Sea surface small target detection method and device based on clutter map and improved YOLOv4
By improving the YOLOv4 target detection network, combining the clutter graph and attention gate mechanism, the problem of small target detection on the sea surface is solved, the detection performance and efficiency are improved, and it is suitable for complex sea clutter environments.
Patent Information
- Application Number
- CN202411185139.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-08-27
AI Technical Summary
In complex sea clutter and strong noise environments, traditional sea surface target detection methods are difficult to effectively detect small targets, and are prone to missed and missed detection, and too many model parameters lead to degradation of detection performance.
Using the improved YOLOv4 object detection method based on clutter graph, the network is enhanced by enhancing the network's feature extraction ability for small targets, simplifying the network structure, reducing the amount of parameters and calculations, using DenseNet as the backbone feature extraction network, and introducing the header before the detection head The attention gate mechanism uses depth-by-depth convolution instead of standard convolution to improve detection performance.
It significantly improves the detection performance of small targets, reduces network parameters and calculation volume, improves training speed, and is suitable for actual engineering application needs, especially for small targets of ship types that are difficult to detect.
Smart Images

Figure CN119068338A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of target detection, and in particular relates to a method and a device for detecting small targets on a sea surface based on a clutter map and an improved YOLOv4. Background Art
[0002] In the task of sea surface target detection, the traditional target detection method is based on the theory of statistical signal processing, and the target detection is achieved by statistical modeling of echoes and binary hypothesis testing. On the one hand, for low-resolution sea clutter, the complex Gaussian model is usually used for modeling; for high-resolution sea clutter, the composite Gaussian model is usually used for modeling; on the other hand, for sea surface targets, point targets or range extended target models are usually used for modeling. These are parameterized models, and the echo data needs to be used to estimate the unknown model parameters. Considering the ease of implementation of the detection method, the model parameters should not be too many. However, due to the complex statistical characteristics of sea clutter in practical applications, it has the characteristics of non-Gaussian, non-stationary, and non-uniform, and a considerable number of sea surface targets have high-speed maneuverability and flickering, and the echo model is space-time varying, making it difficult to model and analyze the echo signal using a general statistical model. At this time, if a simple model is used to model the echo, model mismatch is likely to occur, resulting in a serious decline in detection performance.
[0003] In recent years, machine learning has achieved excellent results in many tasks with its powerful learning ability. In the field of object detection, machine learning and deep learning methods are mainly used to identify and process images or videos, and the category and position of objects are determined by classifying objects and regressing target frames to achieve the purpose of detection. Among them, convolutional neural network (CNN), as an important part of deep learning, has achieved good results in many image object detection tasks. At present, two-stage detection algorithms and single-stage detection algorithms are the two mainstream object detection algorithms based on deep learning algorithms at home and abroad. Among them, the two-stage detection algorithms mainly include R-CNN (Regions with Convolutional Neural Networks), Fast R-CNN (Fast Regions with Convolutional Neural Networks), Faster R-CNN (Faster Regions with Convolutional Neural Networks), etc., and the single-stage detection algorithms mainly include YOLO, SSD (Single Shot MultiBox Detector), etc. Since deep learning technology has a powerful ability to automatically learn features from data, relevant scholars have innovatively combined intelligent target detection technology with sea detection tasks, and have proposed many CNN-based intelligent sea surface target detection algorithms, which have improved the detection speed and accuracy, and have great advantages over traditional target detection algorithms.
[0004] Although experiments have verified that the CNN-based intelligent sea surface target detection algorithm can achieve good detection results, in a strong noise or complex clutter background environment, the echo signal of a small target is generally weak, and the detection difficulty increases sharply. The detection performance of the CNN-based intelligent sea surface target detection algorithm is also unsatisfactory, and it is very easy to miss or misdetect small targets, which needs further improvement. Summary of the invention
[0005] In order to solve the above problems existing in the prior art, the present invention provides a method and device for detecting small targets on the sea surface based on a clutter map and an improved YOLOv4. The technical problem to be solved by the present invention is achieved through the following technical solutions:
[0006] In a first aspect, the present invention provides a method for detecting small targets on the sea surface based on a clutter map and an improved YOLOv4, comprising:
[0007] Acquire radar echo data to be detected, and convert the radar echo data to be detected 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 target detection network model uses the cluttered images with pre-labeled targets as training data sets to train the initial target detection network model.
