CN-121999295-A - YOLOv 11-based lightweight SAR ship detection method

CN121999295ACN 121999295 ACN121999295 ACN 121999295ACN-121999295-A

Abstract

The invention provides a YOLOv 11-based lightweight SAR ship detection method, which relates to the technical field of ship detection, and constructs a lightweight ship detection model based on YOLOv, wherein a Ghost convolution and a lightweight parallel convolution module are used as main feature extraction modules in a YOLOv backbone network, so that repeated extraction of effective features is avoided, calculation resources are saved, feature fusion is performed through a key information enhancement module, and a model prediction result is optimized by using a beta-WIoU loss function. A large amount of significant key features and important feature information of the image are captured simultaneously, so that the detection speed is improved, the convolution step is simplified, and unimportant features are restrained. The method solves the problem of the balance between the convolution complexity and the detection accuracy of the feature extraction in the existing network.

Inventors

ZHU QIAN
GAO TIANXING
Qiu Xinpeng

Assignees

东北大学

Dates

Publication Date: 20260508
Application Date: 20260130

Claims (10)

1. The lightweight SAR ship detection method based on YOLOv is characterized by comprising the following steps of: acquiring a ship SAR image, constructing a data set, and dividing the data set into a training data set and a test data set according to a proportion; Constructing a lightweight ship detection model based on YOLOv < 11 >, training the lightweight ship detection model by using a training data set, and evaluating the lightweight ship detection model by using a test data set to obtain a trained lightweight ship detection model; The light-weight ship detection model comprises a main network, a neck network and a detection head, wherein the main network is used for extracting feature images with different scales and comprises a plurality of light-weight parallel convolution modules; and inputting the ship SAR image to be detected into a trained light ship detection model to obtain a detection result.
2. The method for lightweight SAR ship detection based on YOLOv as set forth in claim 1, wherein the lightweight parallel convolution module comprises a first phantom convolution, a plurality of bottleneck submodules or C3k submodules and a second phantom convolution, wherein the bottleneck submodules or the C3k submodules are used in the lightweight parallel convolution module by setting a C3k selection switch, and when the C3k selection switch is used When using the C3k submodule in a lightweight parallel convolution module, otherwise, when the C3k selector switch In this case, a bottleneck sub-module is used in the lightweight parallel convolution module.
3. The method for detecting the lightweight SAR ships based on YOLOv as set forth in claim 2, wherein the specific method for extracting the feature map with the specified scale by the lightweight parallel convolution module is as follows: and the second part is input into a plurality of bottleneck submodules or C3k submodules to extract deep features, the extracted deep features and the shallow features are spliced, and then grouping convolution and channel compression are carried out through second phantom convolution to obtain the feature map output by the light-weight parallel convolution module.
4. The method for detecting the lightweight SAR ship based on YOLOv as set forth in claim 3, wherein the bottleneck submodule comprises two convolution layers, the feature map input to the bottleneck submodule sequentially passes through the two convolution layers, and the feature map output by the second convolution layer is added with the feature map input to the bottleneck submodule to obtain the feature map output by the bottleneck submodule.
5. The lightweight SAR ship detection method based on YOLOv a is characterized in that the C3k submodule comprises two branches, the first branch comprises a convolution layer, the second branch comprises a first convolution layer, a plurality of bottleneck submodules and a second convolution layer, the feature images of the input C3k submodule pass through the two branches respectively, the first branch convolves the feature images of the input C3k submodule through the convolution layer to obtain feature images of the output of the first branch, the second branch sequentially processes the feature images of the input C3k submodule through the first convolution layer and the bottleneck submodules, the feature images of the output of the last bottleneck submodule are spliced with the feature images of the output of the first branch, and the feature images of the output of the C3k submodule are obtained through the second convolution layer.
6. The YOLOv-based lightweight SAR ship detection method of claim 1, wherein the critical information enhancement module comprises global average pooling, global maximum pooling, one-dimensional rolling and Sigmoid activation functions.
7. The method for detecting the lightweight SAR ship based on YOLOv a in claim 6, wherein the specific method for enhancing the input feature map by the key information enhancement module is as follows: Carrying out global average pooling and global maximum pooling operation on the feature images of the input key information enhancement modules channel by channel, and adding the feature images subjected to global average pooling and global maximum pooling element by element to obtain element vectors of the feature images of the input key information enhancement modules; carrying out one-dimensional convolution on the element vector, and carrying out nonlinear transformation on a one-dimensional convolution output result by utilizing a Sigmoid activation function to obtain the attention weight of each channel; Multiplying the attention weight with the feature map input to the key information enhancement module, and carrying out channel-by-channel weighting to obtain the feature map output by the key information enhancement module; And inputting the feature map output by the key information enhancement module into a corresponding detection head for generating a detection result.
8. The method for lightweight SAR ship detection based on YOLOv a in claim 1, wherein said training of lightweight ship detection model using training dataset uses The loss function optimizes the prediction box.
9. A computer readable storage medium storing executable instructions that when executed cause a processor to perform the YOLOv-based lightweight SAR ship detection method of any one of claims 1 to 8.
10. A computer program product comprising a computer program or instructions which, when executed by a processor, implements the YOLOv-based lightweight SAR ship detection method of any one of claims 1 to 8.

