A rice seedling counting method based on a drone
By improving the YOLOv8 network and utilizing lightweight modules and local window attention mechanisms, the problems of complex backgrounds and high plant density in UAV rice seedling counting were solved, achieving efficient and accurate rice seedling detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ANHUI UNIV
- Filing Date
- 2024-05-21
- Publication Date
- 2026-07-24
AI Technical Summary
Traditional drone-based rice seedling counting methods suffer from insufficient target detection accuracy and robustness under complex backgrounds, high plant density, and variable lighting conditions, making it difficult to accurately identify and count rice seedlings.
An improved YOLOv8 network is adopted, which optimizes the backbone network by introducing a lightweight phantom convolution module and a C2f_Faster module, and combines a local window attention mechanism and a neck network of a repetitive weighted bidirectional feature pyramid network to improve the accuracy and robustness of object detection.
It significantly improves the accuracy and robustness of rice seedling counting, and the model can run efficiently in resource-constrained environments, adapting to complex scenes and occlusion situations, thereby improving the accuracy and speed of target detection.
Smart Images
Figure CN118537754B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning computer vision technology, and in particular to a method for counting rice seedlings based on unmanned aerial vehicles (UAVs). Background Technology
[0002] Accurate rice yield forecasting can help farmers and government policymakers rationally plan planting, adjust agricultural policies, and allocate resources, thereby improving the efficiency and quality of crop production. Furthermore, rice is one of the world's major food crops, and its yield changes directly affect global food security and economic development. Therefore, rice yield forecasting is of great significance for ensuring food supply, maintaining social stability, and promoting economic growth. The number of rice plants is one of the important standards for measuring rice yield.
[0003] Currently, most mainstream rice seedling counting methods rely on target detection using images taken with fixed cameras. However, these methods often suffer from limitations in observation angle and area representation, unpredictable occlusion, and perspective distortion due to fixed cameras, resulting in discrepancies in scale. Drone-based rice seedling counting technology offers significant advantages for agricultural production. Drone aerial photography provides comprehensive and efficient image data of rice fields, offering an aerial view of the entire field and capturing information about rice seedlings from all angles, thus improving the comprehensiveness and coverage of detection. Drones equipped with high-resolution cameras can provide high-quality, high-resolution image data, enabling accurate counting and detection of rice seedlings.
[0004] As modern drones become one of the mainstream tools for data acquisition, traditional object detection methods are facing new challenges. Especially in drone-captured images, the background is often more complex, and vegetation density may be higher. These factors lead to a decline in the counting performance of traditional object detection methods. Modern drones can capture high-resolution images, capturing more details and complex scenes, meaning there may be more interfering factors in the images, making object identification and counting difficult. Secondly, increased vegetation density also poses a challenge to object detection. Under dense vegetation cover, targets may be partially or completely obscured, making them difficult to detect and count accurately. Furthermore, drone-captured images often have different viewpoints and lighting conditions, which may lead to diverse representations of targets in the image, increasing the complexity of object detection algorithms. Summary of the Invention
[0005] To address the challenges posed by more complex backgrounds, higher plant density, and varying viewing angles and lighting conditions in images captured by drones, this invention aims to provide a drone-based rice seedling counting method that improves the accuracy and robustness of target detection tasks and enhances the ability to identify targets.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for counting rice seedlings based on unmanned aerial vehicles (UAVs), the method comprising the following sequential steps:
[0007] (1) Data collection: Using drone equipment to photograph rice paddies to obtain panoramic images of rice;
[0008] (2) Image preprocessing: The panoramic image of rice is cropped into multiple 320*320 JPG images, and the images are preprocessed to obtain the preprocessed rice image;
[0009] (3) Create a dataset: Label the preprocessed rice images with rectangular boxes to generate a rectangular box labeled dataset containing training set, validation set and test set;
[0010] (4) Constructing a rice seedling detection model: The YOLOv8 network was improved to obtain the improved YOLOv8 network, which is the rice seedling detection model;
[0011] (5) Training the rice seedling detection model: Input the rectangular box labeled dataset into the rice seedling detection model for training, verification and testing to obtain the trained rice seedling detection model;
[0012] (6) Rice counting: The rice images to be predicted are preprocessed and input into the trained rice seedling detection model. The trained rice seedling detection model outputs the count results of rice seedlings captured by the UAV.
[0013] In step (2), the preprocessing refers to filtering, rotating, cropping, radiating, adding noise, and enhancing brightness and contrast of the image.
