Sea surface small target detection method and device based on clutter map and improved YOLOv4

Abstract

The application discloses a sea surface small target detection method and device based on a clutter map and an improved YOLOv4, relates to the technical field of target detection, and comprises the following steps: obtaining radar echo data to be detected, and converting the radar echo data to be detected into a clutter image; inputting the clutter image into a trained target detection network model, and sequentially performing feature extraction, feature enhancement, multi-layer feature fusion, target prediction frame regression and target classification to obtain a target detection result of the clutter image; wherein the trained target detection network model is obtained by taking a plurality of clutter images with pre-completed target labels as training data sets and training an initial target detection network model. The application can enhance the feature extraction capability of the network for small targets.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology, specifically relating to a method and apparatus for detecting small targets on the sea surface based on clutter maps and an improved YOLOv4. Background Technology

[0002] In sea surface target detection missions, traditional target detection methods are based on statistical signal processing theory, achieving target detection through statistical modeling of echoes and binary hypothesis testing. On the one hand, for low-resolution sea clutter, complex Gaussian models are typically used; for high-resolution sea clutter, composite Gaussian models are commonly used. On the other hand, for sea surface targets, point target or range-extended target models are usually used. These are all parametric models, requiring the estimation of unknown model parameters using echo data. Considering the ease of implementation of the detection method, the number of model parameters should not be excessive. However, due to the highly complex statistical characteristics of sea clutter in practical applications—characterized by non-Gaussian, non-stationary, and non-uniform features—and the high-speed maneuverability and scintillation of a significant portion of sea surface targets, coupled with the spatiotemporal variation of echo models, it is difficult to model and analyze echo signals using general statistical models. In such cases, using simple models to model the echoes can easily lead to model mismatch, resulting in a significant deterioration in detection performance.

[0003] In recent years, machine learning has achieved outstanding results in numerous tasks thanks to its powerful learning capabilities. In the field of object detection, machine learning and deep learning methods are mainly used to identify and process images or videos. This involves classifying objects and regressing bounding boxes to determine the object's category and location, thus achieving the detection objective. Convolutional Neural Networks (CNNs), as a crucial component of deep learning, have achieved excellent results in many image object detection tasks. Currently, both domestically and internationally, two-stage detection algorithms and single-stage detection algorithms are the two mainstream deep learning-based object detection algorithms. Two-stage detection algorithms mainly include R-CNN (Regions with Convolutional Neural Networks), Fast R-CNN (Fast Regions with Convolutional Neural Networks), and Faster R-CNN (Faster Regions with Convolutional Neural Networks), while single-stage detection algorithms mainly include YOLO and SSD (Single Shot MultiBox Detector). Because deep learning technology has a powerful ability to automatically learn features from data, some scholars have innovatively combined intelligent target detection technology with maritime exploration missions, and have successively proposed many intelligent sea surface target detection algorithms based on CNNs. These algorithms have improved detection speed and accuracy, and have significant advantages over traditional target detection algorithms.

[0004] Although experiments have verified that CNN-based intelligent sea surface target detection algorithms can achieve good detection results, in environments with strong noise or complex clutter, the echo signals of small targets are generally weak, which drastically increases the difficulty of detection. The detection performance of CNN-based intelligent sea surface target detection algorithms is also unsatisfactory, and they are prone to missed detection and false detection of small targets, requiring further improvement and refinement. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a method and apparatus for detecting small targets on the sea surface based on clutter maps and an improved YOLOv4. The technical problem to be solved by this invention is achieved through the following technical solution:

[0006] In a first aspect, the present invention provides a method for detecting small targets on the sea surface based on clutter maps and an improved YOLOv4, comprising:

[0007] Acquire the radar echo data to be detected and convert it into a clutter image;

[0008] The clutter image is input into the trained target detection network model, and feature extraction, feature enhancement, multi-layer feature fusion, target prediction box regression and target classification are performed in sequence to obtain the target detection result of the clutter image;

[0009] The trained object detection network model is obtained by training the initial object detection network model using clutter images with pre-annotated objects as the training dataset.

[0010] Secondly, the present invention also provides a small target detection device for the sea surface based on clutter maps and an improved YOLOv4, comprising:

[0011] The data acquisition module is used to acquire the radar echo data to be detected and convert the radar echo data to be detected into a clutter image;

[0012] The data processing module is used to input clutter images into a trained target detection network model, and sequentially perform feature extraction, feature enhancement, multi-layer feature fusion, target prediction box regression, and target classification to obtain the target detection results of the clutter image;

[0013] The trained object detection network model is obtained by training the initial object detection network model using clutter images with pre-annotated objects as the training dataset.

