A method and system for detecting seed blocking and clearing seed of an air-vibration type precision seeder seed suction disc
Patent Information
- Application Number
- CN202311588434.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-24
- Publication Date
- 2026-10-16
- Estimated Expiration
- 2043-11-24
AI Technical Summary
气振式精密播种机在播种工作过程中，由于连续播种或者吸附带有毛刺的水稻籽粒，会出现吸种盘盘面吸孔堵塞现象，若不及时检测，“卡种”现象会造成连续播种过程中空穴率上升，播种合格率下降
[0037] 1. By using target detection algorithms in the field of deep learning, the system can accurately and quickly identify seed jamming on the surface of the seed suction tray, which can be used for seed cleaning in the production line to improve the efficiency and accuracy of seed jamming detection.
Smart Images
Figure CN117612119B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of agricultural precision seeding technology and deep learning, specifically involving a method, system, and device for detecting and cleaning seeds stuck in the seed suction disc of a pneumatic precision seeder. Background Technology
[0002] Sowing plays a crucial role in agricultural production. With the rapid development of modern agriculture, the requirements for sowing quality and efficiency are becoming increasingly stringent. Pneumatic precision seeders offer advantages such as high sowing accuracy, minimal seed damage, and minimal impact from seed shape. They also meet the technical requirements for precision seeding of super hybrid rice (2±1 seeds / hole) and are widely used in precision sowing, making them highly productive precision sowing devices. However, during the sowing process, continuous sowing or the absorption of rice grains with burrs can cause blockages in the suction holes of the seed suction disc. If not detected promptly, this "seed jamming" phenomenon can lead to an increase in empty holes and a decrease in the sowing qualification rate during continuous sowing.
[0003] Currently, seed jamming on the seed suction plate of pneumatic precision seeders is mainly detected manually. Workers visually inspect the seed suction plate for rice grains, a method that is inefficient and whose accuracy decreases with increasing worker fatigue. While image processing technology has been applied to seed jam detection, and researchers typically use traditional machine vision methods, these methods are too time-consuming to meet real-time detection requirements.
[0004] In recent years, with the development of deep learning, the use of deep learning detection methods can greatly improve the efficiency and accuracy of seed jamming detection in seed suction discs. Therefore, this invention proposes a method and system for detecting and cleaning seed jamming in the seed suction disc of a pneumatic precision seeder. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention proposes a method and system for detecting and cleaning seeds stuck in the seed suction disc of a pneumatic precision seeder, thereby improving detection efficiency and accuracy and further enhancing agricultural automation and mechanization.
[0006] Note that the description of these objectives does not preclude the existence of other objectives. One aspect of the invention does not necessarily require achieving all of the above objectives. The invention achieves the above-mentioned technical objectives through the following technical means.
[0007] A method for detecting and cleaning seeds stuck in the seed suction disc of a pneumatic precision seeder includes the following steps:
[0008] Step S1, Image Acquisition: Acquire images of seed jamming on the seed suction plate of the pneumatic precision seeder, and obtain a set of seed jamming images on the seed suction plate;
[0009] Step S2, Image Preprocessing: Filtering, denoising, distortion correction, and image equalization are performed on the seed suction plate image;
[0010] Step S3, Card Variety Image Dataset Creation: Use the open-source annotation tool Labelimg to annotate the card variety images preprocessed in Step S2 to create a card variety image dataset for the seed sucker.
[0011] Step S4, Network Model Construction and Training: Construct a seed sucker plate detection network model, input the seed sucker plate image dataset made in step S3 into the constructed network model for training and testing, and obtain the trained seed sucker plate detection network model;
[0012] Step S5, Online Image Detection: Deploy the seed suction tray jam detection network model trained in Step S4 onto the industrial control computer to detect seed suction tray jam images in real time. If seed jamming is detected in the seed suction tray, the industrial control computer issues an instruction to clean the seed suction tray through the seed cleaning device.
