CN116343045B - Lightweight SAR image ship target detection method based on YOLO v5 - Google Patents

Abstract

The invention discloses a lightweight SAR image ship target detection method based on YOLO v5, which comprises the following steps: acquiring an SAR image dataset, obtaining an SAR image ship target simulation dataset and a public SAR ship detection dataset through simulation imaging, preprocessing the dataset, and dividing the dataset into a training sample set and a test sample set; establishing an improved lightweight YOLO v5 model; inputting the training data set into an improved lightweight YOLO v5 model for training, and obtaining a trained improved lightweight YOLO v5 model; and inputting the test data set into a trained improved lightweight YOLO v5 model to obtain a detection and identification result. The improved lightweight YOLO v5 model provided by the invention can more accurately identify ships in SAR images, greatly reduce the model size and test time, and remarkably improve the detection precision.

Description

Lightweight SAR image ship target detection method based on YOLO v5

Technical field

The invention belongs to the technical field of radar target detection, and specifically relates to a lightweight SAR image ship target detection method based on YOLO v5.

Background technique

Radar image target detection is an area of research concern for many scholars in recent years. According to the stage of generating candidate frames, mainstream target detection algorithms based on deep learning include single-stage and two-stage. The single-stage target detection algorithm has faster detection speed, but the detection accuracy is relatively low. Recently, with the addition of some improvement schemes, its detection accuracy has been greatly improved and even surpassed the two-stage model.

The YOLO series is a representative of single-stage detection models, and the construction of the model can be completed only through end-to-end training. Due to its real-time advantages, YOLO has gradually become a research focus in the field of radar image target detection. The target detection model of the YOLO series has become more and more powerful with the introduction of YOLO v5. YOLO v5 has the highest inference speed and a very lightweight model size, so YOLO v5 was chosen as the detection framework. Therefore, this invention is improved based on the YOLO v5 algorithm and applied to lightweight SAR image target detection.

In practical applications, ship detection from SAR images often faces complex sea scenes, with ships of various sizes, near-shore environmental interference, and serious coherent noise and background interference. Most of the deep learning networks used in target detection algorithms are very complex, with a large amount of parameters and calculations. The generated models occupy a large amount of memory, which brings difficulties to ocean monitoring. However, lightweight models often cause a significant drop in detection accuracy, so a more efficient target detection algorithm needs to be designed to meet real-time performance.

Contents of the invention

The purpose of this invention is to solve the problems of existing lightweight SAR image detection and recognition methods in different complex sea scenes, such as poor feature generalization ability, greatly reduced ship recognition rate, and insufficient coast and ship feature learning, resulting in the inability to effectively distinguish the coast. In order to solve the technical shortcomings of information and ship target information, a lightweight SAR image ship target detection method based on YOLO v5 is proposed.

The technical solution to achieve the purpose of the present invention is: First, the present invention provides a lightweight SAR image ship target detection method based on YOLO v5, which includes the following steps:

Step 1. Obtain the SAR image data set: obtain the SAR image ship target simulation data set and the public SAR ship detection data set through simulation imaging, preprocess the SAR image and divide it into a training data set and a test data set;

Step 2. Establish an improved lightweight YOLO v5 model, that is, replace the backbone network of YOLO v5 with the RepVGG network structure, and jointly improve it with the representation enhancement module and the feature attention module. The representation enhancement module is improved based on asymmetric convolution, and the features The attention module is based on the SE attention mechanism and replaces the activation function with the SiLU activation function;

Step 3. Input the training data set into the improved lightweight YOLO v5 model for training, and obtain the trained improved lightweight YOLO v5 model;

Step 4. Input the test data set into the trained improved lightweight YOLO v5 model to obtain the detection and recognition results.

In the second aspect, the present invention provides a lightweight SAR image ship target detection system based on YOLO v5, including:

The first module obtains the SAR image ship target simulation data set and the public SAR ship detection data set through simulation imaging, preprocesses the SAR images and divides them into training data sets and test data sets;

The second module is used to establish an improved lightweight YOLO v5 model, that is, replace the backbone network of YOLO v5 with the RepVGG network structure, and jointly improve it with the representation enhancement module and the feature attention module. The representation enhancement module is based on asymmetric convolution. Improved, the feature attention module is based on the SE attention mechanism and replaces the activation function with the SiLU activation function;

The third module is used to input the training data set into the improved lightweight YOLO v5 model for training, and obtain the trained improved lightweight YOLO v5 model;

The fourth module is used to input the test data set into the trained improved lightweight YOLO v5 model to obtain detection and recognition results.

In a third aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method described in the first aspect is implemented. A step of.

In a fourth aspect, the present invention provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the method described in the first aspect are implemented.