[0014] The beneficial effects of this invention are:

[0015] This invention provides a method and apparatus for detecting small targets on the sea surface based on clutter maps and an improved YOLOv4. Considering the difficulty of detecting weak targets on the sea surface, this invention addresses the issue from the perspective of clutter maps. Based on the YOLOv4 target detection network, the internal structure of the network is specifically improved, enhancing the network's feature extraction capability for small targets. The improved network significantly enhances the detection performance of small targets. At the same time, while maintaining detection accuracy, this invention simplifies the network structure to a certain extent, greatly reducing the number of network parameters and computational load, and improving the network's training speed, making it more suitable for practical engineering applications.

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for detecting small targets on the sea surface based on clutter maps and improved YOLOv4 provided in an embodiment of the present invention;

[0018] Figure 2 This is a schematic diagram of a target detection network model provided in an embodiment of the present invention;

[0019] Figure 3This is a schematic diagram of a backbone feature extraction network provided in an embodiment of the present invention;

[0020] Figure 4 This is a schematic diagram of a clutter map provided in an embodiment of the present invention;

[0021] Figure 5 This is a schematic diagram of the annotation of the clutter map provided in an embodiment of the present invention;

[0022] Figure 6 This is a schematic diagram of a dense network connection module provided in an embodiment of the present invention;

[0023] Figure 7 This is a schematic diagram of a head attention gate provided in an embodiment of the present invention;

[0024] Figure 8 This is a schematic diagram of DWConv provided in an embodiment of the present invention;

[0025] Figure 9 This is a schematic diagram of a comparison of target detection performance under different networks provided in an embodiment of the present invention;

[0026] Figure 10 This is a schematic diagram of the clutter map target detection result provided in an embodiment of the present invention. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0028] To address the shortcomings of existing technologies, this invention provides a method and apparatus for detecting small targets on the sea surface based on clutter maps and an improved YOLOv4. Considering the difficulty in detecting weak targets on the sea surface, this invention improves the internal structure of the YOLOv4 target detection network from the perspective of clutter maps, enhancing the network's feature extraction capability for small targets. The improved network significantly enhances the detection performance of small targets. Simultaneously, while maintaining detection accuracy, this invention simplifies the network structure to a certain extent, greatly reducing the number of network parameters and computational load, and improving the network's training speed, making it more suitable for practical engineering applications.

[0029] Please see Figure 1 , Figure 1 This is a flowchart of a method for detecting small sea targets based on clutter maps and improved YOLOv4 provided in an embodiment of the present invention. The method for detecting small sea targets based on clutter maps and improved YOLOv4 provided by the present invention includes:

[0030] S101. Acquire the radar echo data to be detected and convert the radar echo data to be detected into a clutter image.

[0031] S102. Input the clutter image into the trained target detection network model, and perform feature extraction, feature enhancement, multi-layer feature fusion, target bounding box prediction regression and target classification in sequence to obtain the target detection result of the clutter image.

[0032] The trained object detection network model is obtained by training the initial object detection network model using clutter images with pre-annotated objects as the training dataset.

[0033] Specifically, in this embodiment, please refer to Figure 2 and Figure 3 , Figure 2 This is a schematic diagram of a target detection network model provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of a backbone feature extraction network provided in an embodiment of the present invention. Before training the initial target detection network model, it also includes:

[0034] Construct an object detection network model;

[0035] The target detection network model includes a backbone feature extraction network, a pyramid pooling network, a feature fusion network, a head attention gate, and a detection head. The backbone feature extraction network is used to extract features, the pyramid pooling network is used to enhance features, the feature fusion network is used to fuse features, the head attention gate is used to calculate weight coefficients, and the detection head is used for target prediction box regression and target classification of different sizes.

[0036] The backbone feature extraction network includes sequentially connected convolutional layers and multiple combined modules. The combined modules include dense network connection modules and transitional connection modules. The dense network connection modules include multi-layer convolutional layers that are sequentially connected and skipped connections. The transitional connection modules include sequentially connected two-layer convolutional layers.

