until the preset number of training rounds is reached.

[0020] Furthermore, the step 4 is specifically as follows: first, after the forward propagation process of the YOLOv8 network model, the image features are extracted through multi-layer convolution and pooling operations; then, target detection is performed based on the extracted features to determine whether there is a faint target in the image and the location, size and other information of the target: The detection results are presented in the following forms: Target position: expressed in the form of coordinates in the image, with the coordinates of the upper left and lower right corners of the rectangular box representing the position of the target in the image; Confidence: indicates the degree of certainty of the YOLOv8 network model on the detection result, usually a value between 0 and 1.

[0021] This application converts one-dimensional radar data into two-dimensional image data, and then detects weak targets of passive radar through the automatic extraction function of image features through deep learning. The deep learning algorithm has a strong multi-target detection capability and can simultaneously detect and identify multiple targets in a complex signal environment, providing favorable conditions for the application of passive radar in multi-target scenarios. The detection method of this application can improve the accuracy of weak target detection, thereby addressing the problem of low detection probability in passive radar systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a structural schematic diagram of the detection system of the present invention; Figure 2 for Figure 1 Schematic diagram of the improved YOLOv8 network model structure; Figure 3 This is a schematic diagram of the structure of the DeteAux detection head; Figure 4 is a flow chart of the detection method of the present invention; Figure 5 for Figure 3 Flowchart of step 3 in ; Figure 6 Schematic diagram of the image dataset. DETAILED DESCRIPTION

[0023] The following is a specific implementation of the present invention with reference to the accompanying drawings to further illustrate the composition of the present invention.

[0024] In order to facilitate understanding of the embodiments of the present invention, abbreviations and key terms that may be involved in the embodiments of the present invention are first explained or defined.

[0025] Embodiment 1. A passive bistatic radar weak target detection system, such as Figure 1 As shown, it includes: an acquisition module, used to acquire a passive dual-base radar weak target PPI image data set; a preprocessing module, used to load and preprocess the data in the data set obtained by the acquisition module; The improved YOLOv8 network model is used to detect the data processed by the preprocessing module.

[0026] Furthermore, the improved YOLOv8 network model includes Backbone, Head and DeteAux detection heads, and the loss function adopts the WIoU loss function, such as Figure 2 As shown; The Backbone is used to extract the features of the target detection map; the Head is used to perform weighted fusion of multi-scale features at different stages; the WIoU loss function compares the predicted bounding box with the real bounding box to calculate the loss to guide the training of the entire network model; The DetectAux detection head is used for target detection and integrates features at different levels for the final target detection output.

[0027] Furthermore, the Backbone includes a main module one, the output end of the main module one is provided with three main modules two, the output end of the last main module two is a BiFormer attention mechanism module, and the output end of the BiFormer attention mechanism module is an SPPF module; Among them, the main module 1 is composed of two convolutional layers and a C2f module, and the main module 2 is composed of a convolutional layer and a C2f module. The convolutional layer performs convolution operations by sliding the convolution kernel on the image to extract the basic features of the image, such as edges, textures, etc. In this process, the convolutional layer will continuously reduce the spatial resolution of the image and increase the number of channels of the feature map to obtain feature representations at different levels; The C2f module performs feature extraction and fusion operations on the input feature map, refines the information in the feature map, and enhances the expressiveness of the features. Specifically, a series of convolution, normalization, and activation function operations are performed to refine the information in the feature map and enhance the expressiveness of the features. The SPPF module performs a spatial pyramid pooling operation, and obtains a feature map containing multi-scale information by performing pooling operations of different scales on the input feature map and fusing the results. It can process target information of different scales and help the network better detect targets of different sizes. The BiFormer attention mechanism module focuses on the key information in the feature map through its attention mechanism, and can adaptively weight different areas in the feature map to enhance the network's ability to capture target features, so that the network can effectively focus on important target areas and features; First, the input image is sent to the convolution layer, and the feature map output by the convolution layer is sent to the C2f module. The refined feature map output by the C2f module is sent to the SPPF module. The feature map output by the SPPF module is sent to the BiFormer attention mechanism module. The feature map output by the BiFormer attention mechanism module can be used for subsequent target detection tasks.