[0014] Step (4) specifically refers to: The rice seedling detection model includes an input network, a backbone network, a neck network, and a head network. The C2f module in the backbone network of the YOLOv8 network is replaced with the C2f_Faster module. A local window attention mechanism module is added to the backbone network of the YOLOv8 network. The backbone network performs shallow feature extraction on the image output by the input network to obtain a multi-scale first feature map. The neck network includes a repetitive weighted bidirectional feature pyramid network. The first feature map is input into the neck network for a second feature extraction to obtain a multi-scale second feature map. The second feature map is input into the head network for further feature processing and outputs the target detection result. The target detection result includes the location and category information of the rice.
[0015] The backbone network includes two phantom convolution modules, four C2f_Faster modules, three CBS modules, one SPPF module, and one local window attention mechanism module; the four C2f_Faster modules are the first, second, third, and fourth C2f_Faster modules.
[0016] First, two phantom convolutional modules replace the original two convolutional modules of the YOLOv8 network. This extracts image features gradually with less computational cost and gradually reduces the image resolution to obtain a higher level of abstract feature representation. The kernel size of both phantom convolutional modules is 3*3, and the stride is 2. Next, four C2f_Faster modules and three CBS modules are stacked. The four C2f_Faster modules replace the original C2f modules of the YOLOv8 network. The feature response is adaptively adjusted according to the importance of different positions in the input feature map. The first, second, and fourth C2f_Faster modules are repeated 3 times, and the third C2f_Faster module is repeated 6 times. Then, one SPPF module is stacked to perform pooling operations on the input feature map at different scales to capture multi-scale information of the target. The kernel size of the SPPF module is 5*5. Finally, a local window attention mechanism module is stacked to apply the attention mechanism within the local window, enhance the learning of short-range interactions within the local window, and aggregate fine-grained local features.
[0017] The neck network comprises 6 CBS modules, 2 upsampling modules, 5 feature fusion modules, and 5 C2f_Faster modules; the 6 CBS modules are the first, second, third, fourth, fifth, and sixth CBS modules; the 2 upsampling modules are the first and second upsampling modules; the 5 feature fusion modules are the first, second, third, fourth, and fifth feature fusion modules; and the 5 C2f_Faster modules are the fifth, sixth, seventh, eighth, and ninth C2f_Faster modules.
[0018] The P2 layer includes a first CBS module with a kernel size of 3*3 and a stride of 2;
[0019] The P3 layer consists of the following modules in sequence: a second CBS module with a kernel size of 1*1 and a stride of 1, a first feature fusion module, a fifth C2f_Faster module with 3 repetitions, a second feature fusion module, a sixth C2f_Faster module with 3 repetitions, and a third CBS module with a kernel size of 3*3 and a stride of 2.
[0020] Layer P4 consists of the following modules in sequence: the fourth CBS module with a kernel size of 1*1 and a stride of 1; the third feature fusion module; the seventh C2f_Faster module with a repetition count of 3; the first upsampling module; the fourth feature fusion module; the eighth C2f_Faster module with a repetition count of 3; and the fifth CBS module with a kernel size of 3*3 and a stride of 2.
[0021] The P5 layer consists of the sixth CBS module with a kernel size of 1*1 and a stride of 1, the second upsampling module, the fifth feature fusion module, and the ninth C2f_Faster module with a repetition count of 3.
[0022] First, feature maps at three scales, P3, P4, and P5, are selected and fed into the neck network. The three scales correspond to the results of downsampling the original image 3, 4, and 5 times, respectively.
[0023] To align the channel count, the feature maps of P2, P3, P4, and P5 are first input into the four CBS modules of the corresponding layer for convolution. The convolved P5 feature map is then upsampled and fused with the convolved P4 feature map in the third feature fusion module of layer P4. After this fusion, it passes through the seventh C2f_Faster module of P4 and then undergoes another upsampling by the first upsampling module. The output is then convolved with the feature map of P3 and fused in the first feature fusion module of layer P3. This fusion then passes through the fifth C2f_Faster module of layer P3 and is fused with the convolved P2 and P3 feature maps in the second feature fusion module of layer P3. The convolved P3 feature map is then connected to the second feature fusion module. After this fusion, the result 1 is obtained through the sixth C2f_Faster module and then passed through layer P3 again. The third CBS module performs convolution, and the output is fused with the convolved feature map of P4, the output of the seventh C2f_Faster module of P4, and the output of the first upsampling module of P4. These four components are then fused in the fourth feature fusion module of P4. In this case, the convolved feature map of P4 and the output of the seventh C2f_Faster module of P4 are both connected to the fourth feature fusion module. The output after this fusion is passed through the eighth C2f_Faster module of P4 to obtain result 2. Result 2 then passes through the fifth CBS module of P4, and the output is fused with the convolved feature map of P5 in the fifth feature fusion module of P5. In this case, the convolved feature map of P5 is connected to the fifth feature fusion module. The output after this fusion is then passed through the ninth C2f_Faster module of P5 to obtain result 3.