Compared with the existing lightweight detection technology, the significant advantages of this invention are: (1) When the method is applied to identify ships in coastal scenes, it can effectively extract the characteristics of the coast and ships and distinguish the characteristics of the ships at sea. Shore and ship information has a better recognition rate for ships in complex sea scenes; (2) The method described can more accurately identify small, medium and large ships in SAR images, which is consistent with existing algorithms. It has superior performance, has a higher recognition rate for ships, and improves the efficiency of sea detection target recognition.

Description of the drawings

Figure 1 is the schematic diagram of the improved lightweight target detection method based on the YOLO v5 model.

Figure 2 is a schematic diagram of the improved lightweight network that jointly represents the enhancement module and the feature attention module.

Detailed ways

The present invention proposes a SAR image ship lightweight target detection method under complex sea scenes. The steps are as follows: obtain a SAR image data set, and obtain a SAR image ship target simulation data set and a public SAR ship detection data set (actual measurement) through simulation imaging. Data set), divide the data set into a training sample set and a test sample set after preprocessing; establish an improved lightweight YOLO v5 model, that is, replace the backbone network of YOLO v5 with the RepVGG network structure, and jointly characterize the enhancement module and feature attention module Improve it and replace the activation function with the SiLU activation function; input the training data set into the improved lightweight YOLO v5 model for training, and obtain the trained improved lightweight YOLO v5 model; input the test data set into the trained Improve the lightweight YOLO v5 model to obtain detection and recognition results. The improved lightweight YOLO v5 model can more accurately identify ships in SAR images and has higher detection accuracy.

The present invention will be described in further detail below in conjunction with the accompanying drawings.

Step 1. Obtain the SAR image data set: Obtain the SAR image ship target simulation data set and the public SAR ship detection data set (actual measurement data set) through simulation imaging, preprocess the SAR image and divide it into a training data set and a test data set. ;

Step 2. Establish an improved lightweight YOLO v5 model, that is, replace the backbone network of YOLO v5 with the RepVGG network structure, and jointly improve it with the representation enhancement module and the feature attention module. The representation enhancement module is improved based on asymmetric convolution, and the features The attention module is based on the SE (Squeeze-and-Excitation) attention mechanism and replaces the activation function with the SiLU activation function;

Step 3. Input the training data set into the improved lightweight YOLO v5 model for training, and obtain the trained improved lightweight YOLO v5 model;

Step 4. Input the test data set into the trained improved lightweight YOLO v5 model to obtain the detection and recognition results;

Further, in step 2, a lightweight YOLO v5 model based on multiple RepVGG networks based on multi-module fusion is established. That is, the backbone network of YOLO v5 is replaced with the RepVGG network structure. Based on the lightweight model of RepVGG, the network model is divided into two phases: training and inference. The training phase combines the idea of the residual branch and adds a new residual branch. Asymmetric convolution is integrated into this branch, and combined with conventional convolution to form a representation enhancement module based on asymmetric convolution, which can fit richer feature information in the SAR image feature extraction process and has better robustness. It can better focus on ship targets and significantly improve the accuracy of lightweight ship detection in complex sea scenes. In order to improve the attention to ship targets, SE channel attention is added to each layer of convolution and asymmetric convolution in the training phase to form multiple feature attention modules based on SE channel attention, which have stronger nonlinearity. It can improve the ability to fit complex correlations between different channels, better focus on ship targets during the network training process, and can significantly improve the results of ship detection in complex environments. Compared with other lightweight networks, it has better The effect can make the ship target detection accuracy better than other lightweight networks, and it has strong versatility. The inference stage only contains two operations: 3×3 convolution and ReLU activation function. At this stage, all network layer convolutions are converted into 3×3 convolutions through parameter fusion method. Replace the two-stage activation function with the SiLU activation function. The advantage of the SiLU activation function is that it has no upper bound but a lower bound, smoothness, and non-monotone, and its performance is better than the ReLU activation function;

Furthermore, conventional convolution and asymmetric convolution are mixed through the residual method to implement a representation enhancement module based on asymmetric convolution. Asymmetric convolution converts an n×n convolution kernel into three parallel convolution kernels of n×n, 1×n and n×1. At the same time, 1×n and n×1 are expanded through the center of the n×n convolution kernel. After the model training is completed, the n×n convolution kernel is directly fused with the 1×n and n×1 convolution kernels, and a nonlinear activation function is added between the 1×n and n×1 convolution kernels, which improves the non-linearity of the model. Linear. Asymmetric convolution can improve the expressive power of conventional convolution without requiring additional time-consuming, and can improve the robustness of the model to flipped SAR images, and better extract the SAR image sea surface ship detection in complex environments. The characteristics of the ship target reduce the impact of complex nearshore scenes on ship detection and significantly improve the detection accuracy. The representation enhancement module based on asymmetric convolution constructs conventional convolution and asymmetric convolution in the form of residuals. This module can enhance the feature extraction of ships and improve the robustness of the model. In the dual-stage training and inference, the inference stage still consists of a single branch. The representation enhancement module is applied to the training stage to improve the detection accuracy and improve the generalization of the model by improving the ship feature extraction performance;

