Method for detecting low-confidence small targets in radar echo based on hybrid architecture

Abstract

The application belongs to the technical field of radar signal processing, and particularly relates to a radar echo low-confidence small target detection method based on a hybrid architecture. The method first performs spectrum symmetry shifting and dimension reorganization on radar echo data to generate a five-dimensional tensor [B, T, C, H, W] containing time sequence features. Then, the tensor is input into a detection model composed of a Hourglass3D module and a YOLOv8 network. The Hourglass3D module extracts multi-scale space-time features through a structure of three-dimensional convolution downsampling-bottleneck layer-three-dimensional transposed convolution upsampling, and realizes feature fusion through a skip connection. Finally, target detection is completed through the backbone network, neck network and decoupling detection head of the YOLOv8 network. Through joint extraction of space-time features and enhancement of small target features, the detection accuracy and positioning precision of low signal-to-noise ratio small targets in radar echo are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing technology, specifically relating to a method for detecting small targets with low confidence in radar echoes based on a hybrid architecture. Background Technology

[0002] Small target detection in radar echoes is an important research direction in the field of target recognition. Currently, deep learning-based target detection methods (such as YOLO and Faster R-CNN) are widely used in radar signal processing. These methods mainly extract spatial features through two-dimensional convolutional neural networks and combine them with post-processing algorithms such as non-maximum suppression (NMS) to optimize the detection results. Furthermore, for processing time-series signals, some studies use 3D CNN or LSTM networks to extract features from radar echoes in consecutive frames to improve the detection performance of small targets. In existing technologies, single-stage detectors such as YOLOv8 are preferred due to their efficiency, but their default architecture is mainly designed for optical images and has limited adaptability to low signal-to-noise ratio targets in radar echoes.

[0003] However, existing methods still have significant shortcomings in small target detection using radar echoes: on the one hand, traditional 2D convolutions struggle to effectively capture the spatiotemporal correlation of radar echoes, leading to a high false negative rate for low-confidence targets; on the other hand, the backbone structure of the standard YOLOv8 network is insufficient for feature extraction of small targets, and the fixed threshold NMS in the post-processing stage is ill-suited to the dynamic changes in radar target density. Furthermore, existing methods typically do not optimize for the spectral characteristics of radar data, resulting in decreased performance in detecting small targets with low signal-to-noise ratios. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting small targets with low confidence in radar echoes based on a hybrid architecture, in order to solve the problem of how to effectively improve the detection performance and accuracy of small targets in radar echo images that are easily missed due to weak feature information and low model detection confidence.

[0005] The present invention achieves the above objectives through the following technical solutions:

[0006] Firstly, this invention proposes a method for detecting small targets with low confidence in radar echoes based on a hybrid architecture. The method is used to detect targets with a confidence level below a preset threshold, and includes:

[0007] The radar echo data to be tested is subjected to spectrum shifting and dimension reorganization to generate a five-dimensional tensor with temporal characteristics;

[0008] The five-dimensional tensor is input into the detection model to obtain prediction results including the location of the target object prediction box;

[0009] The detection model includes a cascaded Hourglass3D module and a YOLOv8 network. The Hourglass3D module extracts and fuses multi-scale spatiotemporal features from the five-dimensional tensor through three-dimensional convolution and skip connections, and outputs the feature map to the front end of the backbone network of the YOLOv8 network to achieve multi-scale spatiotemporal feature fusion and target classification and localization.

[0010] Furthermore, the dimensional reorganization includes:

[0011] The storage path of the image dataset obtained by symmetrically shifting the spectrum of radar echo data is parsed and decoded into a NumPy array;

[0012] Convert the image data of the NumPy array into RGB channels and store them sequentially.

[0013] Perform tensor dimension rearrangement on the stored data, adjusting the channel dimension from [H,W,C] to [C,H,W];

[0014] Add batch and time dimensions to the tensor to form a five-dimensional tensor with a [B,T,C,H,W] structure;

[0015] Where B is the batch size, T is the time step, C is the number of channels, H is the height, and W is the width.

[0016] Furthermore, the Hourglass3D module includes:

[0017] The encoder consists of multiple levels of three-dimensional convolutional downsampling units, each level containing a three-dimensional convolutional layer, a normalization layer, and an activation function;

[0018] The bottleneck layer uses 3D convolution to keep the number of feature map channels constant.

[0019] The decoder consists of multiple levels of 3D transposed convolutional upsampling units, each of which includes a 3D transposed convolutional layer, a batch normalization layer, and a ReLU activation function layer.

[0020] The skip connection structure fuses the output feature maps of each layer of the encoder with the input feature maps of the corresponding layers of the decoder.

[0021] Furthermore, the three-dimensional convolutional downsampling unit uses a convolutional kernel with a stride greater than 1 to achieve feature map downsampling; the three-dimensional transposed convolutional upsampling unit uses a convolutional kernel with a stride greater than 1 to achieve feature map upsampling; the skip connection structure adds the output of the encoder at level n to the input of the decoder at level (4-n) element by element, where n∈{1,2,3,4}.