[0024] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, by introducing a lightweight phantom convolution module and a C2f_Faster module, the present invention successfully optimizes the backbone network, significantly reduces the number of parameters, and improves the running speed of the task. This improvement reduces the number of GFLOPS parameters from the original 8.1 to 7.1, which provides a significant improvement in the efficiency and performance of the model, making the model more efficient when processing large-scale data, and also providing better support for running in resource-constrained environments. Second, the present invention constructs a novel neck network. Inspired by the repetitive weighted bidirectional feature pyramid network, it adopts a strategy of bidirectional pathways and adding extra edges in the same layer, so that the predicted feature map not only has high semantic information, but also contains rich positional information. This comprehensive feature representation is crucial for object detection tasks, providing the model with a more comprehensive and accurate information foundation, thereby improving the accuracy and robustness of object detection. The design of this neck network not only enhances the model's ability to recognize objects but also makes it perform better in complex scenes and under occlusion conditions. Third, the addition of a local window attention mechanism module, compared to the global self-attention mechanism, only requires attention to a local window of the input image, making it more computationally efficient. Furthermore, the local window attention mechanism module has lower memory requirements, making it better suited for resource-constrained environments. Moreover, because it only considers information within a local window and is not affected by distant locations, it performs better in images of densely packed rice seedlings captured by drones. Attached Figure Description
[0025] Figure 1 This is a flowchart of the method of the present invention;
[0026] Figure 2 This is a schematic diagram of the backbone network structure in this invention;
[0027] Figure 3 This is a schematic diagram of the neck network structure in this invention. Detailed Implementation
[0028] like Figure 1 As shown, a method for counting rice seedlings based on unmanned aerial vehicles (UAVs) includes the following sequential steps:
[0029] (1) Data collection: Using drone equipment to photograph rice paddies to obtain panoramic images of rice;
[0030] (2) Image preprocessing: The panoramic image of rice is cropped into multiple 320*320 JPG images, and the images are preprocessed to obtain the preprocessed rice image;
[0031] (3) Create a dataset: Label the preprocessed rice images with rectangular boxes to generate a rectangular box labeled dataset containing training set, validation set and test set;
[0032] (4) Constructing a rice seedling detection model: The YOLOv8 network was improved to obtain the improved YOLOv8 network, which is the rice seedling detection model;
[0033] (5) Training the rice seedling detection model: Input the rectangular box labeled dataset into the rice seedling detection model for training, verification and testing to obtain the trained rice seedling detection model;
[0034] (6) Rice counting: The rice images to be predicted are preprocessed and input into the trained rice seedling detection model. The trained rice seedling detection model outputs the count results of rice seedlings captured by the UAV.
[0035] In step (2), the preprocessing refers to filtering, rotating, cropping, radiating, adding noise, and enhancing brightness and contrast of the image. The image data is filtered, and a batch of image data with severe reflection, many background impurities, and severe occlusion by branches and leaves is deleted based on lighting conditions, background impurities, and occlusion by branches and leaves.
[0036] Step (4) specifically refers to: The rice seedling detection model includes an input network, a backbone network, a neck network, and a head network. The C2f module in the backbone network of the YOLOv8 network is replaced with the C2f_Faster module. A local window attention mechanism module is added to the backbone network of the YOLOv8 network. The backbone network performs shallow feature extraction on the image output by the input network to obtain a multi-scale first feature map. The neck network includes a repetitive weighted bidirectional feature pyramid network. The repetitive weighted bidirectional feature pyramid network has two feature fusion paths: top-down and bottom-up. Full feature fusion is used during feature fusion. The repetitive weighted bidirectional feature pyramid network inputs the first feature map into the neck network for a second feature extraction to obtain a multi-scale second feature map. The second feature map is input into the head network for further feature processing and outputs the target detection result. The target detection result includes the location and category information of the rice.