[0013] Step S6, Seed Cleaning: After the command is issued in step S5, the seed cleaning brush located between the vibrating seed tray and the production line sweeps the surface of the seed suction tray to clean the seeds stuck in the suction holes.
[0014] In the above scheme, in step S1, an industrial camera with a light source is used to collect images of the seed suction plate stuck on the seed plate on the air vibration precision seeding production line.
[0015] In the above scheme, step S3 involves using a rectangular bounding box to select and label the rice grains that are stuck in the water.
[0016] In the above scheme, the seed detection network model of the seed suction tray in step S4 is an improved YOLOv5 target detection network model; the network structure design is as follows:
[0017] The network structure mainly consists of three parts: Backbone, Neck, and Head. The Backbone, as the backbone feature extraction network, extracts features from the image, transforming the input image into a multi-layer feature map. The main components of the Backbone include the Conv module, C3 module, and SPPF module. The Neck is primarily a PANet structure, containing the Conv module, C3 module, upsampling module, and C3_2_CA module. In the Neck, a small object detection layer is added to improve the algorithm's accuracy in detecting small objects such as rice grains on suction holes. When the input image size is 640×640, convolution and upsampling operations are added after layer 17 of the object detection network. At layer 20, a 160×160 feature map is concatenated with the output feature map of layer 3 in the Backbone network and input into the small object detection layer. A 4x downsampling layer is added to the detection head, increasing detection to four scales and improving the accuracy in detecting small objects such as rice grains.
[0018] Additionally, a coordinate attention module, or CA module, is introduced after the second convolution in the BottleNeck module of the C3 module of the YOLOv5 object detection Neck network. This CA module consists of two steps: coordinate information embedding and coordinate attention generation. Coordinate information embedding transforms global pooling into an encoding operation of two one-dimensional vectors. For the input X, pooling kernels of size (H, 1) and (1, W) are first used to encode each channel along the horizontal and vertical coordinate directions. The feature outputs of the c-th channel, with height h and width w, are as follows:
[0019]
[0020]
[0021] in, This is the feature output of the c-th channel in terms of height; The c-th channel is the feature output in terms of width; H is the feature map height; W is the feature map width; X c This is the input for the c-th channel;
[0022] The coordinate attention generation process is as follows: First, the width and height feature maps are concatenated to obtain a feature layer [C, 1, H+W]. Convolution, normalization, and non-linear activation functions are used for feature transformation. Then, the feature layers are divided into two more feature layers: [C, H, 1] and [C, 1, W]. Two 1x1 convolutions and a sigmoid function are used to obtain attention weights for height and width. Finally, the input features are multiplied and weighted to obtain the output of the coordinate attention module.
[0023] In the Head section, the loss function CIoULoss in the YOLOv5 object detection network is replaced with Focal_EIoULoss to improve convergence speed. The loss function is:
[0024] L Focal-EIOU =IOU γ L EIOU
[0025] in:
[0026]
[0027] In the formula, IOU(A, B) represents the intersection-union ratio of the predicted bounding box and the target bounding box; w c h represents the width of the minimum bounding rectangle of the predicted bounding box and the target bounding box. c b represents the height of the minimum bounding rectangle of the predicted bounding box and the target bounding box; b represents the coordinates of the center point of the predicted bounding box; b gt Represents the coordinates of the center point of the target bounding box; ρ(b, b gt ) represents the Euclidean distance between the center point of the predicted bounding box and the center point of the target bounding box; h represents the height of the predicted bounding box; h gt w represents the height of the target bounding box; w represents the width of the predicted bounding box; w gt Represents the width of the target bounding box; ρ(h, h gt ) represents the difference in height between the predicted bounding box and the target bounding box; ρ(w, w) represents the difference in height between the predicted bounding box and the target bounding box. gt ) represents the difference in width between the predicted bounding box and the target bounding box; γ represents the hyperparameter used to control the curvature of the curve, with a value of 0.5;
[0028] A system for implementing the seed suction disc jamming detection and seed cleaning method of the aforementioned air-vibration precision seeder includes an image acquisition module, an image preprocessing module, a dataset creation module, a network model construction and training module, an online seed suction disc jamming detection module, and a seed cleaning module;
[0029] The image acquisition module is used to acquire images of seeds stuck in the seed suction tray and obtain a set of images of seeds stuck in the seed suction tray.