[0037] Pyramid pooling networks consist of multiple pooling layers arranged in parallel;

[0038] The feature fusion network includes an upsampling structure and a downsampling structure. The upsampling structure includes an upsampling layer and a feature fusion layer connected in sequence, and the downsampling structure includes a downsampling layer and a feature fusion layer connected in sequence.

[0039] The head attention gate consists of convolutional layers and activation function layers;

[0040] The detection head consists of multiple convolutional layers.

[0041] It should be noted that the backbone feature extraction network DenseNet provided in this embodiment differs from the Residual Network (ResNet) in that it is a denser connection network. By establishing connection channels between layers within the network, it maximizes the utilization of feature reuse. The connection relationships between the layers within DenseNet and the overall structure of the DenseNet network are as follows: Figure 3 As shown, DenseBlock and Transition are the basic building blocks of DenseNet.

[0042] In this embodiment, training the initial object detection network model includes:

[0043] Acquire clutter image data, construct a training dataset, and label the targets in the training samples within the training dataset to obtain the corresponding labels for the training dataset. See [link to relevant documentation]. Figure 4 and Figure 5 , Figure 4 This is a schematic diagram of a clutter map provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of clutter image annotation provided in an embodiment of the present invention; the dataset is expanded and the annotation information of the target boxes is improved through image enhancement; it can be understood that the expanded clutter image data is manually annotated image by image using Labelimg software, and the main target types include two types: ships and islands, such as... Figure 5 As shown, ship-type targets occupy a very small proportion of the entire image and their features are not obvious. Coupled with the interference in the surrounding environment, detection is quite difficult.

[0044] The training dataset and its corresponding labels are input into the initial object detection network model for training. Within a preset number of iterations, the loss function is made to converge and reach its minimum value, thus obtaining the parameters of the object detection network model and constructing the trained object detection network model.

[0045] In this embodiment, acquiring clutter image data and constructing a training dataset includes:

[0046] Acquire radar echo data and convert it into training clutter image data;

[0047] Image data augmentation is performed on the training clutter image data to obtain the training dataset; image data augmentation includes rotation, cropping, translation, blurring, and noise addition.

[0048] In this embodiment, the clutter image is input into the trained target detection network model, and feature extraction, feature enhancement, multi-layer feature fusion, target bounding box regression, and target classification are performed sequentially, including:

[0049] The clutter image is input into the trained target detection network model. After passing through the convolutional layer, the first combination module, and the second combination module in the backbone feature extraction network, the first feature is output. The first feature is processed by the third combination module to output the second feature. The second feature is processed by the fourth combination module to output the feature to be enhanced. The feature to be enhanced is processed by the pooling layer in the pyramid pooling network. The enhanced features processed by multiple pooling layers are then connected to output the third feature.

[0050] After the third feature is processed by the upsampling layer in the upsampling structure, it is connected with the second feature to output an intermediate feature. After the intermediate feature is processed by the upsampling layer in the upsampling structure, it is connected with the first feature to obtain the first fused feature. After the first fused feature is processed by the downsampling layer in the downsampling structure, it is connected with the intermediate feature to output the second fused feature. After the second fused feature is processed by the downsampling layer in the downsampling structure, it is connected with the third feature to output the third fused feature.

[0051] The first fused feature and the first feature are processed by a head attention gate to output a first feature matrix. The first feature matrix is ​​input into the corresponding detection head for processing, and the target detection result under the first size feature is output. The second fused feature and the second feature are processed by a head attention gate to output a second feature matrix. The second feature matrix is ​​input into the corresponding detection head for processing, and the target detection result under the second size feature is output. The third fused feature and the third feature are processed by a head attention gate to output a third feature matrix. The third feature matrix is ​​input into the corresponding detection head for processing, and the target detection result under the third size feature is output.

[0052] In this embodiment, please refer to Figure 6 , Figure 6 This is a schematic diagram of a dense connection module provided in an embodiment of the present invention. The dense network connection module includes multiple convolutional layers. The input of the i-th convolutional layer in the dense network connection module is the output of the 1st to i-1th convolutional layers, where i ≥ 2.

