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

Abstract

The application provides a UAV aerial low-visibility sea area target detection method, system and equipment, and belongs to the technical field of computer vision. The method comprises the following steps: acquiring a to-be-detected aerial image photographed by a UAV in a low-visibility sea area; inputting the image into a target detection model constructed based on a YOLOv11 framework to output a target category and a bounding box position. The model adopts an improved StarNet structure in a feature extraction backbone network, integrates a multi-scale convolution kernel and an efficient channel attention mechanism; a deep separable convolution and a high-frequency enhancement residual block HFERB are introduced in a neck network; a parameter-free attention mechanism SimAM is integrated in a classification and regression branch in a detection head. The model is trained by using a joint loss function, and a bounding box regression loss, a distribution focus loss and a classification loss are fused. The application improves the target detection precision and robustness under low-visibility conditions such as foggy days, and simultaneously reduces the computational complexity through lightweight design.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and particularly relates to a UAV aerial low-visibility sea area target detection method, system and device. BACKGROUND

[0002] Maritime search and rescue is an important part of the national emergency system. With the continuous growth of maritime traffic density, the risk of maritime accidents has increased accordingly. The traditional search and rescue mode has inherent limitations in response speed and coverage, and it is urgent to introduce efficient technical means as a supplement. Unmanned aerial vehicles (UAVs) have become an ideal auxiliary platform for maritime rescue due to their strong maneuverability, low cost, and ability to reach remote or dangerous areas, and have shown great application potential in target detection tasks in complex marine environments. However, maritime accidents are often accompanied by low-visibility adverse weather such as fog and night, which poses a serious challenge to target detection technology. Specifically, in low-visibility environments, there are mainly the following technical difficulties: first, the contrast between the target and the background decreases significantly, resulting in serious blurring of the 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 sea waves and water surface reflections); in addition, the scale of the target to be detected varies greatly, further increasing the complexity of detection.

[0003] Currently, deep learning-based target detection methods have been successfully applied in many fields. These methods can be mainly divided into two categories: two-stage detection algorithms and one-stage detection algorithms. The two-stage algorithm represented by the R-CNN series has higher detection accuracy, but its multi-stage processing flow leads to slower inference speed, making it difficult to meet the stringent real-time requirements of the UAV platform. In contrast, one-stage algorithms such as YOLO convert the detection task into a regression problem, achieving higher inference efficiency, and are therefore more suitable for marine target detection scenarios with high real-time requirements. Some research has attempted to apply YOLO series algorithms to marine environments. For example, by introducing attention mechanisms to improve small target detection performance, or by optimizing prior boxes and loss functions to improve the detection accuracy of targets such as ships. Another study combined YOLO with a lightweight network to enhance the model's robustness in complex scenarios. Some work has achieved high-precision detection of water surface targets through network pruning and other strategies.

[0004] However, the above studies mostly take clear images under natural light conditions as the design premise, and when they are directly applied to low-visibility marine environments, the model performance generally decreases significantly. This performance degradation is mainly due to the deterioration of image quality, making the feature extraction and semantic aggregation process more difficult. To address this problem, some studies have proposed solutions. For example, some methods attempt to improve detection accuracy in bad weather by combining image restoration and object detection frameworks, or use multi-task learning paradigms to jointly optimize image restoration and detection tasks. Another study combines an adaptive dehazing module with a detection network to enhance the recognizability of ships in foggy environments.

[0005] However, these existing solutions generally have a common defect: they generally require additional image preprocessing or repair steps before object detection. Such pre-processing modules introduce significant computational overhead, which is limited by the limited hardware resources of unmanned aerial vehicle platforms, making their actual deployment feasibility very low.

[0006] In summary, to achieve efficient and accurate offshore target detection under low-visibility conditions, the key is to design a model that can achieve the best balance between environmental robustness, detection accuracy, and computational resource consumption. How to achieve this balance is a technical problem that needs to be solved in this field. SUMMARY

[0007] The purpose of the present application is to provide a UAV aerial low-visibility sea area target detection method, system and equipment, to solve the technical problems of existing technology in low-visibility sea area environment, such as difficulty in extracting target features, insufficient detection accuracy, high model calculation complexity and difficulty in real-time deployment on UAV platform, and to achieve high-precision and high-efficiency real-time detection of offshore targets under bad weather conditions.

