YOLO-based maritime unmanned aerial vehicle cooperative identification small target detection method and device

Abstract

The invention, which belongs to the technical field of electronic data processing, discloses a YOLO-based detection method and device for cooperatively identifying a small target by a marine unmanned aerial vehicle, and the method comprises the steps: obtaining a marine target region image; inputting the marine target area image into a trained machine learning model to obtain a small target detection result; wherein the trained machine learning model comprises a basic detection layer, a small target detection layer and a bidirectional feature pyramid network; the basic detection layer is used for detecting low-resolution features, and the small target detection layer is used for detecting high-resolution features; and the bidirectional feature pyramid network is used for performing multi-scale feature fusion on the low-resolution features and the high-resolution features. According to the method, the detail information of the small target is reserved in the target detection process, so that the detection result is remarkably improved in the aspects of accuracy, recall rate, average accuracy mean value and the like.

Description

Technical Field

[0001] This invention relates to the field of electronic data processing technology, specifically to a method and apparatus for collaborative identification and detection of small targets by unmanned aerial vehicles (UAVs) based on YOLO. Background Technology

[0002] Before the widespread adoption of deep learning, object detection tasks primarily relied on traditional machine learning algorithms such as AdaBoost and SVM. These algorithms depended on manually extracted features, such as gradient histograms and local binary patterns, to perform object detection. However, this approach often performed poorly in complex and varied image scenarios, with both detection accuracy and speed failing to meet the requirements of practical applications.

[0003] In recent years, two main types of deep learning-based object detection algorithms have emerged in the development of object detection technology: one-stage and two-stage algorithms. One-stage algorithms directly add an object detection branch to the feature extraction network, directly obtaining the object's category and probability; the entire process is single-stage. However, these algorithms can lose important information such as target features, resulting in shortcomings in complex maritime environments, particularly in detecting small targets, occluded targets, and targets in low-light conditions, where they still suffer from poor accuracy and a tendency to miss detections. Summary of the Invention

[0004] This invention provides a method and apparatus for collaborative identification and detection of small targets by unmanned aerial vehicles (UAVs) based on YOLO to improve target detection accuracy and prevent missed detections.

[0005] A first aspect of this invention provides a method for cooperative identification and detection of small targets by unmanned aerial vehicles (UAVs) at sea based on YOLO, the method comprising:

[0006] Acquire images of the target area at sea;

[0007] By inputting images of maritime target areas into a trained machine learning model, small target detection results are obtained.

[0008] The trained machine learning model includes a basic detection layer, a small object detection layer, and a bidirectional feature pyramid network. The basic detection layer is used to detect low-resolution features, and the small object detection layer is used to detect high-resolution features. The bidirectional feature pyramid network is used to perform multi-scale feature fusion of low-resolution and high-resolution features.

[0009] In some optional embodiments of the present invention, the small target detection layer is located in the upsampling path of the neck network of the trained machine learning model and is connected to the high-resolution feature map generation layer to detect small targets in the input image.

[0010] In some alternative embodiments of the present invention, the bidirectional feature pyramid network learns and evaluates the importance of each input feature during training, so as to add an additional weight parameter to each input feature during feature fusion, thereby achieving multi-scale feature fusion.

[0011] In some optional embodiments of the present invention, feature fusion is achieved through the following expression:

[0012]

[0013] Where O represents the output feature after feature fusion; ω i represents the weight of the i-th input feature; o represents a very small positive number.

[0014] In some optional embodiments of the present invention, the trained machine learning model further includes a channel attention module and a spatial attention module;

[0015] The channel attention module is used to perform global max pooling downsampling and global average pooling downsampling on the features of each channel, and then feeds the features of each channel into a fully connected layer for normalization to obtain the first feature map;

[0016] The spatial attention module is used to perform global average pooling and global max pooling on each channel of the first feature map, and then concatenates, convolves and normalizes the two pooling results to obtain the second feature map.

[0017] In some optional embodiments of the present invention, the attention process of the channel attention module is implemented by the following expression:

[0018]

[0019] Where F is the input feature map; M c (F) is the channel attention weight map obtained after the channel attention module processes the feature map F; F′ is the feature map after channel attention weighting.

