Method and system for identifying shallow water target in optical remote sensing image
By enhancing and expanding the shallow water target data set and introducing a variety of attention modules into the YOLOv11 recognition network, the problem of low recognition accuracy in shallow water target recognition is solved, and higher recognition accuracy and stronger detection capabilities are achieved.
Patent Information
- Application Number
- CN202411708704.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-27
- Publication Date
- 2025-03-04
AI Technical Summary
In the recognition of shallow water targets, due to poor image quality and insufficient sample number, the recognition accuracy is low, making it difficult for deep learning technology to effectively identify shallow water targets.
Through data augmentation and data expansion, the original shallow water target data set is integrated to obtain a more comprehensive shallow water target data set, and on this basis, the improved YOLOv11 recognition network is trained, including the introduction of an efficient spatial compression attention module, a multiple attention feature fusion module and a prominent position attention mechanism module to enhance the performance of the recognition network.
The accuracy of shallow water target recognition is improved, the model's detection ability of complex backgrounds and small targets is enhanced, missed and missed detection is reduced, and normalized environmental early warning and situational awareness in shallow water areas are realized.
Smart Images
Figure CN119559509A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of target recognition, and in particular to a method and system for recognizing shallow water targets in optical remote sensing images. Background Art
[0002] In order to maintain the order of our country's territorial waters, it is crucial to accurately locate and identify various types of sea surface and shallow water targets. Compared with underwater sonar detection systems, satellite reconnaissance and surveillance systems have the characteristics of long range, wide coverage, convenience and safety for surface and shallow water target identification. In civilian use, surface target detection and identification can provide information services for activities such as ship entry and exit management, combating illegal maritime smuggling, supervising the development of marine fisheries, integrated coastal zone management, and ship oil spill pollution detection. At present, the data that can be used for surface target detection mainly include infrared images, optical remote sensing images, and synthetic aperture radar images. However, due to the absorption of water bodies and the influence of sea conditions such as high sea conditions and ocean current changes, the image quality of infrared images and synthetic aperture radar images varies greatly, showing an extremely uneven phenomenon, and cannot complete the effective detection of shallow water targets. Therefore, optical remote sensing methods currently have certain exploration and application prospects in shallow water targets.
[0003] Since shallow water targets have a low probability of appearing in remote sensing images and the number of samples is insufficient, the learning efficiency of the recognition algorithm is greatly limited. Although deep learning technology can perform large-scale image feature classification and learning, the probability of shallow water targets appearing is low and there is no image database in this specialized field, which in turn limits the recognition accuracy of the model. Summary of the invention
[0004] The present invention provides a method and system for recognizing shallow water targets in optical remote sensing images, so as to improve the target recognition accuracy.
[0005] In order to solve the above technical problems, the present invention provides a method for identifying shallow water targets in optical remote sensing images, comprising: S1. The original shallow water target dataset is processed by data enhancement and data expansion respectively, and the processed datasets are integrated to obtain the shallow water target dataset; S2. Based on the shallow water target dataset, train the improved YOLOv11 recognition network; S3. Input the optical remote sensing image to be identified into the trained improved YOLOv11 recognition network to obtain the recognition result. According to the above scheme, the data enhancement includes Mosaic data enhancement, MixUp data enhancement, and HideAndSeek data enhancement.
[0006] According to the above scheme, the method for processing the original shallow water target data set by using data expansion is: using a multi-branch conditional generation network to perform data expansion on the original shallow water target data set.
[0007] According to the above scheme, the method for acquiring the original shallow water target data set is: merging the public shallow water target data and manually annotated remote sensing images into the original shallow water target data set.
[0008] According to the above scheme, the improved YOLOv11 recognition network is improved compared to the YOLOv11 recognition network, and its improvements include: enhancing the C3K2 module in the backbone network of the YOLOv11 recognition network by introducing an efficient spatial compression attention module.
[0009] According to the above scheme, the improved YOLOv11 recognition network is improved compared to the YOLOv11 recognition network, and its improvements include: enhancing part of the C3K2 module in the neck network of the YOLOv11 recognition network by introducing a multiple attention feature fusion module.
