An unmanned aerial vehicle-based water surface garbage identification method and system

Abstract

The application discloses a water surface garbage identification method and system based on a UAV, which comprises the following steps: collecting a UAV water surface garbage image containing target information, labeling the garbage, and constructing a UAV water surface garbage image dataset; constructing an LTM-YOLO11 detection model, adding a large-scale convolution kernel LSK attention mechanism between an SPPF module of a backbone layer and a C2PSA module of a neck layer; meanwhile, adding an interactive triple attention mechanism after each C3K2 module of the second column of the neck layer, and connecting the output of the three interactive triple attention mechanisms to a multi-band adaptive dilated convolution module MADC of the head layer; the output of the MADC is subjected to three detection heads to obtain a target detection result; and the detection model is trained by using the UAV water surface garbage image dataset to detect the UAV water surface garbage. The problems that the detection precision and network model lightness cannot be simultaneously satisfied in the detection of the UAV water surface garbage are solved.

Description

Technical Field

[0001] This invention belongs to the field of water surface debris identification technology, specifically relating to a method and system for water surface debris identification based on unmanned aerial vehicles (UAVs). Background Technology

[0002] With the development of deep learning, UAV target detection remains a major challenge in the field, as the scale of the detected targets is extremely imbalanced. Target sizes vary widely, with a high proportion of small targets. Compared to large targets, small targets have limited image information and are easily affected by various factors. Currently, there are two main categories of methods for UAV vision-based target detection: single-stage methods, such as YOLO and SSD, and two-stage methods, such as R-CNN, Fast-RCNN, and Faster-RCNN. Two-stage methods require generating candidate bounding boxes, followed by classification and regression. Single-stage methods, on the other hand, can perform classification and regression without generating candidate bounding boxes.

[0003] Yolo is a single-stage object detection algorithm with advantages such as simplicity, speed, and ease of deployment. Therefore, it is widely used in industrial object detection, object tracking, and object segmentation. The Yolo object detection algorithm is characterized by its fast training speed and high detection accuracy, making it suitable for UAVs to perform real-time small target detection in complex scenarios.

[0004] In drone scenarios, images contain a large proportion of small targets with varying scales. As a result, drone images suffer from false positives and false negatives due to the extremely small size of the targets, failing to meet the requirements of lightweight design while maintaining accuracy and speed.

[0005] Chinese patent CN119887850A discloses a multi-target tracking method in complex scenes based on adaptive association. It uses YOLO11 as the benchmark model and completes multi-target detection by introducing contextual feature extraction and hybrid attention. However, the hybrid attention mechanism includes multi-branch pooling and depooling operations, which have a large computational load and cannot be lightweighted, affecting the real-time detection results of the model. Summary of the Invention

[0006] In view of the aforementioned problems in the prior art, the technical problem to be solved by the present invention is to provide a method and system for identifying surface debris on water based on unmanned aerial vehicles (UAVs). This addresses the technical problem that the prior art cannot simultaneously meet the requirements of detection accuracy and lightweight network model. The present invention uses the YOLO11 model as the baseline network to construct an LTM-YOLO11 detection model, solving the problems of UAV surface debris detection being unable to simultaneously meet the requirements of detection accuracy and lightweight network model.

[0007] The technical solution adopted by the present invention to solve the aforementioned technical problem is as follows:

[0008] A method for identifying surface debris on water based on unmanned aerial vehicles (UAVs), the method comprising the following steps:

[0009] 1) Collect drone images of water surface debris containing target information, label the debris, and construct a drone image dataset of water surface debris;

[0010] 2) Construct an LTM-YOLO11 detection model;

[0011] The LTM-YOLO11 detection model is based on the YOLO11 baseline model. A large-scale convolutional kernel LSK attention mechanism is added between the SPPF module in the backbone layer and the C2PSA module in the neck layer. Simultaneously, an interactive triple attention mechanism is added after each C3K2 module in the second column of the neck layer. The outputs of these three interactive triple attention mechanisms are connected to the multi-band adaptive dilated convolutional module (MADC) in the head layer. The output of the MADC is then processed by three detection heads to obtain the target detection result.

[0012] The specific process of the Multi-Band Adaptive Dilated Convolutional Module (MADC) is as follows: The input features are decomposed into three frequency bands: high, medium, and low. Different dilation rates are learned by controlling the sampling interval of the convolution kernel. In the high-frequency band, a small dilation rate (1) convolution is performed between adjacent pixels; in the medium-frequency band, a medium dilation rate (2) convolution is performed; and in the low-frequency band, a large dilation rate (4) convolution is performed. The convolution outputs of the high, medium, and low frequency bands are weighted and fused to obtain the output of the MADC.