[0008] To achieve the above purpose, the present application provides a UAV aerial low-visibility sea area target detection method, comprising the following steps:

[0009] Obtaining a to-be-detected aerial image taken by a UAV in a low-visibility sea area;

[0010] Inputting the to-be-detected aerial image into a target detection model to output a target detection result, the target detection result including a target category and a bounding box position;

[0011] The target detection model is based on a YOLOv11 framework and includes:

[0012] The feature extraction backbone network is replaced with an improved StarNet structure, which uses convolution kernels of and in the first and second stages, and uses convolution kernels of and The convolutional kernels are optimized, and efficient channel attention (ECA) modules are integrated after the second and fourth stages, respectively.

[0013] 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.

[0014] The detection head integrates the parameter-free attention mechanism SimAM in both the classification and regression branches.

[0015] 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.

[0016] 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.

[0017] Preferably, the SimAM attention mechanism enhances features in the following ways:

[0018] Calculate the energy value corresponding to each neuron in the feature map;

[0019] The energy value is activated using the Sigmoid function to generate attention weights;

[0020] The attention weights are multiplied element-wise with the original feature map to obtain the enhanced features.

[0021] 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.

[0022] Preferably, the formula for calculating the joint loss function is as follows:

[0023] ;

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] wherein, denotes the total loss, denotes the classification loss, denotes the bounding box regression loss, denotes the distribution focal loss, , , denote the weight coefficients assigned to the classification loss, the bounding box regression loss and the distribution focal loss, respectively, denotes the sample index, denotes the total number of samples in a training batch, denotes the true label of the th sample, denotes the probability that the model predicts the presence of an object in the th sample, denotes the complete intersection over union loss, denotes the bounding box regression loss based on the third version of Wise-IoU, , denote the weight coefficients assigned to the and losses, respectively, denotes the intersection over union loss, denotes the length of the diagonal of the minimum enclosing rectangle that simultaneously encloses the predicted box and the ground truth box, denotes the aspect ratio difference between the two, denotes the weighting coefficient that balances this difference, denotes the Euclidean distance between the center points of the predicted box and the ground truth box, denotes the center point of the predicted box, denotes the center point of the ground truth box, and denote the coordinates of the predicted box, and denote the coordinates of the ground truth box, and denote the width and height of the minimum enclosing rectangle, respectively, denotes the power adjustment coefficient, denotes the weight adjustment coefficient, denotes the real probability distribution corresponding to each bounding box, denotes the distribution predicted by the model in the A probability value of a position.

[0030] Preferably, after obtaining the to-be-detected aerial image photographed by the unmanned aerial vehicle in the low-visibility sea area, the method further comprises: pre-processing the to-be-detected aerial image, and normalizing the image size to a uniform resolution.

[0031] The application further provides an unmanned aerial vehicle low-visibility sea area target detection system, comprising:

[0032] an image acquisition module configured to acquire a to-be-detected aerial image photographed by the unmanned aerial vehicle in the low-visibility sea area;

[0033] a detection processing module configured with a target detection model, configured to receive the to-be-detected aerial image and output a target detection result;

[0034] a result output module configured to output the target detection result.

[0035] Preferably, the target detection model is constructed based on a YOLOv11 framework and comprises the following improvements:

[0036] a feature extraction backbone network is replaced by an improved StarNet structure, the improved StarNet uses convolution kernels of and in the first stage and the second stage, uses convolution kernels of and in the third stage and the fourth stage, and integrates an efficient channel attention ECA module after the second stage and the fourth stage, respectively;

[0037] in the neck network, a standard convolution is replaced by a depth separable convolution, and a high-frequency enhancement residual block HFERB is integrated in the C3k2 module;

[0038] the detection head integrates a parameter-free attention mechanism SimAM in the classification branch and the regression branch.

[0039] The application further provides a computer device comprising a memory and a processor; the memory stores a computer program, and the processor implements the steps of the unmanned aerial vehicle low-visibility sea area target detection method when executing the computer program.

