Unmanned aerial vehicle aerial photography low-visibility sea area target detection method, system and equipment

Abstract

The invention provides an unmanned aerial vehicle aerial photography low-visibility sea area target detection method, system and device, and belongs to the technical field of computer vision. The method comprises the following steps: acquiring a to-be-detected aerial image shot by an unmanned aerial vehicle in a low-visibility sea area; and inputting the image into a target detection model constructed based on a YOLOv11 framework, and outputting a target category and a bounding box position. The model adopts an improved StarNet structure in a feature extraction backbone network, and integrates a multi-scale convolution kernel and an efficient channel attention mechanism; introducing a deep separable convolution and a high-frequency enhanced residual block HFERB into the neck network; classification and regression branches in the detection head are both integrated with a parameter-free attention mechanism SimAM. The model is trained by adopting a joint loss function, and bounding box regression loss, distribution focal loss and classification loss are fused. According to the method, the target detection precision and robustness are improved under low-visibility conditions such as foggy days, and meanwhile, the calculation complexity is reduced through lightweight design.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method, system and equipment for detecting targets in low-visibility sea areas by drone aerial photography. Background Technology

[0002] Maritime search and rescue is a crucial component of the national emergency response system. With the continuous increase in maritime traffic density, the risk of maritime accidents is correspondingly rising. Traditional search and rescue methods have inherent limitations in response speed and coverage, necessitating the introduction of efficient technological supplements. Unmanned aerial vehicles (UAVs), with their advantages of high maneuverability, low cost, and ability to reach remote or dangerous areas, have become an ideal maritime rescue auxiliary platform, demonstrating significant application potential in target detection missions in complex marine environments. However, maritime accidents are often accompanied by low-visibility weather conditions such as fog and nighttime, posing severe challenges to target detection technology. Specifically, in low-visibility environments, the following technical difficulties exist: First, the contrast between the target and the background decreases significantly, resulting in severely blurred target edge information; second, the texture features of the target tend to weaken, making it more difficult to distinguish the target from background interference (such as waves and water surface reflections); furthermore, the large range of scale variations of the targets to be detected further increases the complexity of detection.

[0003] Currently, deep learning-based object detection methods have been successfully applied in numerous fields. These methods can be mainly divided into two categories: two-stage detection algorithms and one-stage detection algorithms. While two-stage algorithms, represented by the R-CNN series, offer high detection accuracy, their multi-stage processing leads to slow inference speeds, making it difficult to meet the stringent real-time requirements of UAV platforms. In contrast, one-stage algorithms such as YOLO transform the detection task into a regression problem, achieving higher inference efficiency and thus being more suitable for real-time-critical marine target detection scenarios. Some research has attempted to apply the YOLO series algorithms to marine environments. For example, by introducing attention mechanisms to improve the detection of small targets, or by optimizing prior boxes and loss functions to improve the detection accuracy of targets such as ships. Other research combines YOLO with lightweight networks to enhance the model's robustness in complex scenarios. Still other works have achieved high-precision detection of surface targets through strategies such as network pruning.

[0004] However, most of the aforementioned studies assume clear images under natural light conditions. When directly applied to low-visibility marine environments, the model performance generally declines significantly. This performance degradation mainly stems from the deterioration of image quality, making feature extraction and semantic aggregation more difficult. To address this issue, some studies have proposed solutions. For example, some methods attempt to improve detection accuracy in adverse weather conditions by combining image restoration and object detection frameworks, or employ a multi-task learning paradigm to jointly optimize image restoration and detection tasks. Other studies combine adaptive defogging modules with detection networks to enhance the identifiability of ships in foggy environments.

[0005] Nevertheless, these existing solutions generally share a common drawback: they typically require additional image preprocessing or restoration steps before target detection. Such preprocessing modules introduce significant computational overhead, and their practical deployment feasibility is very low due to the limited hardware resources of UAV onboard platforms.

[0006] In summary, achieving efficient and accurate maritime target detection under low visibility conditions hinges on designing a model that strikes the optimal balance between environmental robustness, detection accuracy, and computational resource consumption. Achieving this balance remains a critical technical challenge in this field. Summary of the Invention

[0007] The purpose of this invention is to provide a method, system, and equipment for detecting targets in low-visibility sea areas using unmanned aerial vehicles (UAVs), in order to solve the technical problems of existing technologies in low-visibility sea environments, such as difficulty in extracting target features, insufficient detection accuracy, and high model computational complexity that makes it difficult to deploy in real time on UAV platforms, and to achieve high-precision and high-efficiency real-time detection of maritime targets even under adverse weather conditions.

