Pneumatic precision reseeding intelligent system based on visual recognition

Abstract

This invention relates to the field of visual recognition and reseeding, and discloses an intelligent system for precise pneumatic reseeding of missing seeds in seed trays based on visual recognition. The system includes: an image acquisition module for real-time image acquisition of the adsorption holes of the seeding roller; a speed detection module for detecting the speed of the conveyor belt; an industrial control computer for matching the current speed to a preset speed level and calling the corresponding pre-stored image acquisition timing parameters to control image acquisition with a lead time trigger; an image processing module for outputting the binarized recognition results of each adsorption hole based on a deep learning model; a control module for generating reseeding instructions based on the recognition results; and a pneumatic reseeding execution module, which includes multiple independent pneumatic reseeding units corresponding one-to-one with the adsorption holes, to perform reseeding for the missing seed holes. This invention achieves precise synchronization through speed discretization and pre-stored timing, eliminates coordinate mapping through physical equal interval design, and achieves real-time accurate recognition through a lightweight deep learning model, making it suitable for seed tray seedling cultivation.

Description

Technical Field

[0001] This invention relates to the field of visual recognition and reseeding, specifically to an intelligent system for precise pneumatic reseeding of seed trays based on visual recognition. Background Technology

[0002] Precision seeding technology is a core component of modern facility agriculture. It involves setting up adsorption holes on a seeding drum that correspond one-to-one with the holes in the seed tray. Negative pressure is used to adsorb the seeds onto the drum surface, and as the drum rotates, the seeds are sown into the holes of the seed tray. However, due to factors such as uneven seed size, clogged adsorption holes, and fluctuations in negative pressure, missed sowing (seeds not being adsorbed by the adsorption holes) is unavoidable in actual operations. This results in some holes in the seed tray having missing seedlings, directly affecting seedling efficiency and the uniformity of subsequent transplanting operations.

[0003] To address the aforementioned issue of missed seeding, various machine vision-based methods for detection and reseeding have been proposed in existing technologies. For example, an industrial camera is used to photograph the adsorption area of ​​the seeding roller, and image processing is used to determine whether seeds are present in each adsorption hole. The corresponding seed tray locations for missing seeds are then marked or reseeded. However, existing solutions still have the following technical shortcomings in practical applications: Firstly, existing solutions typically use continuous speed measurement combined with real-time calculation to determine the time of taking pictures and the time of re-reeling. However, due to factors such as conveyor belt speed fluctuations, industrial control computer calculation delays, and communication transmission time, there is a tendency for deviations to occur between the time of image acquisition and the time when the adsorption hole actually reaches the center of the camera's field of view, resulting in the spatial position of the adsorption hole in the image not accurately corresponding to the recognition result.

[0004] Secondly, the coordinates of the adsorption pores detected by the vision system need to be transformed into physical space to drive the reseeding actuator. Existing technologies typically rely on camera calibration and coordinate transformation algorithms to establish a mapping relationship. Such algorithms have stringent requirements for installation accuracy and require tedious calibration operations before the system is put into operation. Once the equipment vibrates or shifts, it needs to be recalibrated, resulting in high maintenance costs and making it unsuitable for large-scale agricultural production deployment.

[0005] Third, seedlings in plug trays, especially solanaceous seedlings, have slender, irregularly shaped stems and leaves, and their postures at the adsorption pores vary, making them typical targets for micro-target detection. Existing technologies sometimes use morphological or manual contour features to determine the presence of seedlings, which are sensitive to changes in lighting and dust interference, resulting in high false positive and false negative rates. Other solutions directly transplant general target detection models, such as YOLOv5 and Faster-R-CNN, but these models have large parameter sets and slow inference speeds, making real-time processing on industrial control CPU platforms difficult. Furthermore, the default large input size of general models, such as 640×640, results in significant invalid padding for images of slender seedlings, affecting feature extraction efficiency.