[0040] Therefore, the unmanned aerial vehicle low-visibility sea area target detection method, system and device have the following beneficial technical effects:

[0041] (1) The application constructs a special feature processing pipeline for low-visibility scenarios through the deep cooperation of the improved StarNet backbone network, HFERB neck network and SimAM detection head, and realizes the significant improvement of detection accuracy and robustness. The method is not simply stacking advanced modules, but through a systematic cooperation mechanism to form an end-to-end feature enhancement and decoupling process, which produces significant synergistic gain. Specifically, the improved StarNet backbone adaptively captures and strengthens the target channel from local details to global semantics in the early stage of feature extraction through multi-scale convolution kernels and ECA attention mechanism; on this basis, the HFERB neck network accurately connects the multi-scale features output by the backbone, and uses its high-frequency enhancement branch to specially recover and amplify the edge and texture detail information that is easy to lose in low-visibility environment; finally, the SimAM detection head makes full use of the rich features enhanced by the previous two stages, and adaptively allocates appropriate feature weights for classification and regression tasks through a parameter-free attention mechanism, realizing efficient decoupling and avoiding task conflicts. The three core components are closely linked to form a positive cycle: the backbone provides a strong feature basis, the neck supplements fine details, and the head realizes accurate decision-making. Experiments show that this collaborative design makes the model significantly better than the model using only a single or partial improvement in detection accuracy and stability of low-contrast and weak-texture targets under adverse conditions such as foggy and night, and fundamentally solves the core technical problem of feature weakening and blurring in low-visibility environment.

[0042] (2) The application is designed with a forward-looking lightweight architecture, which significantly reduces resource consumption while ensuring real-time performance, successfully solving the deployment bottleneck. While pursuing high performance, the method systematically incorporates lightweight ideas into every aspect of model design: the improved StarNet backbone network uses its efficient star-shaped operation as the core to realize lightweight feature extraction, the neck network uses deep separable convolution to greatly reduce computational complexity, and the detection head structure is optimized to improve feature extraction without adding additional parameters. These lightweight measures form an effective force at the system level, collectively reducing the total number of parameters and floating-point operations of the final MSA-YOLO model. Most importantly, this all-round lightweight design is achieved without sacrificing or even improving detection accuracy, indicating that the model has achieved optimal allocation of computing resources through efficient architecture collaboration. The final model has smaller size, faster inference speed and lower energy consumption demand, and can run smoothly on the embedded platform of the unmanned aerial vehicle with strict limitations on computing resources, fundamentally solving the engineering bottleneck problem that existing complex detection algorithms are difficult to deploy due to excessive computational overhead. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The structure of the MSA-YOLO network;

[0044] Figure 2 ECA structure;

[0045] Figure 3 multi-receptive field StarNet structure, wherein, Figure 3 (a) in (a) is a star operation, Figure 3 (b) in (b) is a multi-receptive field StarNet;

[0046] Figure 4 high-frequency enhancement residual block structure, wherein, Figure 4 (a) in (a) is a HFERB, Figure 4 (b) in (b) is a HFERB-C3k2;

[0047] Figure 5 SinAM structure, wherein, Figure 5 (a) in (a) is a SinAM, Figure 5 (b) in (b) is a MSA-YOLO detection head;

[0048] Figure 6 AFO dataset example, wherein, Figure 6 (a) in (a) contains original images, fog images and weak light images; Figure 6 (b) in (b) contains original images, fog images and weak light images;

[0049] Figure 7 loss curve of MSA-YOLO, wherein, Figure 7 (a) in (a) is a loss change curve of the training process, Figure 7 (b) in (b) is a mAP@0.5 change curve of the training process;

[0050] Figure 8 segmentation results of different networks in the AFO dataset;

[0051] Figure 9 segmentation results using different combination methods, wherein, Figure 9 (a) in (a) is Base, Figure 9 (b) in (b) is Proposal 1, Figure 9 (c) in (c) is Proposal 2, Figure 9 (d) in (d) is Proposal 3;

[0052] Figure 10 evaluation indexes obtained using different loss functions;

[0053] Figure 11 heat maps of YOLO11 and MSA-YOLO on test samples. DETAILED DESCRIPTION

[0054] The technical solutions of the present application are further described below with reference to the accompanying drawings and examples.

