Water surface garbage identification method and system based on unmanned aerial vehicle

Abstract

The invention relates to a water surface garbage identification method and system based on an unmanned aerial vehicle, and the method comprises the following steps: collecting an unmanned aerial vehicle water surface garbage image containing target information, marking the garbage, and constructing an unmanned aerial vehicle water surface garbage image data set; constructing an LTM-YOLO11 detection model, and adding a large-scale convolution kernel LSK attention mechanism between the SPPF module of the trunk layer and the C2PSA module of the neck layer; meanwhile, an interactive triple attention mechanism is added behind each C3K2 module in the second column of the neck layer, and the outputs of the three interactive triple attention mechanisms are connected with a multi-band adaptive expansion convolution module MADC of the head layer; the output of the MADC passes through three detection heads to obtain a target detection result; and training a detection model by using the unmanned aerial vehicle water surface garbage image data set, and detecting unmanned aerial vehicle water surface garbage. The problem that unmanned aerial vehicle water surface garbage detection cannot meet the detection precision and the light weight of a network model at the same time is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of water surface garbage identification, and particularly relates to a water surface garbage identification method and system based on an unmanned aerial vehicle (UAV) system. Background Art

[0002] With the development of deep learning, drone target detection has become a major challenge in the field of target detection. The scales of detected targets are extremely imbalanced. Target sizes range 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 drone vision target detection methods. One is single-stage target detection methods, such as Yolo and SSD. The other is two-stage target detection methods, such as RCNN, Fast-RCNN, and Faster-RCNN. Two-stage target detection methods require the generation of candidate boxes, followed by classification and regression. In contrast, single-stage target detection methods can perform classification and regression without generating candidate 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 target detection, target tracking, and target segmentation. The Yolo target detection algorithm features fast training speed and high detection accuracy, making it suitable for drones to detect small targets in real time in complex scenarios.

[0004] In drone scenarios, images have varying scales and a large proportion of small targets. Therefore, there are problems of false detection and missed detection in drone images due to the extremely small targets, and the lightweight requirements cannot be met while maintaining accuracy and speed.

[0005] Chinese patent publication number CN119887850A discloses a multi-target tracking method in complex scenarios 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 de-pooling operations, which are computationally intensive and cannot be lightweight, affecting the model's real-time detection results. Summary of the Invention

[0006] In view of the above-mentioned problems in the prior art, the present invention aims to provide a method and system for identifying surface debris using drones. This method addresses the existing technical problem of being unable to simultaneously meet the requirements for detection accuracy and network model lightweighting. Using the YOLO11 model as a baseline network, the present invention constructs an LTM-YOLO11 detection model, addressing the problem of drone-based surface debris detection failing to simultaneously meet the requirements for detection accuracy and network model lightweighting.

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

[0008] A method for identifying water surface garbage based on an unmanned aerial vehicle, the method comprising the following steps:

[0009] 1) Collect UAV surface garbage images containing target information, annotate the garbage, and build a UAV surface garbage image dataset;

[0010] 2) Build the LTM-YOLO11 detection model;

[0011] The LTM-YOLO11 detection model is based on the YOLO11 benchmark model. It adds a large-scale convolution kernel LSK attention mechanism between the SPPF module in the backbone layer and the C2PSA module in the neck layer. At the same time, an interactive triple attention mechanism is added after each C3K2 module in the second column of the neck layer. The outputs of the three interactive triple attention mechanisms are connected to the multi-band adaptive dilated convolution module MADC in the head layer. The output of the MADC is passed through three detection heads to obtain the target detection result.

[0012] The specific process of the multi-band adaptive dilated convolution module MADC is as follows: decomposing the input features into three frequency bands: high, medium, and low; learning different dilation rates by controlling the convolution kernel sampling interval; using a small dilation rate convolution with a dilation rate of 1 to convolve between adjacent pixels in the high-frequency band area; using a medium dilation rate convolution with a dilation rate of 2 to process in the medium-frequency band area; and using a large dilation rate convolution with a dilation rate of 4 to process in the low-frequency band area; and obtaining the MADC output by weighted fusion of the convolution outputs of the high, medium, and low frequency bands.

[0013] 3) Model training and testing;

[0014] The drone surface debris image dataset in 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 used to detect drone surface debris and output the recognition results.