[0030] The image preprocessing module is used to filter, denoise, correct distortion, and equalize the acquired seed suction plate image.
[0031] The image dataset creation module is used to annotate the preprocessed card images using the open-source annotation tool Labelimg to create a seed-suction plate card image dataset.
[0032] The network model construction and training module is used to construct a seed sucker plate card detection network model. The seed sucker plate card image dataset is input into the constructed network model for training and testing to obtain a trained seed sucker plate card detection network model.
[0033] The seed suction tray jamming online detection module is used to deploy the trained seed suction tray jamming detection network model on the industrial control computer, detect seed suction tray jamming images in real time, and detect whether there is jamming through the seed suction tray jamming detection network model. If seed jamming is found, the industrial control computer issues an instruction to clean the seed suction tray through the seed cleaning device.
[0034] The seed cleaning module is used to clean seeds stuck on the surface of the seed suction plate. The seed cleaning brush located between the vibrating seed plate and the production line sweeps the surface of the seed suction plate to clean seeds stuck in the suction holes.
[0035] In the above scheme, the image acquisition module includes an industrial camera and a light source; the industrial camera, in conjunction with the light source, acquires images of seeds stuck on the seed suction disc of the pneumatic precision seeder; the industrial camera is connected to a photoelectric sensor for controlling the industrial camera to take pictures.
[0036] The beneficial effects of this invention are:
[0037] 1. By using target detection algorithms in the field of deep learning, the system can accurately and quickly identify seed jamming on the surface of the seed suction tray, which can be used for seed cleaning in the production line to improve the efficiency and accuracy of seed jamming detection.
[0038] 2. An industrial camera with a light source is used to clearly observe whether seeds are stuck on the surface of the seed suction tray.
[0039] 3. An additional small target detection layer was added, which improved the detection accuracy of small rice grains stuck on the suction hole.
[0040] 4. A coordinate attention module is introduced after the second convolution of BottleNeck in the C3 module of the YOLOv5 object detection Neck network to reduce interference from useless background; Focal_EIoU Loss is introduced to accelerate the convergence speed of the model.
[0041] 5. The average accuracy of seed jamming detection on the seed suction disc of the pneumatic precision seeder is 96.8% (mAP@0.5). Attached Figure Description
[0042] Figure 1 This is a general design block diagram of one embodiment of the present invention;
[0043] Figure 2 This is a schematic diagram of an improved YOLOv5 network structure according to an embodiment of the present invention;
[0044] Figure 3 This is a schematic diagram of the CA module structure according to one embodiment of the present invention;
[0045] Figure 4 This is a graph of the Focal_EIoU Loss function according to an embodiment of the present invention;
[0046] Figure 5 This is a detection effect diagram of an embodiment of the present invention. Detailed Implementation
[0047] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0048] Figure 1 The above describes a preferred embodiment of the seed suction disc jamming detection and cleaning method for the pneumatic precision seeder. The method includes the following steps:
[0049] Step S1, Image Acquisition: On the pre-built seed suction plate jamming detection platform, an industrial camera with light source is used to monitor the pneumatic precision seeder production line and acquire a set of seed suction plate jamming images;
[0050] Step S2, Image Preprocessing: Filtering, denoising, distortion correction, and image equalization are performed on the seed suction plate image;
[0051] Step S3, Card Variety Image Dataset Creation: Use the open-source annotation tool Labelimg to annotate the card variety images preprocessed in Step S2 to create a card variety image dataset for the seed sucker.
[0052] Step S4, Network Model Construction and Training: By improving the YOLOv5 target detection network, the detection accuracy of small targets such as rice grains stuck on the suction hole is improved. The image dataset of seed stuck on the suction plate prepared in step S3 is input into the constructed network model for training and testing, and a trained seed stuck on the suction plate detection network model is obtained.