[0055] Unless otherwise defined, the technical terms or scientific terms used in the present application shall have the usual meanings understood by those with ordinary skills in the art to which the present application pertains.

[0056] Example 1

[0057] 1. Data set preparation and preprocessing.

[0058] This example uses the AFO data set (https: / / github.com / airforceofdata / afo-dataset) for model training and verification. The AFO data set contains 3647 images taken by unmanned aerial vehicles on the sea, covering six types of targets: humans, surfboards, sailboats, buoys, small boats, and kayaks. Through cropping and screening, 4216 images with a resolution of 640x640 pixels are obtained, and the channel-based fog synthesis algorithm and night synthesis algorithm are used to construct foggy and nighttime scene images, expanding the low-visibility environment data set (here, the prior art, which will not be described in detail). Figure 6

[0059] The data set is divided into a training set and a verification set in a ratio of 7:3.

[0060] 2. Target detection model construction.

[0061] 2.1 Overall framework.

[0062] This example builds an MSA-YOLO model based on the YOLOv11 framework, and the overall structure is as shown in Figure 1 The model includes three core improved parts: an improved multi-receptive field StarNet backbone network, a lightweight neck network integrating HFERB, and a detection head introducing SimAM.

[0063] 2.2 Multi-receptive field StarNet backbone network.

[0064] This example makes two key improvements to the StarNet backbone network:

[0065] (1) Multi-scale convolution kernel design: based on the receptive field theory, small-scale convolution kernels are used in the shallow stage to capture local details, and large-scale convolution kernels are used in the deep stage to capture global semantic information: the first and second stages use and convolution kernels, and the third and fourth stages use and convolution kernels.

[0066] (2) ECA attention mechanism integration: to address the problem of weak target feature response in low-visibility sea scenes, an efficient channel attention mechanism is introduced after the second and fourth stages, and the structure is as shown in Figure 2 ​​The processing process is as follows:

[0067] Global average pooling is performed on the input feature map;

[0068] The inter-channel dependency is calculated using one-dimensional convolution;

[0069] The channel attention weight is generated by the Sigmoid function;

[0070] The channel attention weight is multiplied with the original feature map channel by channel.

[0071] The improved multi-receptive field StarNet structure is as shown in Figure 3 .

[0072] 2.3, lightweight neck network based on HFERB-C3K2.

[0073] In low visibility environment, the edge and texture information of small scale target is easy to be blurred, which reduces the detection accuracy of the network for small scale target. At the same time, the calculation complexity of the conventional convolution layer in the traditional neck structure is high, which limits the efficiency of the model to some extent. In order to solve these problems, the neck network is optimized in two aspects:

[0074] (1) depth separable convolution: replace the standard convolution with depth separable convolution, which greatly reduces the calculation complexity.

[0075] (2) high frequency enhancement residual block (HFERB): the HFERB structure is as shown in Figure 4 , which includes two parallel branches:

[0076] Local feature extraction branch: sequentially through convolution layer and GELU activation function;

[0077] High frequency enhancement branch: sequentially through max pooling layer, convolution layer and GELU activation function;

[0078] The outputs of the two branches are spliced and fused through convolution, and are connected with the input features in residual connection.

[0079] 2.4, head network based on SimAM.

[0080] Integrate the parameter-free attention mechanism SimAM in the classification and regression branch, and the structure is as shown in Figure 5 . The processing process includes:

[0081] Calculate the energy value of each neuron in the feature map;

[0082] Generate attention weight using Sigmoid function;

[0083] ​The attention weight is multiplied element-wise with the original feature map.

[0084] 2.5, Joint loss function.

[0085] 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 accuracy of the model on the existence of the target, and the form of the BCE loss function is as follows:

[0086] ;

[0087] In the marine scene, the target usually shows large shape and position differences. The traditional IoU only focuses on the overlapping area of the predicted box and the real box, and it is difficult to effectively reflect the differences in position and shape. Therefore, the complete IoU (CIoU) loss is used:

[0088] ;

[0089] Marine targets have uneven spatial distribution, with dense areas near the sea and sparse areas in the open sea. This imbalance can lead to inaccurate annotations, which can negatively affect the model trained using the CIoU-based loss function. To alleviate this training bias, the WIoU loss is introduced, which effectively reduces the impact of low-quality samples on the gradient by assigning different weights to different samples. Finally, the above losses are weighted and summed according to the weights to form a complete bounding box regression loss function, whose expression is as follows:

[0090] ;

[0091] ;

[0092] Distribution Focal Loss (DFL) addresses the uncertainty in bounding box prediction by modeling multiple potential positions of the bounding box coordinates. Therefore, DFL is introduced to further alleviate the error caused by the blurring of the target boundary under low visibility conditions, and its expression is as follows:

[0093] ;

[0094] The total loss of model training can be represented as:

[0095] ;

[0096] where, Ltotal represents the total loss, Lcls represents the classification loss, denotes a bounding box regression loss, denotes a distribution focal loss, 、 、 denote weight coefficients assigned to the classification loss, the bounding box regression loss and the distribution focal loss, respectively, denotes a sample index, denotes the total number of samples in a training batch, denotes the true label of the th sample, denotes the probability that the model predicts the presence of an object in the th sample, denotes a complete intersection over union loss, denotes a bounding box regression loss based on the third version of Wise-IoU, 、 denote weight coefficients assigned to the and losses, respectively, denotes an intersection over union loss, denotes the length of the diagonal of the minimum bounding rectangle that simultaneously encloses the predicted box and the ground truth box, denotes the aspect ratio difference between the two, denotes a weighting coefficient that balances this difference, denotes the Euclidean distance between the center points of the predicted box and the ground truth box, denotes the center point of the predicted box, denotes the center point of the ground truth box, and denote the coordinates of the predicted box, and denote the coordinates of the ground truth box, and denote the width and height of the minimum bounding rectangle, respectively, denotes a power adjustment coefficient, denotes a weight adjustment coefficient, denotes the real probability distribution corresponding to each bounding box, denotes the probability value of the distribution predicted by the model at the th position.

[0097] The evaluation metrics include precision, recall, AP and mAP.

[0098] 3. Model training.

[0099] The experimental configuration is as follows:

[0100] Hardware environment: NVIDIA RTX 3060 GPU (8 GB memory);

[0101] Optimizer: SGD (momentum 0.94, weight decay 0.0005);

[0102] Learning rate: 0.001, with preheating strategy;

[0103] Training period: 200 epochs;

[0104] Batch size: 10.

[0105] The loss and performance indicators during training are shown in Figure 7 , indicating that the model training is stable and effective convergence.

[0106] 4. Experimental results and analysis.

[0107] 4.1. Comparative experiment.

[0108] The comparison results with five advanced lightweight target detection models are shown in Table 1.

[0109] Table 1. Results of different methods on AFO dataset images (%)

[0110]

[0111] MSA-YOLO shows the best performance in multiple key indicators, with a recall rate of 79.71%, an average precision of 86.39%, and a comprehensive detection performance of 52.36, which is better than the other five comparison methods. This indicates that the model has more accurate target recognition ability in low-visibility sea environments. It is particularly noteworthy that although MSA-YOLO is slightly insufficient in precision, its significant advantage in average precision fully proves that the model has more stable detection performance in complex environments. In terms of computational efficiency, MSA-YOLO contains only 4.52M parameters, which is significantly less than the comparison models YOLOX-s, YOLOv7-tiny, YOLOv8-s and YOLOv11-s. At the same time, the model's floating point operation and inference time are also better than YOLOv8-s and YOLOv11-s, indicating that MSA-YOLO can complete the detection task faster under the premise of lower computational resource consumption, achieving a good balance between precision and efficiency.

[0112] Figure 8The detection results of four typical test images in the AFO dataset are shown in the contrast, where the red box identifies false positives (mistaking the background as the target), and the blue box identifies false negatives (mistaking the target as the background). The six comparison methods show obvious performance differences in the detection results. Specifically, EfficientDet, YOLOv7-tiny, and YOLOv8-s have more false positives in the detection results, indicating that these models are prone to misidentifying non-target areas such as sea ripples as the target to be detected. Analysis shows that the errors of EfficientDet and YOLOv7-tiny are mainly due to insufficient semantic expression ability of high-level features, and the background noise interference in low-level features is large, making it difficult to effectively suppress the interference of background information in the feature fusion process; while the error of YOLOv8-s is mainly due to the insufficient fitting ability of the regression branch to the target boundary, causing positioning deviation and resulting in more false detections.