[0008] To achieve the above objectives, this invention provides a method for detecting targets in low-visibility sea areas using unmanned aerial vehicle (UAV) aerial photography, comprising the following steps: Acquire aerial images of the target object taken by a drone in low-visibility sea areas; The aerial image to be detected is input into the target detection model, and the target detection result is output, which includes the target category and bounding box location. The object detection model is built on the YOLOv11 framework and includes: The feature extraction backbone network was replaced with an improved StarNet structure, which was used in stages 1 and 2. and The convolution kernel used in stages 3 and 4 and The convolutional kernels are optimized, and efficient channel attention (ECA) modules are integrated after the second and fourth stages, respectively. In the neck network, standard convolutions are replaced with depthwise separable convolutions, and high-frequency enhancement residual blocks (HFERB) are integrated in the C3k2 module; The detection head integrates the parameter-free attention mechanism SimAM in both the classification and regression branches.

[0009] Preferably, low visibility sea areas include fog conditions and dense fog conditions; where fog conditions refer to an image peak signal-to-noise ratio (PSNR) greater than 30 dB and the corresponding meteorological visibility is 500-1000 meters; dense fog conditions refer to an image peak signal-to-noise ratio (PSNR) less than 30 dB and the corresponding meteorological visibility is less than 500 meters.

[0010] Preferably, the processing procedure for the high-frequency enhancement residual block (HFERB) includes: dividing the input feature map into two sub-features; feeding the first sub-feature into the high-frequency enhancement branch, and sequentially passing it through a max pooling layer, ... The convolutional layer and GELU activation function are used for processing; the second sub-feature is fed into the local feature extraction branch, and so on. The convolutional layers and GELU activation function are used for processing; the outputs of the two branches are concatenated and then processed... After the convolutional layers are fused, they are residually connected with the input feature map to obtain the output of HFERB.

[0011] Preferably, the SimAM attention mechanism enhances features in the following ways: Calculate the energy value corresponding to each neuron in the feature map; The energy value is activated using the Sigmoid function to generate attention weights; The attention weights are multiplied element-wise with the original feature map to obtain the enhanced features.

[0012] Preferably, the object detection model is trained using a joint loss function; the joint loss function is a weighted sum of bounding box regression loss, distribution focal loss, and classification loss; the bounding box regression loss is a weighted combination of CIoU loss and WIoUv3 loss.

[0013] Preferably, the formula for calculating the joint loss function is as follows: ; ; ; ; ; ; in, Indicates the total loss. Represents classification loss. This represents the bounding box regression loss. Indicates distributed focal loss. , , These represent the weight coefficients assigned to the classification loss, bounding box regression loss, and distribution focal loss, respectively. Indicates the sample index. This represents the total number of samples in a training batch. Indicates the first The true label of each sample The model predicts the first... The probability that the target exists in a sample. Indicates the loss of a complete intersection and union. This represents the third version of bounding box regression loss based on Wise-IoU. , They represent the allocations to and The weighting coefficients of the loss, Indicates the loss by intersection and union. This represents the diagonal length of the smallest bounding rectangle that simultaneously encloses the predicted bounding box and the ground truth bounding box. This indicates the difference in aspect ratio between the two. This represents the weighting coefficient that balances the difference. This represents the Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box. Indicates the center point of the prediction box. Indicates the center point of the true bounding box. and Indicates the coordinates of the prediction box. and Represents the coordinates of the true bounding box. and Let represent the width and height of the minimum bounding rectangle, respectively. This represents the power-law adjustment coefficient. This represents the weighting adjustment coefficient. This represents the true probability distribution corresponding to each bounding box. The distribution predicted by the model is in the th order. The probability value of each position.

[0014] Preferably, after acquiring the aerial images to be detected taken by the UAV in low-visibility sea areas, the method further includes: preprocessing the aerial images to be detected and normalizing the image size to a uniform resolution.

[0015] This invention also provides a low-visibility sea target detection system for unmanned aerial vehicles, comprising: The image acquisition module is used to acquire aerial images of the target area taken by the UAV in low-visibility sea areas; The detection and processing module is equipped with a target detection model, which is used to receive the aerial images to be detected and output the target detection results; The results output module is used to output the target detection results.