[0053] Step S5, Online Image Detection: Deploy the seed suction tray jam detection network model trained in Step S4 onto the industrial control computer. Install the industrial camera on the bracket on the production line, located 580mm below the seed suction tray, and capture an image of the seed suction tray surface. Transmit the image to the industrial control computer, which then sends the image to the seed suction tray jam detection network model for processing. If a seed jam is detected in the seed suction tray, the industrial control computer sends a 1 signal to the seed cleaning device, which then cleans the seed suction tray. Otherwise, it sends a 0 signal.
[0054] Step S6, Seed Cleaning: After the command is issued in step S5, the seed cleaning brush located between the vibrating seed tray and the production line sweeps the surface of the seed suction tray to clean the seeds stuck in the suction holes.
[0055] Preferably, in step S3, labeling refers to using a bounding box to select the trapped rice grains and attaching a seed label.
[0056] Preferably, the improvement of the YOLOv5 object detection network in step S4 includes:
[0057] A small target detection layer was added to improve the detection accuracy of small targets such as rice grains stuck on the suction hole;
[0058] A coordinate attention module is added after the second convolution of BottleNeck in the C3 module of the YOLOv5 object detection Neck network to reduce interference from useless background.
[0059] Introducing Focal_EIoU Loss accelerates the model's convergence speed.
[0060] The addition of a small target detection layer specifically improves the algorithm's accuracy in detecting small targets such as rice grains on the suction hole. When the input image size is 640×640, convolution and upsampling operations are added after layer 17 of the target detection network. At layer 20, a 160×160 feature map is concatenated with the output feature map of layer 3 in the backbone network and input into the small target detection layer. A 4x downsampling layer is added to the detection head, increasing detection to four scales, thus improving the accuracy of detecting small targets like rice grains and preventing missed detections due to their small size.
[0061] The addition of the coordinate attention module specifically involves introducing a coordinate attention module, or CA module, after the second convolution in the BottleNeck module of the C3 module of the YOLOv5 object detection Neck network. This module consists of two steps: coordinate information embedding and coordinate attention generation. Coordinate information embedding is an encoding operation that transforms global pooling into two one-dimensional vectors. For the input X, pooling kernels of size (H, 1) and (1, W) are first used to encode each channel along the horizontal and vertical coordinate directions. The feature output of the c-th channel, with height h and width w, is as follows:
[0062]
[0063]
[0064] in, This is the feature output of the c-th channel in terms of height; The c-th channel is the feature output in terms of width; H is the feature map height; W is the feature map width; X c This is the input for the c-th channel;
[0065] The coordinate attention generation process is as follows: First, the width and height feature maps are concatenated to obtain a feature layer [C, 1, H+W]. Convolution, normalization, and non-linear activation functions are used for feature transformation. Then, the feature layers are divided into two more feature layers: [C, H, 1] and [C, 1, W]. Two 1x1 convolutions and a sigmoid function are used to obtain attention weights for height and width. Finally, the input features are multiplied and weighted to obtain the output of the coordinate attention module. By adding a coordinate attention mechanism to the neck network, the detection accuracy of the model can be improved.