[0015] Furthermore, the large-scale convolution kernel LSK attention mechanism combines large convolution kernels with separable convolution. In the LSK attention mechanism, the LSK attention mechanism obtains context 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 in feature maps through different rotation and permutation operations without any information bottleneck, and adopts three different structures to obtain the cross-dimensional interactive calculation attention weights of feature data; the branches of the three different structures are respectively responsible for obtaining the interactive features between the spatial dimension H or W and the channel dimension C.

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

[0018] The interactive inputs between the three dimensions of channel, height, and width are processed by three branches respectively. In the first branch, the output of the first branch is obtained by Z pooling operation, a convolutional layer of size k×k, Sigmoid function, and rotation operation; in the second branch, the output of the second branch is obtained by Z pooling operation, a convolutional layer of size k×k, Sigmoid function, and rotation operation; in the third branch, the output of the third branch is obtained by Z pooling operation, a convolutional layer of size k×k, and Sigmoid function; 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 mAP@0.5 of the LTM-YOLO11 detection model is not less than 40%, and the mAP@0.5-0.95 is 24.9%.

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

[0021] In addition, the present invention also provides a water surface garbage identification system based on a drone, wherein the system performs the steps of the identification method, including:

[0022] UAV image acquisition equipment, used to collect images of water surface garbage;

[0023] Image preprocessing module, used to annotate and enhance the collected images;

[0024] The LTM-YOLO11 detection model is used to detect small and wide-sized debris on the surface of water.

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

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] In drone vision, the problems of false detection and missed detection of small targets lead to low detection accuracy and the inability to meet the lightweight requirements at the same time. The present invention realizes the detection of multiple targets on the water surface in drone scenarios based on the LTM-YOLO11 detection model. By introducing the LSK attention mechanism at the connection position of the backbone layer and the neck layer, the original input is added to the features processed by deep convolution, which not only enhances stability, but also retains the original features and improves the robustness and efficiency of training. At the same time, the present invention introduces an interactive triple attention mechanism after the C3K2 module of the neck layer. Through the three different branches of upper, middle and lower, the channel and spatial and temporal dependencies are modeled in different dimensions, cross-dimensional features are obtained, and the network's ability to understand target features is enhanced. Finally, the present invention introduces a multi-band adaptive dilation convolution module (Multi-band Adaptivel Dilation Convolution, MADC) in the head layer, divides the frequency band into high, medium and low frequency band parts, and uses different dilation rates to complete different convolution operations, which not only retains the global feature detail information but also achieves lightweight. The synergistic effect of the three parts enables the new improved detection model to achieve lightweight and high-precision drone detection of surface debris. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The figure is a flow chart of an embodiment of a method for identifying water surface garbage based on a drone of the present invention.

[0029] Figure 2 Schematic diagram of the overall structure of the LTM-YOLO11 detection model of the present invention.

[0030] Figure 3 Schematic diagram of the LSK attention mechanism.

[0031] Figure 4 Schematic diagram of the structure of the interactive triple attention mechanism.

[0032] Figure 5 Schematic diagram of the structure of the multi-band adaptive dilated convolution module. DETAILED DESCRIPTION

[0033] The present invention will be further explained below with reference to the embodiments and drawings, and the present invention will be fully described. However, these are not intended to limit the scope of protection of the present application.

[0034] The present invention provides a method for identifying water surface garbage based on an unmanned aerial vehicle, comprising the following steps:

[0035] (1) Collect UAV surface garbage images containing target information, then mark the garbage in the collected images, generate a label file of the corresponding position of the garbage in the image, and preprocess the image. The preprocessing process includes normalization and data enhancement to obtain a UAV surface garbage image dataset.

[0036] (2) Build the LTM-YOLO11 detection model;

[0037] The LTM-YOLO11 detection model is based on the YOLO11 benchmark model. It adds a large-scale convolution kernel LSK attention mechanism between the SPPF module in the backbone layer and the C2PSA module in the neck layer. At the same time, an interactive triple attention mechanism is added after each C3K2 module in the second column of the neck layer. The outputs of the three interactive triple attention mechanisms are connected to the multi-band adaptive dilated convolution module MADC in the head layer. The output of the MADC is passed through three detection heads to obtain the target detection result.