[0022] Furthermore, in the Hourglass3D module, the three-dimensional convolutional downsampling unit satisfies the channel number doubling rule: the number of output channels C at the k-th level... k = C0×2 k Where C0 is the initial number of channels, k∈{1,2,3,4}; the three-dimensional transposed convolutional upsampling unit satisfies the channel number reduction rule: the number of output channels at the m-th level C m = C4 / 2 (5-m) Where C4 is the number of bottleneck layer channels, m∈{1,2,3,4}.

[0023] Furthermore, the YOLOv8 network includes:

[0024] The backbone network includes at least two initial convolutional layers and multi-level convolutional modules. It receives the feature maps output by the Hourglass3D module and feeds them into the first two convolutional layers. After the multi-level convolutional modules extract hierarchical representations of local and global features, it outputs multi-scale feature maps.

[0025] The neck network consists of a feature pyramid structure, a path aggregation structure, and a dynamic convolutional layer. The feature pyramid structure fuses deep high-semantic features with shallow high-resolution features through upsampling. The path aggregation structure achieves cross-level feature interaction through downsampling. The dynamic convolutional layer adjusts the feature fusion method according to the adaptive weights of the input features and uses upsampling and downsampling to perform multi-scale fusion of the feature maps output by the backbone network.

[0026] The detection head includes a classification branch and a regression branch: the classification branch outputs the category confidence of the target object through a convolutional layer, and the regression branch outputs the prediction parameters of the target object's bounding box through a separate convolutional layer.

[0027] Furthermore, the method also includes:

[0028] Set dynamic confidence threshold and IoU threshold, and adaptively adjust the threshold parameters according to the density characteristics of the target detection scene;

[0029] An improved nonmaximum suppression algorithm is used to process the predicted bounding boxes, including: feature map interpolation enhancement for predicted bounding boxes with an area smaller than a preset threshold;

[0030] The output includes the detection results containing the target object category, confidence score, and bounding box coordinates. The bounding box parameters include the center point coordinates (x, y) and the width and height dimensions (w, h).

[0031] Furthermore, the method also includes training the cascaded Hourglass3D module and the YOLOv8 network, including:

[0032] The preprocessed five-dimensional tensor [B,T,C,H,W] is input into the cascaded Hourglass3D module and YOLOv8 network;

[0033] The Hourglass3D module processes the input data through four layers of three-dimensional convolutional downsampling units, and the output feature map size is halved layer by layer while the number of channels is doubled layer by layer.

[0034] The feature map output from the last layer of the Hourglass3D module is input into the first two convolutional layers of the YOLOv8 network.

[0035] In the backbone of the YOLOv8 network, the input feature map is processed through multi-level convolutional modules to output feature maps at three scales.

[0036] Feature maps at three scales are input into the neck network for feature fusion.

[0037] The decoupled detection head is used to process the fused feature map, with the classification branch outputting the class probability and the regression branch outputting the bounding box coordinates.

[0038] Calculate the classification loss and regression loss, update the network parameters through backpropagation, and obtain the detection model.

[0039] Secondly, this invention proposes a low-confidence small target detection system based on a hybrid architecture for radar echoes, used to implement the small target detection method described above. The system includes:

[0040] The preprocessing module is used to perform spectrum shifting and dimension reorganization on the radar echo data to be tested, generating a five-dimensional tensor with temporal characteristics;

[0041] The target detection module is used to input the five-dimensional tensor into the detection model to obtain prediction results including the position of the target object prediction box;

[0042] The detection model includes a cascaded Hourglass3D module and a YOLOv8 network. The Hourglass3D module extracts and fuses multi-scale spatiotemporal features from the five-dimensional tensor through three-dimensional convolution and skip connections, and outputs the feature map to the front end of the backbone network of the YOLOv8 network to achieve multi-scale spatiotemporal feature fusion and target classification and localization.

[0043] Furthermore, the system also includes:

[0044] The optimization processing module is used to set dynamic confidence thresholds and IoU thresholds, and adaptively adjusts the threshold parameters according to the density characteristics of the target detection scene; it uses an improved non-maximum suppression algorithm to process the predicted bounding boxes, including: performing feature map interpolation enhancement on predicted bounding boxes with an area smaller than a preset threshold; and outputting detection results containing the target category, confidence level, and bounding box coordinates, where the bounding box parameters include the center point coordinates (x, y) and width and height dimensions (w, h).

[0045] The beneficial effects of this invention are as follows:

[0046] 1. This invention effectively improves the detection performance of small targets in radar echoes by constructing a hybrid architecture through cascading the Hourglass3D module and the YOLOv8 network. The Hourglass3D module uses three-dimensional convolutional operations to process temporal features, and achieves multi-scale spatiotemporal feature extraction and fusion through an encoder-bottleneck layer-decoder structure and skip connections, enhancing the feature representation capability of small targets. The multi-scale feature extraction and FPN+PAN feature fusion mechanism of the YOLOv8 network further optimize the detection effect of targets at different scales.