[0013] 3) Model training and testing;

[0014] The drone surface debris image dataset from step 1) is divided into a training set, a prediction set, and a validation set for training the LTM-YOLO11 detection model to obtain a trained LTM-YOLO11 detection model. The trained LTM-YOLO11 detection model is then used to detect drone surface debris and output the recognition results.

[0015] Furthermore, the large-scale convolution kernel LSK attention mechanism integrates large convolution kernels and separable convolution. In the LSK attention mechanism, the LSK attention mechanism obtains contextual feature information through the large kernel and decomposes the two-dimensional convolution into two one-dimensional convolution operations.

[0016] The interactive triple attention mechanism can effectively extract cross-dimensional features from feature maps through different rotation and arrangement operations without any information bottlenecks. It uses three different structures to obtain cross-dimensional interactive calculation of attention weights for feature data. The branches of the three different structures are respectively responsible for obtaining the interactive features between spatial dimension H or W and channel dimension C.

[0017] Furthermore, the specific process of the interactive triple attention mechanism is as follows:

[0018] The input, which interacts with each other across the three dimensions of channel, height, and width, is processed through three branches. In the first branch, the input undergoes Z-pooling, a k×k convolutional layer, a sigmoid function, and a rotation operation to obtain the output of the first branch. In the second branch, the input undergoes Z-pooling, a k×k convolutional layer, a sigmoid function, and a rotation operation to obtain the output of the second branch. In the third branch, the input undergoes Z-pooling, a k×k convolutional layer, and a sigmoid function to obtain the output of the third branch. The outputs of the three branches are then averaged and aggregated to obtain the output of the interactive triple attention mechanism.

[0019] Furthermore, the detection accuracy of the LTM-YOLO11 detection model is not less than 40% for mAP@0.5 and 24.9% for mAP@0.5-0.95.

[0020] The present invention also protects a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can implement the steps of the identification method.

[0021] Furthermore, the present invention also provides a surface debris identification system based on unmanned aerial vehicles (UAVs), wherein the system performs the identification method by including:

[0022] Unmanned aerial vehicle (UAV) image acquisition equipment is used to capture images of debris on the water surface;

[0023] The image preprocessing module is used to annotate and enhance the acquired images;

[0024] The LTM-YOLO11 detection model is used for target detection of surface debris with a wide range of target sizes and small target sizes.

[0025] The alarm module is used to issue early warnings based on the recognition results of the LTM-YOLO11 detection model. If there is garbage on the water surface, an early warning for cleaning will be issued. If there is no garbage, the drone will update its geographical location for inspection and re-detection.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] In UAV vision, the problems of false detection and false negative detection of small targets lead to low detection accuracy and the inability to simultaneously achieve lightweight design. This invention, based on the LTM-YOLO11 detection model, achieves multi-target detection on water surfaces in UAV scenarios. By introducing an LSK attention mechanism at the connection points of the backbone and neck layers, the original input is added to the features processed by deep convolution, which not only enhances stability but also preserves the original features and improves training robustness and efficiency. Simultaneously, this invention introduces an interactive triple attention mechanism after the C3K2 module in the neck layer. Through three different branches (upper, middle, and lower), it models channel and spatial / temporal dependencies in different dimensions, acquiring cross-dimensional features and enhancing the network's understanding of target features. Finally, this invention introduces a multi-band adaptive dilation convolution (MADC) module in the head layer, dividing the frequency band into high, medium, and low frequency bands and using different dilation rates to perform different convolution operations, preserving global feature details while achieving lightweight design. The synergistic effect of these three components enables the new improved detection model to achieve lightweight and high-precision identification of surface debris by drones. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating one embodiment of the UAV-based method for identifying surface debris in water according to the present invention.

[0029] Figure 2 This is a schematic diagram of the overall structure of the LTM-YOLO11 detection model of the present invention.

[0030] Figure 3 This is a schematic diagram of the LSK attention mechanism.

[0031] Figure 4 This is a schematic diagram of the interactive triple attention mechanism.

[0032] Figure 5 This is a schematic diagram of the structure of a multi-band adaptive dilated convolution module. Detailed Implementation

[0033] The present invention will be further explained below with reference to the embodiments and accompanying drawings, and a complete description of the present invention will be provided, but this is not intended to limit the scope of protection of this application.