[0053] In this embodiment, please refer to Figure 7 , Figure 7 This is a schematic diagram of a head attention gate provided in an embodiment of the present invention. The main idea of ​​the YOLO network is hierarchical detection. It first extracts features x3, x4, and x5 from different receptive field branches, then fuses these features to obtain fused effective feature layers P3, P4, and P5. Finally, three detection heads are responsible for detecting targets of different sizes. The results of the three detection heads are then subjected to non-maximum suppression and bounding box decoding to be used as the final detection result. Figure 7As shown, P3 and YOLO Head_1 are mainly responsible for small-sized target detection, while P4 and YOLO Head_2, and P5 and YOLO Head_5 are mainly responsible for medium-sized and large-sized target detection. Based on this principle, we consider introducing an attention gate mechanism before the detection head. By adding an attention weight to the feature layer of the input detection head, we can distinguish the effective attention of different detection heads to feature information of different sizes. Specifically, the three effective feature layers P3, P4, and P5 output by the feature fusion network fuse features at different scales to a certain extent, which increases the global effectiveness of the feature layers. However, compared with the original three effective feature layers x3, x4, and x5, there is a potential risk that the feature information specific to each scale will be weakened, which will have an adverse effect on target detection. To solve the above problem, we propose to redistribute the attention weights before the final detection, and the basis for weight allocation is the original effective feature layers x3, x4, or x5 of the same size. We call this the Head Attention Gate (Head_AG) mechanism. The principle diagram of Head_AG is shown in the figure. Figure 7 As shown.

[0054] The process of handling head attention gates includes:

[0055] The first fused feature is input into the first convolutional layer of the head attention gate for processing, and the first feature to be processed is output. The first feature is input into the second convolutional layer of the head attention gate for processing, and the second feature to be processed is output. The first feature to be processed and the second feature to be processed are added together, and the result is input into the third convolutional layer of the head attention gate for processing. After processing by the activation function, the first weight coefficient is output. The first weight coefficient is multiplied by the first fused feature to output the first feature matrix.

[0056] The second fused feature is input into the first convolutional layer of the head attention gate for processing, and the third feature to be processed is output. The second feature is input into the second convolutional layer of the head attention gate for processing, and the fourth feature to be processed is output. The third feature to be processed and the fourth feature to be processed are added together, and the result is input into the third convolutional layer of the head attention gate for processing. After processing by the activation function, the second weight coefficient is output. The second weight coefficient is multiplied by the second fused feature to output the second feature matrix.

[0057] The third fused feature is input into the first convolutional layer of the head attention gate for processing, and the fifth feature to be processed is output. The third feature is input into the second convolutional layer of the head attention gate for processing, and the sixth feature to be processed is output. The fifth and sixth features to be processed are added together, and the result is input into the third convolutional layer of the head attention gate for processing. After passing through the activation function, the third weight coefficient is output. The third weight coefficient is multiplied by the third fused feature to output the third feature matrix.

[0058] It is understandable that, such as Figure 7 As shown, x represents the input signal, which consists of the first fused feature, the second fused feature, and the third fused feature, respectively. g represents the gating signal, which consists of the first feature, the second feature, and the third feature, respectively. Both are input into the attention gate to calculate the attention weight coefficients. Then, the calculated weight coefficients are multiplied by the input signal to obtain the final feature matrix after redistributing attention.

[0059] In this embodiment, after the detection head outputs the target detection results under the first size feature, the target detection results under the second size feature, and the target detection results under the third size feature, the following is further included:

[0060] The target detection results under the first size feature, the second size feature, and the third size feature are subjected to nonmaximum suppression and detection box decoding respectively to obtain the target detection results of the clutter image.

[0061] In this embodiment, please refer to Figure 8 , Figure 8 This is a schematic diagram of DWConv provided in an embodiment of the present invention, wherein all convolutional layers with a kernel size of 3×3 are DWConv layers.

[0062] It is understandable that the 3×3 convolutional parts of the backbone feature extraction network and the feature fusion network that are not shown are all DWConv layers provided in this embodiment. They can not only achieve the same feature extraction effect as the standard 3×3 convolution, but also reduce the number of network parameters and computation to a certain extent.

[0063] DWConv, short for Depthwise Convolution, is based on the core idea of ​​grouping convolutions along the channel dimension. Different convolutional kernels perform specific tasks, handling different feature layer outputs. When the number of group convolutions equals the number of input channels, the number of output channels equals the number of input channels. Compared to standard convolution, DWConv significantly reduces the number of convolution operations, thus reducing the network's parameter count and computational cost. Furthermore, to achieve the same output effect as standard convolution in neural network computation, DWConv is typically used in conjunction with 1×1 convolutional layers. The former changes the feature map size, while the latter adjusts the feature map to the ideal channel dimension for output. The principle and structure of DWConv are as follows: Figure 8 As shown. Typically, any ordinary convolutional layer in a neural network with a kernel value other than 1 can be replaced by the corresponding DWConv and 1×1 convolution, which can achieve the effect of lightweighting the network structure.

