Ship target detection method for synthetic aperture radar

Abstract

The invention relates to the technical field of ship detection, in particular to a ship target detection method for a synthetic aperture radar, and the ship target detection method comprises the steps: taking collected synthetic aperture radar data as a detection data source, and taking a target detection neural network as a detection means to detect a ship, the target detection neural network is an improved neural network based on YOLOv8 and comprises a backbone network, a neck network and a detection network, and the backbone network at least comprises a convolutional layer, a C2f structure, an ADown structure and an ISPF structure; according to the method, the ADown structure is introduced into the backbone network of the YOLOv8, the parameter quantity and the calculation quantity of the model can be reduced, deployment of the model on resource-constrained equipment is facilitated, and meanwhile, a new network structure ISPF is provided, so that the problem of detail information loss caused by the SPPF in the YOLOv8 can be relieved.

Description

Technical Field

[0001] The present invention relates to the technical field of ship detection, and in particular to a method for detecting ship targets using a synthetic aperture radar. Background Art

[0002] Ship target detection is a key part of maritime monitoring. The most commonly used technical means for ship target detection is synthetic aperture radar (SAR). SAR generates high-resolution images by emitting electromagnetic waves and receiving returned echo signals. SAR has strong penetration and can obtain information in complex weather scenes such as cloud cover, rain and fog. It is not affected by complex weather, lighting and other factors. SAR can work all day and all weather. Therefore, SAR images can be used to realize ship target recognition.

[0003] Traditional SAR ship detection usually uses binarization, threshold segmentation and other methods to separate the target to be detected from the background image, and then detects the target through the geometric and texture features of the target. In simple scenes, the above methods can achieve good results. However, in bad weather or complex environments, due to the presence of more background interference in the image, texture features need to be added, but texture features will be affected by resolution. In addition, traditional recognition methods require manual feature design. When the image background noise is too large, manual features are difficult to establish, which makes ship recognition more difficult.

[0004] With the continuous development of deep learning, ship target detection algorithms have gradually shifted from traditional methods to target detection methods based on deep learning. Deep learning-based methods no longer require manual design of features, and their deep models can automatically learn effective feature representations from data. This shift has greatly promoted the development of ship target detection technology.

[0005] In the ship detection method based on deep learning, the YOLO series of algorithms are widely used. YOLOv8 is the eighth version of the YOLO series, which mainly consists of three parts: the backbone network, the neck network and the detection network. The backbone network uses the improved CSPDarkNet structure, which consists of convolutional layers, C2f structure and fast pyramid pooling (Spatial Pyramid PoolingFast, SPPF); the neck network part adopts the path aggregation network-feature pyramid network to achieve more effective feature fusion and enhance the model's detection ability for targets of different sizes; the detection network adopts the idea of ​​no anchor box to separate the regression branch and the classification branch, making the network training and reasoning more efficient.

[0006] However, YOLOv8 uses convolutional layers in the backbone network to achieve downsampling. The structure consists of a batch normalization layer, a convolution operation, and an activation function. The structure has a high structural complexity, and the features learned by the structure are relatively simple, which is not conducive to target detection. At the same time, the last structure of the backbone network uses the SPPF structure, which is composed of three maximum poolings in series and a skip connection. By performing a series of pooling operations on the input feature map, it can effectively capture the receptive field information of different scales. Subsequently, by fusing features of different scales, it can effectively improve the detection capability of target feature information of different scales. However, when the feature map is processed using the maximum pooling operation, due to the reduction in resolution during the processing, some detail information of the image will be lost, thereby reducing the detection accuracy. Summary of the invention

[0007] In view of this, the purpose of the present invention is to provide a method for synthetic aperture radar ship target detection, which introduces ADown structure and ISPPF structure in the backbone network, while reducing the number of parameters and calculation amount of the model and alleviating the problem of detail information loss caused by SPPF.

[0008] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a method for synthetic aperture radar ship target detection, the ship target detection method uses collected synthetic aperture radar data as the detection data source, and uses a target detection neural network as the detection means to detect the ship, wherein the target detection neural network is a neural network improved based on YOLOv8, including: a backbone network, a neck network and a detection network, wherein the backbone network includes: a convolutional layer, a C2f structure, an ADown structure and an ISPPF structure.