[0006] In summary, existing precision seeding missed seeding detection and reseeding systems still have some shortcomings in terms of timing synchronization accuracy, ease of coordinate mapping, and the balance between model real-time performance and detection accuracy. There is an urgent need for an intelligent reseeding solution that can achieve high-speed, high-precision, and low-complexity deployment. Summary of the Invention

[0007] The purpose of this invention is to provide a visual recognition-based intelligent system for precise pneumatic reseeding of missing seeds in seed trays, thereby solving at least one of the above-mentioned technical problems.

[0008] The objective of this invention can be achieved through the following technical solutions: The vision-recognition-based intelligent system for precise pneumatic reseeding of seed trays includes: The image acquisition module is used to acquire real-time images of each adsorption hole on the seeding drum and obtain raw image data. The image processing module is used to classify and identify whether each adsorption pore in the original image data has a seed based on a deep learning target detection model, and outputs a binary recognition result array corresponding to each adsorption pore. The control module is used to receive the binarized recognition result array and generate replay control instructions; The pneumatic reseeding execution module includes multiple sets of independently controllable pneumatic reseeding units that correspond one-to-one with each adsorption hole position. Each pneumatic reseeding unit responds to the corresponding sub-instruction in the reseeding control command and performs a reseeding action for the hole position in the seed tray corresponding to the missing adsorption hole position. The image acquisition module performs image acquisition at a trigger time determined by pre-stored shooting timing parameters corresponding to multiple discrete speed levels. The trigger time has an advance relative to the time when the acupoint reaches the center of the field of view of the image acquisition module, and different speed levels correspond to different shooting timing parameters.

[0009] A further embodiment of the proposed solution is a lightweight YOLOv8n model for object detection. The YOLOv8n model employs a three-segment network structure, specifically including: The Backbone network, based on an improved CSPDarknet structure, includes a C2f module and an SPPF fast spatial pyramid pooling layer, which are used to downsample the input image by 8x, 16x and 32x respectively, and output feature maps of different scales in three layers: P3, P4 and P5. The Neck multi-scale fusion network adopts a PAN-FPN bidirectional feature pyramid structure, and all deep and shallow features in the PAN-FPN structure are fused and interacted using C2f modules. The decoupled detection head adopts an anchor-free design, completely separating the classification branch and the coordinate regression branch, and independently predicting the class confidence and bounding box coordinate distribution respectively; The input image size of the YOLOv8n model is a preset small size, which is smaller than the conventional input size of general object detection models. The YOLOv8n model enables a rect rectangular training strategy to adapt to the non-square and slender image features of plug seedlings, thereby reducing invalid calculations and feature interference caused by image filling.

[0010] In a further embodiment, the image processing module is also used to extract edge features from the original image data using the Canny edge detection algorithm before inputting the original image data into the YOLOv8n model, thereby generating an edge feature map; The edge feature map and the original image data are input together into the YOLOv8n model for recognition; In this system, the recognition result of the YOLOv8n model serves as the decision-making entity, while the output of the Canny edge detection algorithm serves as an auxiliary reference feature.

[0011] In a further embodiment, the deep learning target detection model identifies seedling-free adsorption pores based on the following criteria: When no seeds are adsorbed at the adsorption pore, the pore has a regular circular outline; when seeds are adsorbed at the pore, the pore has an irregular shape.

[0012] In a further embodiment, the image acquisition module and the pneumatic seeding execution module are arranged sequentially along the conveyor belt conveying direction. The camera coordinate system of the image acquisition module, the physical coordinates of each hole position in the seeding tray, and the cylinder coordinates of the pneumatic seeding execution module are aligned through an equally spaced physical structure design. Each pneumatic seeding unit of the pneumatic seeding execution module and each adsorption hole position of the seeding roller and each hole position in the seeding tray have a one-to-one equally spaced mapping relationship in physical position, without the need for coordinate mapping transformation at the software level.

[0013] In a further embodiment, the control module includes an industrial computer and an embedded control board; The industrial control computer is communicatively connected to the image acquisition module and is used to receive the raw image data and run the image processing module; The embedded control board is communicatively connected to the industrial computer and is used to receive the binarized recognition result array and generate the replay control command. The embedded control board is also electrically connected to the control terminal of each pneumatic re-seeding unit of the pneumatic re-seeding execution module, and is used to send the re-seeding control command to each pneumatic re-seeding unit.