[0034] This invention relates to a method for identifying surface debris on water using unmanned aerial vehicles (UAVs), comprising the following steps:

[0035] (1) Collect images of surface debris on the water surface by drones containing target information, then label the collected images with debris, generate label files of the corresponding locations of the debris in the images, preprocess the images, including normalization and data augmentation, and obtain a dataset of surface debris images on the water surface by drones.

[0036] (2) Construct an LTM-YOLO11 detection model;

[0037] The LTM-YOLO11 detection model is based on the YOLO11 baseline model. A large-scale convolutional kernel LSK attention mechanism is added between the SPPF module in the backbone layer and the C2PSA module in the neck layer. Simultaneously, an interactive triple attention mechanism is added after each C3K2 module in the second column of the neck layer. The outputs of these three interactive triple attention mechanisms are connected to the multi-band adaptive dilated convolutional module (MADC) in the head layer. The output of the MADC is then processed by three detection heads to obtain the target detection result.

[0038] The specific process of the Multi-band Adaptive Dilated Convolutional Module (MADC) is as follows: the input features are decomposed into three frequency bands: high, medium, and low. Different dilation rates are learned by controlling the sampling interval of the convolution kernel. In the high-frequency band region, a small dilation rate convolution with a dilation rate of 1 is used to perform convolution between adjacent pixels. In the medium-frequency band region, a medium dilation rate convolution with a dilation rate of 2 is used for processing. In the low-frequency band region, a large dilation rate convolution with a dilation rate of 4 is used for processing. The convolution outputs of the high, medium, and low frequency bands are weighted and fused to obtain the output of the MADC.

[0039] The large-scale convolutional kernel LSK attention mechanism integrates large convolutional kernels and separable convolutions, improving the receptive field and efficiency of separable convolutions while reducing computational complexity and memory usage. (See LSK attention mechanism...) Figure 3 In the LSK attention mechanism, contextual feature information is obtained through a large kernel, and the two-dimensional convolution is decomposed into two one-dimensional convolution operations.

[0040] The interactive triple attention mechanism (see TrippleAttention) (see...) Figure 4 This method effectively extracts cross-dimensional features from feature maps through different rotation and arrangement operations, without any information bottlenecks. Three different structures are used to obtain cross-dimensional interactive attention weights from the feature data. The branches of the three different structures are responsible for obtaining the interactive features between the spatial dimension H or W and the channel dimension C, respectively. These branches generate interactive cross-dimensional attention weights by arranging and transforming the input features, followed by Z-pooling and a k×k convolutional layer. The specific process is as follows:

[0041] The input, which interacts with each other across the three dimensions of channel, height, and width, is processed through three branches. In the first branch, the input undergoes Z-pooling, a k×k convolutional layer, a sigmoid function, and a rotation operation to obtain the output of the first branch. In the second branch, the input undergoes Z-pooling, a k×k convolutional layer, a sigmoid function, and a rotation operation to obtain the output of the second branch. In the third branch, the input undergoes Z-pooling, a k×k convolutional layer, and a sigmoid function to obtain the output of the third branch. The outputs of the three branches are then averaged and aggregated to obtain the output of the interactive triple attention mechanism.

[0042] The multi-band adaptive dilated convolution module divides the frequency band into high, medium and low frequency bands and uses different dilation rates to complete different convolution operations, which not only preserves global feature details but also achieves lightweight design.

[0043] (3) Model Training

[0044] The drone surface debris image dataset from step 1) is divided into a training set, a prediction set, and a validation set in a 7:1:2 ratio for training the LTM-YOLO11 detection model, resulting in a trained LTM-YOLO11 detection model. The trained LTM-YOLO11 detection model is then used to detect drone surface debris, and the recognition results are output.

[0045] Example 1

[0046] like Figure 1-4 As shown, this embodiment presents a method for identifying surface debris on water using a drone. It employs the LTM-YOLO11 detection model to improve accuracy while maintaining lightweight performance. Specifically, it includes the following steps:

[0047] Step 1: The drone updates its geolocation and acquires images. It captures images of water surface debris, obtaining accurate images with target information. Preprocessing operations such as normalization are performed on the images. The debris is labeled, generating label files representing the corresponding locations of the debris in the images. Images of debris on the water surface are captured by the drone, and imagelabels are used to label the debris in the images. The labeled rectangles represent the specific pixel locations of the debris in the images. This data constitutes the drone-generated water surface debris dataset. The images in the drone-generated water surface debris dataset are divided into training and testing sets in TXT format. The training, validation, and testing sets are configured in a 7:1:2 ratio.