[0020] The attention process of the spatial attention module is implemented through the following expression:

[0021]

[0022] Where F′ is the feature map after channel attention weighting; M s (F′) is the spatial attention weight map obtained by the spatial attention module after processing the feature map F′; F″ is the final feature map after spatial attention weighting.

[0023] In some optional embodiments of the present invention, the training process of the trained machine learning model includes:

[0024] Collect image data of ships at sea from multiple sources;

[0025] The image data is enhanced to form a dataset.

[0026] A second aspect of this invention provides a YOLO-based collaborative identification and detection device for small targets by unmanned aerial vehicles (UAVs) at sea. This device may include:

[0027] The acquisition module is used to acquire images of target areas at sea;

[0028] The recognition module is used to input images of maritime target areas into a trained machine learning model to obtain small target detection results;

[0029] The trained machine learning model includes a basic detection layer, a small object detection layer, and a bidirectional feature pyramid network. The basic detection layer is used to detect low-resolution features, and the small object detection layer is used to detect high-resolution features. The bidirectional feature pyramid network is used to perform multi-scale feature fusion of low-resolution and high-resolution features.

[0030] A third aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the YOLO-based collaborative identification and small target detection method for maritime unmanned aerial vehicles according to any one of the embodiments of the first aspect.

[0031] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method of YOLO-based collaborative identification and small target detection for maritime unmanned aerial vehicles according to any one of the embodiments of the first aspect.

[0032] This invention adds a small target detection layer and a bidirectional feature pyramid network to the machine learning model, enabling the trained machine learning model to retain detailed information about small targets during the target detection process. This results in a significant improvement in the model's precision, recall, and mean precision, thereby enhancing the detection accuracy of small targets at sea. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the YOLO-based collaborative identification and detection method for small targets by unmanned aerial vehicles (UAVs) at sea, as provided in an embodiment of the present invention.

[0034] Figure 2 This is a schematic diagram of the YOLOv5 network structure provided in an embodiment of the present invention;

[0035] Figure 3 A schematic diagram of the CBS-YOLO network structure provided in an embodiment of the present invention;

[0036] Figure 4 A schematic diagram of the CBAM structure provided in an embodiment of the present invention;

[0037] Figure 5 This is a schematic diagram of the CAM structure provided in an embodiment of the present invention;

[0038] Figure 6 This is a schematic diagram of the SAM structure provided in an embodiment of the present invention;

[0039] Figure 7 This is a schematic diagram of the BiFPN feature fusion structure provided in an embodiment of the present invention;

[0040] Figure 8 A comparison diagram of the feature fusion structures of BiFPN, FPN, PANet, and NAS-FPN provided in the embodiments of the present invention;

[0041] Figure 9 This is a distribution diagram of training set information provided for embodiments of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0043] like Figure 1 As shown, in a first aspect of the present invention, a method for cooperative identification and detection of small targets by unmanned aerial vehicles (UAVs) at sea based on YOLO is provided. This method may include:

[0044] S110: Acquire images of the target area at sea;

[0045] S120: Input the image of the target area at sea into the trained machine learning model to obtain the small target detection results;

[0046] The trained machine learning model includes a basic detection layer, a small object detection layer, and a bidirectional feature pyramid network. The basic detection layer is used to detect low-resolution features, and the small object detection layer is used to detect high-resolution features. The bidirectional feature pyramid network is used to perform multi-scale feature fusion of low-resolution and high-resolution features.

[0047] The improved base model of the machine learning model trained in this embodiment is YOLOv5, whose network structure mainly includes four parts: input, backbone, neck, and head output. Figure 2 The diagram shows the network structure of YOLOv5.

[0048] The backbone network consists of Conv, C3, and SPPF modules, which are mainly responsible for extracting features from the input image and generating feature maps. The neck network adopts the FPN (Feature Pyramid Network) structure, which is responsible for multi-scale feature fusion of the feature maps and passing the fused features to the prediction layer. The head network consists of three parts: Anchors, Classification, and Regression. It performs the final regression prediction through three detectors and outputs the location and category information of the target.