[0028] Furthermore, the BiFormer attention mechanism module includes a BRA module and a 2-layer MLP module with an expansion ratio of e, wherein the BRA module divides the input feature map into non-overlapping regions so that each region contains a feature vector, which is specifically accomplished by reshaping the feature map; The MLP module performs further nonlinear transformation on the features output by the BRA module. The expansion ratio e of the MLP determines the proportional relationship between the number of neurons in its internal hidden layer and the number of neurons in the input layer. It can perform more complex mapping of features and enhance the expressiveness of features. The features processed by the MLP module will be fed back or fused with the output of the BRA module to obtain a feature map processed by the BiFormer attention mechanism module.

[0029] Furthermore, the feature graph reshaping is specifically as follows: query, key, and value tensors are derived using linear projection, involving projection weights of queries, keys, and values; region-to-region routing with directed graphs, that is, constructing a directed graph to find participating relationships, including deriving region-level queries and keys, deriving the adjacency matrix of the region-to-region affinity graph through matrix multiplication, pruning the association graph to obtain a routing index matrix, and applying fine-grained token-to-token attention, collecting key-value tensors and focusing attention on the collected key-value pairs.

[0030] Furthermore, the specific steps of feature map reshaping are as follows: given a two-dimensional input feature map , where X represents the input two-dimensional feature map itself, which is a tensor that contains the feature information of the image after a series of convolution operations; H represents the height of the feature map, which represents the number of pixels or feature vectors in the vertical direction; W represents the width of the feature map, that is, the number of pixels or feature vectors in the horizontal direction; C represents the number of channels of the feature map, which represents the number of different features contained in each pixel or feature vector; R represents a set of real numbers; the feature map X is a three-dimensional tensor, in which each element is a real number; A. Divide the two-dimensional input feature map into Non-overlapping regions, so that each region contains This step is done by taking the feature vector Reshape To complete, as follows; A1. Calculate the area size: the height of each area in the height direction , the width of each area in the width direction ; A2. Reshape: The original feature map Seen as The plane composed of elements will This element is divided into regions, each containing elements; for the feature map , first change its shape from Adjust to , according to the original The elements of the array are rearranged into groups, each group corresponds to a small area divided out, a total of group; then adjust the shape to , get the form we need; B. Use linear projection to derive query, key, and value tensors: , Where: ; ; are the projection weights of query, key, and value respectively; C. Construct a directed graph to find the participating relationships, i.e. the regions that each given region should participate in: C1. Export region-level queries and keys. ; , through and The average value for each region was applied separately; C2. Pass With transposed The matrix multiplication between them derives the adjacency matrix of the region-to-region affinity graph ; , the adjacency matrix The entries in measure the degree to which two regions are semantically related; C3, by keeping only Prune the association graph using row-by-row connections The operator derives a routing index matrix , , No. The line contains the The most relevant areas indicators; C4. Apply fine-grained token-to-token attention.

[0031] Furthermore, the step C1 specifically includes the following steps: C1.1. First, clarify the form of input data: known query tensor and key tensors , their dimensions are in a certain form, here we assume that their dimensions are ,in Some spatial dimension representing the region, such as the number of regions on a two-dimensional plane, Represents the number of feature channels; C1.2. Computational region-level query : Input query tensor Divide by region: Assume that multiple regions are divided in the spatial dimension. For each region, calculate the average value of all elements in the region in the feature dimension. Specifically, for a specific area, the area corresponding to All The eigenvalues ​​are accumulated and then divided by the total number of elements in the region; this process is repeated for all regions to obtain the average eigenvalue of each region, which constitutes the region-level query ; C1.3. Calculate region-level keys : Similar to the computational region-level query, the key tensor Processing, divided by region , for each region, calculate the average value of all elements in the region on the feature dimension; repeat this operation for all regions to obtain the region-level key ; C1.4. Result analysis: The final regional query and zone-level keys All have dimensions ; The result is an averaged representation of each region in the feature space, which is no longer the feature of a single element, but the comprehensive feature of the entire region; This representation can be used to subsequently construct a directed graph to determine the participation relationship between regions. By comparing the region-level queries and keys of different regions, the similarity or correlation between them can be judged, thereby determining which other regions a region should interact with and participate in information transfer.