[0016] Preferably, the object detection model is built on the YOLOv11 framework and includes the following improvements: The feature extraction backbone network was replaced with an improved StarNet structure, which was used in stages 1 and 2. and The convolution kernel used in stages 3 and 4 and The convolutional kernels are optimized, and efficient channel attention (ECA) modules are integrated after the second and fourth stages, respectively. In the neck network, standard convolutions are replaced with depthwise separable convolutions, and high-frequency enhancement residual blocks (HFERB) are integrated in the C3k2 module; The detection head integrates the parameter-free attention mechanism SimAM in both the classification and regression branches.

[0017] The present invention also provides a computer device, including: a memory and a processor; the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method for detecting targets in low-visibility sea areas by drone aerial photography.

[0018] Therefore, the present invention employs the above-mentioned method, system, and equipment for detecting targets in low-visibility sea areas using UAV aerial photography, and the beneficial technical effects are as follows: (1) This invention constructs a dedicated feature processing pipeline for low-visibility scenarios through deep collaboration between the improved StarNet backbone network, the HFERB neck network, and the SimAM detection head, achieving a significant improvement in detection accuracy and robustness. This method is not simply a matter of stacking advanced modules, but rather a systematic collaborative mechanism that forms an end-to-end feature enhancement and decoupling process, generating significant synergistic gains. Specifically, the improved StarNet backbone adaptively captures robust features from local details to global semantics and strengthens the target channel in the early stages of feature extraction through multi-scale convolutional kernels and ECA attention mechanism; on this basis, the HFERB neck network accurately receives the multi-scale features output by the backbone, and uses its high-frequency enhancement branch to specifically recover and amplify edge and texture details that are easily lost in low-visibility environments; finally, the SimAM detection head fully utilizes the rich features after the first two enhancements, and adaptively assigns appropriate feature weights to classification and regression tasks through a parameterless attention mechanism, achieving efficient decoupling and avoiding task conflicts. These three core components are interconnected, forming a positive cycle: the backbone provides a strong feature foundation, the neck network supplements fine details, and the head network achieves quasi-decision. Experiments show that this collaborative design enables the model to significantly outperform models that use only one or partial improvements in terms of detection accuracy and stability for low-contrast, weak-texture targets under adverse conditions such as fog and nighttime. This fundamentally solves the core technical challenge of feature weakening and blurring in low-visibility environments.

[0019] (2) This invention, with its forward-looking lightweight architecture design, significantly reduces resource consumption while ensuring real-time performance, successfully solving the deployment bottleneck. While pursuing high performance, this method systematically integrates lightweight principles into every aspect of model design: the improved StarNet backbone network uses its efficient star-shaped operation as the core to achieve lightweight basic feature extraction; the neck network uses depthwise separable convolutions to significantly reduce computational complexity; and the detection head structure is carefully optimized to improve feature extraction without adding extra parameters. These lightweight measures work together effectively at the system level, significantly reducing the total number of parameters and floating-point operations in the final MSA-YOLO model. Crucially, this comprehensive lightweight design is achieved without sacrificing or even improving detection accuracy, demonstrating that the model achieves optimal allocation of computing resources through efficient architectural collaboration. The final model has a smaller size, faster inference speed, and lower energy consumption, enabling smooth operation on UAV embedded platforms with strictly limited computing resources, fundamentally solving the engineering bottleneck problem of existing complex detection algorithms being difficult to deploy practically due to excessive computational overhead. Attached Figure Description