[0038] The specific process of the multi-band adaptive dilated convolution module MADC is as follows: decomposing the input features into three frequency bands: high, medium and low, and learning different dilation rates by controlling the convolution kernel sampling interval; using a small dilation rate convolution with a dilation rate of 1 to convolve between adjacent pixels in the high-frequency band area, using a medium dilation rate convolution with a dilation rate of 2 to process in the medium-frequency band area, and using a large dilation rate convolution with a dilation rate of 4 to process in the low-frequency band area; the convolution outputs of the high, medium and low frequency bands are weightedly fused to obtain the output of the MADC.

[0039] The large-scale convolution kernel LSK attention mechanism combines large convolution kernels with separable convolution, improves the efficiency of receptive field and separable convolution, and reduces computational complexity and memory usage. Figure 3 ), the LSK attention mechanism obtains contextual feature information through a large kernel and decomposes the two-dimensional convolution into two one-dimensional convolution operations.

[0040] The interactive triple attention mechanism (TrippleAttention) (see Figure 4 ), without any information bottleneck, cross-dimensional features in the feature map can be effectively extracted through different rotation and permutation operations. Three different structures are used to obtain the cross-dimensional interaction of feature data and calculate the attention weight. The three branches with different structures are responsible for obtaining the interactive features between the spatial dimension H or W and the channel dimension C. The three branches with different structures permutate the input features, then generate interactive cross-dimensional attention weights through Z pooling and a k×k convolutional layer. The specific process is:

[0041] The interactive inputs between the three dimensions of channel, height, and width are processed by three branches respectively. In the first branch, the output of the first branch is obtained by Z pooling operation, a convolutional layer of size k×k, Sigmoid function, and rotation operation; in the second branch, the output of the second branch is obtained by Z pooling operation, a convolutional layer of size k×k, Sigmoid function, and rotation operation; in the third branch, the output of the third branch is obtained by Z pooling operation, a convolutional layer of size k×k, and Sigmoid function; 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 perform different convolution operations, thereby retaining global feature detail information while also achieving lightweightness.

[0043] (3) Model training

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

[0045] Example 1

[0046] like Figure 1-4 As shown in FIG, this embodiment is a method for identifying water surface debris based on a drone, which is used to identify debris on the water surface of a drone. The method adopts the LTM-YOLO11 detection model to improve the accuracy while meeting the requirements of lightweight performance. Specifically, the following steps are included:

[0047] Step 1: The drone updates its geographic location and collects images. The drone captures images of water surface debris, obtains accurate target information, and performs preprocessing operations such as normalization. The debris is labeled to generate a label file that indicates the corresponding locations of the debris in the image. The drone collects images of debris on the water surface and then uses imagelabel to annotate the debris within the image. The annotated rectangles represent the specific pixel locations of the debris within the image. This information constitutes the drone water surface debris dataset. The images in the drone water surface debris dataset are divided into training and test sets in txt format. The training, validation, and test sets are arranged in a 7:1:2 ratio.

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

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

[0050] At the same time, an 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 of the head layer.

[0051] Based on the YOLO11 benchmark model, the neck layer introduces the LSK attention mechanism, which uses spatial selection to weight the features processed by the large-core deep convolution kernel and applies a dynamic weight adjustment strategy to automatically select the optimal convolution kernel, thereby improving feature extraction efficiency and reducing the amount of computation, thereby improving the ability of drones to detect surface debris in complex scenarios. The structural diagram of the LSK attention mechanism is shown in the figure below. Figure 3 As shown,

[0052] Decompose the depth convolution DW-Conv into a cascade of horizontal one-dimensional convolution and vertical one-dimensional convolution. 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 kernel Convolution along the height direction. The depth convolution is decomposed into the intermediate feature map of the output of the horizontal one-dimensional convolution and the vertical one-dimensional convolution. for:

[0053]

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

[0055] Decompose the dilated depth convolution DW-D-Conv into horizontal one-dimensional convolution and vertical one-dimensional convolution, and use the intermediate feature map Decomposed as the input of the dilated depth convolution. The horizontal one-dimensional convolution uses the kernel Convolution along the width direction; vertical one-dimensional convolution uses kernel Convolution along the height direction. Dilated depth convolution outputs global feature Z C :

[0056]

[0057] In the global feature Z C Based on this, a 1×1 convolution is performed to form the weight A C , weights and input features F C Perform element-by-element multiplication to output enhanced features