[0009] In particular, the algorithm flow of the ADown structure is as follows: the input data is divided into two branches after average pooling, one branch performs a convolution operation based on the pooled data, and the other branch performs a maximum pooling operation based on the pooled data and then performs a convolution operation. Finally, the data processed by the two branches are fused by Concat and output.

[0010] In particular, the algorithm flow of the ISPPF structure is: (1) Input data X is average pooled and output is X avg , the calculation formula is: ,in, Indicates OK The columns need to be averaged in the pooling area. Number, Indicates the number of elements in the average pooling area; (2) Output data X after average pooling avgDivide into X in a 1:1 ratio 1 With X 2 Two branches, , where Split represents the splitting function; (3) X 1 After two convolutions, the output is X 1b , the calculation formula is: , , Among them, X 1a represents the output after the first convolution, k represents the convolution kernel size, s represents the step size, and d represents the dilation rate; (4) X 2 After pooling and skip connection feature extraction, the second branch calculation result X is output 3 , the formula is: , , , Among them, X 2a Represents X 2 The output after pooling, X 2b Represents X 2a The output after pooling, Skipconnection represents skip connection; (5) The first branch outputs the result X 1b And the second branch outputs the result X 3 After Concat fusion, the output is the output Y after ISPPF operation. .

[0011] Preferably, the training method of the model when the target detection neural network is used as the detection means to detect the ship comprises the following steps: S1, data acquisition, collecting images of ships through synthetic aperture radar; S2. Data labeling: Use the LabelImg tool to label the collected images, and divide the labeled images into training set and validation set in a ratio of 8:2; S3, input the divided image data set into the improved network for training; S4. Use the trained model to detect images with ship targets. The specific calculation method is as follows:

[0012] ,

[0013] ,

[0014] ,

[0015] ,

[0016] Among them, TP represents the number that the model can correctly identify, FP represents the number that is incorrectly identified by the model, FN represents false detection, and Q represents the number of categories.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. The present invention introduces the ADown structure into the backbone network of YOLOv8, which can reduce the number of parameters and the amount of calculation of the model, and is conducive to the deployment of the model on resource-constrained devices.

[0019] 2. The present invention proposes a new network structure ISPPF, which can alleviate the problem of detail information loss caused by SPPF in YOLOv8. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a schematic diagram of the improved network framework of the present invention; Figure 2 ADown structure schematic diagram of the present invention; Figure 3 It is a schematic diagram of the ISPPF structure of the present invention; Figure 4 The present invention is an implementation flow chart. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] The present invention proposes a method for synthetic aperture radar ship target detection based on YOLOv8, which consists of a backbone network, a neck network and a detection network. The backbone network consists of a convolutional layer, C2f, ADown and ISPPF; the neck network consists of an upsampling layer, Concat, C2f and a convolutional layer; the detection network consists of a convolutional layer. The overall structure is as follows Figure 1 As shown, Input represents input, Conv represents convolutional layer, C2f represents bottleneck layer, ADown is the introduced downsampling structure, ISPPF is the enhanced pyramid pooling structure proposed in the present invention, Upsamples represents upsampling, Concat represents merging, and Detect represents detection.

[0023] In the backbone network, the input image first passes through two Conv layers to extract the preliminary features of the image. The Conv layer consists of convolution operations, batch normalization, and SiLU activation functions. The feature map after the Conv layer can be expressed as: , Among them, x represents the sample data input, y represents the corresponding feature processing output of this layer, conv represents the convolution operation, bn represents batch normalization, and act represents the activation function. Batch normalization is expressed as: , , , , in, is the i-th sample in the batch, is the standardized sample, is the batch size, is the mean, is the variance, is a constant used to avoid division by zero, and is the learning parameter. The SiLU activation function is expressed as: , in, is a Sigmoid function. After the Conv layer, it passes through a C2f module to further extract and enhance features. The C2f module optimizes the gradient flow and enhances the performance of the model. Then it passes through an ADown module followed by a C2f module to extract and enhance features and align the P3 layer of the feature pyramid. The structure of ADown is as follows: Figure 2 As shown, Figure 2 Avgpool in the above code represents average pooling, Conv represents convolutional layer, Maxpool represents maximum pooling, and Concat represents merging. It simplifies the model structure by reducing the model parameters, thereby improving the model's computational efficiency. Although ADown is used to reduce the size of feature maps, it also fully considers retaining image information as much as possible during design to ensure that the model can accurately identify the target. After ADown, it passes through C2f to extract and enhance features, corresponding to the P4 layer of the feature pyramid.