[0064] In summary, the sea surface small target detection method based on clutter maps and improved YOLOv4 provided by this invention has the following beneficial effects:

[0065] 1. This invention improves the backbone feature extraction network of YOLOv4 by replacing ResNet with DenseNet as the backbone feature extraction network. Its dense skip connections can better preserve the feature information of the target compared with the original residual network. At the same time, this invention improves the 3×3 standard convolution in the YOLOv4 network by using a method of first performing depthwise convolution and then adjusting the feature dimension. This not only achieves the same feature extraction effect as the 3×3 standard convolution, but also reduces the number of network parameters and computation to a certain extent.

[0066] 2. This invention proposes a Head_AG structure, which adds a gating signal to re-distinguish the attention of the three effective feature layers before the final target detection. This compensates for the risk of weakening of specific scale features due to global feature fusion, allowing the detection heads responsible for different sizes to pay more attention to the feature information at their respective sizes, thereby achieving better target detection results.

[0067] In an optional embodiment of the present invention, the effectiveness of the sea surface small target detection method based on clutter maps and improved YOLOv4 provided in the above embodiment is verified by simulation experiments, specifically as follows:

[0068] Please see Figure 9 and Figure 10 , Figure 9 This is a schematic diagram of a comparison of target detection performance under different networks provided in an embodiment of the present invention. Figure 10 This is a schematic diagram of the clutter map target detection result provided in an embodiment of the present invention. Figure 9As can be seen, the improved YOLOv4 target detection network proposed in this invention performs optimally across all evaluation metrics. For island-type targets, it is... Figure 5 It can be seen that the characteristics are relatively obvious, and the detection performance of each detector is not significantly different; however, for small targets of the ship type, due to the greater difficulty in detection, the detection performance of each detector varies considerably. The improved YOLOv4 target detection network proposed in this invention can better preserve the feature information of small targets, and to a certain extent compensate for the loss of key information about small targets. Figure 9 As shown, the present invention achieves the maximum average precision (AP) for ship type, indicating that the detector proposed in this invention has the best detection performance for small targets among all the detectors mentioned above.

[0069] This embodiment also compares the parameter count, computational cost, and training time of the present invention with those of several common object detection networks. Table 1 shows a comparison of the parameter count, computational cost, and running time of different networks. The improved YOLOv4 object detection network proposed in this invention simplifies the network's internal structure. As shown in Table 1, the present invention has fewer parameters, less computation, and shorter network running time compared to other common object detectors, making it more suitable for practical engineering applications.

[0070] Table 1

[0071]

[0072]

[0073] Based on the same inventive concept, this invention also provides a small target detection device for the sea surface based on clutter maps and improved YOLOv4, used to implement the method provided in the above embodiments of this invention. The method embodiments will not be repeated here. The device includes:

[0074] The data acquisition module is used to acquire the radar echo data to be detected and convert the radar echo data to be detected into a clutter image;

[0075] The data processing module is used to input clutter images into a trained target detection network model, and sequentially perform feature extraction, feature enhancement, multi-layer feature fusion, target prediction box regression, and target classification to obtain the target detection results of the clutter image;

[0076] The trained object detection network model is obtained by training the initial object detection network model using clutter images with pre-annotated objects as the training dataset.