Furthermore, SE channel attention is added to each layer of conventional convolution and asymmetric convolution to construct a feature attention module based on channel attention. In the detection of sea surface ship targets in SAR images in complex environments, SE channel attention can learn the SAR image features of different channels, which can effectively make the feature extraction network pay more attention to the characteristics of the ship, thereby improving the detection accuracy. In SAR images under complex sea scenes, not all areas contribute equally to the detection task. Only areas related to ships need attention. Spatial attention is to find the most important parts of the network for processing. In complex near-shore scenarios, ships often interfere with the shore. Spatial attention will ignore some important information in the SAR image, which will cause the ship to Testing has declined. The feature attention module based on SE channel attention adds SE channel attention to all conventional convolutions and asymmetric convolutions, so that each operation increases the degree of attention to different ship characteristics, and only slightly increases the model The complexity and computational burden effectively reduce the impact of the complex near-shore environment on ship detection and improve the accuracy of model detection.

As shown in Figure 1, the SAR image is preprocessed and input into the detection network. Features are extracted based on the improved lightweight network of the joint representation enhancement module and the feature attention module. Multi-scale feature fusion is achieved through the feature aggregation network, and finally the calculation Loss function to predict the final outcome of the ship.

As shown in Figure 2, this lightweight network combines the representation enhancement module based on asymmetric convolution and the feature attention module based on channel attention to form a new improved lightweight network, which is used as the YOLO v5 The feature extraction module is added to the detection process.

Based on the same inventive concept, the present invention also proposes a lightweight SAR image ship target detection system based on YOLO v5, including:

The first module obtains the SAR image ship target simulation data set and the public SAR ship detection data set through simulation imaging, preprocesses the SAR images and divides them into training data sets and test data sets;

The second module is used to establish an improved lightweight YOLO v5 model, that is, replace the backbone network of YOLO v5 with the RepVGG network structure, and jointly improve it with the representation enhancement module and the feature attention module. The representation enhancement module is based on asymmetric convolution. Improved, the feature attention module is based on the SE attention mechanism and replaces the activation function with the SiLU activation function;

The third module is used to input the training data set into the improved lightweight YOLO v5 model for training, and obtain the trained improved lightweight YOLO v5 model;

The fourth module is used to input the test data set into the trained improved lightweight YOLO v5 model to obtain detection and recognition results.

The specific implementation method of each of the above modules is the same as the aforementioned lightweight SAR image ship target detection method, and will not be described again here.

The lightweight SAR image ship target detection method based on YOLO v5 according to the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

Example 1

This example performs lightweight SAR image ship target detection based on YOLO v5. The platform implemented in this example is CPU: Intel(R) Core(TM) i7-8700CPU@3.20GHz, GPU: NVIDIA GeForce RTX 2080, 32g memory ;Operating system window10; using CUDA10.1 acceleration. The data set is a public SAR ship detection data set. The comparison of lightweight SAR image target detection results based on the YOLO v5 algorithm in complex sea scenes is shown in Table 1.

Table 1 Comparison between this method and other lightweight methods on public SAR ship detection data sets

It can be found from Table 1 that the proposed improved lightweight model based on YOLO v5 has good performance in detecting sea surface ship targets in complex environments, and also has significant advantages in indicators. SAR images of coastal ship targets are easy to detect. Affected by background clutter and coastal buildings, the detection effect of coastal ship targets in SAR images is poor, and the false alarm rate and missed detection rate are high. The accuracy of the lightweight model is seriously reduced, and the lightweight structure proposed in this chapter has been successfully used. Experiments have proven that it also has a higher detection rate under complex backgrounds and near-shore scenes, which is better than other lightweight structures.

Example 2

This example performs lightweight SAR image target detection in complex sea scenes based on the YOLO v5 algorithm. The platform implemented in this example is CPU: Intel(R) Core(TM) i7-8700CPU@3.20GHz, GPU: NVIDIA GeForce RTX2080, 32g memory; operating system window10; using CUDA10.1 acceleration. The data set is a SAR image ship target simulation data set. The comparison of lightweight SAR image target detection results based on YOLO v5 algorithm in complex sea fields is shown in Table 2.