[0024] Finally, after ADown and C2f, and then ISPPF, the feature extraction of the input image is completed.

[0025] The structure of ISPPF is as follows Figure 3As shown in the figure, Avgpool means average pooling, Split means splitting, Conv means convolution layer, Maxpool means maximum pooling, and Concat means merging. The input feature map is first average pooled, and the formula is as follows: , in, represents the output after average pooling, Indicates OK The columns need to be averaged in the pooling area. Number, Indicates the number of elements in the area that need to be averaged.

[0026] Then the input feature map is divided into and Two-part feature map: , Feature Map On the first branch, the features of the image are extracted by passing through a convolution layer with a convolution kernel size of 3, a step size of 1, and a dilation rate of 2 and a convolution layer with a convolution kernel size of 3, a step size of 1, and a dilation rate of 3. represents the size of the convolution kernel, Indicates the size of the step, represents the void ratio, represents the output after the first convolutional layer, Represents the output after the second convolutional layer: , , Feature Map After the second branch, more abstract features are further extracted through two maximum pooling and skip connections. Finally, the two branches are merged to obtain a richer detail feature map, alleviating the problem of partial detail loss in SPPF. represents the output after the first maximum pooling operation, represents the output after the second maximum pooling operation, represents the final output of the second branch, Indicates the output after ISPPF.

[0027] , , , The neck network consists of a path aggregation network-feature pyramid network, which is used to fuse features of different scales and sizes. The path aggregation network upsamples feature maps from different levels and then fuses them with features from other levels. This fusion can retain low-level details and high-level semantic information at the same time. The feature pyramid network constructs and fuses multiple feature representations at different feature levels to enhance the model's understanding of images at multiple scales.

[0028] In the detection network, the idea of ​​decoupling is utilized and two parallel branches are used to independently extract category and position features, so as to separate classification and detection and detect and identify ship targets.

[0029] The complete training and testing process requires Figure 4 Several steps are shown:

[0030] (1) Data acquisition: Collect images of ships through synthetic aperture radar.

[0031] (2) Data labeling: Use the LabelImg tool to label the collected images and divide the labeled images into training and validation sets in a ratio of 8:2.

[0032] (3) The divided image dataset is input into the improved network for training. During training, the input image size is set to 640, the batch_size is set to 64, the initial learning rate is set to 0.01, and a total of 200 rounds of training are performed.

[0033] (4) Use the trained model to detect images containing ship targets.

[0034] Finally, the present invention randomly selects 3000 images from the public dataset sar_ship_dataset to construct a small dataset, and verifies the performance from the aspects of precision (Percision, P), recall (Recall, R), mean average precision (mean Average Percision, mAP), parameter quantity (Params), floating point operations per second (Floating Point Operations PerSecond, FLOPs), etc. The specific calculation method is shown in the formula.

[0035] , , , , Among them, TP represents the number that the model can correctly identify, FP represents the number that is incorrectly identified by the model, FN represents false detection, and Q represents the number of categories.

[0036] The performance of the ISPPF proposed in the present invention is compared with some existing improved pyramid pooling schemes. The comparison results are shown in the following table:

[0037] As can be seen from the table, the accuracy of the present invention is improved by 0.2% compared with the original YOLOv8, and the mAP@0.5 is improved by 0.4% compared with the original YOLOv8, which is the highest among all the comparison schemes. In addition, the present invention also conducts an ablation experiment on the improved network to verify the effectiveness of the network.