[0047] 2. This invention fully preserves the spatiotemporal characteristics of radar echoes through a five-dimensional tensor [B,T,C,H,W] data organization, enabling the network to fully utilize the correlation between consecutive frames. The decoupled detection head design achieves independent optimization of classification and localization tasks, improving the accuracy of small target detection. The dynamic non-maximum suppression strategy and small target compensation mechanism effectively reduce missed detections and false detections. Attached Figure Description

[0048] Figure 1 A flowchart illustrating a low-confidence small target detection method based on a hybrid architecture in radar echoes provided in this application embodiment;

[0049] Figure 2 Another flowchart of a low-confidence small target detection method based on a hybrid architecture in radar echoes provided in this application embodiment;

[0050] Figure 3 This is a schematic diagram of the structure of the Hourglass3D module provided in the embodiments of this application;

[0051] Figure 4 This is a comparison chart of the Precision-Recall curves of the YOLOv8 model and the HG3DCN-YOLOv8 model in the simulation analysis section of this application.

[0052] Figure 5 A comparison of the Precision-Confidence curves of the YOLOv8 model and the HG3DCN-YOLOv8 model in the simulation analysis section of this application;

[0053] Figure 6 This is a schematic diagram of manually annotated original labels in the dataset of the simulation analysis section of this application;

[0054] Figure 7 This is a schematic diagram of the target detection results of the YOLOv8 model in the simulation analysis section of this application;

[0055] Figure 8 This is a schematic diagram of the target detection results performed on the HG3DCN-YOLOv8 model in the simulation analysis section of this application. Detailed Implementation

[0056] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0057] It is worth noting that low-confidence small targets in radar echoes refer to targets that are difficult to detect accurately by traditional methods due to their small cross-sectional area and low signal-to-noise ratio, such as drones or stealth targets. Existing methods based on two-dimensional convolutional neural networks (such as YOLOv8) have significant drawbacks: First, they are unable to effectively capture the spatiotemporal correlation features of radar echoes, leading to missed detections of dynamic targets; second, the network downsampling process easily loses weak features of small targets; third, post-processing algorithms with fixed thresholds cannot adapt to dynamic changes in target density; and finally, the lack of optimized design for radar spectral characteristics further reduces detection performance.

[0058] Example 1

[0059] like Figure 1-3 As shown, to address the aforementioned deficiencies, a specific embodiment of this application proposes a method for detecting small targets with low confidence in radar echoes based on a hybrid architecture. The method is used to detect targets with confidence levels below a preset threshold. The method includes: performing spectrum shifting and dimension reorganization on the radar echo data to be tested to generate a five-dimensional tensor with temporal characteristics; inputting the five-dimensional tensor into a detection model to obtain prediction results including the target's bounding box position; wherein the detection model includes a cascaded Hourglass3D module and a YOLOv8 network. The Hourglass3D module extracts and fuses multi-scale spatiotemporal features from the five-dimensional tensor through three-dimensional convolution and skip connections, and outputs a feature map to the front end of the YOLOv8 network's backbone network to achieve multi-scale spatiotemporal feature fusion and target classification and localization.

[0060] In this application, please refer to Figure 2The Hourglass3D module is an improved design based on the classic Hourglass network architecture, used to extract spatiotemporal joint features of radar echoes. This module employs an encoder-bottleneck-decoder structure. The encoder progressively compresses the spatial dimension and expands the number of channels through multi-level 3D convolutional downsampling. The bottleneck layer maintains the number of channels in the feature map through 3D convolution to capture local contextual information. The decoder restores the spatial resolution through 3D transposed convolutional upsampling and adds it element-wise with the feature map of the corresponding layer of the encoder via skip connections, achieving multi-scale feature fusion.

[0061] Specifically, the dimensionality reorganization includes: using the fftshift tool in MATLAB to symmetrically shift the radar echo data spectrum to obtain an image dataset; parsing the storage path of the image dataset and decoding it into a NumPy array. The NumPy array is a structured numerical container generated by the Python scientific computing library NumPy, used to store the preprocessed radar echo data; converting the image data in the NumPy array into RGB channel sequential storage; performing tensor dimensionality reorganization on the stored data, moving the channel dimension [C] from the last dimension [H,W,C] format to the first dimension [C,H,W] format; then adding two additional dimensions: batch and time; finally, the tensor shape changes from [C, H, W] to [B, T, C, H, W], corresponding to batch size (number of samples processed in parallel during training), time step (number of consecutive radar frames), number of channels (corresponding to radar frequency band / polarization information), height, and width, respectively; H / W represents the spatial resolution.

[0062] It is understood that the five-dimensional tensor [B, T, C, H, W] generated by this invention fully preserves the spatiotemporal characteristics of radar echoes by integrating batch, temporal, channel, and spatial dimensions: the batch dimension (B) supports parallel training, the temporal dimension (T) captures the trajectory of the target object in a continuous frame sequence, the channel dimension (C) stores multi-band / polarization radar features, and the spatial dimensions (H, W) maintain the geometric structure of the target object; this tensor serves as the standardized input of the Hourglass3D module and directly processes the temporal-spatial correlation features through a three-dimensional convolutional kernel, enabling the network to simultaneously analyze the instantaneous scattering characteristics and cross-frame evolution of the target object, especially for tiny targets with a signal-to-noise ratio of less than 10dB and a pixel area of ​​less than 0.5%.