[0048] Step 2: Construct the LTM-YOLO11 detection model

[0049] The LTM-YOLO11 detection model is based on the YOLO11 baseline model and includes a backbone layer, a neck layer, and a head layer. A large-scale convolutional kernel LSK attention mechanism is added between the SPPF module in the backbone layer and the C2PSA module in the neck layer.

[0050] Meanwhile, interactive triple attention mechanism TrippleAttention is added after the C3K2 module in the second column of the neck layer, for a total of three interactive triple attention mechanisms TrippleAttention. The outputs of the three are connected to the multi-band adaptive dilated convolution module MADC in the head layer.

[0051] Based on the YOLO11 baseline model, an LSK attention mechanism is introduced in the neck layer. Spatial selection is used to weight features processed by large-kernel-depth convolutional kernels, and a dynamic weight adjustment strategy is applied to automatically select the optimal convolutional kernel, improving feature extraction efficiency while reducing computational cost and enhancing the ability of UAVs to detect surface debris in complex scenarios. The structure diagram of the LSK attention mechanism is shown below. Figure 3 As shown,

[0052] The depthwise convolution (DW-Conv) is decomposed into a concatenation of horizontal one-dimensional convolutions and vertical one-dimensional convolutions. The input feature map is F∈R. C×H×W Horizontal one-dimensional convolution uses kernel Convolution along the width direction. Vertical one-dimensional convolution uses a kernel. Convolution along the height direction. Depth convolution decomposes into intermediate feature maps from the outputs of horizontal one-dimensional convolution and vertical one-dimensional convolution. for:

[0053]

[0054] Where H, W, and C are the height, width, and number of channels of the input feature map, respectively;

[0055] The dilated depthwise convolution (DW-D-Conv) is decomposed into horizontal one-dimensional convolutions and vertical one-dimensional convolutions, using the intermediate feature map. The input to the dilated depthwise convolution is decomposed. The horizontal one-dimensional convolution uses a kernel. Convolution along the width direction; vertical one-dimensional convolution uses a kernel. Convolution along the height direction. Dilated depthwise convolution outputs the global feature Z. C :

[0056]

[0057] In global feature Z C Based on this, perform 1×1 convolution to form weight A C Weights and input features F C Element-wise multiplication is performed to output the enhanced features.

[0058]

[0059] The LSK attention mechanism decomposes the large kernel depthwise convolution and the dilated depthwise convolution into two one-dimensional convolutions, which reduces the amount of computation while expanding the receptive field and improving the feature extraction capability of the detection model.

[0060] The structure of the interactive triple attention mechanism is as follows: Figure 4 As shown, attention weights are calculated by capturing the interaction relationships between the three dimensions of channel, height, and width through three different structural branches, and then the interdependencies between different dimensions are established by rotating the input tensor and performing residual transformation. The processing procedure of the interactive triple attention mechanism is as follows:

[0061] In the first branch, the channel interacts with the height, and the input tensor is x∈R. C×H×W It is obtained by rotating counterclockwise along the height direction to get x1, and then using Z-pooling operation to compress x1 into two dimensions along the channel dimension to obtain a tensor.

[0062]

[0063] Where MaxPool represents the maximum pooling operation, AvgPool represents the average pooling operation, and Z-pool represents the Z-pooling operation.

[0064] The channel and height weights ω1 are calculated using k×k convolutional layers and batch normalization.

[0065]

[0066] Where σ represents the Sigmoid function;

[0067] Next, the weight ω1 is multiplied by x1, and then rotated 90 degrees clockwise to restore the original shape, thus obtaining the output of the first branch.

[0068] In the second branch, the channels and width interact. The input tensor x is rotated counterclockwise along the width direction to obtain tensor x2. Tensor x2 is then compressed into channel dimensions using a Z-pooling operation.

[0069]

[0070] The channel and width weights are calculated using k×k convolutional layers and batch normalization.

[0071]

[0072] Next, the weight ω2 is multiplied by x2, and then rotated 90 degrees clockwise to restore the original shape, resulting in the output of the second branch.

[0073] In the third branch, height and width interact. The height and width of the input tensor x are selected to form tensor x3, and then the input for spatial attention is obtained using the following formula. With weight ω3:

[0074]

[0075] Next, the weight ω3 is multiplied by x3 to obtain the output of the third branch.