[0010] According to the above scheme, the improved YOLOv11 recognition network is improved compared to the YOLOv11 recognition network, and its improvements include: enhancing part of the C3K2 module in the neck network of the YOLOv11 recognition network by introducing a significant position attention mechanism module.
[0011] According to the above scheme, the improved YOLOv11 recognition network is trained using a loss function, wherein the loss function includes a CIoU loss function, a binary cross entropy function, and a classification loss function added with set weights.
[0012] According to the above scheme, the classification loss function adopts a classifier based on BCE loss.
[0013] The present invention also provides a system for recognizing shallow water targets in optical remote sensing images, comprising: A shallow water target data set acquisition module is used to process the original shallow water target data set by using data enhancement and data expansion, and integrate the processed data sets to obtain the shallow water target data set; The shallow water target recognition module includes a trained improved YOLOv11 recognition network, which is used to obtain recognition results based on the input optical remote sensing image.
[0014] Beneficial effects: The present invention enhances and expands the original shallow water target data set through data enhancement and data expansion, obtains a more comprehensive shallow water target data set, provides a basis for the incremental learning of the subsequent recognition network, and improves the accuracy of the final recognition result. In addition, the present invention improves the original YOLOv11 recognition network, expands and strengthens the original recognition framework, and further enhances the recognition effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1is a flow chart of a method for identifying shallow water targets in optical remote sensing images according to a first embodiment of the present invention; Figure 2 is a flowchart of shallow water target data set acquisition in Embodiment 1 of the present invention; Figure 3 Schematic diagram of a multi-branch conditional generative adversarial network algorithm according to the first embodiment of the present invention; Figure 4 2 is a schematic diagram of a multi-branch conditional generative adversarial network structure according to the first embodiment of the present invention; Figure 5 Schematic diagram of the YOLOv11 recognition network structure of the first embodiment of the present invention; Figure 6 is a schematic diagram of the structure of an efficient spatial compression attention module according to the first embodiment of the present invention; Figure 7 is a schematic diagram of the structure of a multiple attention feature fusion module according to the first embodiment of the present invention; Figure 8 It is a schematic diagram of the structure of the salient position attention mechanism module of Example 1 of the present invention. DETAILED DESCRIPTION
[0016] Embodiment 1: See also Figure 1 , Figure 2 , a method for recognizing shallow water targets in optical remote sensing images, comprising: S1. The original shallow water target dataset is processed by data enhancement and data expansion respectively, and the processed datasets are integrated to obtain the shallow water target dataset; S2. Based on the shallow water target dataset, train the improved YOLOv11 recognition network; S3. Input the optical remote sensing image to be identified into the trained improved YOLOv11 recognition network to obtain the recognition result. Furthermore, the data enhancement includes Mosaic data enhancement, MixUp data enhancement, and HideAndSeek data enhancement; Among them, Mosaic and MixUp data enhancement enrich data diversity and enhance the adaptability of the model by splicing multiple pictures or linearly superimposing different images and labels. Color jitter adjusts parameters such as brightness and contrast to simulate different lighting conditions and improve robustness in complex scenes. Random cropping and scaling help the model better adapt to multi-scale targets, especially for the detection of small targets. In addition, random flipping and rotation increase the direction and perspective diversity of the data. The HideAndSeek technology blocks some areas and forces the model to learn local features, thereby more effectively dealing with complex situations where the target and background overlap.
[0017] Furthermore, the method for acquiring the original shallow water target dataset is: merging public shallow water target data and manually annotated remote sensing images into the original shallow water target dataset; the combination of these enhancement methods significantly improves the performance and versatility of YOLO v11 in remote sensing image target detection.
[0018] Furthermore, the method for processing the original shallow water target data set by using data expansion is: using a multi-branch conditional generation network to perform data expansion on the original shallow water target data set; the multi-branch conditional generation network used in this embodiment is shown in Figure 3 , Figure 4 .