[0014] In a further embodiment, the industrial control computer is also used for: The corresponding photo-taking sequence is matched according to the current running speed of the conveyor belt, and the image acquisition module is controlled to acquire images according to the photo-taking sequence. The image acquisition module uses an industrial camera, which responds to the photo-taking command of the industrial control computer via a soft trigger to acquire images and uploads the acquired BGR format image data to the industrial control computer.

[0015] In a further embodiment, the image processing module runs the deep learning object detection model's recognition process, which includes: The original image data is cropped and then input into the deep learning target detection model for forward inference. The deep learning target detection model outputs the category prediction results for each adsorption pore site, including seedling category and no seedling category. The category prediction results of each adsorption pore site are converted into the binary recognition result array and returned for storage, so as to be called in subsequent replay.

[0016] In a further embodiment, the image processing module generates a recognition failure flag for adsorption wells where the recognition result is failed, and the industrial control computer stops generating and issuing replanting instructions for the adsorption wells based on the recognition failure flag, so as to implement a no-replanting strategy for missed seeds. When a communication timeout occurs between the industrial control computer and the embedded control board, or when the pneumatic re-seeding execution module experiences cylinder jamming, the unfinished re-seeding task corresponding to the current hole position is abandoned, and the re-seeding control process for the next hole position is reset.

[0017] In a further embodiment, the system also includes a light source environment control module, which is used to maintain a constant illumination in the image acquisition environment. The image acquisition module acquires images of each adsorption hole on the seeding roller under preset constant temperature and preset constant illumination conditions.

[0018] A further proposed approach is to use the following loss function for the YOLOv8n model: Loss(total)=λ(box)×Loss(CIoU)+λ(cls)×Loss(BCE)+λ(dfl)×Loss(DFL); Wherein, Loss(CIoU) is the CIoU bounding box regression loss, used to constrain the overlap area, center distance and aspect ratio of the predicted box and the ground truth box; Loss(BCE) is the binary cross-entropy classification loss, used for multi-class seedling identification; Loss(DFL) is the distribution focus loss, used to predict the probability distribution of discrete coordinate intervals to improve the positioning accuracy of small targets, and λ(box), λ(cls), and λ(dfl) are the preset weight coefficients of the corresponding loss functions.

[0019] The beneficial effects of this invention are: (1) This invention obtains the conveyor belt speed in real time by setting a speed detection module. The industrial control computer discretizes the continuous speed value into multiple preset speed ranges and directly calls the pre-stored photo-taking timing parameters according to the matched current speed range. It controls the image acquisition module to take pictures at a trigger time that has an advance relative to the time when the adsorption hole reaches the center of the camera's field of view. Through the synchronization mechanism of speed discretization and pre-stored timing parameter lookup, the cumulative error and communication delay caused by continuous speed measurement and real-time calculation are reduced, ensuring that the image acquisition time always accurately corresponds to the time when the adsorption hole actually reaches the center of the camera's field of view under the condition of conveyor belt speed fluctuation. At the same time, the photo-taking timing parameters switch with the speed range, so there is no need to recalibrate or calculate online when the system speed is adjusted, which greatly improves the dynamic response capability and engineering applicability of the system.

[0020] (2) The present invention adopts the core concept of physical alignment to replace the algorithm target. The image acquisition module and the pneumatic reseeding execution module are arranged sequentially along the conveyor belt direction, and the pneumatic reseeding unit of the pneumatic reseeding execution module, the adsorption hole of the seeding roller, and the hole of the seed tray are physically corresponding to each other at equal intervals. This reduces the complex process of camera calibration, coordinate transformation and pixel-physical space conversion in the traditional solution. After the system is installed, there is no need to perform cumbersome coordinate calibration operations. After the equipment vibrates or the position is finely adjusted, there is no need to recalibrate. This can reduce the difficulty of on-site deployment and long-term maintenance costs. It is especially suitable for harsh environments such as vibration and dust in agricultural sites.

