Remote sensing ship target detection method based on deep learning

Abstract

The invention discloses a remote sensing ship target detection method suitable for deep learning, and belongs to the field of remote sensing ship target detection, and the method comprises the following steps: firstly, adding an EUCB-S upsampling module in a neck network to accurately capture feature information related to a remote sensing ship target; and interference of background noise and irrelevant details is reduced to the minimum. Secondly, a PAN4 feature fusion network with four detection scales is designed to fuse more target feature information, and remote sensing ship targets of different scales can be detected. And finally, providing a new loss function and bidirectional attenuation loss BAIoU on the basis of a KFIOU loss function, quantifying the similarity between a labeling frame and a prediction frame on the basis of an attenuation coefficient, an offset direction and an offset distance, improving the positioning and recognition precision of a rotating frame, and solving the problem that a remote sensing ship target has any rotation angle.

Description

Technical Field

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

[0002] With the rapid development of sensor technology and aerospace platform technology, the quality of remote sensing images has been significantly improved. The acquisition of high-quality remote sensing images provides a solid foundation for subsequent research and applications. Using space platforms such as satellites, high-resolution satellite remote sensing images can be used to monitor the ocean according to different mission requirements, such as maritime transport, port traffic, marine resource development, rescue of ships in distress at sea, and reconnaissance of military targets. Therefore, the detection of ship targets using satellite remote sensing images is of paramount importance to my country's national defense security and economic development.

[0003] Ship target detection involves two aspects: accurate localization of ship targets and classification of ship types. Visible light remote sensing images possess clear location information of ground features, detailed textures and geometric structures, giving them an advantage in ship type and model identification and making them ideal for visualizing the results. However, due to the characteristics of ships themselves and the marine environment, ship identification presents several challenges: background noise and irregular textures can mix with the small size and inconspicuous features of the ship target, making feature extraction difficult. The varying sizes of ships, from small fishing boats to large aircraft carriers, add further complexity to target detection and identification. Ships navigating the ocean may appear in remote sensing images at arbitrary angles, further complicating target detection and identification. Summary of the Invention

[0004] To address the aforementioned problems, this invention aims to provide a remote sensing ship target detection method based on deep learning.

[0005] The technical solution of the present invention is as follows: On the one hand, a remote sensing ship target detection method based on deep learning is provided, which is improved through the following steps: First, an upsampling module is added to the neck network to capture feature information related to remotely sensed ship targets, minimizing interference from background noise and irrelevant details.

[0006] Secondly, a feature fusion network with four detection scales is designed to fuse more target feature information, enabling the detection of remote sensing ship targets at different scales.

[0007] Finally, based on the KFIOU loss function, a new loss function, the bidirectional decay loss BAIoU, is proposed to improve the accuracy of rotating box localization and recognition, and solve the problem of arbitrary rotation angles of remote sensing ship targets.

[0008] As a preferred approach, an attention mechanism is introduced, employing the EUCB-S upsampling module.

[0009] Preferably, the EUCB-S upsampling module consists of an Upsample module, a Depthwise Separable Convolution (DWC) module, a GELU activation function, and an SE module.

[0010] Preferably, the Upsample module enhances the model's ability to recover spatial details and semantic information while maintaining computational efficiency. Depthily separable convolution (DWC) extracts spatial features better, while GELU activation helps the model adapt better to changes in data distribution during training, thereby improving generalization ability, especially when processing remotely sensed ship targets. Using residual-like structures at the input and output preserves original information, alleviates gradient vanishing, and improves detail recovery. The objective function is decomposed into the form H(x) = F(x) + x. The SE module primarily learns feature weights based on the loss, giving greater weight to effective feature maps and smaller weight to ineffective or less effective feature maps.

[0011] As a preferred approach, a feature fusion network with four detection scales is designed, using the PAN_4 feature fusion network.

[0012] Preferably, the PAN_4 feature fusion network adds two lateral skip connections of the same scale to the P3 and P4 layers of YOLOv10. These two newly added lateral skip connections allow direct information exchange between feature maps of the same scale. This connection directly integrates the feature information extracted by the backbone network into the feature map to be detected.

[0013] As a preferred embodiment, the PAN_4 feature fusion network is computed in the P3 layer of YOLOv10 as shown below: , Where `out` represents the output of this network layer, `Conv()` is a convolution operation, `Upsample()` is a linear 2x upsampling operation, and `Contact()` directly sums the channels of all input features. The output features satisfy the following equation with each input feature: , Where I represents the input feature, O represents the output feature, and Channel represents the number of channels.

[0014] As a preferred approach, a rotating target detection loss function is designed, employing bidirectional decay loss BAIoU.