[0037] like Figure 2 As shown, the backbone network includes two phantom convolution modules, four C2f_Faster modules, three CBS modules, one SPPF module, and one local window attention mechanism module; the four C2f_Faster modules are the first, second, third, and fourth C2f_Faster modules.
[0038] First, two phantom convolutional modules replace the original two convolutional modules of the YOLOv8 network. This extracts image features gradually with less computational cost and gradually reduces the image resolution to obtain a higher level of abstract feature representation. The kernel size of both phantom convolutional modules is 3*3, and the stride is 2. Next, four C2f_Faster modules and three CBS modules are stacked. The four C2f_Faster modules replace the original C2f modules of the YOLOv8 network. The feature response is adaptively adjusted according to the importance of different positions in the input feature map. The first, second, and fourth C2f_Faster modules are repeated 3 times, and the third C2f_Faster module is repeated 6 times. Then, one SPPF module is stacked to perform pooling operations on the input feature map at different scales to capture multi-scale information of the target. The kernel size of the SPPF module is 5*5. Finally, a local window attention mechanism module is stacked to apply the attention mechanism within the local window, enhance the learning of short-range interactions within the local window, and aggregate fine-grained local features.
[0039] like Figure 3 As shown, the neck network includes 6 CBS modules, 2 upsampling modules, 5 feature fusion modules, and 5 C2f_Faster modules; the 6 CBS modules are the first, second, third, fourth, fifth, and sixth CBS modules; the 2 upsampling modules are the first and second upsampling modules; the 5 feature fusion modules are the first, second, third, fourth, and fifth feature fusion modules; and the 5 C2f_Faster modules are the fifth, sixth, seventh, eighth, and ninth C2f_Faster modules.
[0040] The P2 layer includes a first CBS module with a kernel size of 3*3 and a stride of 2;
[0041] The P3 layer consists of the following modules in sequence: a second CBS module with a kernel size of 1*1 and a stride of 1, a first feature fusion module, a fifth C2f_Faster module with 3 repetitions, a second feature fusion module, a sixth C2f_Faster module with 3 repetitions, and a third CBS module with a kernel size of 3*3 and a stride of 2.
[0042] Layer P4 consists of the following modules in sequence: the fourth CBS module with a kernel size of 1*1 and a stride of 1; the third feature fusion module; the seventh C2f_Faster module with a repetition count of 3; the first upsampling module; the fourth feature fusion module; the eighth C2f_Faster module with a repetition count of 3; and the fifth CBS module with a kernel size of 3*3 and a stride of 2.
[0043] The P5 layer consists of the sixth CBS module with a kernel size of 1*1 and a stride of 1, the second upsampling module, the fifth feature fusion module, and the ninth C2f_Faster module with a repetition count of 3.
[0044] First, feature maps at three scales, P3, P4, and P5, are selected and fed into the neck network. The three scales correspond to the results of downsampling the original image 3, 4, and 5 times, respectively.
[0045] To align the channel count, the feature maps of P2, P3, P4, and P5 are first input into the four CBS modules of the corresponding layer for convolution. The convolved P5 feature map is then upsampled and fused with the convolved P4 feature map in the third feature fusion module of layer P4. After this fusion, it passes through the seventh C2f_Faster module of P4 and then undergoes another upsampling by the first upsampling module. The output is then convolved with the feature map of P3 and fused in the first feature fusion module of layer P3. This fusion then passes through the fifth C2f_Faster module of layer P3 and is fused with the convolved P2 and P3 feature maps in the second feature fusion module of layer P3. The convolved P3 feature map is then connected to the second feature fusion module. After this fusion, the result 1 is obtained through the sixth C2f_Faster module and then passed through layer P3 again. The third CBS module performs convolution, and the output is fused with the convolved feature map of P4, the output of the seventh C2f_Faster module of P4, and the output of the first upsampling module of P4. These four components are then fused in the fourth feature fusion module of P4. In this case, the convolved feature map of P4 and the output of the seventh C2f_Faster module of P4 are both connected to the fourth feature fusion module. The output after this fusion is passed through the eighth C2f_Faster module of P4 to obtain result 2. Result 2 then passes through the fifth CBS module of P4, and the output is fused with the convolved feature map of P5 in the fifth feature fusion module of P5. In this case, the convolved feature map of P5 is connected to the fifth feature fusion module. The output after this fusion is then passed through the ninth C2f_Faster module of P5 to obtain result 3. During the training process, i.e. in step (5), a list of feature maps is obtained by predicting results 1, 2, and 3; during the counting process, i.e. in step (6), a list of feature maps is obtained by predicting results 1, 2, and 3, and then a prediction box is generated. The prediction boxes are counted to obtain the rice seedling count results.