[0021] (3) The present invention uses the YOLOv8n lightweight model as the decision-making body. By limiting the input image size to a preset small size and enabling the rect rectangular training strategy, it effectively adapts to the non-square and slender image features of the seedlings in the plug tray, reducing invalid filling calculations. At the same time, the Canny edge detection algorithm is used to extract edge feature maps, which are input into the YOLOv8n model together with the original image data to participate in recognition. The inherent physical prior of the adsorption holes without seedlings presenting a regular circle due to the physical structure of the air nozzle, and presenting an irregular shape when there are seedlings, is utilized to enhance the model's ability to perceive the features of tiny seedlings.

[0022] (4) The present invention sets up differentiated processing logic for abnormal working conditions such as identification failure, communication timeout, and cylinder jamming: For adsorption holes that fail to be identified, a failure mark is generated and the replanting command is stopped, and the missing seeding does not replanting strategy is implemented to reduce damage to adjacent qualified holes caused by incorrect replanting; For communication timeout or cylinder jamming, the current unfinished replanting task is abandoned and the system is reset to enter the next replanting cycle to ensure that single point failure does not affect the continuous operation of the overall system. Attached Figure Description

[0023] The invention will now be further described with reference to the accompanying drawings.

[0024] Figure 1 This is a system structure block diagram of the present invention; Figure 2 These are actual test images of test sample 1; Figure 3 These are actual test images of test sample 2; Figure 4 The actual test image is for test sample 3; Figure 5 The actual test image is for test sample 4; Figure 6 The image shown is from the actual test sample 5. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figure 1 As shown, the intelligent pneumatic precision reseeding system for seed trays based on visual recognition provided by this invention comprises an image acquisition module, a seeding roller, and a pneumatic reseeding execution module arranged sequentially along the conveyor belt conveying direction. The surface of the seeding roller is evenly distributed with several adsorption holes, each connected to a vacuum generator via a negative pressure channel, used to adsorb seeds from the hopper during the rotation of the seeding roller. The seed tray is located on the conveyor belt and moves along the conveying direction with the belt. The seed tray has seed tray holes that correspond one-to-one in number and position to the adsorption holes on the seeding roller. The seeding roller sows the adsorbed seeds into the corresponding seed tray holes as it rotates.

[0027] An image acquisition module is positioned in front of the seeding roller to capture real-time images of each adsorption hole on the seeding roller after the seeding roller has adsorbed the seeds and before the seeds are sown into the seed trays. A pneumatic reseeding module is positioned behind the image acquisition module to reseed any missed seeding holes in the seed trays after the seeding roller has finished sowing. A speed detection module is also installed on the conveyor belt to monitor its operating speed in real time.

[0028] The system also includes an industrial computer and an embedded control board. The industrial computer is communicatively connected to the image acquisition module, the speed detection module, and the embedded control board, respectively. The embedded control board is electrically connected to the control terminals of each pneumatic reseeding unit of the pneumatic reseeding execution module.

[0029] Under continuous conveyor belt operation, each hole in the seeding tray sequentially passes through the seeding roller, image acquisition module, and pneumatic supplementary seeding execution module. To ensure that the image acquisition module can accurately capture images of the designated adsorption holes, the triggering time of image acquisition must be strictly controlled.

[0030] In this embodiment, the speed detection module uses a rotary encoder, which is installed on the drive roller of the conveyor belt to detect the current running speed of the conveyor belt in real time and output the speed signal to the industrial control computer.

[0031] The industrial control computer pre-stores multiple speed ranges and corresponding image capture timing parameters for each speed range. The speed ranges are obtained by discretizing the continuous operating speed range of the conveyor belt. For example, the first speed range corresponds to 0-30 mm / s, the second speed range to 30-60 mm / s, the third speed range to 60-90 mm / s, and so on. Each speed range corresponds to a set of pre-stored image capture timing parameters, which include an advance trigger amount relative to the moment when the hole position 4 of the image acquisition module reaches the center of the field of view.