[0066] The introduction of Focal_EIoU Loss specifically involves replacing the loss function CIoULoss in the YOLOv5 object detection network with Focal_EIoU Loss to improve convergence speed. The loss function is:
[0067] L Focal-EIOU =IOU γ L EIOU
[0068] in:
[0069]
[0070] In the formula, IOU(A, B) represents the intersection-union ratio of the predicted bounding box and the target bounding box; w c h represents the width of the minimum bounding rectangle of the predicted bounding box and the target bounding box. c b represents the height of the minimum bounding rectangle of the predicted bounding box and the target bounding box; b represents the coordinates of the center point of the predicted bounding box; b gt Represents the coordinates of the center point of the target bounding box; ρ(b, b gt ) represents the Euclidean distance between the center point of the predicted bounding box and the center point of the target bounding box; h represents the height of the predicted bounding box; h gt w represents the height of the target bounding box; w represents the width of the predicted bounding box; w gt Represents the width of the target bounding box; ρ(h, h gt ) represents the difference in height between the predicted bounding box and the target bounding box; ρ(w, w) represents the difference in height between the predicted bounding box and the target bounding box. gt ) represents the difference in width between the predicted bounding box and the target bounding box; γ represents the hyperparameter used to control the curvature of the curve, with a value of 0.5; replacing the CIoU Loss loss function in the YOLOv5 object detection network with Focal_EIoU Loss improves the convergence speed and optimizes the sample imbalance problem in the bounding box regression task, that is, reducing the optimization contribution of a large number of anchor boxes with little overlap with the target box to the BBox regression, so that the regression process focuses on high-quality anchor boxes and improves the prediction accuracy.
[0071] A system for implementing the seed suction disc jamming detection and seed cleaning method of the aforementioned air-vibration precision seeder includes an image acquisition module, an image preprocessing module, a dataset creation module, a network model construction and training module, an online seed suction disc jamming detection module, and a seed cleaning module;
[0072] The image acquisition module is used to acquire images of the seed suction plate and obtain a clear set of images of the seed suction plate and seed.
[0073] The image preprocessing module is used to filter, denoise, correct distortion, and equalize the acquired seed suction plate image.
[0074] The image dataset creation module is used to annotate the preprocessed card images using the open-source annotation tool Labelimg to create a seed-suction plate card image dataset.
[0075] The network model construction and training module is used to construct a seed sucker plate card detection network model. The seed sucker plate card image dataset is input into the constructed network model for training and testing to obtain a trained seed sucker plate card detection network model.
[0076] The seed suction tray jamming online detection module is used to deploy the trained seed suction tray jamming detection network model on the industrial control computer, detect seed suction tray jamming images in real time, and detect whether there is jamming through the seed suction tray jamming detection network model. If seed jamming is found, the industrial control computer issues an instruction to clean the seed suction tray through the seed cleaning device.
[0077] The seed cleaning module is used to clean seeds stuck on the surface of the seed suction plate. The seed cleaning brush located between the vibrating seed plate and the production line sweeps the surface of the seed suction plate to clean seeds stuck in the suction holes.
[0078] Preferably, the image acquisition module includes an industrial camera and a light source; the industrial camera, in conjunction with the light source, acquires images of seeds stuck on the seed suction disc of the pneumatic precision seeder; the industrial camera is connected to a photoelectric sensor for controlling the industrial camera to take pictures. Specific implementation examples:
[0080] According to this embodiment, preferably, the model training parameters are set as follows: Batch size is set to 8, number of iterations is set to 200, initial learning rate is set to 0.01, and learning rate momentum is set to 0.937.
[0081] The overall flowchart of the seed detection and cleaning method for the seed suction disc of the pneumatic precision seeder described in this invention is as follows: Figure 1 As shown, it includes the following operation steps:
[0082] Step S101: Image Acquisition: An image acquisition device is built using an industrial camera, a light source, a camera controller, and a light source controller. Preferably, the image acquisition device is located 580mm below the seed suction plate to acquire images of seed jamming on the seed suction plate of the pneumatic precision seeder, thus obtaining a set of seed jamming images.
[0083] Step S102: Image preprocessing: The seed suction plate image obtained in step S101 is filtered, denoised, distorted, and equalized.
[0084] Step S103: Card variety image dataset creation: The card variety images preprocessed in step S102 are labeled using the open-source annotation tool Labelimg to create a card variety image dataset for seed suckers;
[0085] Step S104: Network Model Construction and Training: By improving the YOLOv5 target detection network, the detection accuracy of small targets such as rice grains stuck on the suction hole is improved, making it more suitable for seed suction plate jam detection; the seed suction plate jam image dataset prepared in step S103 is input into the constructed network model for training and testing, and a trained seed suction plate jam detection network model is obtained.
