Method and system for detecting adsorption condition of corn seeds in air-suction type seed-metering device

Abstract

The invention provides a method and system for detecting the adsorption condition of corn seeds in an air-suction type seed sowing device, and relates to the technical field of intelligent agriculture. A camera is installed above a seed sowing disc, videos on the upper surface of the seed sowing disc are collected through the camera, and a YOLOv8n model is improved; according to the improved YOLOv8n model, a C2fCIB convolution module is introduced into a neck network of an original YOLOv8n model to replace an original C2f module of a P5 layer, a RepHead structure is adopted at a detection head to replace an original detection head, an internal intersection-to-union ratio and a shape intersection-to-union ratio are adopted for a loss function to replace an original complete intersection-to-union ratio, the improved YOLOv8n model is trained through a training set, and a target detection model is obtained. The target detection model is verified by using the test set, the target detection model is used for identifying the video, collected by the camera, of the upper surface of the seed-metering plate, and the adsorption condition of the corn seeds sown each time is obtained, so that the problem that the adsorption condition of the corn seeds in an existing air-suction type seed-metering device is not accurate is solved, the problems of miss-seeding and repeated seeding are found in time, and the seeding efficiency is improved. The device is suitable for seeding detection of the air-aspiration type seed-metering device.

Description

Technical Field

[0001] The present invention relates to the field of smart agricultural technology, and in particular to a method and system for detecting the adsorption status of corn seeds in an air-suction seed metering device. Background Art

[0002] The corn no-till planter is an important achievement in the process of agricultural modernization. In no-till planting, the planter directly sows seeds into uncultivated soil, effectively saving arable land costs, reducing soil erosion, and increasing the content of soil organic matter.

[0003] As the core component of the seeder, the performance of the seed meter plays a decisive role in the sowing accuracy. At present, the seed meter is mainly divided into air suction type. Air suction seed meter is widely used because of its advantages such as no damage to seeds and strong adaptability. However, it is very easy to miss and re-seed during operation due to various factors such as air chamber pressure, machine structure and rotation speed. Missed seeding causes waste of land and seed resources, and re-seeding leads to uneven seed distribution. When the seedlings grow, they compete fiercely for nutrients, water and light, resulting in an imbalance in corn planting density, which seriously affects the growth and development of crops, and ultimately causes a decrease in corn yield and quality. Precision sowing is the key to improving corn yield and quality. During the sowing process, it is of great significance to grasp the adsorption status of corn seeds in the air suction seed meter in real time and discover missed and re-seeded problems in time, which is of great significance to improving the quality and efficiency of no-till sowing and ensuring the stable growth of corn yield.

[0004] In order to detect the seeding performance of the air-suction seed metering device, for example, the Chinese invention patent with publication number CN116649055A discloses a multi-row infrared seed drop detection system suitable for a no-till seeder, in which infrared seed drop detection sensors are installed at both ends of the bottom of the seed guide tubes of each row of the no-till seeder, and infrared transmitting and receiving ends are respectively provided to detect seed drop signals using infrared rays. The row missed seeding judgment module in the missed seeding judgment module receives the seed drop signal and determines the seed drop time, and transmits it to the seed drop detection controller for corresponding processing, such as updating the seed drop time interval, recording the seed drop time, displaying the missed seeding information, and controlling the abnormal alarm module to alarm. The technical solution recorded in this invention patent has a wide range of applicability. The missed seeding detection at the seed guide tube has a time delay, which makes it difficult to provide sufficient time for replanting operations.

[0005] For another example, a Chinese invention patent with publication number CN 118176889 A discloses an air-suction corn seed meter and a missed seeding reseeding method based on seed status recognition. A detection area is formed in the seed meter according to a through-beam photoelectric sensor. If the seeds in the chute are normal seeds, there is an obstruction in the detection area. When the seeds in the chute are damaged seeds, there is no obstruction in the detection area. The through-beam photoelectric sensor converts whether an obstruction occurs or not into an electrical signal and transmits it to a control unit to determine the type of seeds in the chute, thereby driving a stepper motor for reseeding.

[0006] However, the above two methods cannot accurately detect whether the seeds are damaged. For example, the damaged part of the damaged seeds is not large enough, the damaged seeds are large in size, or the damaged part of the damaged seeds is facing or facing away from the sensor when passing through. These situations will lead to detection errors. Summary of the invention

[0007] Technical problem solved by the present invention: The present invention provides a method system for detecting the adsorption status of corn seeds in an air-suction seed metering device, which solves the problem of inaccurate adsorption status of corn seeds in the existing air-suction seed metering device, thereby timely discovering missed seeding and reseeding problems.