[0019] Furthermore, YOLO v11 adopts an improved backbone architecture (C3K2 module) and an optimized neck structure (SpatialPyramid Feature Fusion, SPFF module), which can effectively extract details and semantic features under complex backgrounds, especially for the detection of small targets. Given that YOLO v11 can ensure a faster detection speed of the network while also ensuring better detection accuracy of the target, which is very important for target detection in remote sensing images, YOLOv11 is selected as the basic detection framework. See also Figure 5 Compared with the YOLOv11 recognition network, the improved YOLOv11 recognition network has the following improvements: 1) See Figure 6 By introducing the Efficient Spatial Reduction Attention (ESRA) module to enhance the C3K2 module in the backbone network of the YOLOv11 recognition network, the detection performance can be significantly improved; the core of the efficient spatial compression attention is to compress the key and value of the multi-head self-attention, thereby reducing the number of parameters and computational complexity while retaining key spatial information; this feature makes it very suitable for processing high-resolution remote sensing images, especially in resource-limited edge device environments, and can achieve efficient detection; applying the efficient spatial compression attention module to the backbone network can efficiently compress redundant features in the early stage, making downstream feature extraction more accurate, thus laying the foundation for detecting potential targets; In practical applications, the efficient spatial compression attention module has multiple advantages: first, it can effectively improve the feature extraction capability, and can efficiently capture the regional features of complex features and interweaving of targets and backgrounds in remote sensing images, making the model more robust in complex scenes; second, efficient spatial compression attention enhances the detection capability of small targets and multi-scale targets; through the precise allocation of attention, efficient spatial compression attention can better capture the detailed information of small targets, while adapting to multi-scale changes, and significantly improve the detection accuracy of specific targets such as ships and buildings; in addition, efficient spatial compression attention plays an important role in the background interference problem unique to remote sensing images; by paying attention to the target area and effectively suppressing background information, efficient spatial compression attention can significantly reduce the interference of irrelevant information such as complex terrain and clouds on detection; these characteristics comprehensively improve the performance of the model in complex remote sensing tasks, making it particularly suitable for the detection of shallow water targets, ships and other specific targets; 2) See Figure 7 , by introducing the Multi-Attention Feature Fusion (MAFF) module to enhance some C3K2 modules in the neck network of the YOLOv11 recognition network; it can effectively improve the performance of YOLOv11 in complex remote sensing tasks, especially in the need for precise detection of small targets such as shallow water targets and ships; Integrating the multi-attention feature fusion module into the neck structure of YOLO v11 can significantly improve the performance of complex remote sensing image tasks, especially the detection ability of specific targets such as shallow water targets and ships; the multi-attention feature fusion module effectively fuses features of different scales through the synergy of multiple attention mechanisms (efficient attention gating, efficient channel attention and spatial attention), enhancing the model's understanding of complex backgrounds and target areas; this feature enables the model to more accurately extract the features of small and dense targets, while effectively suppressing background interference, making it more robust in remote sensing images; The multi-attention feature fusion module connects low-level and high-level features to ensure that detail information loss is reduced in feature fusion, thereby retaining the edge and detail features of the target in the target detection task and improving detection accuracy. In addition, the multi-attention feature fusion module reduces the computational complexity through grouped convolution and residual connection, while ensuring performance improvement without significantly increasing the model parameters and hardware burden, which is particularly suitable for resource-constrained edge device environments. The efficient feature fusion strategy of multi-attention feature fusion works in conjunction with the YOLO v11 feature pyramid structure, which helps to improve the model's detection performance for multi-scale targets, thereby significantly improving the detection accuracy of long-distance targets, shallow-water targets, and ships in remote sensing images. This enhanced feature enables YOLO v11 to perform well in task scenarios with complex background interference, rich target detail information, and limited hardware resources. 