Table 2 Comparison between this method and other lightweight methods on the SAR image ship target simulation data set

As can be seen from Table 2, compared with the original YOLO v5 model, the test duration of the proposed model and the compared lightweight model is increased by about three times, which can better meet the real-time performance of surface ship target detection. From the above results, it can be concluded that although the model size of the proposed new lightweight model is slightly higher than that of other lightweight models, the accuracy of other lightweight models has dropped significantly, and this model has demonstrated excellent performance in multi-category ship detection. Advantages: Whether it is a large aircraft carrier, a small fishing boat or other ships, it maintains a high accuracy. In complex environments, offshore ships and ships under different sea conditions maintain a high recognition rate. Its various indicators are better than the original model.

Claims (4)

1. A method for detecting a ship target of a lightweight SAR image based on YOLO v5 is characterized by comprising the following steps:

step 1, acquiring SAR image data sets: obtaining a SAR image ship target simulation data set and a public SAR ship detection data set through simulation imaging, preprocessing the SAR image and dividing the SAR image into a training data set and a test data set;

step 2, an improved lightweight YOLO v5 model is built, namely a backbone network of YOLO v5 is replaced by a RepVGG network structure, the model is improved by combining a characterization enhancement module and a feature attention module, the characterization enhancement module is improved based on asymmetric convolution, the feature attention module is formed based on an SE attention mechanism, and an activation function is replaced by a SiLU activation function;

a lightweight YOLO v5 model of a plurality of RepVGG networks based on multi-module fusion is established, namely a backbone network of the YOLO v5 is replaced by a RepVGG network structure, a new residual branch is newly added in a training stage on the basis of training reasoning double-stage lightweight of the RepVGG, and an asymmetric convolution-based characterization enhancement module, a SE channel attention-based feature attention module and a SiLU activation function are fused;

mixing conventional convolution and asymmetric convolution in a residual mode to realize a characterization enhancement module based on the asymmetric convolution; the asymmetric convolution is to transform an n×n convolution kernel into three parallel convolution kernels of n×n, 1×n, and n×1; and unwrapping 1 xn and n 1 through the center of the n xn convolution kernel; after training, directly fusing the n×n convolution kernels with the 1×n and n×1 convolution kernels, and adding a nonlinear activation function between the 1×n and n×1 convolution kernels;

adding SE channel attention to each layer of conventional convolution and asymmetric convolution to construct a characteristic attention module based on the channel attention;

step 3, inputting the training data set into an improved lightweight YOLO v5 model for training, and obtaining a trained improved lightweight YOLO v5 model;

and 4, inputting the test data set into a trained improved lightweight YOLO v5 model to obtain a detection and identification result.

2. A lightweight SAR image ship target detection system based on YOLO v5 is characterized by comprising:

the SAR image ship target simulation data set and the public SAR ship detection data set are obtained through simulation imaging, and the SAR image is preprocessed and divided into a training data set and a testing data set;

the second module is used for building an improved lightweight YOLO v5 model, namely replacing a backbone network of YOLO v5 with a RepVGG network structure, combining a characterization enhancement module and a feature attention module to improve the model, wherein the characterization enhancement module is improved based on asymmetric convolution, the feature attention module is formed based on an SE attention mechanism, and an activation function is replaced with a SiLU activation function;

a lightweight YOLO v5 model of a plurality of RepVGG networks based on multi-module fusion is established, namely a backbone network of the YOLO v5 is replaced by a RepVGG network structure, a new residual branch is newly added in a training stage on the basis of training reasoning double-stage lightweight of the RepVGG, and an asymmetric convolution-based characterization enhancement module, a SE channel attention-based feature attention module and a SiLU activation function are fused;

mixing conventional convolution and asymmetric convolution in a residual mode to realize a characterization enhancement module based on the asymmetric convolution; the asymmetric convolution is to transform an n×n convolution kernel into three parallel convolution kernels of n×n, 1×n, and n×1; and unwrapping 1 xn and n 1 through the center of the n xn convolution kernel; after training, directly fusing the n×n convolution kernels with the 1×n and n×1 convolution kernels, and adding a nonlinear activation function between the 1×n and n×1 convolution kernels;

adding SE channel attention to each layer of conventional convolution and asymmetric convolution to construct a characteristic attention module based on the channel attention;

the third module is used for inputting the training data set into the improved lightweight YOLO v5 model for training, and obtaining a trained improved lightweight YOLO v5 model;

and the fourth module is used for inputting the test data set into the trained improved lightweight YOLO v5 model to obtain a detection and identification result.

3. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the method of claim 1 when the program is executed by the processor.

4. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the steps of the method according to claim 1.