[0063] Furthermore, the Hourglass3D module includes an encoder, a bottleneck layer, a decoder, and a skip connection structure. The encoder consists of multiple levels of 3D convolutional downsampling units, each level containing a 3D convolutional layer, a normalization layer, and an activation function. The bottleneck layer uses 3D convolution to maintain the number of feature map channels. The decoder consists of multiple levels of 3D transposed convolutional upsampling units, each upsampling unit containing a 3D transposed convolutional layer, a batch normalization layer, and a ReLU activation function layer. The skip connection structure fuses the output feature maps of each layer of the encoder with the corresponding feature map inputs of the decoder.

[0064] Among them, the 3D convolutional downsampling unit uses a convolutional kernel with a stride greater than 1 to achieve feature map downsampling (corresponding to downsampling); the 3D transposed convolutional upsampling unit uses a convolutional kernel with a stride greater than 1 to achieve feature map upsampling (corresponding to upsampling); the skip connection structure adds the output of the nth level of the encoder to the input of the (4-n)th level of the decoder element by element, where n∈{1,2,3,4}.

[0065] As a preferred solution, in the Hourglass3D module, the 3D convolutional downsampling unit satisfies the channel number doubling rule: the number of output channels C at the k-th level... k = C0×2 k Where C0 is the initial number of channels, k∈{1,2,3,4}; the 3D transposed convolutional upsampling unit satisfies the channel number reduction rule: the number of output channels at the m-th level C m = C4 / 2 (5-m) C4 is the number of bottleneck layer channels (i.e., the final stage of the encoder, C4 = C0 × 2). 4 ), m∈{1,2,3,4}. Where k∈{1,2,3,4}, it corresponds to the 4th level downsampling unit of the encoder (i.e., k=1 is the first level downsampling, k=4 is the fourth level downsampling); similarly, m corresponds to the 4th level upsampling unit of the decoder.

[0066] In this application, the Hourglass3D module achieves multi-scale spatiotemporal feature extraction through a cascaded structure of three-dimensional convolutional downsampling units and three-dimensional transposed convolutional upsampling units.

[0067] Specifically, the 3D convolutional downsampling unit uses a 3×3×3 convolutional kernel with a stride greater than 1 to downsample the feature map. With each downsampling stage, the spatial size (H, W) of the feature map is halved, while the number of channels is increased according to C. k = C0×2 kThe rule of doubling (k∈{1,2,3,4}) is used. For example, when the input feature map size is [8,5,64,256,256], after the first level of downsampling, it becomes [8,5,128,128,128]. This design enhances the expressive power of features while compressing spatial information. The corresponding 3D transposed convolution upsampling unit uses a 3×3×3 transposed convolution kernel with a stride greater than 1 to upsample the feature map, gradually restoring the spatial size to the original resolution. At the same time, the number of channels follows C... m = C4 / 2 (5-m) The rule decreases (m∈{1,2,3,4}), for example, the feature map [8,5,512,16,16] output by the bottleneck layer becomes [8,5,256,32,32] after the first-level upsampling.

[0068] Specifically, a skip connection structure is used to add the nth-level output of the encoder to the (4-n)th-level input of the decoder element-wise (n∈{1,2,3,4}). For example, the feature map [8,5,128,128,128] from the first-level output of the encoder is added to the feature map [8,5,128,128,128] from the third-level output of the decoder, achieving complementary advantages between low-level detailed features and high-level semantic features. This symmetrical encoding and decoding structure, combined with a cross-layer feature fusion mechanism, effectively solves the problem of feature vanishing caused by downsampling of small targets in radar echoes.

[0069] As a preferred option, the YOLOv8 network includes:

[0070] The backbone network includes at least two initial convolutional layers and multi-level convolutional modules. It receives the feature maps output by the Hourglass3D module and feeds them into the first two convolutional layers. After the multi-level convolutional modules extract hierarchical representations of local and global features, it outputs multi-scale feature maps.

[0071] The neck network consists of a feature pyramid structure, a path aggregation structure, and a dynamic convolutional layer. The feature pyramid structure fuses deep high-semantic features with shallow high-resolution features through upsampling. The path aggregation structure achieves cross-level feature interaction through downsampling. The dynamic convolutional layer adjusts the feature fusion method according to the adaptive weights of the input features and uses upsampling and downsampling to perform multi-scale fusion of the feature maps output by the backbone network.

[0072] The detection head includes a classification branch and a regression branch: the classification branch outputs the category confidence of the target object through a convolutional layer, and the regression branch outputs the prediction parameters of the target object's bounding box through a separate convolutional layer.

[0073] In this application, the YOLOv8 network is used as the core detection module, and the following optimized design achieves efficient detection of small targets by radar:

[0074] After receiving the five-dimensional feature map output by the Hourglass3D module, the backbone network first performs channel dimension adaptation and shallow feature extraction through two initial convolutional layers (kernel size 3×3, stride 1), keeping the output feature map size consistent with the input. Subsequently, deep feature extraction is performed through a residual structure composed of multi-level CBS modules (Conv-BN-SiLU). Each level uses convolutional kernels of different lengths (a combination of 1×1 and 3×3) to achieve hierarchical capture of local details and global semantics, finally outputting feature maps at three scales (e.g., when the input is 640×640, the output has three resolutions: 80×80, 40×40, and 20×20), corresponding to the detection requirements of different sized targets.