[0010] In a second aspect, the present invention further provides a sea surface small target detection device based on a clutter map and an improved YOLOv4, comprising:
[0011] A 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;
[0012] The data processing module is used to input the clutter image into the trained target detection network model, and perform feature extraction, feature enhancement, multi-layer feature fusion, target prediction box regression and target classification in sequence to obtain the target detection result of the clutter image;
[0013] The trained target detection network model uses the cluttered images with pre-labeled targets as training data sets to train the initial target detection network model.
[0014] Beneficial effects of the present invention:
[0015] The present invention provides a method and device for detecting small targets on the sea surface based on a clutter map and improved YOLOv4. Considering the difficulty in detecting weak small targets on the sea surface, the present invention starts from the perspective of the clutter map and is based on a YOLOv4 target detection network. The internal structure of the network is improved in a targeted manner, and the feature extraction capability of the network for small targets is enhanced. The detection performance of the improved network for small targets is significantly improved. At the same time, while retaining the detection accuracy, the present invention simplifies the structure of the network to a certain extent, greatly reduces the number of network parameters and the amount of calculation, and the training speed of the network is also improved, which is more in line with the needs of actual engineering applications.
[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 It is a flow chart of a method for detecting small targets on the sea surface based on a clutter map and improved YOLOv4 provided by an embodiment of the present invention;
[0018] Figure 2 is a schematic diagram of a target detection network model provided by an embodiment of the present invention;
[0019] Figure 3is a schematic diagram of a backbone feature extraction network provided by an embodiment of the present invention;
[0020] Figure 4 is a schematic diagram of a clutter map provided by an embodiment of the present invention;
[0021] Figure 5 It is a schematic diagram of the annotation of the clutter map provided by an embodiment of the present invention;
[0022] Figure 6 is a schematic diagram of a dense network connection module provided by an embodiment of the present invention;
[0023] Figure 7 is a schematic diagram of a head attention gate provided by an embodiment of the present invention;
[0024] Figure 8 is a schematic diagram of DWConv provided by an embodiment of the present invention;
[0025] Fig. 9 This is a schematic diagram of a comparison chart of target detection performance under different networks provided by an embodiment of the present invention;
[0026] Fig.10 It is a schematic diagram of a clutter image target detection result provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0027] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.
[0028] In view of the defects in the prior art, the present invention provides a method and device for detecting small targets on the sea surface based on a clutter map and an improved YOLOv4. Taking into account the difficulty in detecting weak and small targets on the sea surface, from the perspective of the clutter map, based on the YOLOv4 target detection network, the internal structure of the network is improved in a targeted manner, and the feature extraction capability of the network for small targets is enhanced. The detection performance of the improved network for small targets is significantly improved. At the same time, while retaining the detection accuracy, the present invention simplifies the structure of the network to a certain extent, greatly reduces the number of network parameters and the amount of calculation, and the training speed of the network is also improved, which is more in line with the needs of actual engineering applications.
[0029] See also Figure 1 , Figure 1 The present invention provides a flow chart of a method for detecting small sea targets based on a clutter map and an improved YOLOv4. The present invention provides a method for detecting small sea targets based on a clutter map and an improved YOLOv4, including:
[0030] S101, obtaining radar echo data to be detected, and converting the radar echo data to be detected into a clutter image.
[0031] S102, inputting the clutter image into the trained target detection network model, and sequentially performing feature extraction, feature enhancement, multi-layer feature fusion, target box prediction regression and target classification to obtain the target detection result of the clutter image.
[0032] The trained target detection network model uses the cluttered images with pre-labeled targets as training data sets to train the initial target detection network model.