[0032] After receiving the current operating speed from the speed detection module, the industrial control computer matches the current operating speed with multiple preset speed ranges to determine the current speed range. Then, it calls the pre-stored image capture timing parameters corresponding to that speed range. Finally, it sends a trigger signal to the image acquisition module according to these timing parameters, controlling the image acquisition module to perform image acquisition at a trigger moment that is ahead of the time the image reaches the center of the image acquisition module's field of view relative to the hole position. When a change in conveyor belt speed causes a speed range switch, the industrial control computer automatically switches and calls the corresponding image capture timing parameters.

[0033] Preferably, the industrial computer and the image acquisition module are connected via a GPIO interface. The industrial computer sends a soft trigger signal to the image acquisition module through the GPIO interface according to the shooting timing parameters. Before sending the soft trigger signal, the industrial computer also checks the level status of the Pin14-GPIO port. When a high level is detected for a sustained duration greater than 200ms, the trigger condition is confirmed to be met, and the soft trigger is executed. The image acquisition module uses an industrial camera, preferably a Hikvision industrial camera, which responds to the soft trigger signal to perform image acquisition and uploads the acquired image data to the industrial computer.

[0034] Under the control of the industrial computer, the image acquisition module takes pictures of each adsorption hole on the seeding roller, acquiring raw image data containing each adsorption hole and its corresponding seed. Preferably, the image acquisition module performs image acquisition under preset constant temperature and light conditions. The system also includes a light source environment control module to maintain constant light intensity in the image acquisition environment, preventing light changes from interfering with subsequent image recognition. The light source environment control module preferably uses a ring-shaped LED light source in conjunction with a constant current drive power supply to ensure that illuminance fluctuations are controlled within ±5%.

[0035] After the image acquisition module uploads the acquired raw image data to the industrial control computer, the industrial control computer runs the image processing module to process the raw image data.

[0036] The image processing module first preprocesses the raw image data. The preprocessing steps include: Image cropping: Based on the distribution of each adsorption hole on the seeding drum, the original image data is cropped to extract the effective area image containing all adsorption holes, in order to remove background interference and reduce the amount of subsequent calculations.

[0037] Image Enhancement: Before inputting the original image data into the deep learning object detection model, the Canny edge detection algorithm is used to extract edge features from the original image data, generating an edge feature map. The edge feature map generation process is as follows: Gaussian filtering is applied to the original image data for noise reduction, then the gradient magnitude and gradient direction of each pixel in the image are calculated, and finally, non-maximum suppression and double threshold detection are used to extract the edge contour information in the image, forming an edge feature map.

[0038] Preferably, the high threshold of the Canny edge detection algorithm is set to 150, and the low threshold is set to 50. The edge point determination rule is as follows: if the gradient magnitude of a pixel is higher than the high threshold, it is determined to be a strong edge point and retained; if it is lower than the low threshold, it is discarded; if it is between the high and low thresholds, it is retained only if the pixel is adjacent to a strong edge point, otherwise it is discarded. The above thresholds can be adaptively adjusted according to the actual image contrast. The adjustment principle is to ensure that the overlap between the detected edge contour and the actual edge of the adsorption hole and its seed reaches the preset standard.

[0039] Data fusion: The original image data and the edge feature map generated by the Canny edge detection algorithm are concatenated channel by channel and used together as input data for the deep learning object detection model. Specifically, the original image data is a three-channel RGB image, and the edge feature map is a single-channel grayscale image. The two are concatenated to form four-channel input data, which is then input into the deep learning object detection model for forward inference.

[0040] In this embodiment, the deep learning target detection model is preferably the lightweight YOLOv8n model. This model is a pre-trained neural network model used to classify and identify whether each adsorption pore 11 has a seed or not.

[0041] The specific model structure is as follows: The YOLOv8n model adopts a three-segment network structure, specifically including a backbone network, a Neck multi-scale fusion network, and a decoupled detection head.