[0020] Figure 1 The structure of the MSA-YOLO network; Figure 2 It is an ECA structure; Figure 3 For the multi-receptive field StarNet structure, in which, Figure 3 (a) in the diagram represents the star schema operation. Figure 3 (b) in the diagram represents StarNet with multiple receptive fields; Figure 4 It is a high-frequency enhanced residual block structure, in which, Figure 4 (a) in the text is HFERB. Figure 4 (b) in the example is HFERB-C3k2; Figure 5 It is a SinAM structure, where, Figure 5 In this context, (a) represents SinAM. Figure 5 (b) in the image represents the MSA-YOLO detection head; Figure 6 Here is an example of the AFO dataset, where, Figure 6 (a) in the image contains the original image, the fogged image, and the low-light image; Figure 6 (b) includes the original image, the fogged image, and the low-light image; Figure 7 Here is the loss curve for MSA-YOLO, where... Figure 7 In the figure, (a) is the loss change curve during the training process. Figure 7 (b) in the figure shows the mAP@0.5 change curve during the training process; Figure 8 The segmentation results using different networks in the AFO dataset; Figure 9 The results show the segmentation outcomes using different combinations of methods, where... Figure 9 (a) in the text is Base. Figure 9 (b) in the text refers to Proposal 1. Figure 9 (c) in the text refers to Proposal 2. Figure 9 (d) in the text refers to Proposal 3; Figure 10 Evaluation metrics obtained using different loss functions; Figure 11 Heatmaps of YOLO11 and MSA-YOLO on the test samples. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] Example 1 1. Dataset preparation and preprocessing.

[0024] This embodiment uses the AFO dataset ( Figure 6 The model was trained and validated using the AFO dataset, which contains 3,647 images of the sea taken by drones, covering six categories of targets: humans, surfboards, sailboats, buoys, small boats, and kayaks. 4,216 images were obtained through cropping and filtering. The image is pixel-based, and fog and night scene images are constructed using channel-based fog synthesis and night synthesis algorithms to expand the low visibility environment dataset (this is existing technology and will not be elaborated further).

[0025] The dataset is divided into training and validation sets in a 7:3 ratio.

[0026] 2. Target detection model construction.

[0027] 2.1 Overall Framework.

[0028] This embodiment constructs an MSA-YOLO model based on the YOLOv11 framework, and its overall structure is as follows: Figure 1 As shown, the model comprises three core improvements: an improved multi-receptive-field StarNet backbone, a lightweight neck network integrating HFERB, and a detection head incorporating SimAM.

[0029] 2.2. StarNet backbone network with multiple receptive fields.

[0030] This embodiment makes two key improvements to the StarNet backbone network: (1) Multi-scale convolutional kernel design: Based on receptive field theory, small-scale convolutional kernels are used to capture local details in shallow layers, and large-scale convolutional kernels are used to capture global semantic information in deep layers: First stage and second stage: using and Convolution kernel third and fourth stages: using and Convolution kernel.

[0031] (2) ECA attention mechanism integration: To address the issue of weak target feature response in low-visibility maritime scenarios, an efficient channel attention mechanism is introduced after the second and fourth stages, with the structure as follows: Figure 2 As shown. The processing procedure is as follows: Perform global average pooling on the input feature map; One-dimensional convolution is used to calculate inter-channel dependencies; Channel attention weights are generated using the Sigmoid function; Multiply the channel attention weights with the original feature map channel by channel.

[0032] The improved multi-receptive field StarNet structure is as follows: Figure 3 As shown.

[0033] 2.3 Lightweight neck network based on HFERB-C3K2.

[0034] In low-visibility environments, the edges and textures of small-scale targets are easily blurred, reducing the network's detection accuracy for such targets. Simultaneously, the conventional convolutional layers in the traditional neck structure have high computational complexity, limiting the model's efficiency to some extent. To address these issues, the neck network undergoes optimization in two aspects: (1) Depthwise separable convolution: replacing standard convolution with depthwise separable convolution greatly reduces computational complexity.

[0035] (2) High Frequency Enhancement Residual Block (HFERB): The structure of the HFERB is as follows Figure 4 As shown, it contains two parallel branches: Local feature extraction branch: sequentially through Convolutional layers and the GELU activation function; High-frequency enhancement branch: sequentially through the max pooling layer, Convolutional layers and the GELU activation function; The outputs of the two branches are concatenated and then passed. Convolutional fusion is performed, and residual connections are made with the input features.

[0036] 2.4. SimAM-based head network.

[0037] SimAM, a parameter-free attention mechanism, is integrated into the classification and regression branches, with the structure as follows: Figure 5 As shown. The processing procedure includes: Calculate the energy value of each neuron in the feature map; Use the Sigmoid function to generate attention weights; The attention weights are multiplied element-wise with the original feature map.

[0038] 2.5 Joint Loss Function.