[0033] Specifically, in this embodiment, see Figure 2 and Figure 3 , Figure 2 is a schematic diagram of a target detection network model provided by an embodiment of the present invention, Figure 3 is a schematic diagram of a backbone feature extraction network provided by an embodiment of the present invention, and before training the initial target detection network model, it also includes:
[0034] Build a target detection network model;
[0035] Among them, 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 strengthen 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 a plurality of combination modules, the combination module includes a dense network connection module and a transition connection module, the dense network connection module includes sequentially connected and skip-connected multi-layer convolutional layers, and the transition connection module includes sequentially connected double-layer convolutional layers;
[0037] The pyramid pooling network includes 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 includes convolutional layer and activation function layer;
[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 is different from the residual network (Residual Network, ResNet) in that it is a more densely connected network, which can maximize the feature reuse rate by establishing connection channels between the layers inside the network; the connection relationship between the layers inside DenseNet and the overall structure of the DenseNet network are as follows Figure 3 As shown in the figure, DenseBlock and Transition are the basic components of DenseNet.
[0042] In this embodiment, the initial target detection network model is trained, including:
[0043] Obtain clutter image data, construct a training data set, and annotate the objects of the training samples in the training data set to obtain the labels corresponding to the training data set. Figure 4 and Figure 5 , Figure 4 is a schematic diagram of a clutter map provided by an embodiment of the present invention, Figure 5 is a schematic diagram of the annotation of the clutter image provided by the embodiment of the present invention; the data set is expanded and the annotation information of the target frame is improved by image enhancement; it can be understood that the clutter image data after expansion is manually annotated with targets one by one using the Lableimg software, and the main target types include two types: ships and islands, such as Figure 5 As shown in the figure, the ship-type target occupies a very small proportion of the entire image and its features are not obvious. Coupled with the interference from the surrounding environment, it is difficult to detect.
[0044] The training data set and the labels corresponding to the training data set are input into the initial target detection network model for training. Within the preset number of iterations, the value of the loss function converges and reaches the minimum, the parameters of the target detection network model are obtained, and the trained target detection network model is constructed.
[0045] In this embodiment, clutter image data is obtained and a training data set is constructed, including:
[0046] Acquire radar echo data, and convert the radar echo data into training clutter image data;
[0047] The training clutter image data is subjected to image data enhancement to obtain a training data set; wherein the image data enhancement includes rotation, cropping, displacement, 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 prediction box regression and target classification are performed in sequence, including:
[0049] The clutter image is input into the trained target detection network model, and is processed by the convolution layer, the first combination module and the second combination module in the backbone feature extraction network in sequence, and the first feature is output. The first feature is processed by the third combination module, and the second feature is output. The second feature is processed by the fourth combination module, and the feature to be enhanced is output. The feature to be enhanced is processed by the pooling layer in the pyramid pooling network, and then the enhanced features processed by multiple pooling layers are 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 the 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 the head attention gate to output the 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 the head attention gate to output the 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 the head attention gate to output the 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, see Figure 6 , Figure 6 It 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, and i≥2.
[0053] In this embodiment, see Figure 7 , Figure 7 It is a schematic diagram of the head attention gate provided by an embodiment of the present invention. The main idea of the YOLO network is hierarchical detection. It first extracts the features x3, x4 and x5 of different receptive field branches, and then fuses them to obtain the fused effective feature layers P3, P4 and P5. Finally, the three detection heads are responsible for the detection of targets of different sizes respectively. The results of the three detection heads are then used as the final detection results after non-maximum suppression and detection frame decoding. Figure 7As shown in the figure, P3 and YOLO Head_1 are mainly responsible for small-sized target detection, while P4 and YOLO Head_2 as well as P5 and YOLO Head_5 are mainly responsible for medium-sized and large-sized target detection. Based on the above principles, it is considered to introduce an attention gate mechanism before the detection head, and to achieve effective attention distinction of different detection heads to feature information of different sizes by adding an attention weight to the feature layer of the input detection head; specifically, the three effective feature layers P3, P4 and P5 output by the feature fusion network have integrated features of different scales to a certain extent, which has increased the global effectiveness of the feature layer to a certain extent, but compared with the original three effective feature layers x3, x4 and x5, the unique feature information of their respective scales has the potential risk of being weakened, which will have an adverse effect on target detection. In order to solve the above problems, we propose to re-allocate the attention weights before the final detection, and the basis for weight allocation is the original effective feature layer x3, x4 or x5 of the same size. We call this the head attention gate mechanism (HeadAttention Gate, Head_AG), and the principle diagram of Head_AG is shown in the figure below. Figure 7 shown.