[0046] This invention makes three improvements to the YOLOv8 network:
[0047] First, the C2f module in the backbone of the YOLOv8 network is replaced with the C2f_Faster module. Since the C2f_Faster module adds candidate box regression and region proposal network, the accuracy and speed of detection are improved. The Conv convolution in layers 0 and 1 is replaced with ghostconv. By combining a small number of convolution kernels with cheaper linear transformation operations to replace the conventional convolution method, the learning cost of non-key features is reduced.
[0048] Second, a local window attention mechanism module is added to the encoding and decoding of the YOLOv8 network to enhance the learning of short-range interactions within local windows and aggregate fine-grained local features.
[0049] Third, a repetitive weighted bidirectional feature pyramid network is introduced in the neck network. When performing feature fusion, nodes with only one input edge are deleted, additional edges are added to the input and output of the same layer, and weighted feature fusion is used. This not only takes into account the accurate location information and the recognition of small targets, but also reduces the amount of computation and parameters compared to PAnet, i.e., the path aggregation network. It is also more regular, interpretable, and modifiable than the NAS-FPN network.
[0050] In summary, this invention successfully optimized the backbone network by introducing a lightweight phantom convolution module and a C2f_Faster module, significantly reducing the number of parameters and improving the task's running speed. This improvement reduced the number of GFLOPS parameters from 8.1 to 7.1, providing a significant boost to the model's efficiency and performance, making the model more efficient when processing large-scale data, and also providing better support for running in resource-constrained environments. This invention constructs a novel neck network, inspired by the repetitive weighted bidirectional feature pyramid network, and adopts a strategy of bidirectional pathways and adding extra edges in the same layer, so that the predicted feature maps not only possess high semantic information but also contain rich positional information. This comprehensive feature representation is crucial for object detection tasks, providing the model with a more comprehensive and accurate information foundation, thereby improving the accuracy and robustness of object detection. The design of this neck network not only enhances the model's ability to recognize objects but also makes it perform better in complex scenes and under occlusion conditions. The addition of a local window attention mechanism module, compared to the global self-attention mechanism, only requires attention to a local window of the input image, making it more computationally efficient. Furthermore, the local window attention mechanism has lower memory requirements, making it better suited for resource-constrained environments. Moreover, because it only considers information within a local window and is not affected by distant locations, it demonstrates better performance in images of densely packed rice seedlings captured by drones.
Claims
1. A method for counting rice seedlings based on unmanned aerial vehicles (UAVs), characterized in that: The method includes the following steps in sequence: (1) Data collection: Using drone equipment to photograph rice paddies to obtain panoramic images of rice; (2) Image preprocessing: The panoramic image of rice is cropped into multiple 320*320 JPG images, and the images are preprocessed to obtain the preprocessed rice image; (3) Create a dataset: Label the preprocessed rice images with rectangular boxes to generate a rectangular box labeled dataset containing training set, validation set and test set; (4) Constructing a rice seedling detection model: The YOLOv8 network was improved to obtain the improved YOLOv8 network, which is the rice seedling detection model; (5) Training the rice seedling detection model: Input the rectangular box labeled dataset into the rice seedling detection model for training, verification and testing to obtain the trained rice seedling detection model; (6) Rice counting: The rice images to be predicted are preprocessed and input into the trained rice seedling detection model. The trained rice seedling detection model outputs the rice seedling counting results captured by the UAV. Step (4) specifically refers to: the rice seedling detection model includes an input network, a backbone network, a neck network, and a head network. The C2f module in the backbone network of the YOLOv8 network is replaced with the C2f_Faster module. A local window attention mechanism module is added to the backbone network of the YOLOv8 network. The backbone network performs shallow feature extraction on the image output by the input network to obtain a multi-scale first feature map. The neck network includes a repetitive weighted bidirectional feature pyramid network. The first feature map is input into the neck network for a second feature extraction to obtain a multi-scale second feature map. The second feature map is input into the head network for further feature processing and outputs the target detection result. The target detection result includes the location and category information of the rice. The backbone network includes two phantom convolution modules, four C2f_Faster modules, three CBS modules, one SPPF module, and one local window attention mechanism module; the four C2f_Faster modules are the first, second, third, and fourth C2f_Faster modules. First, two phantom convolutional modules replace the original two convolutional modules of the YOLOv8 network. This extracts image features gradually with less computational cost and gradually reduces the image resolution to obtain a higher level of abstract feature representation. The kernel size of both phantom convolutional modules is 3*3, and the stride is 2. Next, four C2f_Faster modules and three CBS modules are stacked. The four C2f_Faster modules replace the original C2f modules of the YOLOv8 network. The feature response is adaptively adjusted according to the importance of different positions in the input feature map. The first, second, and fourth C2f_Faster modules are repeated 3 times, and the third C2f_Faster module is repeated 6 times. Then, one SPPF module is stacked to perform pooling operations on the input feature map at different scales to capture multi-scale information of the target. The kernel size of the SPPF module is 5*5. Finally, a local window attention mechanism module is stacked to apply the attention mechanism within the local window, enhance the learning of short-range interactions within the local window, and aggregate fine-grained local features.