[0113] At the same time, YOLOX-s has more false negatives in the detection results, indicating that this model has obvious missing detection problems. This may be due to insufficient decoupling of the classification and regression branches, resulting in insufficient use of features for difficult targets. In contrast, although the proposed method still has some errors in individual cases (which may be due to the need to further improve the model's sensitivity to image noise or pixel changes), the overall number of false positives and false negatives is significantly less than other comparison methods. Especially in low-visibility scenarios such as foggy and night, the proposed method enhances the ability of detail feature extraction and semantic expression, and shows more accurate detection results in all test images. Qualitative analysis results show that the comprehensive detection performance of the proposed method on the AFO dataset is better than other comparison algorithms.

[0114] 4.2, Ablation experiment.

[0115] To verify the contribution of each module, an ablation experiment is designed as shown in Table 2.

[0116] Base: Use StarNet as the backbone network.

[0117] Proposal 1: Replace the traditional StarNet with a multi-receptive field StarNet.

[0118] Proposal 2: On the basis of Proposal 1, replace the standard convolution in the neck network with a depth separable convolution, and add a high-frequency enhancement residual block branch based on C3k2.

[0119] Proposal 3: On the basis of Proposal 2, further add a SimAM attention mechanism to the head structure part.

[0120] Table 2 Evaluation indexes (%) obtained by different combination methods

[0121]

[0122] Figure 9 The segmentation results obtained using different combination methods are shown. It can be observed that as each module is added step by step, the false positive rate and the false negative rate gradually decrease, which is reflected in the change in the number of red and blue borders. Specifically, compared with Base, Proposal 1 contains fewer blue borders and more red borders, indicating that the introduction of the multi-scale receptive field StarNet effectively enhances the detection capability of targets of different scales. However, relying solely on the multi-scale receptive field StarNet, while detecting more targets, it is difficult to accurately constrain the target area, resulting in an increase in false positives. Next, after introducing HFERB, the results of Proposal 2 compared to Proposal 1 contain fewer red borders, indicating that HFERB has a positive effect on reducing false positives. Further, after introducing SimAM to the head structure, the results of Proposal 3 compared to Proposal 2 contain fewer red and blue borders, indicating the effectiveness of feature optimization. Therefore, each module is effective and makes a positive contribution to detection performance to varying degrees.

[0123] 4.3, Loss function comparison.

[0124] To verify the applicability of the CIoU-WIoUv3 bounding box loss function in the MSA-YOLO model, a comparative experiment was designed, using the EIOU loss and the WIoUv3 loss as substitutes for training and verification. The experimental results show that Figure 10 ), on the AFO dataset, the CIoU-WIoUv3 loss is superior to the other two loss functions in terms of key indicators such as mAP@0.5, mAP@0.5:0.95, and recall rate. Detailed analysis shows that although the EIoU loss optimizes geometric matching by directly constraining the width-height difference, its linear square penalty may lead to excessive correction of the predicted box when dealing with extreme aspect ratio targets; while the WIoUv3 loss improves training stability through dynamic weight adjustment, but due to the lack of explicit geometric constraints, its optimization ability for bounding box accuracy is limited. In contrast, the CIoU-WIoUv3 loss effectively combines the geometric constraint advantage of the CIoU loss and the stability advantage of the WIoUv3 loss, while maintaining training stability, it significantly improves the positioning accuracy, and is more suitable for the detection needs of the MSA-YOLO model in complex sea environments.

[0125] 4.4, Comparison of backbone networks.

[0126] The comparison results of the multi-receptive field StarNet and other lightweight backbone networks are shown in Table 3.