[0049] The CBS-YOLO model, as an improved framework of YOLOv5, retains the advantages of YOLOv5 while further optimizing and enhancing it. To improve the model's accuracy in detecting small objects, a small object detection layer is added. Simultaneously, a CBAM attention mechanism is introduced at the end of the backbone network, allowing the model to focus more on key regions and salient features in the image. Furthermore, a BiFPN structure is integrated into the neck network to fully utilize and fuse feature information from different levels. The loss function is changed to SIoU to simplify computation, improve efficiency, and enhance bounding box localization accuracy. The overall structure of CBS-YOLO is as follows: Figure 3 As shown.

[0050] The original YOLOv5 model cleverly utilizes detectors at three scales: 20×20, 40×40, and 80×80. This design allows the model to capture some detail on targets at different scales. However, in practical applications, extremely small targets are frequently encountered, such as distant pedestrians or tiny ships at sea. For these small targets, the original detectors fall short in terms of resolution and feature extraction capabilities, resulting in poor detection performance and frequent false positives and false negatives. Furthermore, when the target size changes significantly, the original detectors may not adapt well. This not only affects the accuracy of target detection but also limits the application of the YOLOv5 model in more complex scenarios.

[0051] To address the aforementioned issues, a small object detection layer P4 is proposed to be added to the YOLOv5 model. Object detection is performed directly on the high-resolution feature map, thus preserving the detailed information of small objects to the maximum extent. This process is mainly achieved by adding a detection head and an anchor box layer. This detection layer is located in the upsampling path of the Neck part, immediately following the high-resolution feature map generation layer. By inserting an additional detection head here, small objects in the input image are specifically detected. This new detection head has a higher resolution, namely a 160×160 scale, meaning the model can capture more image details. This is crucial for small object detection, as the feature information of small objects is often very weak and can only be effectively extracted by a high-resolution detection head. Experimental results also show that the newly added small object detection layer significantly improves the detection capability for small objects, especially performing excellently in dense ship scenes and complex marine backgrounds.

[0052] CBAM (Convolutional Block Attention Module) is an attention mechanism in deep learning designed to enhance the ability of convolutional neural networks to model and represent image features. It improves model performance by applying attention weights to the channels and spatial dimensions of the image. In object detection tasks, CBAM helps the network accurately locate and identify objects at different scales. Its structure is as follows: Figure 4 As shown, CBAM consists of two key components: the Channel Attention Module and the Spatial Attention Module.

[0053] Depend on Figure 4 As can be seen, assuming the input feature map is: F∈R C×H×W CBAM sequentially derives the one-dimensional channel attention diagram M. c ∈R C×1×l And two-dimensional spatial attention diagram M s ∈R 1×H×W The entire attention process can be summarized as follows:

[0054]

[0055] The Channel Attention Module (CAM) structure is as follows: Figure 5 As shown, it first performs global max pooling downsampling and global average pooling downsampling on the features of each channel of the input feature map F, then feeds the features of each channel into a fully connected layer (MLP), and then uses the sigmoid activation function to normalize the activation of each channel, thus obtaining the final feature map M. c(Channel Attention). The entire process can be expressed by the following formula:

[0056] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) (3)

[0057] Where σ is the sigmoid function.

[0058] The structure of the Spatial Attention Module (SAM) is as follows: Figure 6 As shown. It first performs global average pooling and global max pooling on each channel of the feature map F′ obtained in the previous step. Global max pooling will yield... Figure 6 The blue feature map, obtained by global average pooling. Figure 6 The orange feature map is then concatenated with the two pooling results and passed through a convolutional layer. Finally, the activation of each channel is normalized using the sigmoid activation function to obtain the final feature map M. s (SpatialAttention). The entire process can be expressed by the following formula:

[0059] M s (F)=σ(f 7×7 / [AvgPool(F);MaxPool(F)])) (4)

[0060] Where σ is the sigmoid function, f 7×7 This indicates that the convolution operation uses a 7×7 convolution kernel.

[0061] In summary, CBAM is a lightweight and versatile module that can be easily integrated into the YOLOv5 network architecture as a plug-and-play component, with almost no additional computational burden. Introducing the CBAM module into the YOLOv5 backbone and feature fusion network (Neck) demonstrates significant advantages in feature representation capabilities, detection accuracy, and robustness.