[0039] The loss function in object detection consists of two main components: classification loss and bounding box loss. In the classification loss part, the Binary Cross-Entropy (BCE) loss is used to measure the model's accuracy in detecting the presence of the target. The BCE loss function takes the following form: ; In maritime scenarios, targets often exhibit significant differences in shape and location. Traditional IoU only focuses on the overlap area between the predicted and ground truth bounding boxes, which fails to effectively reflect the differences in their location and shape. Therefore, Complete IoU (CIoU) loss is employed: ; Maritime targets exhibit an uneven spatial distribution, with high density in nearshore areas and low density in offshore areas. This imbalance can lead to inaccurate annotations, negatively impacting models trained using CIoU-based loss functions. To mitigate this training bias, WIoU loss is introduced. By assigning differentiated weights to different samples, it effectively reduces the influence of low-quality samples on the gradient. Finally, the above losses are weighted and summed to form the complete bounding box regression loss function, as expressed below: ; ; Distribution Focal Loss (DFL) addresses the uncertainty in bounding box prediction by modeling multiple potential locations of the bounding box coordinates. Therefore, DFL is introduced to further mitigate the error caused by target boundary blurring under low visibility conditions. Its expression is as follows: ; The total loss during model training can be expressed as: ; in, Indicates the total loss. Represents classification loss. This represents the bounding box regression loss. Indicates distributed focal loss. , , These represent the weight coefficients assigned to the classification loss, bounding box regression loss, and distribution focal loss, respectively. Indicates the sample index. This represents the total number of samples in a training batch. Indicates the first The true label of each sample The model predicts the first... The probability that the target exists in a sample. Indicates the loss of a complete intersection and union. This represents the third version of bounding box regression loss based on Wise-IoU. , They represent the allocations to and The weighting coefficients of the loss, Indicates the loss by intersection and union. This represents the diagonal length of the smallest bounding rectangle that simultaneously encloses the predicted bounding box and the ground truth bounding box. This indicates the difference in aspect ratio between the two. This represents the weighting coefficient that balances the difference. This represents the Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box. Indicates the center point of the prediction box. Indicates the center point of the true bounding box. and Indicates the coordinates of the prediction box. and Represents the coordinates of the true bounding box. and Let represent the width and height of the minimum bounding rectangle, respectively. This represents the power-law adjustment coefficient. This represents the weighting adjustment coefficient. This represents the true probability distribution corresponding to each bounding box. The distribution predicted by the model is in the th order. The probability value of each position.

[0040] Evaluation metrics include precision, recall, AP, and mAP.

[0041] 3. Model training.

[0042] The experimental setup is as follows: Hardware environment: NVIDIA RTX 3060 GPU (8GB RAM); Optimizer: SGD (momentum 0.94, weight decay 0.0005); Learning rate: 0.001, using a warm-up strategy; Training cycle: 200 epochs; Batch size: 10.

[0043] Loss and performance metrics changes during training, such as Figure 7 As shown, this indicates that the model training is stable and converges effectively.

[0044] 4. Experimental Results and Analysis.

[0045] 4.1 Comparative experiment.

[0046] The comparison results with five advanced lightweight target detection models are shown in Table 1.

[0047] Table 1. Results (%) of different methods on AFO dataset images

[0048] MSA-YOLO demonstrated superior performance across multiple key metrics, achieving recall, average precision, and overall detection performance of 79.71%, 86.39%, and 52.36%, respectively, comprehensively outperforming the other five comparison methods. This indicates that the model possesses more accurate target identification capabilities in low-visibility marine environments. Notably, while MSA-YOLO lags slightly behind in precision, its significant advantage in average precision fully demonstrates its more stable detection performance in complex environments. In terms of computational efficiency, MSA-YOLO uses only 4.52M parameters, significantly fewer than the comparison models YOLOX-s, YOLOv7-tiny, YOLOv8-s, and YOLOv11-s. Furthermore, its floating-point computation and inference time are also superior to YOLOv8-s and YOLOv11-s, indicating that MSA-YOLO can complete detection tasks faster while maintaining lower computational resource consumption, achieving a good balance between accuracy and efficiency.