3) See Figure 8 , by introducing the Salient Position Attention Block (SPAB) to enhance some C3K2 modules in the neck network of the YOLOv11 recognition network; the feature fusion mechanism of YOLO v11 is designed to better combine the underlying and high-level features to improve the target detection performance; however, in the process of deep feature transmission, the long path may cause the feature information of small targets to be lost, resulting in missed detection and false detection. In order to solve this problem, the Salient Position Attention Block module is introduced; this module is based on the Salient Position Selection (SPS) algorithm, which selects a small number of salient positions for calculation, thereby reducing the computational complexity and memory usage, and enhancing the feature expression ability; The working principle of SPAB is to first convert the feature map extracted by the backbone network into a query matrix (Q) and a numerical matrix (V) through a two-dimensional convolution to prepare for attention calculation; then, the SPS algorithm is used to select the salient positions according to the square sum of the query matrix, which can reduce the positions that need to be paid attention to and reduce the computational overhead; then, the attention matrix is calculated using the salient positions, normalized through softmax, and then matrix multiplied with the numerical matrix to aggregate important contextual information; finally, the features are updated through a 1×1 convolution, and the updated features are added to the input features to form a new output feature map; SPAB is particularly suitable for application in the low layers of the network, because low-level feature maps usually contain more spatial details, and SPAB can effectively extract these significant information and improve the ability to express details. After integrating SPAB into the feature fusion module of YOLO v11, the network is more efficiently utilized when processing the underlying information, significantly reducing the loss of deep feature information. This improvement not only improves the detection accuracy of small and multi-scale targets, but also effectively reduces missed detections and false detections, thereby significantly enhancing the network's target detection capabilities. Overall, by introducing SPAB, YOLO v11 performs better in target detection in complex scenarios, especially in multi-scale target detection and detail feature extraction.
[0020] Further, the improved YOLOv11 recognition network is trained using a loss function, wherein the loss function includes a CIoU loss function, a binary cross entropy function, and a classification loss function added with set weights; The CIoU loss function: The position regression branch of the bounding box considers the IoU, position relationship, aspect ratio, etc. of the bounding box at the same time, and adopts CIoU-Loss, which is only calculated for the assigned positive samples; The binary cross entropy function: The main task of the confidence branch of the bounding box is to distinguish whether there is a target to be detected in the bounding box, which can be regarded as a binary classification problem. Therefore, the binary cross entropy (BCE) loss function can be used, and considering the quality of the positive samples amplified during model training, the labels are weighted based on IoU. The loss is calculated for all predicted boxes; the standard BCE loss is:
[0021] Among them, y is the label value and p is the predicted probability; The classification loss function: For single-category target detection, this branch is ignored; for general multi-classification problems, it is more common to train a classifier based on Softmax loss, but its calculation results are mutually exclusive and cannot be used for multi-label classification tasks, so it is converted to training C (C is the number of categories) classifiers based on BCE loss, which also has reliable performance in multi-label tasks. This loss is only calculated for the assigned positive samples; In summary, the loss function of the improved YOLOv11 recognition network is:
[0022] In the above formula, α, γ, and δ are weight coefficients of the CIoU loss function, the binary cross entropy function, and the classification loss function, which are used to correct the numerical relationship and the backpropagation gradient to prevent the training model from being dominated by a certain branch; in this embodiment, the default values of α, γ, and δ are 0.05, 1.0, and 0.5, respectively.
[0023] It should be noted that IOU (Intersection Over Union) is a commonly used concept in the field of target detection, called the intersection over union ratio, which is used to simply measure the performance of the model algorithm and determine the degree of overlap between the predicted candidate box (candidate bound) and the original annotation box (ground truth bound), that is, the ratio of the intersection area to the union area of the two, where candidate is the area of the candidate box and ground truth is the area of the original annotation box; in the most ideal state, IOU should be 1, indicating that the candidate box completely overlaps with the original annotation box; In early object detection algorithms, the bounding box was regressed directly on the center point of the bounding box and its height and width using loss functions such as Mean Squared Error (MSE) loss and SmoothL1, which resulted in a slow model convergence speed. Deviations at the same distance contributed equally to the loss. Since the scales of the targets to be measured are different, their tolerance for errors at the same distance varies greatly, which can easily lead to inaccurate positioning of small targets. Based on the above, the existing technology proposes IOU Loss to regress the bounding box as a whole, taking the correlation of each vertex of the bounding box into consideration, and defines IOU Loss as follows:
[0024] IOU can reflect the target detection effect, but there are still some problems when using IOU to calculate gradients for regression. First, if the two bounding boxes do not intersect, according to the definition of IOU, IOU is equal to 0. At this time, the distance between the two boxes cannot be reflected, and because the loss is 0, learning and training cannot be performed. Second, the IOU indicator cannot accurately reflect the degree of overlap between the two boxes. When IOU is a fixed value, there may be multiple situations, and the regression effects between different situations are far apart. Therefore, this solution is not adopted in this embodiment.