[0062] Bi-directional Feature Pyramid Network (BiFPN) is a multi-level feature pyramid network that optimizes the traditional Feature Pyramid Network (FPN). By constructing bidirectional channels to achieve cross-scale connections, it directly fuses features from the feature extraction network with relatively large features from the bottom-up path. This allows for simple and fast multi-scale feature fusion and aims to solve the problems of feature propagation and information flow in object detection.

[0063] like Figure 7The diagram illustrates the BiFPN feature fusion architecture, where P3, P4, P5, P6, and P7 represent the output layers of the backbone network. Each output layer has a corresponding output feature (containing information such as the number of channels and the feature size). For example, the output feature size of P3 is half the resolution of the input image. 3 The output feature size of P4 is half the resolution of the input image. 4 Similarly, the output feature size of P7 is half the resolution of the input image. 7 When fusing features from different resolutions, a common practice is to first uniformly adjust the size of these features to the same resolution, and then perform an addition operation. However, given that input features at different resolutions often contribute differently to the final output features, an additional weight parameter is proposed for each input feature. This strategy allows the network to learn and evaluate the importance of each input feature through the training process. This invention proposes Fast normalized fusion, the expression of which is as follows:

[0064]

[0065] In the formula, in each ω i Adding the ReLU activation function afterwards ensures that ω i ≥0, ε≥0.0001 are used to avoid numerical instability. Finally, based on bidirectional cross-scale connectivity and fast normalization fusion, the final BiFPN is obtained. For example, Figure 7 The feature fusion expression for layer 6 is as follows:

[0066]

[0067] in, This is an intermediate feature of the 6th layer of the top-down path. This represents the output features of layer 6 in the bottom-up path. Feature fusion is performed using Conv depthwise separable convolutions, with batch normalization and activation functions added after each convolution.

[0068] BiFPN's feature fusion structure is compared with FPN, PANet, and NAS-FPN, for example... Figure 8As shown, FPN only includes a top-down path for feature fusion. In contrast, PANet adds a bottom-up path. NAS-FPN, on the other hand, uses a neural architecture search technique to automatically discover irregular feature network topologies and repeatedly apply the same network blocks to build its architecture. Compared to PANet, BiFPN is more lightweight, with faster inference speed and fewer parameters. While NAS-FPN may have better performance, its complex search process typically requires more computational resources. Furthermore, BiFPN employs an adaptive feature adjustment mechanism that learns weights to adjust features at different levels to better match the needs of different tasks.

[0069] In summary, BiFPN's modular design makes it easy to embed into various deep neural network architectures. Using BiFPN in the neck network further optimizes the multi-scale feature fusion process and makes full use of feature information at different levels to improve the model's detection performance on targets of various sizes.

[0070] To verify the effectiveness of the improved algorithm and the rationality of ship identification, experiments and verifications were conducted based on a self-built dataset specifically designed for maritime ship identification.

[0071] This self-built dataset collects image data of ships at sea from multiple sources, including drone footage, surveillance camera captures, and professional photography equipment. It includes three categories: ships, civilian ships, and warships, totaling 6548 images. The training set contains 5537 images, the validation set 203 images, and the test set 808 images. Information about the training set is as follows... Figure 9 As shown.

[0072] This self-built dataset also employs a data augmentation module, specifically designed for experiments improving the YOLOv5 algorithm. The aim is to simulate images under different lighting and noise conditions through a series of image transformation techniques, including image scaling, flipping, center cropping, random changes to brightness, contrast, and saturation, and the addition of Gaussian noise, salt noise, and pepper noise. This data augmentation method effectively expands the original dataset, increasing the diversity and quantity of training samples, thereby mitigating overfitting and improving the model's robustness and generalization ability.

[0073] This self-built dataset was carefully designed and collected specifically for maritime vessel identification, thus better reflecting the challenges and complexities of real-world application scenarios. The dataset helps to more accurately evaluate and improve algorithm performance, ensuring that the algorithms perform well in practical applications.