[0058]

[0059] The LSK attention mechanism decomposes the large-kernel depth convolution and the dilated depth 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 in the figure, the interactive relationship between the three dimensions of channel, height, and width is captured through three different structural branches to calculate the attention weights, and the interdependence between different dimensions is established by rotating the input tensor and the residual transformation. The processing process of the interactive triple attention mechanism is as follows:

[0061] In the first branch, the channels interact with the height and the input tensor is x∈R C×H×W , which is rotated counterclockwise along the height direction to obtain x1, and x1 is then compressed to two dimensions along the channel dimension using the Z pooling operation to obtain a tensor

[0062]

[0063] Among them, MaxPool represents the maximum pooling operation, AvgPool represents the average pooling operation; Z-pool represents the Z pooling operation;

[0064] The channel and height weights ω1 are calculated through the k×k convolution layer and batch normalization:

[0065]

[0066] Where σ represents the Sigmoid function;

[0067] Then the weight ω1 is multiplied by x1 and rotated 90 degrees clockwise to restore the original shape to obtain the output of the first branch

[0068] In the second branch, the channel and width interact, and the input tensor x is rotated counterclockwise along the width direction to obtain the tensor x2, which is compressed by the channel dimension through the Z pooling operation.

[0069]

[0070] The channel and width weights are calculated through the k×k convolution layer and batch normalization:

[0071]

[0072] Then the weight ω2 is multiplied by x2 and rotated 90 degrees clockwise to restore the original shape to obtain the output of the second branch

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

[0074]

[0075] Then the weight ω3 is multiplied by x3 to get the output of the third branch

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

[0077]

[0078] Without adding other parameters to the network, the interval size between convolution kernels is controlled according to the local frequency information to increase the overall receptive field of the network. The specific steps of frequency adaptive dilated convolution are as follows:

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

[0080]

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

[0082]

[0083] Transform the time domain high frequency into frequency domain decomposition and then weight it.

[0084]

[0085] Among them A b ∈R H×W is the weight map of each frequency band; X b (i, j) is the decomposed frequency band. This method suppresses the high-frequency background and improves the model's ability to perceive the target.

[0086] The present invention optimizes and proposes a multi-band adaptive dilation convolution module (Multi-band Adaptive Dilation Convolution, MADC) based on frequency adaptive dilation convolution (see Figure 5 ). The specific steps are:

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

[0088] The training times of all models in this invention are 200 epochs. Through the training model of the data set, the model detects whether there is garbage on the water surface. If there is garbage, it will issue a cleaning warning and generate a PDF report. If there is no garbage, the drone will update the geographical location patrol and re-detect.

[0089] Step 3: After the detection model is trained, seven evaluation indicators such as the number of layers, the number of parameters, and the accuracy are used to evaluate and compare the performance of each detection model to verify whether the improved model of the present invention is the optimal model. The evaluation indicators for model verification include Layers, Parameters, GFLOPS, mAP@0.5, mAP@0.5-0.95, R, and P. Layers is the total number of layers in the network, and Parameters is the total parameters contained in the model. GFLOPS is a giga-floating-point operation per second, which measures the mathematical operation speed of the device. mAP@0.5 is the average precision when IOU=0.5, mAP@0.5-0.95 is the average precision under multiple thresholds of IOU between 0.5 and 0.95, R is the recall rate (Recall), P is the precision (Precision), and the average precision (MeanAverage Precision, mAP).

[0090]

[0091] Among them, P (precision) is the number of predicted positive samples that are true positive samples. R (recall) is the proportion of positive samples predicted as positive samples to the total positive samples. AP (average) is the quality of each category. The average of the average precision. TP (True Positive) is the true label predicted correctly, while FN (False Negative) is the true label that was not predicted, FP (False Positive) is the true label predicted incorrectly, and TN is the true negative label that was predicted. The verification experiment conducted on the dataset of 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 can be seen from Table 1, from a high-precision perspective, the LTM-YOLO11 detection module of the present invention is optimal in every performance indicator. Although the number of layers is slightly more than the baseline model, the difference in the number of parameters is not significant. YOLO11-ABC is the detection model mentioned in patent CN119887850A, with an mAP@0.5 accuracy of 34.7%. The detection accuracy mAP@0.5 of the LTM-YOLO11 detection model of the present invention is 41.9%, which is 7.2% higher than that of similar detection models. Therefore, the LTM-YOLO11 detection model of the present invention meets the requirements of lightweight while also having higher recognition accuracy than other surface garbage multi-target recognition algorithms. It is the best among similar surface garbage multi-target recognition models and can meet the needs of drone surface garbage recognition, verifying the effectiveness of this method.