[0042] The Backbone backbone network is an improved structure based on CSPDarknet, including a C2f module and a SPPF fast spatial pyramid pooling layer. The input image first passes through the Backbone backbone network for feature extraction, undergoing downsampling at 8x, 16x, and 32x sequentially, corresponding to three feature maps of different scales: P3, P4, and P5. The P3 feature map has the highest resolution and smallest receptive field, used for detecting small targets, such as tiny eggplant seedlings; the P5 feature map has the lowest resolution and largest receptive field, used for detecting large targets; and the P4 feature map falls between the two. The C2f module reduces the number of parameters while maintaining feature extraction capabilities through cross-layer connections and feature reuse.

[0043] The Neck multi-scale fusion network employs a PAN-FPN bidirectional feature pyramid structure, with all deep and shallow features fused using C2f modules. The PAN-FPN structure fuses the P3, P4, and P5 feature maps output from the backbone network at multiple scales through top-down and bottom-up paths, ensuring that each feature map contains both deep and shallow feature information, thus enhancing the model's ability to detect targets at different scales.

[0044] The decoupled detection head employs an anchor-free design, completely separating the classification branch from the coordinate regression branch, which independently predicts the category confidence score and bounding box coordinate distribution. The classification branch outputs the confidence score of each adsorption well site belonging to the seedling-bearing or seedless category; the coordinate regression branch outputs the center coordinates and width and height of the bounding box.

[0045] Preferably, the YOLOv8n model has a total of 224 network layers. Specifically, the backbone network contains 80 layers, the Neck multi-scale fusion network contains 72 layers, and the decoupled detection head contains 72 layers.

[0046] Preferably, the input image size of the YOLOv8n model is a preset small size, which is smaller than the conventional input size of general object detection models. The conventional input size is usually 640×640 pixels, for example, in this embodiment, the preset small size is 250×250 pixels. Using a preset small input size can significantly reduce the amount of computation during model inference and adapt to the computing power limitations of industrial control computer CPU platforms. At the same time, the YOLOv8n model enables a rect rectangular training strategy, that is, during the training process, the image is not forcibly scaled to a square, but the original aspect ratio of the image is maintained, and only the short side is filled to the preset size to adapt to the non-square and slender image features of plug seedlings, reducing the invalid computation and feature interference caused by image filling.

[0047] The model training process is as follows: The YOLOv8n model is pre-trained using the following steps: Step S1, Sample Collection and Labeling: Collect image samples of the adsorption holes on the seeding roller. The samples include two categories: those with seedlings and those without, for a total of 3000 sample images. Label each adsorption hole in each sample image, including the location coordinates of the adsorption hole and the category label (with seedlings / without seedlings).

[0048] Step S2, Dataset Partitioning: Divide the labeled sample images into training set, validation set and test set according to the ratio, for example, in an 8:1:1 ratio.

[0049] Step S3, Model Initialization: Load the pre-trained weights of YOLOv8n and train using a transfer learning strategy. The optimizer uses SGD stochastic gradient descent, with an initial learning rate set to 0.001 and a final learning rate decayed to 0.001.

[0050] Step S4, Model Training: Use the `model.train()` interface, set the training epochs to 100, the batch size to 16, and enable validation in each epoch. Data augmentation strategies used during training include: low-amplitude HSV color enhancement, and HSV... h =0.1, hsv s =0.2, hsv v =0.2, automatic randaugment enhancement, disabling mosaic enhancement. Image input size set to 250×250 pixels, enabling rectangle training.

[0051] The data augmentation strategies employed during training included: low-amplitude HSV color augmentation, automatic randaugment augmentation, and disabling mosaic augmentation. Specifically, HSV color augmentation randomly perturbs the training images across the three dimensions of color space: hue, saturation, and value, to enhance the model's robustness to different lighting conditions and color deviations. h hsv represents the maximum enhancement value in the hue dimension. s hsv represents the maximum enhancement value in the saturation dimension. v The maximum enhancement value for the brightness dimension is indicated. In this embodiment, the three values ​​are preferably set to 0.1, 0.2, and 0.2, respectively. Mosaic enhancement is a data augmentation method that randomly crops four training images and stitches them into a single synthetic image for training. In this embodiment, mosaic enhancement is turned off to avoid excessive truncation and distortion of the adsorption pores and their boundaries in the stitched image, thereby maintaining the integrity and authenticity of the original morphological characteristics of the seedlings.