[0074] The hardware environment for the experiment was as follows: Windows 10 (64-bit) operating system, Intel(R) Xeon(R) Gold 6248 CPU @ 2.50Hz processor, 256GB of RAM, and four NVIDIA GeForce RTX 3090 GPUs, each with 24GB of VRAM. The software environment consisted of Python 3.8, PyTorchl.10.1, and CUDA 11.1. All four GPUs were used simultaneously during training, along with the SGD optimizer.

[0075] Evaluation metrics used in the YOLOv5 experiments are key elements for measuring model performance, comprehensively reflecting the model's accuracy, robustness, and generalization ability in object detection tasks. Precision (P), recall (R), average precision (AP), and mean average precision (mAP) were used as evaluation metrics to comprehensively and deeply evaluate the performance of the CBS-YOLO algorithm in object detection. The following is a detailed analysis of the evaluation metrics used in the YOLOv5 experiments:

[0076] Precision (P) is the probability that a classifier correctly identifies a positive class as a positive class; it is used to evaluate the accuracy of predictions.

[0077] Recall (R) is the probability that a classifier can find all true positives, i.e., the evaluated recall rate.

[0078] Average precision (AP) is the accuracy and recall of a model in detecting targets at different confidence thresholds. It is obtained by calculating the area under the PR curve.

[0079] Mean precision (mAP) is the average AP across different intersection-over-union (IoU) thresholds. The IoU threshold represents the degree of overlap between the detected bounding box and the ground truth bounding box, where mAP... 50 This represents the average detection accuracy for all target categories when the IoU threshold is 0.50; mAP 50：95 This refers to calculating the average precision (mAP) for all classes within an IoU threshold range of 0.5 to 0.95 (with values ​​in increments of 0.05, for a total of 10 thresholds). By calculating the mAP across multiple IoU thresholds and averaging the results, the model's performance can be evaluated more comprehensively.

[0080] The formulas for calculating P, R, AP, and MAP are as follows:

[0081]

[0082] In this context, TP represents True Positive, TN represents True Negative, FP represents False Positive, and FN represents False Negative. P(R) refers to the precision value at a specific recall rate R, and n represents the total number of classes. APk refers to the average precision corresponding to the k-th class.

[0083] The hyperparameters involved in the experiment are shown in Table 1.

[0084] Table 1 Hyperparameter settings for model training

[0085] Tab.1 Model training hyperparameter settings

[0086]

[0087] To further explore the potential advantages of the improved algorithm in specific scenarios or tasks and to quantitatively evaluate its performance improvement, comparative experiments were conducted with classic object detection methods. These classic methods include Faster R-CNN, SSD, and other algorithms in the YOLO family, while the evaluation metrics mainly focused on mAP (mean accuracy) and FPS (frames per second). The results are shown in Table 2.

[0088] Table 2 Training results of different object detection methods on a self-built ship dataset.

[0089]

[0090] As shown in Table 2, compared to the classic Faster R-CNN and SSD algorithms, as well as other algorithms in the YOLO series, the CBS-YOLO algorithm achieves higher mAP. 50 With mAP 50：95 Improvements have been achieved in all aspects. Although the feature enhancement module introduced into the model increases the computational load, it currently basically meets the needs of UAV aerial photography detection. In summary, the improved algorithm proposed in this patent demonstrates significant advantages in the field of target detection, achieving remarkable improvements in detection accuracy and real-time performance, and has broad application prospects in multiple potential application scenarios such as maritime UAV reconnaissance.

[0091] Images with characteristics such as occlusion, small targets, and low light were selected from a self-built maritime vessel test set for testing. Table 3 shows the mAP of the YOLOv5s and CBS-YOLO algorithms on different categories. 50 With mAP 50：90 Comparison results.

[0092] Table 3 Comparison of the effects of the improved model and the original model in each category.

[0093]

[0094] Analysis of Table 3 reveals that the CBS-YOLO algorithm surpasses the basic YOLOv5 algorithm in both accuracy and precision for target detection across all categories, with particularly significant improvements. Comparing the number of detected targets and all targets in the image scene, and further comparing the confidence values ​​of the detection boxes, it is evident that the CBS-YOLO algorithm excels in recognizing occluded, dimly lit, blurred, small, and small-area targets, especially in terms of false positive and false negative rates. This significantly improves the accuracy of ship target detection and also demonstrates superior generalization ability.