[0127] The multi-receptive field StarNet is superior to StarNet in recall rate, mAP@0.5 and mAP@0.5:0.95. In particular, the multi-receptive field StarNet improves the recall rate by 3.22%, indicating that the multi-scale convolution kernel introduced effectively enhances the model's detection ability for targets of different sizes. In addition, although the multi-receptive field StarNet is lower than StarNet in precision, the higher mAP@0.5 and mAP@0.5:0.95 indicate that it is more advantageous in detection stability. On the other hand, using GhostNet is superior to MobileNetv2 in precision, mAP@0.5 and mAP@0.5:0.95, which may be because the Ghost convolution is more sufficient in channel information interaction compared to the depth separable convolution, thereby being able to extract more rich features. At the same time, the multi-receptive field StarNet is superior to GhostNet and MobileNetV2 in recall rate, mAP@0.5, mAP@0.5:0.95 and parameter amount, indicating that it has higher feature extraction capability while maintaining lower memory, thereby realizing more efficient detection.

[0128] Table 3 Evaluation indicators obtained by different encoding networks (%)

[0129]

[0130] 4.5, Feature visualization analysis.

[0131] In order to more intuitively visualize the attention distribution of MSA-YOLO in the image, Grad-CAM technology is used to generate attention heat maps. In these heat maps, red pixels represent areas of high model attention, while blue pixels represent areas of lower attention. Figure 11 The heat maps of YOLO11 and MSA-YOLO on the same test sample are shown, from which it can be seen that compared with YOLO11, the heat map generated by MSA-YOLO is more accurate in alignment with the actual target area, and pays less attention to noise. In addition, under various environmental conditions, the heat map generated by MSA-YOLO is always superior to YOLO11, further proving the excellent detection performance and robustness of MSA-YOLO.

[0132] Embodiment Two

[0133] A low-visibility sea area target detection system for a UAV, comprising:

[0134] An image acquisition module for acquiring a to-be-detected aerial image taken by a UAV in a low-visibility sea area;

[0135] The detection processing module is configured with a target detection model, and is used for receiving a to-be-detected aerial photograph image and outputting a target detection result.

[0136] The result output module is used for outputting the target detection result.

[0137] The target detection model is constructed based on a YOLOv11 framework and includes the following improvements:

[0138] The feature extraction backbone network is replaced with an improved StarNet structure, and the improved StarNet uses and convolution kernels in the first and second stages, uses and convolution kernels in the third and fourth stages, and integrates an efficient channel attention ECA module after the second and fourth stages, respectively;

[0139] In the neck network, the standard convolution is replaced with a depth separable convolution, and a high-frequency enhancement residual block HFERB is integrated in the C3k2 module;

[0140] The detection head integrates a parameter-free attention mechanism SimAM in the classification branch and the regression branch.

[0141] If the above functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application or the parts that essentially contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0142] The logic and / or steps represented in flow diagrams or otherwise described herein, for example, can be considered as a sequence of executable instructions, and can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For purposes of this specification, a "computer-readable medium" can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-readable medium can be, for example but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber (optical), and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example via an optical scanner, then compiled, interpreted, or otherwise processed, and stored in a computer memory in a form that can be later executed by a computer. In this context, the computer- readable medium can be paper or another suitable medium that can be human-readably presented with the program depicted on it, and the like.

[0143] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber (optical), and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example via an optical scanner, then compiled, interpreted, or otherwise processed, and stored in a computer memory in a form that can be later executed by a computer.

[0144] It is worth noting that the contents not elaborated in the present application are all prior art known to those skilled in the art.

[0145] Therefore, the unmanned aerial vehicle aerial low-visibility sea area target detection method, system and device of the present application use the improved StarNet backbone, HFERB neck and SimAM detection head to perform deep cooperation, not only improve the detection accuracy and robustness in foggy weather, night and other harsh conditions, but also greatly reduce the model complexity and resource consumption through system-level lightweight design, successfully solving the core problem of real-time deployment of high-precision detection models on resource-limited unmanned aerial vehicle platforms.