[0008] The present invention solves the above technical problems by adopting a technical solution: a method for detecting the adsorption status of corn seeds in an air-suction seeding device, wherein the air-suction seeding device comprises a seeding disc, and the seeding disc is provided with a plurality of shaped holes, and the shaped holes are used to adsorb corn seeds. The method comprises the following steps:

[0009] S1. Install a camera above the seeding tray, collect the video of the upper surface of the seeding tray through the camera, and obtain the image corresponding to each frame;

[0010] S2, marking the targets in the image to obtain a marked image, wherein the targets include mold holes, single-grain complete corn seeds, multiple-grain corn seeds, and damaged corn seeds;

[0011] S3, dividing the annotated images into a training set and a test set;

[0012] S4. Establish an improved YOLOv8n model, wherein the improved YOLOv8n model includes introducing a C2fCIB convolution module in the neck network of the original YOLOv8n model to replace the original C2f module in the P5 layer, adopting a RepHead structure in the detection head to replace the original detection head, and adopting an internal intersection-and-union ratio and a shape intersection-and-union ratio to replace the original complete intersection-and-union ratio in the loss function;

[0013] S5. Using the training set to train the improved YOLOv8n model to obtain a target detection model, and using the test set to verify the target detection model;

[0014] S6. Use the target detection model to identify the upper surface video of the seed tray collected by the camera to obtain the adsorption status of the corn seeds for each sowing.

[0015] Furthermore, in S2, LabelImg is used to label the target in the image.

[0016] Furthermore, each sowing is determined by tracking the target in continuous frames of the video on the upper surface of the seed tray.

[0017] Furthermore, the adsorption conditions of corn seeds for each sowing include shaped holes, damaged corn seeds, single intact corn seeds and multiple corn seeds. If the adsorption condition of the corn seeds is shaped holes or damaged corn seeds, it is determined as missed sowing; if the adsorption condition of the corn seeds is a single intact corn seed, it is determined as normal sowing; if the adsorption condition of the corn seeds is multiple corn seeds, it is determined as re-sowing.

[0018] The beneficial effects of the present invention are as follows: the present invention provides a method and system for detecting the adsorption status of corn seeds in an air-suction seed metering device. A camera is installed above the seed metering disk, the upper surface video of the seed metering disk is collected by the camera, and the YOLOv8n model is improved. The improved YOLOv8n model includes introducing a C2fCIB convolution module in the neck network of the original YOLOv8n model to replace the original C2f module of the P5 layer, adopting a RepHead structure to replace the original detection head in the detection head, and adopting an internal intersection-and-union ratio and a shape intersection-and-union ratio to replace the original complete intersection-and-union ratio in the loss function. The improved YOLOv8n model is trained by using a training set to obtain a target detection model, and the target detection model is verified by using a test set. The target detection model is used to identify the upper surface video of the seed metering disk collected by the camera to obtain the adsorption status of the corn seeds for each sowing, thereby solving the problem of inaccurate adsorption status of corn seeds in the existing air-suction seed metering device, thereby timely discovering missed sowing and reseeding problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic flow chart of a method for detecting the adsorption status of corn seeds in an air-suction seed metering device provided by the present invention;

[0020] Figure 2 It is a structural diagram of the improved YOLOv8n model in the present invention, wherein Conv represents a convolutional layer, C2f represents a CSP bottleneck structure with two convolutions and feature fusion, SPPF represents fast spatial pyramid pooling, Concat represents a splicing layer, Upsample represents an upsampling layer, C2fCIB represents a convolutional module combining C2f and CIB, and CIB represents a cross-stage interaction module. DETAILED DESCRIPTION

[0021] Aiming at the problem that the adsorption status of corn seeds in the existing air-suction seeding device is inaccurate, the present invention provides a method for detecting the adsorption status of corn seeds in the air-suction seeding device. The air-suction seeding device comprises a seeding disc, and a plurality of shaped holes are arranged on the seeding disc. The shaped holes are used to adsorb corn seeds. The method, as shown in FIG. Figure 1 As shown, the following steps are included:

[0022] S1. Install a camera above the seed tray, collect the video of the upper surface of the seed tray through the camera, and obtain the image corresponding to each frame.

