A lightweight remote sensing ship target detection method based on improved YOLOv10n

Abstract

The application discloses a kind of light-weight remote sensing ship target detection methods based on improved YOLOv10n, method obtains first by the following mode is proposed a kind of light-weight remote sensing ship detection algorithm based on YOLOv10n solves the problem that target distribution scale is wide in remote sensing ship detection, feature is complex and various and needs to handle large batch data set.The global attention mechanism is introduced into the main network to enhance the extraction and fusion ability of remote sensing ship features.Secondly, the lightweight network structure GhostNet is introduced into the residual calculation module to improve the feature extraction efficiency of the model and reduce the computational complexity.Finally, the Focaler-Shape-IoU loss function is designed to make the model pay more attention to the shape and scale factors of the bounding box, compensate for the impact of sample difficulty distribution on bounding box regression, and improve the convergence speed and generalization performance of the model.

Description

Technical Field

[0001] This invention relates to the field of remote sensing ship target detection technology, and in particular to a lightweight remote sensing ship target detection method based on the improved YOLOv10n. Background Technology

[0002] With the rapid development of deep learning technology, the YOLO series of algorithms has become the mainstream framework in the field of natural image object detection due to its excellent detection speed and accuracy. The recently proposed YOLOv10n has further optimized the model architecture design and training strategy, significantly improving object detection performance in general scenarios. However, directly applying YOLOv10n to remote sensing ship target detection still faces many challenges: ship targets in remote sensing images have a wide distribution scale and complex and diverse features, and practical application scenarios such as sea surface monitoring often require processing large datasets, which places higher demands on the model's feature extraction capabilities, computational efficiency, and generalization performance.

[0003] Existing YOLOv10n-based detection methods for remote sensing ship detection suffer from the following shortcomings: First, the backbone network has limited ability to extract and fuse small-scale, weak-feature targets such as ships in remote sensing images, making it difficult to fully capture global contextual information. This results in insufficient feature information for small targets, which are easily classified as background, affecting detection accuracy. Second, the residual calculation module has a high number of parameters and computational complexity, leading to severe model redundancy when processing high-resolution remote sensing images. This increases the difficulty of training and deployment, hindering the algorithm's embedding into mobile platforms and its lightweight migration. Furthermore, existing bounding box regression loss functions fail to effectively consider the shape and scale factors of the bounding box itself and do not adequately address the imbalance in sample difficulty distribution, affecting the regression accuracy of rotated boxes and the model's convergence speed. This limits the detection performance of small targets and targets in arbitrary directions in complex remote sensing scenarios.

[0004] In summary, how to enhance the extraction and fusion capabilities of remote sensing ship features while maintaining the high efficiency of YOLOv10n, reduce the computational complexity of the model to achieve network lightweighting, and improve the accuracy and convergence speed of bounding box regression remains an unsolved problem in the field of remote sensing image target detection. Summary of the Invention

[0005] Therefore, the purpose of this invention is to provide a lightweight remote sensing ship target detection method based on the improved YOLOv10n, which solves the problems of difficulty in detecting small targets and redundancy in detection models.

[0006] To achieve the above objectives, the present invention provides the following solution: A lightweight remote sensing method for ship target detection based on an improved YOLOv10n includes the following steps: First, an attention mechanism is introduced to improve the backbone network of the YOLOv10n model, enhancing feature extraction for remotely sensed ship targets at different resolutions; Then, the GhostNet network was introduced, and the original C2f module was replaced with a newly constructed C2fGhost module to achieve network lightweighting without losing feature information.

[0007] Finally, a Focaler-Shape-IoU loss function was designed to make the model pay more attention to the shape and scale factors of the bounding box itself, compensate for the influence of the sample difficulty distribution on the bounding box regression, and improve the convergence speed and generalization performance of the model. The aforementioned lightweight remote sensing ship target detection method based on improved YOLOv10n introduces an attention mechanism to improve the YOLOv10n model backbone network. The steps include: In the improved YOLOv10n network model, the backbone is the part used for feature extraction. Considering the importance of adjusting features globally, GAM is inserted in the backbone to focus on important parts of the image. By extracting the channel and spatial features of the image, global attention is introduced between feature maps at different levels, and long-distance links are established to enhance feature fusion.

[0008] In the CA module, the input feature map is first transposed through the channel dimension, then processed by a linear layer of a multilayer perceptron (MLP) and a sigmoid activation function to obtain the channel attention weights. Finally, the attention weights are multiplied by the feature map to obtain M(F), which contains the channel weight information of the image of interest.