[0086] In this embodiment, to improve the detection accuracy of small rice grain targets, the target detection network was improved in three aspects: the detection head, the attention mechanism, and the loss function. When the input image size is 640×640, convolution and upsampling operations are added after layer 17 of the target detection network. At layer 20, a 160×160 feature map is concatenated with the output feature map of layer 3 in the backbone network and input into the small target detection layer. A 4x downsampling layer is added to the detection head, increasing the detection to four scales, thus improving the detection accuracy of small rice grain targets. The improved YOLOv5 network structure is as follows: Figure 2 As shown.
[0087] In the YOLOv5 object detection Neck network C3 module, a coordinate attention module, or CA module, is introduced after the second convolution of BottleNeck, as follows: Figure 3 As shown. Figure 3 (a) is a schematic diagram of the coordinate attention mechanism. Figure 3 (b) C3 module after attention is added. The CA module consists of two steps: coordinate information embedding and coordinate attention generation. Coordinate information embedding is an encoding operation that replaces global pooling with two one-dimensional vectors. For input X, firstly, pooling kernels of size (H, 1) and (1, W) are used to encode each channel along the horizontal and vertical coordinate directions. The feature outputs of the c-th channel with height h and width w are as follows:
[0088]
[0089]
[0090] in, This is the feature output of the c-th channel in terms of height; The c-th channel is the feature output in terms of width; H is the feature map height; W is the feature map width; X c This is the input for the c-th channel;
[0091] The coordinate attention generation process is as follows: First, the width and height feature maps are concatenated to obtain a feature layer [C, 1, H+W]. Convolution, normalization, and non-linear activation functions are used for feature transformation. Then, the feature layers are divided into two more feature layers: [C, H, 1] and [C, 1, W]. Two 1x1 convolutions and a sigmoid function are used to obtain attention weights for height and width. Finally, the input features are multiplied and weighted to obtain the output of the coordinate attention module. By adding a coordinate attention mechanism to the neck network, the detection accuracy of the model can be improved.
[0092] Replace the CIoU Loss function in the YOLOv5 object detection network with Focal_EIoU Loss, as follows: Figure 4 As shown, to improve the convergence speed, the loss function is:
[0093] L Focal-EIOU =IOU γ L EIOU
[0094] in:
[0095]
[0096] To optimize the sample imbalance problem in bounding box regression tasks, we need to reduce the contribution of a large number of anchor boxes with little overlap with the target box to the optimization of BBox regression, so that the regression process focuses on high-quality anchor boxes and improves prediction accuracy.
[0097] Step S105: Online Image Detection: Deploy the seed suction tray jam detection network model trained in Step S104 onto the industrial control computer. Mount an industrial camera on a bracket on the production line, positioned 580mm below the seed suction tray, to capture images of the seed suction tray surface and transmit them to the industrial control computer. The industrial control computer sends the images to the seed suction tray jam detection network model for processing. If seed jamming is detected in the seed suction tray, the industrial control computer sends a 1 signal to the seed cleaning device, which then cleans the seed suction tray. Otherwise, it sends a 0 signal. The detection results are as follows: Figure 5 As shown in the figure, this method achieves the purpose of detecting seed jamming on the seed suction plate.
[0098] Step S106: Seed cleaning: The seed cleaning brush located between the vibrating seed tray and the production line sweeps the surface of the seed suction tray to clean the seeds stuck in the suction holes.
[0099] Therefore, through the above steps, the average accuracy mAP@0.5 of the improved YOLOv5 target detection model for detecting seed jamming on the seed suction disc of the pneumatic precision seeder is 96.8%. The preferred hardware and software platform settings are shown in Table 1.
[0100] Table 1
[0101]
[0102]
[0103] The method for detecting and cleaning seeds stuck in the seed suction disc of a pneumatic precision seeder described in this invention can improve the ability to detect and clean seeds stuck in the seed suction disc of a pneumatic precision seeder, meeting the requirements for detection speed and accuracy in agricultural production.