[0077] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0078] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0079] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for detecting small targets on the sea surface based on clutter maps and an improved YOLOv4, characterized in that, include: Acquire radar echo data to be detected and convert the radar echo data to be detected into a clutter image; The clutter image is input into a trained target detection network model, and feature extraction, feature enhancement, multi-layer feature fusion, target prediction box regression, and target classification are performed sequentially to obtain the target detection result of the clutter image. The trained target detection network model is obtained by training the initial target detection network model using several clutter images with pre-annotated targets as the training dataset. The target detection network model includes a backbone feature extraction network, a pyramid pooling network, a feature fusion network, a head attention gate, and a detection head. The backbone feature extraction network is used to extract features, the pyramid pooling network is used to enhance features, the feature fusion network is used to fuse features, the head attention gate is used to calculate weight coefficients, and the detection head is used for target bounding box regression and target classification of different sizes. The backbone feature extraction network includes sequentially connected convolutional layers and multiple combination modules. The combination modules include dense network connection modules and transition connection modules. The dense network connection modules include multiple convolutional layers that are sequentially connected and skip-connected, and the transition connection modules include sequentially connected double convolutional layers. The pyramid pooling network includes multiple pooling layers arranged in parallel. The feature fusion network includes an upsampling structure and a downsampling structure. The upsampling structure includes an upsampling layer and a feature fusion layer connected in sequence, and the downsampling structure includes a downsampling layer and a feature fusion layer connected in sequence. The head attention gate includes multiple convolutional layers and activation function layers; The detection head includes multiple convolutional layers; The clutter image is input into a trained object detection network model, where feature extraction, feature enhancement, multi-layer feature fusion, bounding box prediction and regression, and object classification are performed sequentially, including: The clutter image is input into the trained target detection network model. After passing through the convolutional layer, the first combination module, and the second combination module in the backbone feature extraction network, a first feature is output. The first feature is then processed by the third combination module to output a second feature. The second feature is then processed by the fourth combination module to output a feature to be enhanced. The feature to be enhanced is then processed by the pooling layer in the pyramid pooling network. Finally, the enhanced features processed by the multiple pooling layers are connected to output a third feature. The third feature, after being processed by the upsampling layer in the upsampling structure, is connected to the second feature to output an intermediate feature. The intermediate feature, after being processed by the upsampling layer in the upsampling structure, is connected to the first feature to obtain a first fused feature. The first fused feature, after being processed by the downsampling layer in the downsampling structure, is connected to the intermediate feature to output a second fused feature. The second fused feature, after being processed by the downsampling layer in the downsampling structure, is connected to the third feature to output a third fused feature. The first fused feature and the first feature are processed by the head attention gate to output a first feature matrix. The first feature matrix is ​​input into the corresponding detection head processing to output a target detection result under the first size feature. The second fused feature and the second feature are processed by the head attention gate to output a second feature matrix. The second feature matrix is ​​input into the corresponding detection head processing to output a target detection result under the second size feature. The third fused feature and the third feature are processed by the head attention gate to output a third feature matrix. The third feature matrix is ​​input into the corresponding detection head processing to output a target detection result under the third size feature.

2. The method for detecting small sea surface targets based on clutter maps and improved YOLOv4 according to claim 1, characterized in that, Training the initial object detection network model includes: Acquire clutter image data, construct a training dataset, and label the targets of the training samples in the training dataset to obtain the labels corresponding to the training dataset; The training dataset is input into the initial object detection network model for training. Within a preset number of iterations, the value of the loss function converges and reaches its minimum, thereby obtaining the parameters of the object detection network model and constructing the trained object detection network model.

3. The method for detecting small sea surface targets based on clutter maps and improved YOLOv4 according to claim 2, characterized in that, The acquisition of clutter image data and the construction of a training dataset include: Acquire radar echo data and convert the radar echo data into clutter image data; The clutter image data is augmented to obtain a training dataset; wherein, the image data augmentation includes rotation, cropping, displacement, blurring, and noise addition.

4. The method for detecting small sea surface targets based on clutter maps and improved YOLOv4 according to claim 1, characterized in that, The dense network connection module includes multiple convolutional layers. The dense network connection module contains multiple convolutional layers. The input to the nth convolutional layer is the nth convolutional layer The output of each convolutional layer .

5. The method for detecting small sea surface targets based on clutter maps and improved YOLOv4 according to claim 1, characterized in that, The head attention gate processing procedure includes: The first fused feature is input into the first convolutional layer of the head attention gate for processing, and the first feature to be processed is output. The first feature is input into the second convolutional layer of the head attention gate for processing, and the second feature to be processed is output. The first feature to be processed and the second feature to be processed are added together, and the result is input into the third convolutional layer of the head attention gate for processing. After processing by the activation function, the first weight coefficient is output. The first weight coefficient is multiplied by the first fused feature to output the first feature matrix. The second fused feature is input into the first convolutional layer of the head attention gate for processing, and the third feature to be processed is output. The second feature is input into the second convolutional layer of the head attention gate for processing, and the fourth feature to be processed is output. The third feature to be processed and the fourth feature to be processed are added together, and the result is input into the third convolutional layer of the head attention gate for processing. After processing by the activation function, the second weight coefficient is output. The second weight coefficient is multiplied by the second fused feature to output the second feature matrix. The third fused feature is input into the first convolutional layer of the head attention gate for processing, and the fifth feature to be processed is output. The third feature is input into the second convolutional layer of the head attention gate for processing, and the sixth feature to be processed is output. The fifth feature to be processed and the sixth feature to be processed are added together, and the result is input into the third convolutional layer of the head attention gate for processing. After processing by the activation function, the third weight coefficient is output. The third weight coefficient is multiplied by the third fused feature to output the third feature matrix.