[0049] Figure 8 The comparison of detection results for four typical test images from the AFO dataset is presented, with red boxes indicating false positives (misclassifying background as the target) and blue boxes indicating false negatives (misclassifying the target as background). The six methods exhibit significant performance differences in the detection results. Specifically, EfficientDet, YOLOv7-tiny, and YOLOv8-s show a high number of false positives, indicating that these models are prone to misidentifying non-target areas such as sea surface ripples as the target. Analysis suggests that the errors of EfficientDet and YOLOv7-tiny mainly stem from insufficient semantic representation of high-level features and significant background noise interference in low-level features, making it difficult to effectively suppress background information interference during feature fusion. The errors of YOLOv8-s are primarily due to its regression branch's insufficient ability to fit the target boundary, resulting in localization bias and thus more false detections.

[0050] Meanwhile, YOLOX-s showed a significant number of false negatives, indicating a clear problem of missed detections. This may be due to insufficient decoupling of features between its classification and regression branches, leading to inadequate utilization of features for difficult targets. In contrast, while our proposed method still exhibits some errors in individual cases (possibly stemming from the need for further improvement in the model's sensitivity to image noise or pixel variations), the overall number of false positives and false negatives is significantly lower than other comparative methods. Particularly in low-visibility scenarios such as fog and nighttime, our method demonstrates more accurate detection results across all test images by enhancing detailed feature extraction and semantic expression capabilities. Qualitative analysis results show that our method outperforms other comparative algorithms in overall detection performance on the AFO dataset.

[0051] 4.2 Ablation experiment.

[0052] To verify the contribution of each module, an ablation experiment was designed as shown in Table 2.

[0053] Base: StarNet is used as the backbone network.

[0054] Proposal 1: Replace the traditional StarNet with a multi-receptive-field StarNet.

[0055] Proposal 2: Based on Proposal 1, the standard convolutions in the neck network are replaced with depthwise separable convolutions, and high-frequency enhanced residual block branches are added to the C3k2 network.

[0056] Proposal 3: Building upon Proposal 2, it further incorporates the SimAM attention mechanism into the head structure.

[0057] Table 2 Evaluation indicators (%) obtained by different combination methods

[0058] Figure 9 The segmentation results obtained using different combinations of methods are presented. It can be observed that the false positive and false negative rates gradually decrease with the gradual addition of each module, which is reflected in the change in the number of red and blue borders. Specifically, compared to Base, Proposal 1 contains fewer blue borders and more red borders, indicating that the introduced multi-scale receptive field StarNet effectively enhances the detection capability for targets at different scales. However, relying solely on the multi-scale receptive field StarNet, while detecting more targets, makes it difficult to accurately constrain the target regions, leading to an increase in false positives. Next, after introducing HFERB, the result of Proposal 2 contains fewer red borders compared to Proposal 1, indicating that HFERB has a positive effect on reducing false positives. Furthermore, after introducing SimAM into the head structure, the result of Proposal 3 contains fewer red and blue borders compared to Proposal 2, indicating the effectiveness of feature optimization. Therefore, each module is effective and contributes positively to the detection performance to varying degrees.

[0059] 4.3 Comparison of loss functions.

[0060] To verify the applicability of the CIoU-WIoUv3 bounding box loss function in the MSA-YOLO model, a comparative experiment was designed, using EIOU loss and WIoUv3 loss as alternatives for training and validation. Experimental results show ( Figure 10On the AFO dataset, the CIoU-WIoUv3 loss outperforms the other two loss functions in key metrics such as mAP@0.5, mAP@0.5:0.95, and recall. Specifically, the analysis shows that while the EIoU loss optimizes geometric matching by directly constraining the aspect ratio difference, its linear squared penalty may lead to over-correction of the predicted bounding box when dealing with targets with extreme aspect ratios. While the WIoUv3 loss improves training stability through dynamic weight adjustment, its ability to optimize bounding box accuracy is limited due to the lack of explicit geometric constraints. In contrast, the CIoU-WIoUv3 loss effectively combines the geometric constraint advantages of the CIoU loss with the stability advantages of the WIoUv3 loss, significantly improving localization accuracy while maintaining training stability, making it more suitable for the detection needs of the MSA-YOLO model in complex marine environments.

[0061] 4.4 Comparison of backbone networks.

[0062] The comparison results of StarNet with multiple receptive fields and other lightweight backbone networks are shown in Table 3.