[0032] Furthermore, the step C4 specifically includes the following steps: C41. Collect key-value tensors, i.e. in ; is a key-value tensor; C42. Focus on the collected key-value pairs: for the area For each query token in , it will focus on the For index However, implementing this step efficiently is non-trivial, as these routing regions are expected to be scattered across the feature map, and modern GPUs rely on coalescing memory operations that load chunks of dozens of contiguous bytes at a time, as shown below: .

[0033] Using BRA as the basic building block, BiFormer is constructed, which includes the following steps: In stage In the first stage, overlapping patch embedding is used, and patch merging modules are used in the second to fourth stages to reduce the input spatial resolution while increasing the number of channels, and then Successive BiFormer blocks to transform features: In each BiFormer block, a 3×3 depthwise convolution is used at the beginning to implicitly encode relative position information; Then, a BRA module and a 2-layer MLP module with an expansion ratio of e are applied sequentially for cross-position relationship modeling and position-by-position embedding.

[0034] Furthermore, the Head is improved based on the existing technology by replacing the neck network with a BiFPN network. The BiFPN network performs bidirectional cross-scale connection and weighted feature fusion, and introduces learnable weights to learn the importance of different input features. The BiFPN_Add2 and BiFPN_Add3 modules are used to repeatedly apply top-down and bottom-up fusion of multi-scale features from different stages of the Backbone. The BiFPN network enables the neural network to better understand and interpret multi-scale information, thereby improving the accuracy and efficiency of faint target detection.

[0035] Furthermore, the BiFPN network combines bidirectional cross-scale connections and fast normalization fusion: First, we remove nodes with only one input edge based on FPN; Second, when the original input and output nodes are at the same level, an extra edge is added between the original input and output nodes to fuse more features without increasing too much cost; Third, each bidirectional path is regarded as a feature network layer, and the same layer is repeated multiple times to achieve more advanced feature fusion; fast normalization fusion solves the problem that different input features contribute unequally to the output features at different resolutions; Adding an extra weight to each input allows the network to learn the importance of each input feature: the formula is: , Where: is through each Then apply Relu to ensure that, is a small value to avoid numerical instability, are input features.

[0036] Furthermore, the Head includes a convolutional layer (Conv): used for feature extraction and dimension adjustment. In the given configuration, there are multiple convolutional layers, which adjust the dimensions of feature maps at different stages respectively; Upsampling layer (nn.Upsample): used to restore the feature map to a higher resolution so that it can be fused with features from other levels; C2f module: a feature fusion and enhancement module for further processing and fusing features at different levels; BiFPN_Add2 and BiFPN_Add3 modules: used for bidirectional cross-scale connections and weighted feature fusion, introducing learnable weights to learn the importance of different input features.

[0037] Furthermore, the two fusion features of the BiFPN level 6 are: , Where: The intermediate result obtained after specific processing is an intermediate feature map; : Indicates the convolution operation, which performs convolution on the result in the brackets to obtain the final ; : Weight value, used to weight different input feature maps; : The input feature map comes from a feature map of a specific level in the previous stage. “in” means “input” input; :right Adjust the size so that it fits Perform appropriate operations; : Input feature map; , used to avoid numerical instability; , Where: : The final output feature map, the result obtained after a series of operations, "out" means "output" output; : New weight value, used to weight different input feature maps; :right Resize it so that it can be operated with other feature maps; is another output feature map that may come from a different stage; is the intermediate feature of the sixth layer of the top-down path, is the output feature of the 6th layer of the bottom-up path; all other features are constructed in a similar way.

[0038] To further improve efficiency, depthwise separable convolution is used for feature fusion, and batch normalization and activation are added after each convolution; BiFPN enables the neural network to better understand and interpret multi-scale information through bidirectional connection and fast normalization fusion, thereby improving the accuracy and efficiency of weak target detection.