[0054] The processing of the head attention gate includes:
[0055] Input the first fused feature to the first convolution layer in the head attention gate for processing, output the first feature to be processed, input the first feature to the second convolution layer in the head attention gate for processing, output the second feature to be processed, add the first feature to be processed and the second feature to be processed, input the addition result to the third convolution layer in the head attention gate for processing, and then process it through the activation function to output the first weight coefficient, multiply the first weight coefficient by the first fused feature, and output the first feature matrix;
[0056] Input the second fused feature to the first convolution layer in the head attention gate for processing, output the third feature to be processed, input the second feature to the second convolution layer in the head attention gate for processing, output the fourth feature to be processed, add the third feature to be processed and the fourth feature to be processed, input the addition result to the third convolution layer in the head attention gate for processing, and then process it through the activation function to output the second weight coefficient, multiply the second weight coefficient by the second fused feature, and output the second feature matrix;
[0057] The third fused feature is input into the first convolution layer in the head attention gate for processing, and the fifth feature to be processed is output; the third feature is input into the second convolution layer in 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, and the addition result is input into the third convolution layer in the head attention gate for processing, and then processed by the activation function to output the third weight coefficient; the third weight coefficient is multiplied by the third fused feature to output the third feature matrix.
[0058] It is understandable that if Figure 7 As shown, x is the input signal, which is the first fusion feature, the second fusion feature and the third fusion feature, respectively; g is the gate signal, which is the first feature, the second feature and the third feature, respectively. The two are input into the attention gate to calculate the attention weight coefficient, and then the calculated weight coefficient is multiplied by the input signal to obtain the feature matrix after the final redistribution of attention.
[0059] In this embodiment, after the detection head outputs the target detection result under the first size feature, the target detection result under the second size feature, and the target detection result under the third size, the method further includes:
[0060] 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 are respectively subjected to non-maximum suppression and detection frame decoding to obtain the target detection results of the clutter image.
[0061] In this embodiment, see Figure 8 , Figure 8 is a schematic diagram of DWConv provided in an embodiment of the present invention, and the convolution layers with a convolution kernel size of 3×3 are all DWConv layers.
[0062] It can be understood that the 3×3 convolution parts not shown in the backbone feature extraction network and the feature fusion network are all DWConv layers provided in this embodiment, which can not only achieve the same feature extraction effect as the 3×3 standard convolution, but also reduce the network parameters and calculation amount to a certain extent.
[0063] The full name of DWConv is Depthwise Convolution. Its core idea is to perform group convolution in the channel dimension. Different convolution kernels perform their respective duties and are responsible for different feature layer outputs. When the number of group convolutions is equal to the number of input channels, the number of output channels is equal to the number of input channels. Compared with standard convolution, DWConv reduces the number of convolution operations, so the number of network parameters and the amount of calculation are also greatly reduced. At the same time, in the calculation of neural networks, in order to achieve the same output effect as standard convolution, DWConv is usually used in combination with a convolution layer with a convolution kernel size of 1×1. The former is used to change the size of the feature map and the latter is used to adjust the feature map to the ideal channel dimension output. The principle structure of DWConv is as follows: Figure 8 Generally, any ordinary convolution layer in a neural network with a convolution kernel other than 1 can be replaced by the corresponding DWConv and 1×1 convolution, which can achieve the effect of lightweight network structure.
[0064] In summary, the method for detecting small targets on the sea surface based on clutter map and improved YOLOv4 provided by the present invention has the following beneficial effects:
[0065] 1. The present invention improves the backbone feature extraction network of YOLOv4, and uses DenseNet instead of ResNet as the backbone feature extraction network. Compared with the original residual network, the dense jump connection inside it can better retain the feature information of the target; at the same time, the present invention improves the 3×3 standard convolution in the YOLOv4 network, and replaces it by first performing depth-by-depth convolution and then adjusting the feature dimension. This can not only achieve the same feature extraction effect as the 3×3 standard convolution, but also reduce the network parameters and calculation amount to a certain extent.
[0066] 2. The present invention proposes a Head_AG structure, which re-differentiates the attention of the three effective feature layers by adding a gating signal before the final target detection, so as to compensate for the risk of weakening the specific scale features that may be brought about by the global feature fusion, so that the detection heads responsible for different sizes can pay more attention to the feature information at this size, thereby achieving better target detection effect.