[0095] In addition, this application also conducted ablation experiments, as shown in Table 2. These experiments also show that the detection model of the present invention has the ability to detect multiple targets of surface debris using drones. In practical application scenarios, it meets the requirements of lightweight and high precision. The synergistic effect of the three modules significantly improves the effect, raising the detection accuracy to over 41%.

[0096] Table 2

[0097]

[0098] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A method for identifying water surface garbage based on drones, characterized in that: The identification method comprises the following steps: 1) Collect UAV surface garbage images containing target information, annotate the garbage, and build a UAV surface garbage image dataset; 2) Build the LTM-YOLO11 detection model; The LTM-YOLO11 detection model is based on the YOLO11 benchmark model. It adds a large-scale convolution kernel LSK attention mechanism between the SPPF module in the backbone layer and the C2PSA module in the neck layer. At the same time, an interactive triple attention mechanism is added after each C3K2 module in the second column of the neck layer. The outputs of the three interactive triple attention mechanisms are connected to the multi-band adaptive dilated convolution module MADC in the head layer. The output of the MADC is passed through three detection heads to obtain the target detection result. The specific process of the multi-band adaptive dilated convolution module MADC is as follows: decomposing the input features into three frequency bands: high, medium, and low; learning different dilation rates by controlling the convolution kernel sampling interval; using a small dilation rate convolution with a dilation rate of 1 to convolve between adjacent pixels in the high-frequency band area; using a medium dilation rate convolution with a dilation rate of 2 to process in the medium-frequency band area; and using a large dilation rate convolution with a dilation rate of 4 to process in the low-frequency band area; and obtaining the MADC output by weighted fusion of the convolution outputs of the high, medium, and low frequency bands. 3) Model training and testing; The drone surface debris image dataset in 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 used to detect drone surface debris and output the recognition results.

2. The identification method according to claim 1, characterized in that The large-scale convolution kernel LSK attention mechanism combines large convolution kernels with separable convolution. In the LSK attention mechanism, the LSK attention mechanism obtains context feature information through the large kernel and decomposes the two-dimensional convolution into two one-dimensional convolution operations. The interactive triple attention mechanism can effectively extract cross-dimensional features in feature maps through different rotation and permutation operations without any information bottleneck, and adopts three different structures to obtain the cross-dimensional interactive calculation attention weights of feature data; the branches of the three different structures are respectively responsible for obtaining the interactive features between the spatial dimension H or W and the channel dimension C.

3. The identification method according to claim 1, characterized in that The specific process of the interactive triple attention mechanism is: The interactive inputs between the three dimensions of channel, height, and width are processed by three branches respectively. In the first branch, the output of the first branch is obtained by Z pooling operation, a convolutional layer of size k×k, Sigmoid function, and rotation operation; in the second branch, the output of the second branch is obtained by Z pooling operation, a convolutional layer of size k×k, Sigmoid function, and rotation operation; in the third branch, the output of the third branch is obtained by Z pooling operation, a convolutional layer of size k×k, and Sigmoid function; the outputs of the three branches are then averaged and aggregated to obtain the output of the interactive triple attention mechanism.

4. The identification method according to claim 1, wherein: The detection accuracy mAP@0.5 of the LTM-YOLO11 detection model is not less than 40%, and the mAP@0.5-0.95 is 24.9%.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the identification method described in any one of claims 1 to 4 can be implemented.

6. A surface garbage identification system based on drone, characterized in that: The system performs the steps of the identification method according to any one of claims 1 to 4, including: UAV image acquisition equipment, used to collect images of water surface garbage; Image preprocessing module, used to annotate and enhance the collected images; The LTM-YOLO11 detection model is used to detect small and wide-sized debris on the surface of water. The alarm module is used to issue an early warning based on the recognition results of the LTM-YOLO11 detection model. If there is garbage on the water surface, a cleaning warning will be issued. If there is no garbage, the drone will update its geographic location patrol and re-detect.

Images