[0009] The SA module primarily uses convolution to process the input features (F2), and utilizes the Softmax function to convert the dot product into weights to characterize the importance of the input vectors, thereby capturing the global dependencies between the input vectors and obtaining the most important regional information in the image.

[0010] The aforementioned lightweight remote sensing ship target detection method based on improved YOLOv10n introduces the GhostNet network, replacing the original C2f module with a newly constructed C2fGhost module. The steps include: A C2fGhost architecture was designed to achieve lightweight model implementation. By introducing a Ghost Module, the convolutional kernel is divided into Ghost and Cheap parts. The Ghost convolutional kernel extracts the main information, while the Cheap convolutional kernel extracts additional information, thereby significantly reducing the amount of computation and parameters and improving computational efficiency.

[0011] The aforementioned lightweight remote sensing ship target detection method based on the improved YOLOv10n designs a Focaler-Shape-IoU loss function, and the steps include: By introducing the Shape-IoU loss function to replace the CIoU loss function in YOLOv10n, better bounding box regression results are obtained. The Shape-IoU loss function can be expressed as follows: Where: w and h are the weight coefficients in the horizontal and vertical directions, respectively, and their values ​​are related to the shape of the ground truth box; s is the scaling factor, which is related to the scale of the target in the dataset. The smaller the target, the greater the influence of the shape factor; b and These are the center points of the predicted bounding box and the ground truth bounding box, respectively; w, h, and These represent the width and height of the predicted bounding box and the ground truth bounding box, respectively. These represent the horizontal and vertical distance components between the center point of the predicted bounding box and the center point of the ground truth bounding box, respectively. It is the square of the diagonal length of the minimum bounding box; The weighted normalized distance to the center point; For shape loss; Control the degree of attention given to shape loss, avoiding excessive focus on shape loss that reduces the movement of the predicted bounding box. The value is 4.

[0012] To focus on different regression samples in different detection tasks and compensate for the impact of sample difficulty distribution on bounding box regression, the Focaler-IoU concept is introduced. This method uses linear interval mapping to reconstruct the IoU loss, thereby improving the bounding box regression performance. The Focaler-IoU loss function is expressed as follows: In the formula: For the reconstructed Focaler-IoU; The original IoU value; By adjusting the values ​​of d and u, it is possible to make Focusing on different regression samples, and combining the ideas of both loss functions, Focaler-Shape-IoU is designed to improve model robustness, expressed as... . Attached Figure Description

[0013] Figure 1 An improved YOLOv10n network model structure; Figure 2 For CA, SA, and GAM modules; Figure 3 It is a Ghost Module structure; Figure 4 For the C2fGhost module; Detailed Implementation The experiment used Microsoft Windows 11 Professional operating system, paired with an Intel Core™ i7-10700 CPU@2.90 GHz processor and 48GB of video memory, and was equipped with an NVIDIA GeForce RTX 4090 GPU. Python 3.10 was used as the programming language, CUDA version 11.7 was used, and the model was trained based on the deep learning framework PyTorch 2.0.1.

[0014] Therefore, the purpose of this invention is to provide a lightweight remote sensing ship target detection method based on the improved YOLOv10n, which solves the problems of difficulty in detecting small targets and redundancy in detection models.

[0015] The experiment selected the best model after 300 training rounds to ensure that the model could fully learn the key features of the data during training. To comprehensively evaluate the performance, scale, and efficiency of the object detection model, accuracy (P), recall (R), mean average precision (mAP), number of model parameters, floating-point operations per second (GFLOPs), weight file size, and detection speed were selected as evaluation metrics.

[0016] To achieve the above objectives, the present invention provides the following solution: A lightweight remote sensing method for ship target detection based on an improved YOLOv10n includes the following steps: First, an attention mechanism is introduced to improve the backbone network of the YOLOv10n model, enhancing feature extraction for remotely sensed ship targets at different resolutions; Then, the GhostNet network was introduced, and the original C2f module was replaced with a newly constructed C2fGhost module to achieve network lightweighting without losing feature information.

[0017] Finally, a Focaler-Shape-IoU loss function was designed to make the model pay more attention to the shape and scale factors of the bounding box itself, compensate for the influence of the sample difficulty distribution on the bounding box regression, and improve the convergence speed and generalization performance of the model. The aforementioned lightweight remote sensing ship target detection method based on improved YOLOv10n introduces an attention mechanism to improve the YOLOv10n model backbone network. The steps include: In the improved YOLOv10n network model, the backbone is used for feature extraction. Considering the importance of adjusting features globally, a Global Aspect-Oriented Model (GAM) is inserted into the backbone to focus on important parts of the image. By extracting channel and spatial features of the image, global attention is introduced between feature maps at different levels, establishing long-distance links to enhance feature fusion. The C2f structure is the core module for learning residual features, and its structure determines the model's running speed.