[0067] In an optional embodiment of the present invention, the effect of the sea surface small target detection method based on the clutter map and the improved YOLOv4 provided in the above embodiment is verified by simulation experiments, specifically:
[0068] See also Fig. 9 and Fig.10 , Fig. 9 is a schematic diagram of a comparison chart of target detection performance under different networks provided by an embodiment of the present invention, Fig.10 is a schematic diagram of a clutter map target detection result provided by an embodiment of the present invention, from Fig. 9It can be seen that the improved YOLOv4 target detection network proposed in this invention performs best in various evaluation indicators. For island type targets, Figure 5 It can be seen that the characteristics are relatively obvious, and the detection effects of each detector are not much different; for small targets of the ship type, due to its greater detection difficulty, the detection performance of each detector is quite different. The improved YOLOv4 target detection network proposed in this invention can better retain the feature information of small targets and make up for the loss of key information of small targets to a certain extent. Fig. 9 As shown, the present invention achieves the maximum value in the average precision (AP) of the ship type, which means that the detector proposed by the present invention has the best detection performance for small targets among all the above detectors.
[0069] This embodiment also compares the parameter amount, calculation amount and training time of the present invention with several common target detection networks. The comparison of the parameter amount, calculation amount and running time of different networks is shown in Table 1. The improved YOLOv4 target detection network proposed in the present invention simplifies the internal structure of the network. It can be seen from Table 1 that the present invention has fewer parameters, calculation amount and network running time than other common target detectors, which is more conducive to the realization of practical engineering applications.
[0070] Table 1
[0071]
[0072]
[0073] Based on the same inventive concept, the present invention also provides a sea surface small target detection device based on a clutter map and an improved YOLOv4, which is used to implement the method provided by the above embodiment of the present invention. The method embodiment is not repeated here. The device includes:
[0074] A 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;
[0075] The data processing module is used to input the clutter image into the trained target detection network model, and perform feature extraction, feature enhancement, multi-layer feature fusion, target prediction box regression and target classification in sequence to obtain the target detection result of the clutter image;
[0076] The trained target detection network model uses the cluttered images with pre-labeled targets as training data sets to train the initial target detection network model.
[0077] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the term "include", "comprise" or any other variant is intended to cover non-exclusive inclusion, so that the article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed. In the absence of more restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the article or device including the elements. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The orientation or position relationship indicated by "up", "down", "left", "right", etc. is based on the orientation or position relationship shown in the accompanying drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention.
[0078] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.
[0079] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.
Claims
1. A method for detecting small targets on the sea surface based on clutter map and 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; Inputting the clutter image into a trained target detection network model, performing feature extraction, feature enhancement, multi-layer feature fusion, target prediction box regression and target classification in sequence, and obtaining a target detection result of the clutter image; The trained target detection network model is obtained by training an initial target detection network model using a number of clutter images with target annotations completed in advance as training data sets.
2. The method for detecting small targets on the sea surface based on clutter map and improved YOLOv4 according to claim 1, characterized in that: Before training the initial target detection network model, it also includes: Build a target detection network model; 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 strengthen 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. The backbone feature extraction network includes sequentially connected convolutional layers and a plurality of combination modules, the combination module includes a dense network connection module and a transition connection module, the dense network connection module includes multiple convolutional layers that are sequentially connected and skip-connected, and the transition connection module includes a double-layer convolutional layer that is sequentially connected; The pyramid pooling network includes multiple pooling layers arranged in parallel; The feature fusion network includes an upsampling structure and a downsampling structure, wherein 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 a plurality of convolutional layers.
3. The method for detecting small targets on the sea surface based on clutter map and improved YOLOv4 according to claim 2, characterized in that: Training the initial target detection network model includes: Acquire clutter image data, construct a training data set, and mark the targets of training samples in the training data set to obtain labels corresponding to the training data set; The training data set is input into the initial target detection network model for training. Within a preset number of iterations, the value of the loss function converges and reaches a minimum, the parameters of the target detection network model are obtained, and a trained target detection network model is constructed.