[0015] Preferably, the BAIoU loss function is: , , , , , in, and The values ​​of hyperparameters can be obtained separately through projection formulas. The value is 1.75. and For the actual width and height, and To predict the width and height of the bounding box, and These are the vectors corresponding to the true value and the longer side of the predicted bounding box, respectively. It is the inner product of vectors, and the absolute value of the inner product is taken to ensure that the included angle is less than 90° during the calculation process.

[0016] On the other hand, a deep learning-based remote sensing ship target detection method is provided, which uses any of the methods described above to perform detection operations on ship targets in remote sensing images.

[0017] This invention introduces an EUCB-S upsampling module to accurately capture feature information related to remotely sensed ship targets, minimizing interference from background noise and irrelevant details. Secondly, a BiFP4 feature fusion network with four detection scales is designed to fuse more target feature information, enabling the detection of remotely sensed ship targets at different scales. Finally, the loss function is improved to enhance the accuracy of bounding box localization and recognition, solving the problem of arbitrary rotation angles for remotely sensed ship targets. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the improved YOLOv10-BA network structure of the present invention; Figure 2 This is a schematic diagram of the network structure of the EUCB-S module of the present invention; Figure 3 This is a schematic diagram of the PAN_4 feature fusion network; Figure 4 A schematic diagram showing the farthest attenuation distance (short side direction) of the bidirectional attenuation loss BAIoU; Figure 5 This is a schematic diagram showing the farthest attenuation distance (long side direction) of the bidirectional attenuation loss BAIoU; Detailed Implementation The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and technical features described in this application can be combined with each other. It should also be pointed out that, unless otherwise indicated, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terms "comprising" or "including" and similar words used in this invention refer to elements or objects preceding the word that encompass the elements or objects listed following the word and their equivalents, without excluding other elements or objects.

[0020] On the one hand, a remote sensing ship target detection method based on deep learning is provided, such as... Figure 1 As shown, the model is obtained through the following steps: First, an upsampling module is added to the neck network to capture feature information related to the remote sensing target, minimizing interference from background noise and irrelevant details.

[0021] Secondly, a feature fusion network with four detection scales is designed to fuse more target feature information, enabling the detection of remote sensing ship targets at different scales.

[0022] Finally, based on the KFIOU loss function, a new loss function, the bidirectional decay loss BAIoU, is proposed to improve the accuracy of rotating box localization and recognition, and solve the problem of arbitrary rotation angles of remote sensing ship targets.

[0023] EUCB-S module such as Figure 2As shown, this approach enhances the model's ability to recover spatial details and semantic information while maintaining computational efficiency. Depthily separable convolution (DWC) extracts spatial features better, while GELU activation helps the model adapt to changes in data distribution during training, thus improving generalization ability, especially when processing remotely sensed ship targets. Using residual-like structures at the input and output preserves original information, alleviates gradient vanishing, and improves detail recovery. The objective function is decomposed into the form H(x) = F(x) + x. The SE module primarily learns feature weights based on the loss, giving greater weight to effective feature maps and smaller weight to ineffective or less effective feature maps.

[0024] The PAN_4 feature fusion network directly integrates the feature information extracted by the backbone network into the feature map to be detected, enabling the feature map to acquire more contextual information. This is especially beneficial for small targets, significantly enhancing the surrounding context and making them easier for the network to capture. Simultaneously, the presence of lateral skip connections better preserves high-resolution details. In object detection tasks, high-resolution feature maps are crucial for capturing fine-grained features such as edges and textures. This improvement does not significantly increase the network's computational complexity. Although two lateral skip connections are added, they only exchange information between feature maps of the same scale, thus having no significant impact on the overall computational cost. This makes the improvement strategy highly feasible and practical in real-world applications. Figure 3 This is a schematic diagram of the PAN_4 feature fusion network.

[0025] The PAN_4 feature fusion network adds two lateral skip connections of the same scale to the P3 and P4 layers of YOLOv10. These two new lateral skip connections allow direct information exchange between feature maps of the same scale. This connection directly integrates the feature information extracted by the backbone network into the feature map to be detected.

[0026] The relevant computations of the PAN_4 feature fusion network in the P3 layer of YOLOv10 are shown below: , Where `out` represents the output of this network layer, `Conv()` is a convolution operation, `Upsample()` is a linear 2x upsampling operation, and `Contact()` directly sums the channels of all input features. The output features satisfy the following equation with each input feature: , Where I represents the input feature, O represents the output feature, and Channel represents the number of channels.