[0075] Specifically, the first-level convolutional module retains the spatiotemporal feature dimensions of the Hourglass3D output, while subsequent modules gradually compress temporal information and enhance spatial features.

[0076] The neck network employs an improved FPN+PAN dual-pathway structure: the FPN pathway fuses deep high-semantic features (e.g., 20×20) with shallow high-resolution features (e.g., 40×40) element-wise through a 2x upsampling; the PAN pathway achieves inverse feature enhancement through 3×3 convolution downsampling with a stride of 2. Dynamic convolutional layers automatically adjust kernel weights based on the energy distribution of the feature maps; for example, larger receptive field kernels (5×5 instead of 3×3) are used for densely populated regions of small objects (determined through threshold segmentation).

[0077] The detection head adopts a decoupled design: the classification branch outputs the class confidence through 1×1 convolution, and the regression branch uses 3×3 depthwise separable convolution to predict the bounding box parameters (the center coordinates x, y are normalized by sigmoid, and the width and height w, h are predicted based on the logarithmic offset of the predicted box). The positioning accuracy is optimized by CIoU Loss.

[0078] As a preferred embodiment, the method further includes: setting a dynamic confidence threshold and an IoU threshold, and adaptively adjusting the threshold parameters according to the density characteristics of the target object detection scene; using an improved non-maximum suppression algorithm to process the predicted bounding boxes, including: performing feature map interpolation enhancement on the predicted bounding boxes with an area smaller than a preset threshold; and outputting detection results containing the target object category, confidence level, and bounding box coordinates, wherein the bounding box parameters include the center point coordinates (x, y) and width and height dimensions (w, h).

[0079] This application employs a dynamic optimization strategy in the post-processing stage to improve the detection performance of small targets. Specifically, the dynamic confidence threshold is automatically adjusted based on the spatial density distribution of the targets: by statistically analyzing the clustering of the predicted bounding boxes in the current frame, for example, the confidence threshold is increased from a baseline of 0.4 to 0.5-0.6 for dense regions (such as target spacing less than 10 pixels), and decreased to 0.3-0.35 for sparse regions, effectively balancing false negatives and false positives. The improved non-maximum suppression algorithm is specifically optimized for small targets: firstly, for predicted bounding boxes with an area less than 32×32 pixels, bilinear interpolation enhancement (weighted fusion of 4×4 neighborhood feature values) is performed at their corresponding feature map positions, significantly improving the feature response of weak targets; subsequently, a dynamic IoU threshold strategy is adopted, adaptively adjusting the suppression threshold according to the target density. The bounding box parameters are output using normalized coordinates.

[0080] As a preferred approach, the method further includes training a cascaded Hourglass3D module and a YOLOv8 network, including: inputting a preprocessed five-dimensional tensor [B,T,C,H,W] into the cascaded Hourglass3D module and the YOLOv8 network; the Hourglass3D module processes the input data through four layers of three-dimensional convolutional downsampling units, with the output feature map size halved layer by layer and the number of channels doubled layer by layer; the feature map output from the last layer of the Hourglass3D module is input into the first two convolutional layers of the YOLOv8 network; in the backbone network of the YOLOv8 network, the input feature map is processed through multi-level convolutional modules to output feature maps at three scales; the feature maps at the three scales are input into the neck network for feature fusion; the fused feature map is processed using a decoupled detection head, with the classification branch outputting class probabilities and the regression branch outputting bounding box coordinates; the classification loss and regression loss are calculated, and the network parameters are updated through backpropagation to obtain the detection model.

[0081] In practice, the joint training process of the cascaded Hourglass3D module and the YOLOv8 network is achieved through the following specific steps:

[0082] During training, the preprocessed 3D radar echo sequence (a five-dimensional tensor [B,T,C,H,W]) is input into the network. The Hourglass3D module first performs downsampling processing through four layers of 3D convolutions, each layer using a 3×3×3 convolution kernel (stride 2, zero padding 1), combined with a channel multiplication rule, to compress the feature map size from [8,5,64,256,256] to [8,5,1024,16,16] step by step, while retaining the feature maps of each layer through skip connections. The spatiotemporal features output by this module are then adjusted for the number of channels by a 1×1×1 convolution before being input into the first two convolutional layers of the YOLOv8 network (kernel size 3×3, stride 1). This is integrated with the native YOLOv8 backbone network. The backbone network processes data through a four-level CBS module (Conv-BN-SiLU), outputting feature maps at three scales: 20×20, 40×40, and 80×80, corresponding to target detection tasks of different sizes. The neck network adopts an improved FPN+PAN structure, where the FPN pathway fuses deep semantic features through bilinear interpolation upsampling, and the PAN pathway enhances localization accuracy through 3×3 convolutions (stride 2). The dynamic convolutional layer automatically adjusts the fusion weights based on the feature energy distribution. In the decoupled detection head, the classification branch uses 1×1 convolutions to output class probabilities and uses Focal Loss to address sample imbalance; the regression branch predicts bounding boxes through 3×3 depthwise separable convolutions and uses CIoU Loss to optimize localization, where the center coordinates (x, y) are normalized by sigmoid, and the width and height (w, h) are calculated based on the logarithmic offset of the anchor box size.