[0076] The total output obtained by averaging the outputs of the different branches obtained through the above steps is:

[0077]

[0078] Without adding other parameters to the network, the overall receptive field of the network is increased by controlling the spacing between convolutional kernels based on local frequency information. The specific steps of frequency-adaptive dilated convolution are as follows:

[0079] Based on the relationship between the local feature X at position p and the convolution weights, the dilation rate of each pixel is adaptively adjusted to achieve global optimum. The feature map output value at pixel position p is:

[0080]

[0081] Next, the convolution weights are divided into high-frequency weights. and low-frequency weights The two branches dynamically adjust their proportions: in the high-frequency region, small expansion protects detailed features; in the low-frequency region, large expansion increases the receptive field. l , λ h These are dynamic weight parameters. The adjusted weights are:

[0082]

[0083] The high-frequency transformation in the time domain is converted to the frequency domain and then weighted.

[0084]

[0085] Where A b ∈R H×W Weighted graph for each frequency band; X b (i,j) represents the decomposed frequency bands. This method suppresses background noise in high-frequency regions, improving the model's ability to perceive targets.

[0086] This invention proposes a multi-band adaptive dilation convolution (MADC) module based on frequency-adaptive dilation convolution (FAD) (see [link]). Figure 5 The specific steps are as follows:

[0087] The input features are decomposed into three frequency bands: high, medium, and low. Different dilation rates are learned by controlling the sampling interval of the convolutional kernels. In the high-frequency band, a small dilation rate (d1=1) is used for convolution between adjacent pixels; in the mid-frequency band, a medium dilation rate (d2=2) is used, equivalent to a 3×3 kernel covering a 5×5 region; and in the low-frequency band, a large dilation rate (d3=4) is used, equivalent to a 3×3 kernel covering a 9×9 region. The outputs of the high, medium, and low-frequency convolutional kernels are weighted and fused to output the final features for detection. A small dilation rate in the high-frequency band extracts detailed features, a medium dilation rate in the mid-frequency band expands the receptive field, and a high dilation rate in the low-frequency band extracts global features. The output of the MADC then passes through different detection heads to output the target object.

[0088] All models in this invention are trained for 200 epochs. By training the models on the dataset, the models detect whether there is trash on the water surface. If trash is found, a cleanup warning is issued, and a PDF report is generated. If no trash is found, the drone updates its geographical location for inspection and re-detects the trash.

[0089] Step 3: After training the detection model, the performance of each detection model is evaluated and compared using seven evaluation metrics, including the number of layers, the number of parameters, and accuracy, to verify whether the improved model of this invention is the optimal model. The evaluation metrics for model validation include Layers, Parameters, GFLOPS, mAP@0.5, mAP@0.5-0.95, R, and P. Layers represents the total number of layers in the network, and Parameters represents the total number of parameters in the model. GFLOPS is gigabit floating-point operations per second, measuring the device's mathematical operation speed. mAP@0.5 is the average precision when IOU = 0.5, mAP@0.5-0.95 is the average precision at multiple thresholds between 0.5 and 0.95 for IOU, R is recall, P is precision, and mean average precision (mAP) is the average precision.

[0090]

[0091] Where P (precision) represents the percentage of predicted positive samples that are actually positive samples. R (recall) is the proportion of positive samples that are predicted as positive samples out of the total number of positive samples. AP (mean) represents the goodness or badness for each class. Mean precision is the average of the mean precision. TP (True Positive) represents the correctly predicted true label, FN (False Negative) represents the unpredicted true label, FP (False Positive) represents the incorrectly predicted true label, and TN represents the correctly predicted true negative label. The validation experiments conducted on the dataset used in this invention yielded the following results:

[0092] Table 1

[0093] Model version Layers Parameters GFLOPS P R mAP@0.5 mAP@0.5-0.95 YOLO11 319 2591400 6.4 18.9 44 33.9 24.2 YOLO11-ABC - - - - - 34.7 - LTM-YOLO11 408 2656502 6.6 29.8 45.8 41.9 24.9

[0094] As shown in Table 1, from a high-precision perspective, the LTM-YOLO11 detection module of this invention is optimal in every performance metric. Although it has slightly more layers than the benchmark model, the difference in the number of parameters is not significant. YOLO11-ABC, the detection model mentioned in patent CN119887850A, achieves an accuracy of 34.7% at mAP@0.5. The detection accuracy of the LTM-YOLO11 detection model of this invention is 41.9% at mAP@0.5, which is 7.2% higher than similar detection models. Therefore, the LTM-YOLO11 detection model of this invention satisfies the requirements of lightweight design while also achieving higher accuracy than other multi-target recognition algorithms for water surface debris. It is the optimal model among similar multi-target recognition models for water surface debris, and can meet the needs of UAV water surface debris recognition, verifying the effectiveness of this method.