[0018] GAM adds a Spatial Attention (SA) module to the Channel Attention (CA) module to establish long-distance links of feature information, enabling the model to focus on important regions of the detected target globally. The CA module consists of two linear layers and an activation function. In this part, the input feature map is first transposed by the channel dimension, then processed by a linear layer of a multilayer perceptron (MLP) and a sigmoid activation function to obtain the channel attention weights. Finally, the attention weights are multiplied by the feature map to obtain M(F), which contains the channel weight information of the image of interest.

[0019] The SA module contains convolutional layers and batch normalization layers. This part mainly uses convolution to process the input features (F2) and uses the Softmax function to convert the dot product into weights to represent the importance of the input vectors, so as to capture the global dependencies between the input vectors and obtain the most important regional information in the image.

[0020] The aforementioned lightweight remote sensing ship target detection method based on improved YOLOv10n introduces the GhostNet network, replacing the original C2f module with a newly constructed C2fGhost module. The steps include: In practical remote sensing ship target detection, the detection set contains a large number of images with inherent repetitive features. Traditional convolutional modules perform individual convolutions on channels at different scales to obtain feature maps, which not only introduces a large amount of redundant information but also leads to high cost in feature map acquisition. Furthermore, neural network learning requires even more feature maps. To reduce the cost of feature map acquisition and obtain more feature maps with fewer parameters and lower computational cost, we borrowed the GhostNet network structure and introduced the C2f residual calculation module from YOLOv10n to design the C2fGhost structure, achieving model lightweighting. GhostNet is a lightweight network structure that uses the idea of ​​spatial convolution. By introducing the Ghost Module, it divides the convolutional kernel into Ghost and Cheap parts. The Ghost convolutional kernel extracts the main information, while the Cheap convolutional kernel extracts additional information, thereby significantly reducing computational cost and parameter count and improving computational efficiency.

[0021] Based on the Ghost Module structure, a Ghost Bottleneck module was designed to replace the original Bottleneck in YOLOv10n. This module consists of two stacked Ghost Modules. By expanding the number of channels and reducing the number of output channels, it better captures the diversity of input features and reduces the computational burden, thus promoting the network to learn more compact and efficient representations. Residual connections accelerate gradient flow, improving training stability and deep feature representation. The newly designed module C2fGhost uses Ghost Bottleneck to replace the original Bottleneck, reducing the large number of 3×3 ordinary convolutions in the original structure, significantly compressing the number of model parameters and computational cost, making the model more suitable for deployment on mobile devices and easier to implement large-scale remote sensing edge detection.

[0022] The aforementioned lightweight remote sensing ship target detection method based on the improved YOLOv10n designs a Focaler-Shape-IoU loss function, and the steps include: YOLOv10n uses CIoU as the bounding box regression loss function, primarily employing overlap area, center point distance, and bounding box aspect ratio as factors in the loss calculation. However, CIoU only considers the geometric relationship between the ground truth and predicted bounding boxes, neglecting the shape and scale factors of the bounding boxes themselves. Therefore, the Shape-IoU loss function is introduced to replace the CIoU loss function, achieving better bounding box regression results. The Shape-IoU loss function can be expressed as... Where: w and h are the weight coefficients in the horizontal and vertical directions, respectively, and their values ​​are related to the shape of the ground truth box; s is the scaling factor, which is related to the scale of the target in the dataset. The smaller the target, the greater the influence of the shape factor; b and These are the center points of the predicted bounding box and the ground truth bounding box, respectively; w, h, and These represent the width and height of the predicted bounding box and the ground truth bounding box, respectively. These represent the horizontal and vertical distance components between the center point of the predicted bounding box and the center point of the ground truth bounding box, respectively. It is the square of the diagonal length of the minimum bounding box; The weighted normalized distance to the center point; For shape loss; Control the degree of attention given to shape loss, avoiding excessive focus on shape loss that reduces the movement of the predicted bounding box. The value is 4.