[0083] Specifically, during model training, the data is divided into training and validation sets in an 8:2 ratio, with a batch size of 8 and 100 iterations. An early stopping mechanism is implemented, using the patience parameter: if the monitored metric mAP@0.5:0.95 shows no improvement over 30 consecutive training iterations, training is stopped to avoid overfitting. After each training iteration, the validation set is used to evaluate the model's performance (HG3D is short for Hourglass3D, CN is short for Cascade Network), monitoring the loss and accuracy on the validation set.

[0084] In addition, multiple metrics were calculated, including mean precision (mAP), F1 score, precision (P), and recall (R). Precision-confidence curves and precision-recall curves were plotted, and the dataset samples and detection results were visualized to determine their performance in real-world scenarios. Specifically, mean precision is an important metric for measuring the accuracy of multi-class object detection; the F1 score, a weighted average of precision and recall, measures the overall performance and stability of the model; precision measures the proportion of samples predicted as positive by the model to actual positive samples, reflecting the accuracy of the model's predictions; and recall measures the proportion of correctly predicted positive samples out of all actual positive samples, reflecting the model's ability to capture positive examples.

[0085] Based on the above embodiments, the working principle of the present invention is as follows:

[0086] This invention proposes a core concept based on joint extraction of spatiotemporal features and multi-scale fusion. Through symmetrical spectral shifting and five-dimensional tensor reconstruction, it fully preserves the temporal dynamic characteristics and spatial structure information of radar echoes. The Hourglass3D module adopts a symmetrical encoding and decoding structure. The encoder progressively compresses the spatial dimension and expands the number of channels through three-dimensional convolutional downsampling, enhancing feature representation while preserving the target's motion trajectory. The decoder restores resolution through transposed convolution and combines skip connections to fuse multi-scale features, effectively solving the feature loss problem of small targets during downsampling. The YOLOv8 network extracts multi-level features through the backbone network, and the neck network uses a bidirectional feature pyramid to achieve complementary features between deep and shallow layers. The dynamic convolutional layer adaptively adjusts the receptive field according to the target distribution. The decoupled design of the detection head promotes the mutual promotion of classification and localization tasks, and the post-processing process is optimized with a dynamic threshold strategy. The entire system achieves collaborative optimization of spatiotemporal features through end-to-end training, enabling the network to simultaneously capture the instantaneous scattering characteristics and cross-frame evolution patterns of targets, significantly improving the detection performance of small targets in low signal-to-noise ratio environments.

[0087] Example 2

[0088] Based on the same inventive concept, a specific embodiment of this application proposes a low-confidence small target detection system based on a hybrid architecture for radar echoes, used to implement the small target detection method proposed in Embodiment 1. The system includes a preprocessing module, a target detection module, and an optimization processing module. The preprocessing module is used to perform spectrum shifting and dimension reorganization on the radar echo data to be tested, generating a five-dimensional tensor with temporal characteristics. The target detection module is used to input the five-dimensional tensor into the detection model to obtain prediction results including the predicted bounding box position of the target. The detection model includes a cascaded Hourglass3D module and a YOLOv8 network. The Hourglass3D module extracts and fuses multi-scale spatiotemporal features in the five-dimensional tensor through three-dimensional convolution and skip connections, and outputs the feature map to the front end of the backbone network of the YOLOv8 network to achieve multi-scale spatiotemporal feature fusion and target classification and localization.

[0089] The optimization processing module is used to set dynamic confidence thresholds and IoU thresholds, and adaptively adjusts the threshold parameters according to the density characteristics of the target detection scene; it uses an improved non-maximum suppression algorithm to process the predicted boxes, including: performing feature map interpolation enhancement on predicted boxes with an area smaller than a preset threshold; and outputting detection results containing target category, confidence, and bounding box coordinates, where the bounding box parameters include center point coordinates (x, y) and width and height dimensions (w, h).

[0090] For specific limitations regarding the low-confidence small target detection system based on a hybrid architecture in radar echoes, please refer to the limitations of the low-confidence small target detection method based on a hybrid architecture in radar echoes mentioned above, which will not be repeated here. It should be noted that each module in the above detection system corresponds to a step in implementing the above detection method. Multiple modules and their corresponding steps are implemented in the same instances and application scenarios, but are not limited to the content disclosed in Embodiment 1 above.

[0091] Understandably, the radar echo low-confidence small target detection system proposed in this invention adopts a three-level processing architecture: the preprocessing module is responsible for the spectrum shifting and five-dimensional tensor transformation of radar data, reorganizing the original echo data into a [B,T,C,H,W] structure containing temporal features; the target detection module achieves feature extraction and recognition through cascaded Hourglass3D and YOLOv8 networks, where the Hourglass3D module uses three-dimensional convolution to mine spatiotemporal correlation features, and the YOLOv8 network completes multi-scale target detection; the optimization processing module adopts a dynamic threshold strategy, automatically adjusting the confidence and IoU thresholds according to the target density, and performing feature enhancement on the predicted bounding boxes of small targets.