[0038]

[0039] As shown in the table, the improved network improves mAP@0.5 by 0.8% compared to the original YOLOv8, and the number of parameters, computational complexity, and model size are all reduced by 5%.

[0040] In summary, the SAR ship image detection method based on improved YOLOv8 proposed in the present invention can reduce the number of model parameters and the amount of calculation compared with the traditional method, so that the model can be more conveniently deployed on resource-constrained devices. At the same time, it can alleviate the problem of detail information loss caused by SPPF and improve the detection accuracy.

[0041] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting ship targets using synthetic aperture radar, characterized in that: The ship target detection method uses the collected synthetic aperture radar data as the detection data source and uses the target detection neural network as the detection means to detect the ship, wherein: The target detection neural network is a neural network improved based on YOLOv8, including: a backbone network, a neck network and a detection network, wherein the backbone network at least includes: Convolutional layer, C2f structure, ADown structure and ISPPF structure; The algorithm flow of the ADown structure is as follows: the input data is divided into two branches after average pooling, one branch performs a convolution operation based on the pooled data, and the other branch performs a maximum pooling operation based on the pooled data and then performs a convolution operation. Finally, the data processed by the two branches are fused by Concat and output.

2. The method for synthetic aperture radar ship target detection according to claim 1, characterized in that: The neck network consists of at least upsampling layers, Concat, C2f, and convolutional layers.

3. The method for synthetic aperture radar ship target detection according to claim 2, characterized in that: The detection network consists of convolutional layers.

4. The method for synthetic aperture radar ship target detection according to claim 3, characterized in that: In the backbone network, the input image first passes through two Conv layers to extract the preliminary features of the image. The Conv layer consists of convolution operations, batch normalization, and SiLU activation functions. The feature map is represented by the Conv layer: , where x represents the sample data input, y represents the corresponding feature processing output of this layer, conv represents the convolution operation, bn represents batch normalization, act represents the activation function, and batch normalization is expressed as: , ， ， ， in, is the i-th sample in the batch, is the standardized sample, is the batch size, is the mean, is the variance, is a constant used to avoid division by zero, and is the learning parameter, and the SiLU activation function is expressed as: ， in, It is a Sigmoid function. After passing through the Conv layer, it passes through a C2f module to extract and enhance features. The C2f module enhances the performance of the model by optimizing the gradient flow. Then it passes through an ADown followed by a C2f module to extract and enhance features and align the P3 layer of the feature pyramid.

5. The method for synthetic aperture radar ship target detection according to claim 4, characterized in that: The algorithm flow of the ISPPF structure is as follows: (1) Input data X is average pooled and output is X avg , the calculation formula is: ,in, Indicates OK The columns need to be averaged in the pooling area. Number, Indicates the number of elements in the average pooling area; (2) Output data X after average pooling avg It is divided into two branches, X1 and X2, in a 1:1 ratio. , where Split represents the splitting function; (3) After two convolutions, X1 outputs X 1a and X 1b , the calculation formula is: ， ， Among them, X 1a represents the output after the first convolution, X 1b represents the output after the second convolution, k represents the convolution kernel size, s represents the step size, and d represents the dilation rate; (4) After pooling and skip connection feature extraction, X2 outputs the second branch calculation result X3, the formula is: ， ， ， Among them, X 2a represents the output after X2 pooling, X 2b Represents X 2a The output after pooling, Skipconnection represents skip connection; (5) The first branch outputs the result X 1b After being fused with the output result X3 of the second branch through Concat, the output Y after the ISPPF operation is output. .

6. The method for synthetic aperture radar ship target detection according to claim 5, characterized in that: The training method of the model when the target detection neural network is used as the detection means to detect the ship includes the following steps: S1, data acquisition, collecting images of ships through synthetic aperture radar; S2. Data labeling: Use the LabelImg tool to label the collected images, and divide the labeled images into training set and validation set in a ratio of 8:2; S3, input the divided image data set into the improved network for training; S4. Use the trained model to detect images with ship targets. The specific calculation method is as follows: ， ， ， ， Among them, TP represents the number that the model can correctly identify, FP represents the number that is incorrectly identified by the model, FN represents false detection, and Q represents the number of categories.

Images