[0027] Traditional loss functions fail to reflect the skew IoU variation caused by the directional deviation of the center point position, resulting in inconsistencies between the loss and the actual situation. The bidirectional attenuation loss method reflects the angular error of the predicted value by calculating the KFIoU result. Instead of using center point loss, it attenuates the approximate value based on the offset direction and distance, reflecting different offset scenarios. This method multiplies the KFIoU result by an attenuation coefficient α, making the loss variation closely resemble the variation of skew IoU. The attenuation coefficient is based on the projection length of the center point offset onto the true value in both the length and width directions. The area of ​​the rectangular frame is attenuated to approximately the size of the actual skew IoU. This bidirectional attenuation loss method can simultaneously reflect the skew IoU variation caused by both the predicted value offset error and the angular error.

[0028] The new loss function expression is: , , , in: Functions representing BAIoU, such as ,1-BAIoU or . and These are the projection lengths of the line vector *u* connecting the center point of the ground truth box to the center point of the predicted bounding box along the long and short sides of the ground truth, respectively, and are used as independent variables in the attenuation coefficient. Hyperparameters The value is 1.75.

[0029] In remote sensing ship target detection tasks, remotely sensed ship targets are characterized by a large range in size distribution and varying aspect ratios. To better reflect the different size information of specific ship targets, this paper introduces a scaling factor. and The scaling factor is used to design the relative attenuation of the loss in two directions for the length and width of a specific remotely sensed ship target. Its values ​​represent the farthest attenuation distance of the predicted bounding box in the long and short sides, respectively, i.e., the center point offset distance when the skew IoU attenuation is 0. For example... Figure 4 and Figure 5 As shown. The specific calculation process is as follows: , , , , in, and For the actual width and height, and To predict the width and height of the bounding box, and These are the vectors corresponding to the true value and the longer side of the predicted bounding box, respectively. It is the inner product of vectors, and the absolute value of the inner product is taken to ensure that the included angle is less than 90° during the calculation process.

[0030] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A remote sensing ship target detection method based on deep learning, characterized in that, The method includes: First, an upsampling module is added to the neck network to capture feature information related to remotely sensed ship targets, minimizing interference from background noise and irrelevant details; Secondly, a feature fusion network with four detection scales is used to fuse more target feature information to detect remote sensing ship targets at different scales. Finally, based on the KFIOU loss function, a bidirectional decay loss BAIoU is proposed to improve the accuracy of rotating box positioning and recognition.

2. The remote sensing ship target detection method based on deep learning according to claim 1, characterized in that, An attention mechanism is introduced, employing the EUCB-S upsampling module.

3. The remote sensing ship target detection method based on deep learning according to claim 2, characterized in that, The EUCB-S upsampling module consists of an Upsample module, a depthwise separable convolution module, a GELU activation function, and an SE module.

4. The remote sensing ship target detection method based on deep learning according to claim 3, characterized in that, The Upsample module enhances the model's ability to recover spatial details and semantic information while maintaining computational efficiency; depthwise separable convolution extracts spatial features, while GELU activation better adapts to changes in data distribution during training, improving generalization ability, especially when processing remotely sensed ship targets; similar residual structures are used at the input and output to alleviate gradient vanishing and improve detail recovery; the objective function is decomposed into the form H(x) = F(x) + x; the SE module learns feature weights based on the loss, so that effective feature maps have larger weights and ineffective or less effective feature maps have smaller weights.

5. The remote sensing ship target detection method based on deep learning according to claim 1, characterized in that, Design a feature fusion network with four detection scales, using the PAN_4 feature fusion network.

6. The remote sensing ship target detection method based on deep learning according to claim 5, characterized in that, The PAN_4 feature fusion network adds two lateral skip connections of the same scale to the P3 and P4 layers of YOLOv10. These two newly added lateral skip connections allow direct information exchange between feature maps of the same scale. This connection directly integrates the feature information extracted by the backbone network into the feature map to be detected.

7. The remote sensing ship target detection method based on deep learning according to claim 6, characterized in that, The relevant computations of the PAN_4 feature fusion network in the P3 layer of YOLOv10 are shown below: Where `out` represents the output of this network layer, `Conv()` is a convolution operation, `Upsample()` is a linear 2x upsampling operation, and `Contact()` directly sums the channels of all input features. The output features satisfy the following equation with each input feature: ， Where I represents the input feature, O represents the output feature, and Channel represents the number of channels.

8. The remote sensing ship target detection method based on deep learning according to claim 1, characterized in that, Design a rotating target detection loss function, using bidirectional decay loss BAIoU.

9. The remote sensing ship target detection method based on deep learning according to claim 8, characterized in that, The BAIoU loss function is: ， ， ， ， ， in, and The values ​​of hyperparameters can be obtained separately through projection formulas. The value is 1.75; and For the actual width and height, and To predict the width and height of the bounding box, and These are the vectors corresponding to the true value and the long side of the predicted bounding box, respectively. It is the inner product of vectors, and the absolute value of the inner product is taken to ensure that the included angle is less than 90° during the calculation process.

Images