[0092] In practical applications, the above-mentioned detection system is suitable for real-time detection scenarios such as border monitoring and drone early warning.

[0093] To more clearly illustrate the present invention and its advantages, the method provided by the present invention will be further explained below in conjunction with specific simulation experiments and related partial figures.

[0094] (1) The training time, mAP@0.5 and F1 score of the YOLOv8 model and the HG3DCN-YOLOv8 model in this application are shown in Table 1.

[0095] Table 1. Results of training time, mAP@0.5, and F1 score.

[0096] ;

[0097] The improved model in this application takes more than ten times less time to train than the original YOLOv8 model, resulting in a significant improvement in computational efficiency. Simultaneously, mAP@0.5 also increased by 2.89%, indicating that the improved model has enhanced accuracy in target localization and reduced the false negative rate for small targets.

[0098] (2) Figure 4 The figure compares the precision-recall curves of the YOLOv8 model and the HG3DCN-YOLOv8 model, with the horizontal axis representing recall and the vertical axis representing precision. The figure indicates that the HG3DCN-YOLOv8 model has higher precision than the original YOLOv8 model, especially when the recall R is lower. This means that under more challenging object detection conditions, the improved model in this application achieves more accurate object detection, i.e., better reliability.

[0099] In this application, Hourglass3D's skip connection mechanism preserves the low-level features of small objects, while the dynamic FPN-PAN structure optimizes multi-scale feature fusion.

[0100] (3) Figure 5 The precision-confidence curves of the YOLOv8 model and the HG3DCN-YOLOv8 model are compared, with the horizontal axis representing confidence level and the vertical axis representing precision. The curves show that the HG3DCN-YOLOv8 model not only improves precision at a confidence level of 0.5, but also demonstrates better precision across the entire confidence range from 0 to 1.

[0101] (4) Figure 6 , Figure 7 and Figure 8 These are the original labels manually labeled in the dataset, the detection results using YOLOv8, and the detection results using the improved model, respectively. Figure 7The results show that YOLOv8 misdetects a single small target as two neighboring targets, because its two-dimensional convolution has difficulty distinguishing the spatiotemporal characteristics of dense targets. Figure 8 The reason why HG3DCN-YOLOv8 correctly detected the target object is that: 3D convolution captures the motion consistency of the target object in consecutive frames; dynamic NMS suppresses redundant boxes; and feature interpolation enhances the response signal of small target objects.

[0102] In summary, this application employs an improved Hourglass3D module to capture motion features between consecutive frames, combined with a dynamically optimized YOLOv8 detection network to achieve multi-scale target recognition. Feature enhancement and adaptive thresholding strategies significantly improve the detection reliability of small targets. Simulation results demonstrate that, compared to traditional methods, this scheme exhibits significant advantages in detection accuracy, training efficiency, and robustness, making it particularly suitable for complex radar monitoring scenarios with low signal-to-noise ratios and small target sizes.

[0103] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0104] In addition, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0105] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for detecting small targets with low confidence in radar echoes based on a hybrid architecture, the method being used to detect targets with a confidence level below a preset threshold, characterized in that... The method includes: The radar echo data to be tested is subjected to spectrum shifting and dimension reorganization to generate a five-dimensional tensor with temporal characteristics; The dimensional restructuring includes: The storage path of the image dataset obtained by symmetrically shifting the spectrum of radar echo data is parsed and decoded into a NumPy array; Convert the image data of the NumPy array into RGB channels and store them sequentially. Perform tensor dimension rearrangement on the stored data, adjusting the channel dimension from [H,W,C] to [C,H,W]; Add batch and time dimensions to the tensor to form a five-dimensional tensor with a [B,T,C,H,W] structure; Where B is the batch size; T is the time step, used to characterize the number of consecutive radar frames; C is the number of channels, corresponding to different frequency bands or polarization information of the radar echo; H is the height; and W is the width. The five-dimensional tensor is input into the detection model to obtain prediction results including the location of the target object prediction box; The detection model includes a cascaded Hourglass3D module and a YOLOv8 network. The Hourglass3D module extracts and fuses multi-scale spatiotemporal features from the five-dimensional tensor through three-dimensional convolution and skip connections, and outputs the feature map to the front end of the backbone network of the YOLOv8 network to achieve multi-scale spatiotemporal feature fusion and target classification and localization.

2. The method for detecting small targets with low confidence in radar echoes based on a hybrid architecture according to claim 1, characterized in that, The Hourglass3D module includes: The encoder consists of multiple levels of three-dimensional convolutional downsampling units, each level containing a three-dimensional convolutional layer, a normalization layer, and an activation function; The bottleneck layer uses 3D convolution to keep the number of feature map channels constant. The decoder consists of multiple levels of 3D transposed convolutional upsampling units, each of which includes a 3D transposed convolutional layer, a batch normalization layer, and a ReLU activation function layer. The skip connection structure fuses the output feature maps of each layer of the encoder with the input feature maps of the corresponding layers of the decoder.