[0146] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application but not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: the technical solutions of the present application can still be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for detecting targets in low-visibility sea areas by UAV aerial photography, characterized in that, The method comprises the following steps: obtaining a to-be-detected aerial image photographed by a UAV in a low-visibility sea area; inputting the to-be-detected aerial image into a target detection model to output a target detection result, the target detection result comprising a target category and a bounding box position; the target detection model is constructed based on a YOLOv11 framework and comprises: The feature extraction backbone network is replaced with an improved StarNet structure, the improved StarNet uses and convolution kernels in the first stage and the second stage, uses and convolution kernels in the third stage and the fourth stage, and respectively integrates an efficient channel attention ECA module after the second stage and the fourth stage; in the neck network, a standard convolution is replaced by a depth separable convolution, and a high-frequency enhancement residual block HFERB is integrated in a C3k2 module; the detection head integrates a parameter-free attention mechanism SimAM in both a classification branch and a regression branch; the target detection model is trained by a joint loss function; the joint loss function is a weighted sum of a bounding box regression loss, a distribution focus loss and a classification loss; the bounding box regression loss is a weighted combination of a CIoU loss and a WIoUv3 loss. 2.The UAV aerial low-visibility sea area target detection method of claim 1, wherein, The low-visibility sea area comprises a fog condition and a thick fog condition; the fog condition refers to an image peak signal-to-noise ratio PSNR greater than 30 dB and a corresponding meteorological visibility of 500-1000 meters; the thick fog condition refers to an image peak signal-to-noise ratio PSNR less than 30 dB and a corresponding meteorological visibility less than 500 meters. 3.The UAV aerial low-visibility sea area target detection method of claim 1, wherein, 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 of claim 1, wherein, The SimAM attention mechanism enhances the features in the following manner: calculating an energy value corresponding to each neuron in a feature map; activating the energy value by using a Sigmoid function to generate an attention weight; performing element-wise multiplication of the attention weight and the original feature map to obtain enhanced features.

5. The method of claim 1, wherein, The calculation formula of the joint loss function is as follows: ； ； ； ； ； ； wherein, denotes the total loss, denotes the classification loss, denotes the bounding box regression loss, denotes the distribution focal loss, , , denote the weight coefficients assigned to the classification loss, the bounding box regression loss and the distribution focal loss, respectively, denotes the sample index, denotes the total number of samples in a training batch, denotes the true label of the th sample, denotes the probability that the model predicts the presence of an object in the th sample, denotes the complete intersection over union loss, denotes the bounding box regression loss based on the third version of Wise-IoU, , denote the weight coefficients assigned to the and losses, respectively, denotes the intersection over union loss, denotes the length of the diagonal of the minimum enclosing rectangle that simultaneously encloses the predicted box and the ground truth box, denotes the aspect ratio difference between the two, denotes the weighting coefficient that balances the difference, denotes the Euclidean distance between the center points of the predicted box and the ground truth box, denotes the center point of the predicted box, denotes the center point of the ground truth box, and denote the coordinates of the predicted box, and denote the coordinates of the ground truth box, and denote the width and height of the minimum enclosing rectangle, respectively, denotes the power adjustment coefficient, denotes the weight adjustment coefficient, denotes the real probability distribution corresponding to each bounding box, denotes the probability value of the distribution predicted by the model at the th position. 6.The UAV aerial low-visibility sea area target detection method of claim 1, wherein, After obtaining the to-be-detected aerial image photographed by the UAV in the low-visibility sea area, the method further comprises: pre-processing the to-be-detected aerial image to normalize the image size to a unified resolution.

7. A UAV low-visibility sea area target detection system, characterized in that, The method for detecting targets in a low-visibility sea area by a UAV according to any one of claims 1-6 comprises: an image acquisition module for acquiring a to-be-detected aerial image photographed by a UAV in a low-visibility sea area; a detection processing module configured with a target detection model for receiving the to-be-detected aerial image and outputting a target detection result; a result output module for outputting the target detection result.

8. The UAV low-visibility sea area target detection system of claim 7, wherein, The target detection model is constructed based on a YOLOv11 framework and comprises the following improvements: The feature extraction backbone network is replaced with an improved StarNet structure, the improved StarNet uses and convolution kernels in the first and second stages, uses and convolution kernels in the third and fourth stages, and respectively integrates an efficient channel attention ECA module after the second and fourth stages; in the neck network, a standard convolution is replaced by a depth separable convolution, and a high-frequency enhancement residual block HFERB is integrated in a C3k2 module; the detection head integrates a parameter-free attention mechanism SimAM in both a classification branch and a regression branch.

9. A computer device comprising: a memory and a processor; the memory stores a computer program, and the processor executes the computer program to implement the steps of the target detection method for a UAV aerial low-visibility sea area according to any one of claims 1-6.

Images