2. The method for counting rice seedlings based on unmanned aerial vehicles according to claim 1, characterized in that: In step (2), the preprocessing refers to filtering, rotating, cropping, radiating, adding noise, and enhancing brightness and contrast of the image.
3. The method for counting rice seedlings based on unmanned aerial vehicles according to claim 1, characterized in that: The neck network comprises 6 CBS modules, 2 upsampling modules, 5 feature fusion modules, and 5 C2f_Faster modules; the 6 CBS modules are the first, second, third, fourth, fifth, and sixth CBS modules; the 2 upsampling modules are the first and second upsampling modules; the 5 feature fusion modules are the first, second, third, fourth, and fifth feature fusion modules; and the 5 C2f_Faster modules are the fifth, sixth, seventh, eighth, and ninth C2f_Faster modules. The P2 layer includes a first CBS module with a kernel size of 3*3 and a stride of 2; The P3 layer consists of the following modules in sequence: a second CBS module with a kernel size of 1*1 and a stride of 1, a first feature fusion module, a fifth C2f_Faster module with 3 repetitions, a second feature fusion module, a sixth C2f_Faster module with 3 repetitions, and a third CBS module with a kernel size of 3*3 and a stride of 2. Layer P4 consists of the following modules in sequence: the fourth CBS module with a kernel size of 1*1 and a stride of 1; the third feature fusion module; the seventh C2f_Faster module with a repetition count of 3; the first upsampling module; the fourth feature fusion module; the eighth C2f_Faster module with a repetition count of 3; and the fifth CBS module with a kernel size of 3*3 and a stride of 2. The P5 layer consists of the sixth CBS module with a kernel size of 1*1 and a stride of 1, the second upsampling module, the fifth feature fusion module, and the ninth C2f_Faster module with a repetition count of 3. First, feature maps at three scales, P3, P4, and P5, are selected and fed into the neck network. The three scales correspond to the results of downsampling the original image 3, 4, and 5 times, respectively. To align the channel count, the feature maps of P2, P3, P4, and P5 are first input into the four CBS modules of the corresponding layer for convolution. The convolved P5 feature map is then upsampled and fused with the convolved P4 feature map in the third feature fusion module of layer P4. After this fusion, it passes through the seventh C2f_Faster module of P4 and then undergoes another upsampling by the first upsampling module. The output is then convolved with the feature map of P3 and fused in the first feature fusion module of layer P3. This fusion then passes through the fifth C2f_Faster module of layer P3 and is fused with the convolved P2 and P3 feature maps in the second feature fusion module of layer P3. The convolved P3 feature map is then connected to the second feature fusion module. After this fusion, the result 1 is obtained through the sixth C2f_Faster module and then passed through layer P3 again. The third CBS module performs convolution, and the output is fused with the convolved feature map of P4, the output of the seventh C2f_Faster module of P4, and the output of the first upsampling module of P4. These four components are then fused in the fourth feature fusion module of P4. In this case, the convolved feature map of P4 and the output of the seventh C2f_Faster module of P4 are both connected to the fourth feature fusion module. The output after this fusion is passed through the eighth C2f_Faster module of P4 to obtain result 2. Result 2 then passes through the fifth CBS module of P4, and the output is fused with the convolved feature map of P5 in the fifth feature fusion module of P5. In this case, the convolved feature map of P5 is connected to the fifth feature fusion module. The output after this fusion is then passed through the ninth C2f_Faster module of P5 to obtain result 3.