[0095] It should be understood that in the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this description, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate different embodiments or examples described in this specification, as well as some features of different embodiments or examples.

[0096] Of course, those skilled in the art can make various corresponding changes and modifications based on the present invention without departing from its spirit and essence, but such changes and modifications should all fall within the protection scope of the claims of the present invention.

Claims

1. A YOLO-based small target detection method for cooperative identification of maritime unmanned aerial vehicles, characterized in that, The method comprises: acquiring a sea target area image; inputting the sea target area image into a trained machine learning model to obtain a small target detection result; wherein the trained machine learning model comprises a basic detection layer, a small target detection layer, and a bidirectional feature pyramid network; the basic detection layer is used for detecting low-resolution features, the small target detection layer is used for detecting high-resolution features, and the bidirectional feature pyramid network is used for multi-scale feature fusion of the low-resolution features and the high-resolution features.

2. The YOLO-based small target detection method for cooperative identification of offshore UAVs according to claim 1, characterized in that, The small target detection layer is located in an up-sampling path of a neck network of the trained machine learning model and is connected with a high-resolution feature map generation layer to detect small targets in the input image. 3.The YOLO-based small target detection method for cooperative identification of maritime UAVs according to claim 1, characterized in that, The bidirectional feature pyramid network learns and evaluates the importance of each input feature during the training process to add an additional weight parameter to each input feature in the feature fusion process, thereby realizing multi-scale feature fusion.

4. The YOLO-based small target detection method for cooperative identification of offshore UAVs according to claim 3, characterized in that, The feature fusion is realized by the following expression: wherein O denotes the output feature after feature fusion; ω i denotes the weight of the i-th input feature; o represents a very small positive number.

5. The YOLO-based small target detection method for cooperative identification of offshore UAVs according to claim 1, characterized in that, The trained machine learning model further comprises a channel attention module and a spatial attention module; The channel attention module is used for global maximum pooling down-sampling and global average pooling down-sampling of features of each channel, normalizing the features of each channel in a fully connected layer to obtain a first feature map; The spatial attention module is used for global average pooling and global maximum pooling of each channel of the first feature map, and the two pooling results are spliced, convolved, and normalized to obtain a second feature map.

6. The YOLO-based small target detection method for cooperative identification of offshore UAVs according to claim 5, characterized in that, The attention process of the channel attention module is realized by the following expression: wherein, F is the input feature map; M c (F) is the channel attention weight map obtained after the channel attention module processes the feature map F; F' is the feature map after channel attention weighting; The attention process of the spatial attention module is realized by the following expression: Wherein, F' is the feature map after channel attention weighting; M s (F') is the spatial attention weight map obtained after the spatial attention module processes the feature map F'; F" is the final feature map after spatial attention weighting.

7. The YOLO-based small target detection method for cooperative identification of offshore UAVs according to any one of claims 1-6, characterized in that, The training process of the trained machine learning model comprises: collecting image data of sea vessels from multiple sources; performing enhancement processing on the image data to form a data set.

8. A YOLO-based offshore unmanned aerial vehicle cooperative small target detection device, characterized in that, The method comprises: an acquisition module configured to acquire a sea target area image; an identification module configured to input the sea target area image into a trained machine learning model to obtain a small target detection result; wherein the trained machine learning model comprises a basic detection layer, a small target detection layer, and a bidirectional feature pyramid network; the basic detection layer is used for detecting low-resolution features, the small target detection layer is used for detecting high-resolution features, and the bidirectional feature pyramid network is used for multi-scale feature fusion of the low-resolution features and the high-resolution features.

9. A computer device comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program comprises instructions that, when executed by the processor, cause the processor to perform the method of any one of claims 1-8. The processor executes the computer program to implement the steps of the small target detection method based on YOLO for sea unmanned aerial vehicle cooperative identification according to any one of claims 1-7.

10. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the small target detection method based on YOLO for sea unmanned aerial vehicle cooperative identification according to any one of claims 1-7.

Images