[0104] This invention improves the YOLOv5 target detection network to address the problem of seed jamming on the seed suction plate of a pneumatic precision seeder. First, a small target detection layer is added to improve the detection accuracy of small targets such as rice grains stuck on the suction hole. Second, a coordinate attention module is added to reduce interference from useless background. Finally, Focal_EIoU Loss is introduced to accelerate the model's convergence speed. Final testing shows that this invention achieves an average accuracy (mAP@0.5) of 96.8% for seed jamming on the seed suction plate of a pneumatic precision seeder, enhancing the detection model's ability to detect seed jamming on the seed suction plate and laying a foundation for seed cleaning operations in pneumatic precision seeders.
[0105] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent methods or modifications that do not depart from the technology of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for detecting and cleaning seeds stuck in the seed suction disc of a pneumatic precision seeder, characterized in that, Includes the following steps: Step S1, Image Acquisition: Acquire images of seed jamming on the seed suction plate of the pneumatic precision seeder, and obtain a set of seed jamming images on the seed suction plate; Step S2, Image Preprocessing: Filtering, denoising, distortion correction, and image equalization are performed on the seed suction plate image; Step S3, Card Variety Image Dataset Creation: Use the open-source annotation tool Labelimg to annotate the card variety images preprocessed in Step S2 to create a card variety image dataset for the seed sucker. Step S4, Network Model Construction and Training: Construct a seed sucker plate detection network model, input the seed sucker plate image dataset made in step S3 into the constructed network model for training and testing, and obtain the trained seed sucker plate detection network model; The seed suction disc detection network model adopts an improved YOLOv5 target detection network model, and its specific structure design is as follows: The network structure consists of three parts: Backbone, Neck, and Head. The Backbone, as the main feature extraction network, extracts features from the image and transforms the input image into a multi-layer feature map. The Backbone contains Conv, C3, and SPPF modules. The Neck structure is PANet, containing Conv, C3, upsampling, and C3_2_CA modules. In the Neck part, a small target detection layer is added to improve the algorithm's detection accuracy for small targets such as rice grains on suction holes. When the input image size is 640×640, convolution and upsampling operations are added after layer 17 of the target detection network. At layer 20, a 160×160 feature map is concatenated with the output feature map of layer 3 in the Backbone network and input into the small target detection layer. A 4x downsampling layer is added to the detection head part, increasing the detection to four scales and improving the detection accuracy for small targets such as rice grains. Additionally, a coordinate attention module, or CA module, is introduced after the second convolution in the Bottleneck section of the C3 module of the YOLOv5 object detection Neck network, forming the C3_2_CA module. This module consists of two steps: coordinate information embedding and coordinate attention generation. Coordinate information embedding is an encoding operation that replaces global pooling with two one-dimensional vectors for the input... First, use a size of ( ,1) and (1, The pooling kernel encodes each channel along the horizontal and vertical coordinates, with a height of [missing information]. and width are The The feature outputs for each channel are as follows: in, For the first in terms of height The characteristic output of the channel; For the width of the first The characteristic output of the channel; For feature map height; The width of the feature map; For the first Channel input; The coordinate attention generation process is as follows: First, the width and height feature maps are concatenated to obtain a feature layer [C,1,H+W], and feature transformation is performed using convolution, normalization, and non-linear activation functions; then, it is divided into two feature layers [C,H,1] and [C,1,W], and attention weights on the height and width are obtained using two 1x1 convolutions and a sigmoid function; finally, the input features are multiplied and weighted to obtain the output of the coordinate attention module. In the Head section, the loss function CIoU Loss in the YOLOv5 object detection network is replaced with Focal_EIoULoss to improve the convergence speed. The loss function is: in: In the formula, in the formula, This represents the intersection-union ratio (IU) between the predicted bounding box and the target bounding box. This represents the width of the smallest bounding rectangle of the predicted bounding box and the target bounding box; This represents the height of the minimum bounding rectangle between the predicted bounding box and the target bounding box; Indicates the coordinates of the center point of the prediction box; Indicates the coordinates of the center point of the target box; This represents the Euclidean distance between the center point of the predicted bounding box and the center point of the target bounding box; Indicates the height of the prediction box; Indicates the height of the target bounding box; Indicates the width of the prediction box; Indicates the width of the target bounding box; This represents the difference in height between the predicted bounding box and the target bounding box; This represents the difference in width between the predicted bounding box and the target bounding box; This represents the hyperparameter used to control the curvature of the curve, with a value of 0.5; Step S5, Online Image Detection: Acquire images of the seed suction plate in real time, and perform real-time seed detection using the trained seed suction plate seed detection network model; Step S6, Seed Cleaning: If seed jamming is found in the seed suction tray, perform a seed cleaning operation on the seed suction tray.