6. The method for detecting small sea surface targets based on clutter maps and improved YOLOv4 according to claim 1, characterized in that, After the detection head outputs the target detection results under the first size feature, the target detection results under the second size feature, and the target detection results under the third size feature, the method further includes: The target detection results under the first size feature, the second size feature, and the third size feature are subjected to non-maximum suppression and detection box decoding respectively to obtain the target detection results of the clutter image.

7. The method for detecting small sea surface targets based on clutter maps and improved YOLOv4 according to claim 2, characterized in that, All convolutional layers with a kernel size of 3×3 are DWConv layers.

8. A small target detection device for the sea surface based on clutter maps and an improved YOLOv4, characterized in that, include: The data acquisition module is used to acquire radar echo data to be detected and convert the radar echo data to be detected into a clutter image; The data processing module is used to input the clutter image into the trained target detection network model, and sequentially perform feature extraction, feature enhancement, multi-layer feature fusion, target prediction box regression and target classification to obtain the target detection result of the clutter image; The trained target detection network model is obtained by training an initial target detection network model using clutter images with pre-annotated targets as the training dataset. The target detection network model includes a backbone feature extraction network, a pyramid pooling network, a feature fusion network, a head attention gate, and a detection head. The backbone feature extraction network extracts features, the pyramid pooling network enhances features, the feature fusion network fuses features, the head attention gate calculates weight coefficients, and the detection head performs target bounding box regression and target classification of different sizes. The backbone feature extraction network includes sequentially connected convolutional layers and multiple combination modules. The combination modules include dense network connection modules and transition connection modules. The dense network connection modules include multiple convolutional layers with sequential and skip connections, and the transition connection modules include sequentially connected double convolutional layers. The pyramid pooling network includes multiple pooling layers arranged in parallel. The feature fusion network includes an upsampling structure and a downsampling structure. The upsampling structure includes an upsampling layer and a feature fusion layer connected in sequence, and the downsampling structure includes a downsampling layer and a feature fusion layer connected in sequence. The head attention gate includes multiple convolutional layers and activation function layers; The detection head includes multiple convolutional layers; The clutter image is input into a trained target detection network model, where feature extraction, feature enhancement, multi-layer feature fusion, target bounding box prediction and regression, and target classification are performed sequentially, including: The clutter image is input into the trained target detection network model. After passing through the convolutional layer, the first combination module, and the second combination module in the backbone feature extraction network, a first feature is output. The first feature is then processed by the third combination module to output a second feature. The second feature is then processed by the fourth combination module to output a feature to be enhanced. The feature to be enhanced is then processed by the pooling layer in the pyramid pooling network. Finally, the enhanced features processed by the multiple pooling layers are connected to output a third feature. The third feature, after being processed by the upsampling layer in the upsampling structure, is connected to the second feature to output an intermediate feature. The intermediate feature, after being processed by the upsampling layer in the upsampling structure, is connected to the first feature to obtain a first fused feature. The first fused feature, after being processed by the downsampling layer in the downsampling structure, is connected to the intermediate feature to output a second fused feature. The second fused feature, after being processed by the downsampling layer in the downsampling structure, is connected to the third feature to output a third fused feature. The first fused feature and the first feature are processed by the head attention gate to output a first feature matrix. The first feature matrix is ​​input into the corresponding detection head processing to output a target detection result under the first size feature. The second fused feature and the second feature are processed by the head attention gate to output a second feature matrix. The second feature matrix is ​​input into the corresponding detection head processing to output a target detection result under the second size feature. The third fused feature and the third feature are processed by the head attention gate to output a third feature matrix. The third feature matrix is ​​input into the corresponding detection head processing to output a target detection result under the third size feature.

Images