[0039] Furthermore, the calculation method of the WIoU loss function is as follows: 1) Calculate the loU score of the predicted box and the true box: Let the predicted box , real frame , Calculate the intersection area: , Compute the area of ​​the union: , Calculate the IoU score: ; 2) Calculate the area between the two boxes: Use the center points of the predicted box and the real box to calculate the distance between them, and use this distance as the maximum distance between the two boxes to calculate the area between the two boxes: Prediction box The coordinates of the center point are: , the coordinates of the center point of the real frame are , the distance between the two center points , by Construct a circular region for the radius, the area between the two boxes is: ; 3) According to the area between the two boxes, the weight coefficient is calculated, which measures the relationship between the two boxes and is used to weight the loU score: the area between the two boxes is , the weight coefficient is calculated using the following formula: , Where: is a hyperparameter used to control the shape of the function; 4) By introducing the area and weight coefficient between boxes, WIOU can more accurately evaluate the target detection results and avoid the deviation problem of traditional IoU; calculation , Where: Indicates the number of object boxes, Indicates The coordinates of the object box, Indicates The coordinates of the ground-truth box of the object, Indicates The loU value between the object box and the real annotation box, Indicates the weight value.

[0040] In WIOU, the weight value of each object box depends on its degree of overlap with the true annotation box. The object box with a greater degree of overlap has a higher weight, and the object box with a smaller degree of overlap has a lower weight. The dynamic non-monotonic focusing mechanism uses "outlier degree" to evaluate the quality of the anchor box and provides a wise gradient gain allocation strategy. This strategy not only reduces the competitiveness of high-quality anchor boxes, but also reduces the harmful gradients generated by low-quality examples. This enables WIoU to process anchor boxes of ordinary quality more centrally, thereby improving the overall performance of the detector. Using WIoU for faint target detection improves the detection performance and training accuracy of target detection, and enhances the network's ability to locate the detected target.

[0041] Furthermore, if Figure 3 As shown, the DeteAux detection head includes: feature extraction layer: feature extraction is performed through a series of convolutional layers to extract higher-level feature representations from the input feature map; these convolutional layers can gradually extract feature information at different levels to capture different feature attributes of the target; and convert the original feature map into a feature representation with more semantic information; Supervision module: The supervision module interacts with the extraction results of the feature extraction layer, compares the extracted features, the features fused by the fusion module and the real annotations, calculates the loss function, and calculates the gradient based on the loss function. The gradient is used for back propagation and adjusts the parameters of the detection head to improve the detection performance; Fusion module: Fusion of feature information of different scales or levels, fusion of features from different layers to obtain a more comprehensive target representation, improve detection accuracy and generalization ability, the output of the fusion module is the supervision module; Output module: Receives the fused feature information from the fusion module, performs convolution and classification to generate the passive radar weak target detection results as the final output.

[0042] The feature map enters the head from the output of the backbone: For the first stage: a convolution layer (Conv) adjusts the feature map dimension to 512; then the upsampling layer (nn.Upsample) restores the feature map to a higher resolution; then it is connected to a specific layer of the backbone (here is the 6th layer, corresponding to P4), and the feature fusion is performed through the BiFPN_Add2 module. The fused result enters the C2f module for further processing; For the second stage: the feature map dimension is adjusted to 256 through a convolution layer (Conv); the feature map is restored to a higher resolution through an upsampling layer (nn.Upsample); it is connected to a specific layer of the backbone (here is the 4th layer, corresponding to P3), and feature fusion is performed through the BiFPN_Add2 module. The fused result enters the C2f module for further processing; For the third stage: the feature map dimension is adjusted to 512 through a convolution layer (Conv), and then down-sampling is achieved through convolution operation; it is connected with the previous specific layers in the head (layers 14 and 6, corresponding to P4) and the specific layers of the backbone, and feature fusion is performed through the BiFPN_Add3 module. The fused results enter the C2f module for further processing; For the fourth stage: the feature map dimension is adjusted to 512 through a convolution layer (Conv), and then down-sampling is achieved through convolution operation; it is connected to the previous specific layer in the head (the 11th layer, corresponding to P5), and feature fusion is performed through the BiFPN_Add2 module. The fused result enters the C2f module for further processing; Finally, the features at different stages (18th, 21st, 24th, 14th, and 10th layers) are input into the DetectAux module for target detection, and the detection results are output.