[0063] Multi-receptive-field StarNet outperforms StarNet in recall, mAP@0.5, and mAP@0.5:0.95. Specifically, Multi-receptive-field StarNet improves recall by 3.22%, indicating that the introduced multi-scale convolutional kernels effectively enhance the model's ability to detect targets of different sizes. Furthermore, although Multi-receptive-field StarNet has lower precision than StarNet, its higher mAP@0.5 and mAP@0.5:0.95 suggest a greater advantage in detection stability. On the other hand, GhostNet outperforms MobileNetv2 in precision, mAP@0.5, and mAP@0.5:0.95. This is likely because Ghost convolutions, compared to depthwise separable convolutions, are more efficient in terms of channel information interaction, thus extracting richer features. Meanwhile, in terms of recall, mAP@0.5, mAP@0.5:0.95, and number of parameters, StarNet with multiple receptive fields outperforms GhostNet and MobileNetV2, indicating that it has higher feature extraction capabilities while maintaining lower memory usage, thus achieving more efficient detection.

[0064] Table 3 Evaluation metrics (%) obtained from different coding networks

[0065] 4.5 Feature visualization analysis.

[0066] To more intuitively visualize the attention distribution of MSA-YOLO in images, attention heatmaps were generated using the Grad-CAM technique. In these heatmaps, red pixels represent areas of high model attention, while blue pixels represent areas of lower attention. Figure 11 The heatmaps of YOLO11 and MSA-YOLO on the same test sample are shown. It can be seen that, compared to YOLO11, the heatmap generated by MSA-YOLO is more accurately aligned with the actual target area and pays less attention to noise. Furthermore, under various environmental conditions, the heatmap generated by MSA-YOLO consistently outperforms that of YOLO11, further demonstrating the superior detection performance and robustness of MSA-YOLO.

[0067] Example 2 A low-visibility marine target detection system for unmanned aerial vehicles (UAVs) includes: The image acquisition module is used to acquire aerial images of the target area taken by the UAV in low-visibility sea areas; The detection and processing module is equipped with a target detection model, which is used to receive the aerial images to be detected and output the target detection results; The results output module is used to output the target detection results.

[0068] The object detection model is built on the YOLOv11 framework and includes the following improvements: The feature extraction backbone network was replaced with an improved StarNet structure, which was used in stages 1 and 2. and The convolution kernel used in stages 3 and 4 and The convolutional kernels are optimized, and efficient channel attention (ECA) modules are integrated after the second and fourth stages, respectively. In the neck network, standard convolutions are replaced with depthwise separable convolutions, and high-frequency enhancement residual blocks (HFERB) are integrated in the C3k2 module; The detection head integrates the parameter-free attention mechanism SimAM in both the classification and regression branches.

[0069] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0071] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0072] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0073] Therefore, the present invention adopts the above-mentioned UAV aerial photography low-visibility sea target detection method, system and equipment, and deeply coordinates the improved StarNet backbone, HFERB neck and SimAM detection head, which not only improves the detection accuracy and robustness under adverse conditions such as fog and night, but also significantly reduces the model complexity and resource consumption through system-level lightweight design, and successfully solves the core problem of real-time deployment of high-precision detection models on resource-constrained UAV platforms.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for detecting targets in low-visibility sea areas using unmanned aerial vehicle (UAV) aerial photography, characterized in that: Includes the following steps: Acquire aerial images of the target object taken by a drone in low-visibility sea areas; The aerial image to be detected is input into the target detection model, and the target detection result is output, which includes the target category and bounding box location. The object detection model is built on the YOLOv11 framework and includes: The feature extraction backbone network was replaced with an improved StarNet structure, which was used in stages 1 and 2. and The convolution kernel used in stages 3 and 4 and The convolutional kernels are optimized, and efficient channel attention (ECA) modules are integrated after the second and fourth stages, respectively. In the neck network, standard convolutions are replaced with depthwise separable convolutions, and high-frequency enhancement residual blocks (HFERB) are integrated in the C3k2 module; The detection head integrates the parameter-free attention mechanism SimAM in both the classification and regression branches.

2. The method for detecting targets in low-visibility sea areas by unmanned aerial vehicle (UAV) aerial photography according to claim 1, characterized in that, Low visibility sea areas include fog conditions and dense fog conditions; fog conditions refer to image peak signal-to-noise ratio (PSNR) greater than 30 dB and corresponding meteorological visibility of 500-1000 meters; dense fog conditions refer to image peak signal-to-noise ratio (PSNR) less than 30 dB and corresponding meteorological visibility less than 500 meters.