[0023] To focus on different regression samples in different detection tasks and compensate for the impact of sample difficulty distribution on bounding box regression, the Focaler-IoU concept is introduced. This method uses linear interval mapping to reconstruct the IoU loss, thereby improving the bounding box regression performance. The Focaler-IoU loss function is expressed as follows: In the formula: For the reconstructed Focaler-IoU; The original IoU value; By adjusting the values ​​of d and u, it is possible to make Focusing on different regression samples, and combining the ideas of both loss functions, Focaler-Shape-IoU is designed to improve model robustness, expressed as... .

[0024] To verify the feature detection performance of the GAM module, experiments were conducted by adding other commonly used attention modules to the original YOLOV10n algorithm for comparison.

[0025] Table 1 Comparison of experimental results for different attention mechanisms As shown in Table 2, the model with GAM has the highest detection accuracy. Compared with the original model, mAP@0.5 and mAP@0.5:0.95 are improved by 3.7% and 0.3% respectively. This indicates that GAM has the strongest feature extraction capability for remote sensing ships and can better focus on the features of remote sensing ships.

[0026] Table 2 Comparison of different loss functions Table 2 shows the different performances of the loss functions CIoU, SIoU, EIoU, WIoU, and Focaler-Shape IoU on the NWPUVHR-10 dataset, using mAP50 as the performance evaluation metric. As shown in Table 2, the designed loss function Focaler-Shape-IoU performs better than the other loss functions, with mAP values ​​exceeding ClolU, SloU, EloU, and WloU by 3.3, 3.0, 3.0, and 2.7 percentage points, respectively, indicating that the improved loss function has better convergence speed and performance.

[0027] In summary, the improved YOLOv10n algorithm of this invention achieves excellent results in both accuracy and recall. Compared with existing technologies, this invention represents a significant advancement.

[0028] The above are merely specific embodiments of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing specific embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing specific embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the specific embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A lightweight remote sensing method for ship target detection based on an improved YOLOv10n, characterized in that, The method was improved through the following steps: An attention mechanism was introduced to improve the backbone network of the YOLOv10n model, enhancing the feature extraction of remotely sensed ship targets at different resolutions. The hostNet network replaces the original C2f module with a newly built C2fGhost module to achieve network lightweighting without losing feature information; Using the Focaler-Shape-IoU loss function makes the model pay more attention to the shape and scale factors of the bounding box itself, making up for the impact of the sample difficulty distribution on the bounding box regression, and improving the convergence speed and generalization performance of the model.

2. The lightweight remote sensing ship target detection method based on improved YOLOv10n according to claim 1, characterized in that... When introducing an attention mechanism, a hybrid self-attention module of GAM is used.

3. The lightweight remote sensing ship target detection method based on improved YOLOv10n as described in claim 2 is characterized in that... The GAM adds a spatial attention module to the channel attention module to establish long-distance links of feature information, enabling the model to focus on important regions of the detected target globally.

4. The lightweight remote sensing ship target detection method based on improved YOLOv10n according to claim 2, characterized in that... In the GAM, for the input feature map F, it first passes through the CA module to learn the importance of the feature map in the channel dimension, and compresses the feature information by reducing the number of channels to obtain a weighted combination of channel features; Then, through the SA module, the importance of feature maps in the spatial dimension is learned, and the spatial location of channel features is weighted to enhance feature representation.

5. The lightweight remote sensing ship target detection method based on improved YOLOv10n according to claim 4, characterized in that... The GAM module enables cross-dimensional interaction of feature information, reduces information diffusion, and allows the network to effectively build global dependencies on deeper remote sensing ship target features, significantly improving recognition performance and making it more suitable for detecting complex and diverse remote sensing ship targets.

6. The lightweight remote sensing ship target detection method based on improved YOLOv10n according to claim 1, characterized in that... YOLOv10n consists of four parts: Input, Backbone, Neck, and Head.

7. The lightweight remote sensing ship target detection method based on improved YOLOv10n according to claim 6, characterized in that... The Input uses the GAM channel attention mechanism, and the Neck uses the GhostNet network structure.

8. The lightweight remote sensing ship target detection method based on improved YOLOv10n according to claim 1, characterized in that... The Focaler-IoU loss function is expressed as follows: In the formula The reconstructed Focaler-IoU; Original IoU value .

9. The lightweight remote sensing ship target detection method based on improved YOLOv10n according to claim 8, characterized in that... By adjusting the values ​​of d and u, it is possible to make Focusing on different regression samples, Focaler-Shape-IoU was designed to improve model robustness, denoted as... .

10. A lightweight remote sensing ship target detection method based on improved YOLOv10n, characterized in that... The detection is performed using the lightweight remote sensing ship target method based on the improved YOLOv10n as described in any one of claims 1-9.

Images