[0025] The beneficial effects of applying the method described in this embodiment include: targeting the recognition needs of targets such as boats, landforms, and artificial objects in shallow water areas, and combining with hot spots of interest to obtain near-real-time remote sensing images in advance. Although shallow water targets are less likely to appear, there is no image database for this specialized field. However, this method can be used to identify shallow water targets, use limited shallow water targets for incremental learning, and use deep learning technology to expand and strengthen the recognition framework to achieve normalized near-coastal regional environmental warnings and enhance wide-area situational awareness; it can also be used for early detection of external areas to form a regional situation map, providing environmental information for environmental perception and near-shore security in shallow water areas.
[0026] Embodiment 2: This embodiment provides a system for recognizing shallow water targets in optical remote sensing images, including: A shallow water target data set acquisition module is used to process the original shallow water target data set by using data enhancement and data expansion, and integrate the processed data sets to obtain the shallow water target data set; The shallow water target recognition module includes a trained improved YOLOv11 recognition network, which is used to obtain recognition results based on the input optical remote sensing image.
[0027] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, and two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.
[0028] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for identifying shallow water targets in optical remote sensing images, characterized in that: include: S1. The original shallow water target dataset is processed by data enhancement and data expansion respectively, and the processed datasets are integrated to obtain the shallow water target dataset; S2. Based on the shallow water target dataset, train the improved YOLOv11 recognition network; S3. Input the optical remote sensing image to be identified into the trained improved YOLOv11 recognition network to obtain the recognition result.
2. The method for identifying shallow water targets in optical remote sensing images according to claim 1, characterized in that: The data enhancement includes Mosaic data enhancement, MixUp data enhancement, and HideAndSeek data enhancement.
3. The method for recognizing shallow water targets in optical remote sensing images according to claim 1, characterized in that: The method for processing the original shallow water target data set by using data expansion is: using a multi-branch conditional generation network to perform data expansion on the original shallow water target data set.
4. The method for recognizing shallow water targets in optical remote sensing images according to claim 1, characterized in that: The method for acquiring the original shallow water target data set is: merging public shallow water target data and manually annotated remote sensing images into the original shallow water target data set.
5. The method for recognizing shallow water targets in optical remote sensing images according to claim 1, characterized in that: Compared with the YOLOv11 recognition network, the improved YOLOv11 recognition network includes: enhancing the C3K2 module in the backbone network of the YOLOv11 recognition network by introducing an efficient spatial compression attention module.
6. The method for recognizing shallow water targets in optical remote sensing images according to claim 1, characterized in that: Compared with the YOLOv11 recognition network, the improved YOLOv11 recognition network includes: enhancing part of the C3K2 module in the neck network of the YOLOv11 recognition network by introducing a multiple attention feature fusion module.
7. The method for recognizing shallow water targets in optical remote sensing images according to claim 1, characterized in that: Compared with the YOLOv11 recognition network, the improved YOLOv11 recognition network includes: enhancing part of the C3K2 module in the neck network of the YOLOv11 recognition network by introducing a salient position attention mechanism module.
8. The method for recognizing shallow water targets in optical remote sensing images according to claim 1, characterized in that: The improved YOLOv11 recognition network is trained using a loss function, wherein the loss function includes a CIoU loss function, a binary cross entropy function, and a classification loss function added with set weights.
9. The method for recognizing shallow water targets in optical remote sensing images according to claim 8, characterized in that: The classification loss function adopts a classifier based on BCE loss.
10. An optical remote sensing image shallow water target recognition system, characterized in that: include: A shallow water target data set acquisition module is used to process the original shallow water target data set by using data enhancement and data expansion, and integrate the processed data sets to obtain the shallow water target data set; The shallow water target recognition module includes a trained improved YOLOv11 recognition network, which is used to obtain recognition results based on the input optical remote sensing image.