3. The method for detecting targets in low-visibility sea areas by unmanned aerial vehicle (UAV) aerial photography according to claim 1, characterized in that, The processing steps of the High Frequency Enhancement Residual Block (HFERB) include: dividing the input feature map into two sub-features; feeding the first sub-feature into the high frequency enhancement branch, and sequentially passing it through a max pooling layer, ... The convolutional layer and GELU activation function are used for processing; the second sub-feature is fed into the local feature extraction branch, and so on. The convolutional layers and GELU activation function are used for processing; the outputs of the two branches are concatenated and then processed... After the convolutional layers are fused, they are residually connected with the input feature map to obtain the output of HFERB.

4. The method for detecting targets in low-visibility sea areas by unmanned aerial vehicle (UAV) aerial photography according to claim 1, characterized in that, SimAM's attention mechanism enhances features in the following ways: Calculate the energy value corresponding to each neuron in the feature map; The energy value is activated using the Sigmoid function to generate attention weights; The attention weights are multiplied element-wise with the original feature map to obtain the enhanced features.

5. The method for detecting targets in low-visibility sea areas by unmanned aerial vehicle (UAV) aerial photography according to claim 1, characterized in that, The object detection model is trained using a joint loss function; the joint loss function is a weighted sum of bounding box regression loss, distribution focal loss, and classification loss; the bounding box regression loss is a weighted combination of CIoU loss and WIoUv3 loss.

6. The method for detecting targets in low-visibility sea areas by unmanned aerial vehicle (UAV) aerial photography according to claim 5, characterized in that, The formula for calculating the joint loss function is as follows: ； ； ； ； ； ； in, Indicates the total loss. Represents classification loss, This represents the bounding box regression loss. Indicates distributed focal loss. , , These represent the weight coefficients assigned to the classification loss, bounding box regression loss, and distribution focal loss, respectively. Indicates the sample index. This represents the total number of samples in a training batch. Indicates the first The true label of each sample The model predicts the first... The probability that the target exists in a sample. Indicates the loss of a complete intersection and union. This represents the third version of bounding box regression loss based on Wise-IoU. , They represent the allocations to and The weighting coefficients of the loss, Indicates the loss by intersection and union. This represents the diagonal length of the smallest bounding rectangle that simultaneously encloses the predicted bounding box and the ground truth bounding box. This indicates the difference in aspect ratio between the two. This represents the weighting coefficient that balances the difference. This represents the Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box. Indicates the center point of the prediction box. Indicates the center point of the true bounding box. and Indicates the coordinates of the prediction box. and Represents the coordinates of the true bounding box. and Let represent the width and height of the minimum bounding rectangle, respectively. This represents the power-law adjustment coefficient. This represents the weighting adjustment coefficient. This represents the true probability distribution corresponding to each bounding box. The distribution predicted by the model is in the th order. The probability value of each position.

7. The method for detecting targets in low-visibility sea areas by unmanned aerial vehicle (UAV) aerial photography according to claim 1, characterized in that, After acquiring the aerial images to be tested taken by the drone in low-visibility sea areas, the process also includes: preprocessing the aerial images to be tested and normalizing the image size to a uniform resolution.

8. A low-visibility sea target detection system for unmanned aerial vehicles (UAVs), characterized in that: include: The image acquisition module is used to acquire aerial images of the target area taken by the UAV in low-visibility sea areas; The detection and processing module is equipped with a target detection model, which is used to receive the aerial images to be detected and output the target detection results; The results output module is used to output the target detection results.

9. The UAV low-visibility sea target detection system according to claim 8, characterized in that, The object detection model is built on the YOLOv11 framework and includes the following improvements: The feature extraction backbone network was replaced with an improved StarNet structure, which was used in stages 1 and 2. and The convolution kernel used in stages 3 and 4 and The convolutional kernels are optimized, and efficient channel attention (ECA) modules are integrated after the second and fourth stages, respectively. In the neck network, standard convolutions are replaced with depthwise separable convolutions, and high-frequency enhancement residual blocks (HFERB) are integrated in the C3k2 module; The detection head integrates the parameter-free attention mechanism SimAM in both the classification and regression branches.

10. A computer device, comprising: Memory and processor; The memory stores a computer program, characterized in that when the processor executes the computer program, it implements the steps of the method for detecting targets in low-visibility sea areas by drone aerial photography as described in any one of claims 1-7.

Images