[0023] Specifically, to ensure the video effect captured by the camera, an industrial camera is adopted and installed in a camera box equipped with an observation port and a light source. The camera box is installed on the outer shell of the air-suction type seed metering device with an observation port. The industrial camera placed inside the camera box can collect the video of the upper surface of the seed plate through the observation port. To prevent dust from interfering with the camera lens, a glass sheet is used to seal the observation port to ensure that the image acquisition work is not affected by dust.

[0024] S2. Label the targets in the image to obtain the labeled image. The targets include type holes, single intact corn seeds, multiple corn seeds, and damaged corn seeds.

[0025] Specifically, LabelImg is used to label the targets in the image. LabelImg belongs to an existing open-source tool for image annotation.

[0026] S3. Divide the labeled image into a training set and a test set.

[0027] S4. Establish an improved YOLOv8n model. The improved YOLOv8n model includes introducing a C2fCIB convolution module in the neck network of the original YOLOv8n model to replace the original C2f module in the P5 layer, using a RepHead structure in the detection head to replace the original detection head, and using the internal intersection over union and shape intersection over union as the loss function to replace the original complete intersection over union.

[0028] Specifically, the P5 layer is the high-semantic feature layer. The structure of the improved YOLOv8n model is as Figure 2 shown. Conv represents the convolutional layer, which is used to extract image features; C2f represents the CSP bottleneck structure with two convolutions and feature fusion, which is used to reduce computational redundancy while enhancing the multi-branch feature fusion ability, thereby improving the efficiency and accuracy of the model; SPPF represents the fast spatial pyramid pooling, which is used to maintain the multi-scale context information fusion and enhance the robustness of the model to object size changes; Concat represents the concatenation layer, which is used to concatenate multiple feature maps together, and Upsample represents the upsampling layer; C2fCIB represents the convolution module combining C2f and CIB, and CIB represents the cross-stage interaction module. Compared with the C2f module, the C2fCIB convolution module is more accurate in global perception and local detail extraction, which helps to enhance the feature expression ability while improving the computational efficiency, and effectively improves the balance between the model accuracy and speed.

[0029] The original detection head has a weak ability to represent dynamic targets when dealing with fast-moving targets, which limits the detection efficiency. Therefore, the RepHead structure is adopted to replace the original detection head. The RepHead uses a decoupled head architecture to independently process the classification and regression tasks, enabling the model to learn corresponding features for different tasks more efficiently. The RepHead structure is built on the RepConv module as the basic building unit. The RepConv module uses the convolutional reparameterization strategy and adopts a multi-branch structure, including a direct connection branch, a 1×1 convolution branch, and a 3×3 convolution branch. Each branch is equipped with a batch normalization operation. This design greatly improves the feature extraction efficiency and enhances the model's adaptability to the complex characteristics of fast-moving targets. When entering the inference stage, through reparameterization means, the multi-branch structure is integrated into a single 3×3 convolution, achieving a significant improvement in detection efficiency while ensuring detection accuracy, and achieving the detection goal of high efficiency and high accuracy. After introducing the RepHead, the improved YOLOv8n model can more accurately and quickly identify and locate fast-moving targets in the fast-moving target detection task.

[0030] The loss function uses the internal intersection over union and the shape intersection over union to replace the original complete intersection over union, which better improves the model's ability to locate the target boundary and recognition accuracy.

[0031] S5. Train the improved YOLOv8n model using the training set to obtain a target detection model, and verify the target detection model using the test set.

[0032] Specifically, the resolution (image_size) of the input image is 640×640 pixels, the initial learning rate (learning_rate) is set to 0.01, the batch size (batch_size) is set to 16, and the number of iterations (epochs) is 100 to train the improved YOLOv8n model.

[0033] After comparison, the detection frame rate (FPS, frames per second) of the improved model reaches 303.03 frames per second, which is 139.096 frames per second higher than the original YOLOv8n model. The mean average precision mAP50 is stably maintained at 99.5%. Although the recall rate (R) has decreased by 3.9%, the improvement in precision (P) has reached 4%. This indicates that the model can still maintain a high detection accuracy while significantly improving the detection speed, and at the same time, the number of floating-point operations (GFLOPs, Giga Floating-point Operations Per Second) has not increased significantly, still meeting the lightweight requirements.

[0034] S6. Use the target detection model to identify the upper surface video of the seed tray collected by the camera to obtain the adsorption status of the corn seeds for each sowing.