[0095] Furthermore, this application also conducted ablation experiments, as shown in Table 2. These experiments demonstrate that the detection model of this invention possesses the capability for multi-target identification of surface debris by unmanned aerial vehicles (UAVs). In practical applications, it meets the requirements of lightweight design and high accuracy. The synergistic effect of the three modules significantly improves its performance, increasing the detection accuracy to over 41%.

[0096] Table 2

[0097]

[0098] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for identifying surface debris on water based on unmanned aerial vehicles (UAVs), characterized in that, The identification method includes the following steps: 1) Collect drone images of water surface debris containing target information, label the debris, and construct a drone image dataset of water surface debris; 2) Construct an LTM-YOLO11 detection model; The LTM-YOLO11 detection model is based on the YOLO11 baseline model. A large-scale convolutional kernel LSK attention mechanism is added between the SPPF module in the backbone layer and the C2PSA module in the neck layer. Simultaneously, an interactive triple attention mechanism is added after each C3K2 module in the second column of the neck layer. The outputs of these three interactive triple attention mechanisms are connected to the multi-band adaptive dilated convolutional module (MADC) in the head layer. The output of the MADC is then processed by three detection heads to obtain the target detection result. The specific process of the Multi-Band Adaptive Dilated Convolutional Module (MADC) is as follows: The input features are decomposed into three frequency bands: high, medium, and low. Different dilation rates are learned by controlling the sampling interval of the convolution kernel. In the high-frequency band, a small dilation rate (1) convolution is performed between adjacent pixels; in the medium-frequency band, a medium dilation rate (2) convolution is performed; and in the low-frequency band, a large dilation rate (4) convolution is performed. The convolution outputs of the high, medium, and low frequency bands are weighted and fused to obtain the output of the MADC. The large-scale convolution kernel LSK attention mechanism integrates large convolution kernels and separable convolution. In the LSK attention mechanism, the large kernel is used to obtain contextual feature information and decompose the two-dimensional convolution into two one-dimensional convolution operations. The interactive triple attention mechanism can effectively extract cross-dimensional features from the feature map through different rotation and arrangement operations without any information bottlenecks. It uses three different structures to obtain cross-dimensional interactive calculation of attention weights for feature data. The branches of the three different structures are respectively responsible for obtaining the interactive features between spatial dimension H or W and channel dimension C. The specific process of the interactive triple attention mechanism is as follows: The input, which interacts with each other across the three dimensions of channel, height, and width, is processed through three branches. In the first branch, each dimension undergoes Z-pooling and a pool of size... The first branch outputs are obtained through a convolutional layer, a sigmoid function, and a rotation operation; in the second branch, the outputs undergo Z-pooling and a rotation operation of size 1. The second branch outputs are obtained through a convolutional layer, a sigmoid function, and a rotation operation; in the third branch, the outputs undergo Z-pooling and a rotation operation of size 1. The convolutional layer and sigmoid function obtain the output of the third branch; the outputs of the three branches are then averaged and aggregated to obtain the output of the interactive triple attention mechanism. 3) Model training and testing; The drone surface debris image dataset from step 1) is divided into a training set, a prediction set, and a validation set for training the LTM-YOLO11 detection model to obtain a trained LTM-YOLO11 detection model. The trained LTM-YOLO11 detection model is then used to detect drone surface debris and output the recognition results.

2. The identification method according to claim 1, characterized in that, The detection accuracy of the LTM-YOLO11 detection model is no less than 40% for mAP@0.5 and 24.9% for mAP@0.5-0.

95.

3. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program can implement the steps of the identification method described in claim 1 or 2.

4. A drone-based water surface debris identification system, characterized in that, The system performs the steps of the identification method according to claim 1 or 2, including: Unmanned aerial vehicle (UAV) image acquisition equipment is used to capture images of debris on the water surface; The image preprocessing module is used to annotate and enhance the acquired images; The LTM-YOLO11 detection model is used for target detection of surface debris with a wide range of target sizes and small target sizes. The alarm module is used to issue early warnings based on the recognition results of the LTM-YOLO11 detection model. If there is garbage on the water surface, an early warning for cleaning will be issued. If there is no garbage, the drone will update its geographical location for inspection and re-detection.

Images