4. The method for detecting small targets on the sea surface based on clutter map and improved YOLOv4 according to claim 3, characterized in that: The obtaining of clutter image data and constructing a training data set includes: Acquire radar echo data, and convert the radar echo data into clutter image data; The clutter image data is subjected to image data enhancement to obtain a training data set; wherein the image data enhancement includes rotation, cropping, displacement, blurring and noise addition.
5. The method for detecting small targets on the sea surface based on clutter map and improved YOLOv4 according to claim 3, characterized in that: The clutter image is input into the trained target detection network model, and feature extraction, feature enhancement, multi-layer feature fusion, target box prediction regression and target classification are performed in sequence, including: The clutter image is input into the trained target detection network model, and after being processed by the convolution 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 processed by the third combination module, a second feature is output, and the second feature is processed by the fourth combination module to output a feature to be enhanced, the features to be enhanced are respectively processed by the pooling layers in the pyramid pooling network, and then the enhanced features processed by the multiple pooling layers are connected to output a third feature; After the third feature is processed by the upsampling layer in the upsampling structure and connected with the second feature, an intermediate feature is output; after the intermediate feature is processed by the upsampling layer in the upsampling structure and connected with the first feature, a first fused feature is obtained; after the first fused feature is processed by the downsampling layer in the downsampling structure and connected with the intermediate feature, a second fused feature is output; after the second fused feature is processed by the downsampling layer in the downsampling structure and connected with the third feature, a third fused feature is output; 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 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 the 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 the 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.
6. The method for detecting small sea targets based on clutter map and improved YOLOv4 according to claim 5, characterized in that: The dense network connection module includes multiple convolutional layers, and 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, i≥2.
7. The method for detecting small targets on the sea surface based on clutter map and improved YOLOv4 according to claim 5, characterized in that: The processing process of the head attention gate includes: Input the first fused feature to the first convolution layer in the head attention gate for processing, output the first feature to be processed, input the first feature to the second convolution layer in the head attention gate for processing, output the second feature to be processed, add the first feature to be processed and the second feature to be processed, input the addition result to the third convolution layer in the head attention gate for processing, and then process it through the activation function to output the first weight coefficient, multiply the first weight coefficient by the first fused feature, and output the first feature matrix; Input the second fused feature to the first convolution layer in the head attention gate for processing, output the third feature to be processed, input the second feature to the second convolution layer in the head attention gate for processing, output the fourth feature to be processed, add the third feature to be processed and the fourth feature to be processed, input the addition result to the third convolution layer in the head attention gate for processing, and then process it through the activation function to output the second weight coefficient, multiply the second weight coefficient by the second fused feature, and output the second feature matrix; The third fused feature is input into the first convolution layer in the head attention gate for processing, and the fifth feature to be processed is output; the third feature is input into the second convolution layer in 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, and the addition result is input into the third convolution layer in the head attention gate for processing, and then processed by the activation function to output the third weight coefficient; the third weight coefficient is multiplied by the third fused feature, and the third feature matrix is output.
8. The method for detecting small targets on the sea surface based on clutter map and improved YOLOv4 according to claim 5, characterized in that: After the detection head outputs the target detection result under the first size feature, the target detection result under the second size feature, and the target detection result under the third size, the method further includes: Non-maximum suppression and detection frame decoding are performed on the target detection result under the first size feature, the target detection result under the second size feature, and the target detection result under the third size, respectively, to obtain the target detection result of the clutter image.
9. The method for detecting small targets on the sea surface based on clutter map and improved YOLOv4 according to claim 3, characterized in that: The convolution layers with a convolution kernel size of 3×3 are all DWConv layers.
10. A small target detection device on the sea surface based on clutter map and improved YOLOv4, characterized in that: include: A data acquisition module, used for acquiring radar echo data to be detected, and converting the radar echo data to be detected into a clutter image; A data processing module is used to input the clutter image 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 a target detection result of the clutter image; The trained target detection network model is obtained by training the initial target detection network model using the clutter images with pre-labeled targets as training data sets.
Citation Information
Patent Citations
Sea-surface infrared small object detection method
CN105931264A
Improved Yolo underwater image small target and occlusion target detection method
CN112733929A
Road surface information identification method and system based on YOLO v4
CN113255524A
SAR image ship target detection method based on multistage feature fusion and mixed attention
CN116469002A
Substation grounding wire state target detection method based on improved YOLOv5
CN117253188A