[0035] Specifically, each sowing is determined by tracking the target in continuous frames in the video of the upper surface of the seed plate. Each sowing is from the time when the corn seeds appear in the image corresponding to a certain frame to the time when the corn seeds disappear in the image corresponding to the subsequent frame of the certain frame, or the corn seeds do not appear in a preset number of continuous frames. The existing YOLOv8n model has the function of tracking targets.

[0036] The adsorption conditions of corn seeds for each sowing include shaped holes, damaged corn seeds, single intact corn seeds and multiple corn seeds. If the adsorption condition of the corn seeds is shaped holes or damaged corn seeds, it is judged as missed sowing; if the adsorption condition of the corn seeds is a single intact corn seed, it is judged as normal sowing; if the adsorption condition of the corn seeds is multiple corn seeds, it is judged as re-sowing.

[0037] The present invention also provides a system for detecting the adsorption status of corn seeds in an air-suction seed meter, comprising a camera and a target detection model. The camera is installed above a seeding tray and is used to collect a video of the upper surface of the seeding tray. The target detection model is used to identify the video of the upper surface of the seeding tray collected by the camera to obtain the adsorption status of the corn seeds for each sowing. The target detection model is an improved YOLOv8n model after training. The improved YOLOv8n model includes introducing a C2fCIB convolution module in the neck network of the original YOLOv8n model to replace the original C2f module, adopting a RepHead structure in the detection head to replace the original detection head, and adopting an internal intersection-and-union ratio and a shape intersection-and-union ratio to replace the original complete intersection-and-union ratio in the loss function.

Claims

1. Detection method for adsorption condition of corn seeds in air-suction type seed metering device, the air-suction type seed metering device includes a seed metering disc, multiple type holes are arranged on the seed metering disc, and the type holes are used for adsorbing corn seeds, characterized in that, The method comprises the following steps: S1. Install a camera above the seeding tray, collect the video of the upper surface of the seeding tray through the camera, and obtain the image corresponding to each frame; S2, marking the targets in the image to obtain a marked image, wherein the targets include mold holes, single-grain complete corn seeds, multiple-grain corn seeds, and damaged corn seeds; S3, dividing the annotated images into a training set and a test set; S4. Establish an improved YOLOv8n model, wherein the improved YOLOv8n model includes introducing a C2fCIB convolution module in the neck network of the original YOLOv8n model to replace the original C2f module in the P5 layer, adopting a RepHead structure in the detection head to replace the original detection head, and adopting an internal intersection-and-union ratio and a shape intersection-and-union ratio to replace the original complete intersection-and-union ratio in the loss function; S5. Using the training set to train the improved YOLOv8n model to obtain a target detection model, and using the test set to verify the target detection model; S6. Use the target detection model to identify the upper surface video of the seed tray collected by the camera to obtain the adsorption status of the corn seeds for each sowing.

2. The method for detecting the adsorption state of corn seeds in the air-suction type seed metering device according to claim 1, wherein, In S2, LabelImg is used to label the objects in the image.

3. The method for detecting the adsorption state of corn seeds in the air-suction type seed metering device according to claim 1, characterized in that, Each sowing is determined by tracking the target in consecutive frames of the video on the upper surface of the seed disk.

4. The method for detecting the adsorption state of corn seeds in a pneumatic suction type seed metering device according to claim 1, characterized in that, The corn seed adsorption conditions of each sowing include shaped holes, damaged corn seeds, single intact corn seeds and multiple corn seeds. If the corn seed adsorption condition is shaped holes or damaged corn seeds, it is determined as missed sowing; If the corn seed adsorption condition is a single complete corn seed, it is judged as normal sowing; if the corn seed adsorption condition is multiple corn seeds, it is judged as re-sowing.

5. Detection system for adsorption state of corn seeds in air-suction type seed metering device, characterized in that, The invention comprises a camera and a target detection model. The camera is installed above a seeding disk and is used to collect the video of the upper surface of the seeding disk. The target detection model is used to identify the video of the upper surface of the seeding disk collected by the camera to obtain the adsorption status of corn seeds sown each time. The target detection model is an improved YOLOv8n model after training. The improved YOLOv8n model comprises introducing a C2fCIB convolution module in the neck network of the original YOLOv8n model to replace the original C2f module, adopting a RepHead structure in the detection head to replace the original detection head, and adopting an internal intersection-and-union ratio and a shape intersection-and-union ratio to replace the original complete intersection-and-union ratio in the loss function.

Images