[0043] Embodiment 2. A passive bistatic radar weak target detection method, the process of which is as follows Figure 4 As shown: the following steps are included: Step 1: Obtain a passive bistatic radar weak target PPI image dataset; Step 2: preprocessing and dividing the passive bistatic radar weak target image data obtained in step 1; Step 3: Input the training set in step 2 into the improved YOLOv8 network model for iterative training. When the improved YOLOv8 network model iterates to a preset number of training rounds, weak target detection of the passive dual-base radar is achieved. Step 4: Input the data in the test set of step 2 into the YOLOv8 network model of step 3 for detection.

[0044] Furthermore, the image data set in step 1 is usually presented in the form of a set of two-dimensional images, and the grayscale value or color value of each pixel reflects the corresponding radar signal feature, such as Figure 6 shown.

[0045] Furthermore, the preprocessing and partitioning in step 2 include data normalization, data enhancement and data partitioning, wherein the data normalization: normalizing the pixel values ​​of the image data so that the pixel values ​​are distributed in the range of [0, 1] or [-1, 1] to improve the training efficiency and stability of the model and avoid the adverse effects of too large or too small values ​​on the model training; the data enhancement: randomly rotating, flipping, scaling the normalized data, adding noise, increasing the diversity of the data and improving the generalization ability of the model; The data division is as follows: the enhanced data set is divided into a training set, a validation set and a test set. The training set is used for model training, the validation set is used to evaluate the performance of the model and adjust the hyperparameters during the training process, and the test set is used to finally evaluate the generalization ability of the model.

[0046] Further, the specific process of step 3 is as follows Figure 5 As shown, the following steps are included: Step 3.1, forward propagation: The image in the training set of step 2 enters the forward propagation stage, and through multi-layer convolution and pooling operations, the prediction result is finally generated, that is, the prediction of the YOLOv8 network model for the input image, that is, the location information of the target; Step 3.2, calculate the loss: Determine the performance of the model by comparing the difference between the model's prediction results and the actual label; the loss value is calculated by the loss function, which represents the gap between the prediction of the YOLOv8 network model and the actual situation. The smaller this value is, the closer the prediction of the YOLOv8 network model is to the actual situation; Step 3.3, back propagation: Calculate the gradient of the model parameters based on the loss value, and back propagate this gradient information back to the model to update the parameters; Step 3.4, parameter update: Use the gradient descent optimization algorithm to update the parameters of the YOLOv8 network model so that the YOLOv8 network model can be continuously optimized to improve the detection accuracy of weak targets; Steps 3.1 to 3.4 are repeated until a preset number of training rounds is reached. In this embodiment, the preset number of training rounds is 200.

[0047] Furthermore, the step 4 is specifically as follows: first, after the forward propagation process of the YOLOv8 network model, image features are extracted through multi-layer convolution and pooling operations; Then, target detection is performed based on the extracted features to determine whether there is a faint target in the image and the target's position, size and other information: The detection results are presented in the following forms: Target position: expressed in the form of coordinates in the image, with the upper left corner and lower right corner coordinates of the rectangular box representing the position of the target in the image; Confidence: Indicates the degree of certainty of the YOLOv8 network model about the detection result, usually a value between 0 and 1.

[0048] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A passive bistatic radar weak target detection system, characterized in that: include Acquisition module, used to acquire passive bistatic radar weak target PPI image data set; A preprocessing module, used for loading and preprocessing the data in the data set obtained by the acquisition module; The improved YOLOv8 network model is used to detect the data processed by the preprocessing module.

2. A passive bistatic radar weak target detection system according to claim 1, characterized in that: The improved YOLOv8 network model includes Backbone, Head and DeteAux detection heads, and the loss function adopts WIoU loss function; The Backbone is used to extract features of the target detection graph; The Head is used to perform weighted fusion of multi-scale features at different stages; The WIoU loss function: The predicted bounding box is compared with the true bounding box to calculate the loss to guide the training of the entire network model; The DetectAux detection head is used for target detection and integrates features at different levels for the final target detection output.