Description

YOLOv 11-based lightweight SAR ship detection method Technical Field The invention belongs to the technical field of ship detection, and particularly relates to a YOLOv 11-based lightweight SAR ship detection method. Background Synthetic aperture radar (SYNTHETIC APERTURE RADAR, SAR) is used as an active microwave remote sensing technology to acquire high resolution images of ground or marine targets by transmitting electromagnetic wave pulses and receiving echo signals. Compared with the traditional optical remote sensing, the SAR has the advantages of overcoming cloud layers, heavy fog and other severe weather environments, continuously outputting high-resolution images of targets, and providing effective data support for positioning, tracking, monitoring and predicting sea and land targets. This makes it an indispensable tool in marine monitoring. Due to the fact that the radar echo signals of the ship and the sea surface interference background have obvious differences, the SAR image can accurately position the ship target under the complex marine environment or the severe weather condition. In recent years, the deep learning accelerates the development of SAR image ship target detection but also faces new challenges, and because the deep learning network not only needs a large amount of computing resources in the process of training the model, but also needs to consume a large amount of learning time, the learning cost of the model is increased. Complex network models are difficult to apply in the face of computing resource-poor edge devices or scenarios requiring monitoring of real-time. Therefore, the deep learning network model needs to be optimized in a light-weight manner, so that the deep learning network model still has the capability of accurately extracting the ship target characteristics and improving the detection efficiency in equipment with limited computing resources The existing lightweight technology for SAR ship detection algorithm has the problems of lacking characteristic generalization capability and reduced ship recognition precision when facing to the interference of complex background on low-calculation-force edge equipment. Mainly because of its model's weak learning ability for new data features. The method is characterized in that the general model is light and improved, only high speed and little memory are focused, the parameter quantity and the calculation quantity of the model are less considered, the convolution form often uses traditional convolution, the adjacent feature graphs extracted by the traditional convolution mode are highly similar, and the limited computational power resource is wasted on the feature extraction level of the information. Disclosure of Invention Aiming at the defects of the prior art, the invention provides the lightweight SAR ship detection method based on YOLOv11, which can save computing resources and improve ship detection accuracy, can capture the significant key features and important feature information of the image at the same time of improving detection speed and simplifying convolution steps, can inhibit unimportant features, and can solve the problem of the balance between the convolution complexity and detection accuracy of feature extraction in the existing network. In a first aspect, the invention provides a YOLOv-based lightweight SAR ship detection method, which comprises the following steps: acquiring a ship SAR image, constructing a data set, and dividing the data set into a training data set and a test data set according to a proportion; Constructing a lightweight ship detection model based on YOLOv < 11 >, training the lightweight ship detection model by using a training data set, and evaluating the lightweight ship detection model by using a test data set to obtain a trained lightweight ship detection model; The light-weight ship detection model comprises a main network, a neck network and a detection head, wherein the main network is used for extracting feature images with different scales and comprises a plurality of light-weight parallel convolution modules; and inputting the ship SAR image to be detected into a trained light ship detection model to obtain a detection result. Further, the lightweight parallel convolution module comprises a first phantom convolution, a plurality of bottleneck sub-modules or C3k sub-modules and a second phantom convolution, wherein the bottleneck sub-modules or the C3k sub-modules are used in the lightweight parallel convolution module by setting a C3k selection switch, and when the C3k selection switch is usedWhen using the C3k submodule in a lightweight parallel convolution module, otherwise, when the C3k selector switchIn this case, a bottleneck sub-module is used in the lightweight parallel convolution module. Further, the specific method for extracting the feature map with the specified scale by the lightweight parallel convolution module comprises the following steps: and the second