3. The method for detecting small targets with low confidence in radar echoes based on a hybrid architecture according to claim 2, characterized in that, The three-dimensional convolutional downsampling unit uses a convolutional kernel with a stride greater than 1 to achieve feature map downsampling; the three-dimensional transposed convolutional upsampling unit uses a convolutional kernel with a stride greater than 1 to achieve feature map upsampling; the skip connection structure adds the output of the encoder at level n to the input of the decoder at level (4-n) element by element, where n∈{1,2,3,4}.

4. The method for detecting small targets with low confidence in radar echoes based on a hybrid architecture according to claim 3, characterized in that, In the Hourglass3D module, the three-dimensional convolutional downsampling unit satisfies the channel number doubling rule: the number of output channels C at the k-th level... k = C0×2 k Where C0 is the initial number of channels, k∈{1,2,3,4}; the three-dimensional transposed convolutional upsampling unit satisfies the channel number reduction rule: the number of output channels at the m-th level C m = C4 / 2 (5-m) Where C4 is the number of bottleneck layer channels, m∈{1,2,3,4}.

5. The method for detecting small targets with low confidence in radar echoes based on a hybrid architecture according to claim 1, characterized in that, The YOLOv8 network includes: The backbone network includes at least two initial convolutional layers and multi-level convolutional modules. It receives the feature maps output by the Hourglass3D module and feeds them into the first two convolutional layers. After the multi-level convolutional modules extract hierarchical representations of local and global features, it outputs multi-scale feature maps. The neck network consists of a feature pyramid structure, a path aggregation structure, and a dynamic convolutional layer. The feature pyramid structure fuses deep high-semantic features with shallow high-resolution features through upsampling. The path aggregation structure achieves cross-level feature interaction through downsampling. The dynamic convolutional layer adjusts the feature fusion method according to the adaptive weights of the input features and uses upsampling and downsampling to perform multi-scale fusion of the feature maps output by the backbone network. The detection head includes a classification branch and a regression branch: the classification branch outputs the category confidence of the target object through a convolutional layer, and the regression branch outputs the prediction parameters of the target object's bounding box through a separate convolutional layer.

6. The method for detecting low-confidence small targets in radar echoes based on a hybrid architecture according to claim 1, characterized in that, The method further includes: Set dynamic confidence threshold and IoU threshold, and adaptively adjust the threshold parameters according to the density characteristics of the target detection scene; An improved nonmaximum suppression algorithm is used to process the predicted bounding boxes, including: feature map interpolation enhancement for predicted bounding boxes with an area smaller than a preset threshold; The output includes the detection results containing the target object category, confidence score, and bounding box coordinates. The bounding box parameters include the center point coordinates (x, y) and the width and height dimensions (w, h).

7. The method for detecting small targets with low confidence in radar echoes based on a hybrid architecture according to claim 1, characterized in that, The method also includes training the cascaded Hourglass3D module and the YOLOv8 network, including: The preprocessed five-dimensional tensor [B,T,C,H,W] is input into the cascaded Hourglass3D module and YOLOv8 network; The Hourglass3D module processes the input data through four layers of three-dimensional convolutional downsampling units, and the output feature map size is halved layer by layer while the number of channels is doubled layer by layer. The feature map output from the last layer of the Hourglass3D module is input into the first two convolutional layers of the YOLOv8 network. In the backbone of the YOLOv8 network, the input feature map is processed through multi-level convolutional modules to output feature maps at three scales. Feature maps at three scales are input into the neck network for feature fusion. The decoupled detection head is used to process the fused feature map, with the classification branch outputting the class probability and the regression branch outputting the bounding box coordinates. Calculate the classification loss and regression loss, update the network parameters through backpropagation, and obtain the detection model.

8. A low-confidence small target detection system based on a hybrid architecture radar echo, characterized in that, For implementing the small target detection method as described in any one of claims 1-7, the system comprises: The preprocessing module is used to perform spectrum shifting and dimension reorganization on the radar echo data to be tested, generating a five-dimensional tensor with temporal characteristics; The target detection module is used to input the five-dimensional tensor into the detection model to obtain prediction results including the position of the target object prediction box; The detection model includes a cascaded Hourglass3D module and a YOLOv8 network. The Hourglass3D module extracts and fuses multi-scale spatiotemporal features from the five-dimensional tensor through three-dimensional convolution and skip connections, and outputs the feature map to the front end of the backbone network of the YOLOv8 network to achieve multi-scale spatiotemporal feature fusion and target classification and localization.

9. The low-confidence small target detection system based on a hybrid architecture in radar echoes according to claim 8, characterized in that, The system also includes: The optimization processing module is used to set dynamic confidence thresholds and IoU thresholds, and adaptively adjusts the threshold parameters according to the density characteristics of the target detection scene; it uses an improved non-maximum suppression algorithm to process the predicted bounding boxes, including: performing feature map interpolation enhancement on predicted bounding boxes with an area smaller than a preset threshold; and outputting detection results containing the target category, confidence level, and bounding box coordinates, where the bounding box parameters include the center point coordinates (x, y) and width and height dimensions (w, h).

Images