3. A passive bistatic radar weak target detection system according to claim 2, characterized in that: The Backbone includes a main module one, the output end of the main module one is provided with three main modules two, the output end of the last main module two is a BiFormer attention mechanism module, and the output end of the BiFormer attention mechanism module is an SPPF module; The main module 1 is composed of two convolutional layers and a C2f module, and the main module 2 is composed of a convolutional layer and a C2f module. The convolutional layer performs convolution operations by sliding the convolution kernel on the image to extract basic features of the image, such as edges, textures, etc.; The C2f module performs feature extraction and fusion operations on the input feature map, refines the information in the feature map, and enhances the expressiveness of the features; The SPPF module performs a spatial pyramid pooling operation, and obtains a feature map containing multi-scale information by performing pooling operations of different scales on the input feature map and fusing the results; The BiFormer attention mechanism module focuses on the key information in the feature map through its attention mechanism, can adaptively weight different areas in the feature map, enhance the network's ability to capture target features, and enable the network to effectively focus on important target areas and features.

4. A passive bistatic radar weak target detection system according to claim 3, characterized in that: The BiFormer attention mechanism module includes a BRA module and a 2-layer MLP module with an expansion ratio of e, wherein the BRA module divides the input feature map into non-overlapping regions so that each region contains a feature vector, which is specifically achieved by reshaping the feature map; The MLP module performs further nonlinear transformation on the features output by the BRA module. The expansion ratio e of the MLP determines the proportional relationship between the number of neurons in its internal hidden layer and the number of neurons in the input layer. It can perform more complex mapping of features and enhance the expressiveness of features. The features processed by the MLP module will be fed back or fused with the output of the BRA module to obtain a feature map processed by the BiFormer attention mechanism module.

5. A passive bistatic radar weak target detection system according to claim 4, characterized in that: The feature map reshaping is specifically as follows: Use linear projection to derive query, key, and value tensors, involving the projection weights of query, key, and value; Region-to-region routing with directed graphs, i.e., constructing a directed graph to find participating relations, includes deriving region-level queries and keys, deriving the adjacency matrix of the region-to-region affinity graph by matrix multiplication, pruning the association graph to obtain the routing index matrix, and applying fine-grained token-to-token attention, collecting key-value tensors and focusing attention on the collected key-value pairs.

6. A passive bistatic radar weak target detection system according to any one of claims 1 to 5, characterized in that: The Head is improved based on the existing technology, and the neck network is replaced by the BiFPN network. The BiFPN network performs bidirectional cross-scale connection and weighted feature fusion, and introduces learnable weights to learn the importance of different input features. Use BiFPN_Add2 and BiFPN_Add3 modules to repeatedly apply top-down and bottom-up fusion of multi-scale features from different stages of Backbone; The BiFPN network combines bidirectional cross-scale connections and fast normalization fusion: First, we remove nodes with only one input edge based on FPN; Second, when the original input and output nodes are at the same level, an extra edge is added between the original input and output nodes to fuse more features without increasing too much cost; Third, each bidirectional path is regarded as a feature network layer, and the same layer is repeated multiple times to achieve more advanced feature fusion; Fast normalization fusion solves the problem that different input features contribute unequally to the output features at different resolutions; Adding an extra weight to each input allows the network to learn the importance of each input feature: the formula is: , Where: is through each Then apply Relu to ensure that, , are input features.

7. A passive bistatic radar weak target detection system according to claim 6, characterized in that: The calculation method of the WIoU loss function is as follows: 1) Calculate the loU score of the predicted box and the real box: Set prediction box , real frame , Calculate the intersection area: , Compute the area of ​​the union: , Calculate the IoU score: ; 2) Calculate the area between the two boxes: Use the center points of the predicted box and the real box to calculate the distance between them, and use this distance as the maximum distance between the two boxes to calculate the area between the two boxes: Prediction box The coordinates of the center point are: , the coordinates of the center point of the real frame are , the distance between the two center points , by Construct a circular region for the radius, the area between the two boxes is: ; 3) Calculate the weight coefficient based on the area between the two boxes: The area between the two boxes is , the weight coefficient is calculated using the following formula: , Where: is a hyperparameter used to control the shape of the function; 4) By introducing the area between the boxes and the weight coefficient, calculate , Where: Indicates the number of object boxes, Indicates The coordinates of the object box, Indicates The coordinates of the ground-truth box of the object, Indicates The loU value between the object box and the real annotation box, Indicates the weight value.