2. The method for detecting and cleaning seeds stuck in the seed suction disc of a pneumatic precision seeder according to claim 1, characterized in that, In step S1, an industrial camera with a light source is used to capture images of the seed suction plate stuck on the seed in the air-vibration precision seeding production line.
3. The method for detecting and cleaning seeds stuck in the seed suction disc of a pneumatic precision seeder according to claim 1, characterized in that, In step S3, the rice grains that are selected and trapped are marked with a rectangular bounding box.
4. A seed-catching detection and cleaning device for a pneumatic precision seeder, characterized in that, The device, which deploys the seed suction plate jam detection network model trained in step S4 of any one of claims 1-3, is an industrial control computer; it acquires images of the seed suction plate in real time and performs real-time seed jam detection using the trained seed suction plate jam detection network model; if seed jamming is detected in the seed suction plate, the device issues a seed clearing command to clear the seed suction plate.
5. The seed detection and cleaning device for the seed suction disc of a pneumatic precision seeder according to claim 4, characterized in that, The seed cleaning operation is as follows: after issuing the seed cleaning command, the seed cleaning brush located between the vibrating seed tray and the production line is controlled to sweep the surface of the seed suction tray to clean the seeds stuck in the suction holes.
6. A system for implementing the seed suction disc jamming detection and seed cleaning method of any one of claims 1-3 in a pneumatic precision seeder, characterized in that, include: Image acquisition module, image preprocessing module, dataset creation module, network model construction and training module, seed suction tray online detection module, and seed cleaning module; The image acquisition module is used to acquire images of seeds stuck in the seed suction tray and obtain a set of images of seeds stuck in the seed suction tray; The image preprocessing module is used to perform filtering, noise reduction, distortion correction, and image equalization processing on the acquired seed suction plate image. The image dataset creation module is used to annotate the preprocessed card images using the open-source annotation tool Labelimg to create a seed-suction plate card image dataset. The network model construction and training module is used to construct a seed sucker plate card detection network model, input the seed sucker plate card image dataset into the constructed network model for training and testing, and obtain a trained seed sucker plate card detection network model. The seed suction tray jamming online detection module deploys the trained seed suction tray jamming detection network model on the industrial control computer, detects seed suction tray jamming images in real time, and detects whether seed is jammed through the seed suction tray jamming detection network model. If seed jamming is detected, the industrial control computer issues an instruction to clear the seed suction tray through the seed clearing module.
7. The system according to claim 6, characterized in that, The seed cleaning module is used to clean seeds stuck on the surface of the seed suction plate. The seed cleaning brush between the vibrating seed plate and the production line is used to sweep the surface of the seed suction plate to clean the seeds stuck in the suction holes.
8. The system according to claim 6, characterized in that, The image acquisition module includes an industrial camera and a light source; the industrial camera, in conjunction with the light source, acquires images of seeds stuck on the seed suction disc of the pneumatic precision seeder; the industrial camera is connected to a photoelectric sensor for controlling the industrial camera to take pictures.
9. The system according to claim 8, characterized in that, The industrial camera is mounted on a bracket on the production line, located 580mm below the seed suction plate, to capture images of the seed suction plate surface.