8. A passive bistatic radar weak target detection system according to claim 7, characterized in that: The DeteAux detection head includes: Feature extraction layer: Feature extraction is performed through a series of convolutional layers to extract higher-level feature representations from the input feature map; these convolutional layers can gradually extract feature information at different levels to capture different characteristic attributes of the target; and transform the original feature map into a feature representation with more semantic information; Supervision module: The supervision module interacts with the extraction results of the feature extraction layer, compares the extracted features, the features fused by the fusion module and the real annotations, calculates the loss function, and calculates the gradient based on the loss function. The gradient is used for back propagation and adjusts the parameters of the detection head to improve the detection performance; Fusion module: Fusion of feature information of different scales or levels, fusion of features from different layers to obtain a more comprehensive target representation, improve detection accuracy and generalization ability, the output of the fusion module is the supervision module; Output module: Receives the fused feature information from the fusion module, performs convolution and classification to generate the passive radar weak target detection results as the final output.

9. A method for detecting weak targets using a passive bistatic radar using the system of any one of claims 1 to 8, characterized in that: The following steps are involved: Step 1: Obtain a passive bistatic radar weak target PPI image dataset; Step 2: preprocessing and dividing the passive bistatic radar weak target image data obtained in step 1; Step 3: Input the training set in step 2 into the improved YOLOv8 network model for iterative training. When the improved YOLOv8 network model iterates to a preset number of training rounds, weak target detection of the passive dual-base radar is achieved. Step 4: Input the data in the test set of step 2 into the YOLOv8 network model of step 3 for detection.

10. A passive bistatic radar weak target detection method according to claim 9, characterized in that: The preprocessing and division in step 2 include data normalization, data enhancement and data division, wherein the data normalization: normalizing the pixel values ​​of the image data so that the pixel values ​​are distributed in the range of [0, 1] or [-1, 1]; The data enhancement: randomly rotating, flipping, scaling the image and adding noise to the normalized data; The data division: the enhanced data set is divided into a training set, a validation set and a test set, the training set is used for model training, the validation set is used to evaluate the performance of the model and adjust the hyperparameters during the training process, and the test set is used to finally evaluate the generalization ability of the model; The step 3 specifically includes the following steps: Step 3.1, forward propagation: The image in the training set of step 2 enters the forward propagation stage, and through multi-layer convolution and pooling operations, the prediction result is finally generated, that is, the prediction of the YOLOv8 network model for the input image, that is, the location information of the target; Step 3.2, calculate the loss: Determine the performance of the model by comparing the difference between the model's prediction results and the actual label; the loss value is calculated by the loss function, which represents the gap between the prediction of the YOLOv8 network model and the actual situation. The smaller this value is, the closer the prediction of the YOLOv8 network model is to the actual situation; Step 3.3, back propagation: Calculate the gradient of the model parameters based on the loss value, and back propagate this gradient information back to the model to update the parameters; Step 3.4, parameter update: Use the gradient descent optimization algorithm to update the parameters of the YOLOv8 network model so that the YOLOv8 network model can be continuously optimized to improve the detection accuracy of weak targets; Repeat steps 3.1 to 3.4 until the preset number of training rounds is reached; The step 4 is specifically as follows: first, after the forward propagation process of the YOLOv8 network model, image features are extracted through multi-layer convolution and pooling operations; Then, target detection is performed based on the extracted features to determine whether there is a faint target in the image and the target's location, size, and other information: The test results are presented in the following form: Target position: expressed in the form of coordinates in the image, with the coordinates of the upper left corner and lower right corner of the rectangular box representing the position of the target in the image; Confidence: Indicates the degree of certainty of the YOLOv8 network model about the detection result, usually a value between 0 and 1.

Images