[0052] Step S5, Loss Function Optimization: The YOLOv8n model is optimized using a three-term weighted loss function, and the total loss is expressed as: Loss(total)=λ(box)×Loss(CIoU)+λ(cls)×Loss(BCE)+λ(dfl)×Loss(DFL); Among them, Loss(CIoU) is the CIoU bounding box regression loss, which simultaneously constrains the overlap area, center distance, and aspect ratio of the predicted box and the ground truth box to improve the regression accuracy of irregular seedling outlines; Loss(BCE) is the binary cross-entropy classification loss, used for binary classification of seedlings / no seedlings; Loss(DFL) is the distribution focus loss, used to predict the probability distribution of discrete coordinate intervals, transforming the prediction of bounding box coordinates into probability distribution prediction, effectively improving the localization accuracy of small targets such as seedling growth points and stem and leaf tips. λ(box), λ(cls), and λ(dfl) are the preset weight coefficients for the corresponding loss functions. In this embodiment, they are preferably set to λ(box) = 7.5, λ(cls) = 0.5, and λ(dfl) = 1.5.

[0053] Step S6, Early Stopping and Model Saving: During training, monitor the validation set metrics. If the validation set metrics do not improve for 50 consecutive rounds, automatically terminate training. After training, save the optimal weights and the last set of weights.

[0054] The model inference process is as follows: In actual operation, the image processing module inputs the preprocessed image data—that is, the four-channel data resulting from the concatenation of the original image data and the edge feature map—into the trained YOLOv8n model for forward inference. The YOLOv8n model outputs the category prediction results for each adsorption pore site, including the seedling-bearing category and the seedless category, along with their corresponding confidence scores.

[0055] In this embodiment, the recognition result of the YOLOv8n model serves as the decision-making body. The output of the Canny edge detection algorithm serves as an auxiliary reference feature, which enhances the model's ability to perceive the features of adsorption pore edges and seed contours, but is not used independently as a judgment criterion. When the recognition result of the YOLOv8n model and the output of the Canny edge detection algorithm are logically inconsistent, the recognition result of the YOLOv8n model shall prevail.

[0056] The criteria for determining whether a seed-free adsorption hole is present are as follows: when no seed is adsorbed at the adsorption hole, the nozzle structure exposed on the roller surface is a regular circular outline due to manufacturing tolerances, appearing as a closed circular edge in the image; when a seed is adsorbed at the adsorption hole, the seed covers the nozzle, and its outline is an irregular elliptical or random shape, significantly different from the regular circular nozzle outline. The YOLOv8n model automatically learns this morphological difference feature during training, thereby achieving accurate classification of seed presence or absence.

[0057] A core concept of this invention is to replace software-level coordinate mapping transformation with a physically equidistant structural design, thereby reducing the cumbersome process of camera calibration and coordinate transformation in traditional solutions.

[0058] Specifically, the camera coordinate system of the image acquisition module, the physical coordinates of each hole in the seed tray, and the cylinder coordinates of each pneumatic seeding unit in the pneumatic seeding execution module are all mapped at equal intervals through mechanical structure design. That is, the adsorption holes on the seeding roller are evenly spaced along the axial direction of the seeding roller, the holes on the seed tray are evenly spaced along the transverse direction, and the pneumatic seeding units of the pneumatic seeding execution module are evenly spaced along the transverse direction, with all three intervals being equal. Simultaneously, the installation positions of the image acquisition module and the seeding roller, as well as the installation positions of the pneumatic seeding execution module and the seeding roller, are precisely fixed using positioning fixtures to ensure that the relative positions of each component are calibrated at the factory.

[0059] Due to the aforementioned physically equidistant structural design, each pneumatic reseeding unit of the pneumatic reseeding execution module has a one-to-one, equidistant mapping relationship with each adsorption hole of the seeding roller and each hole of the seed tray. Therefore, in the binary recognition result array output by the image processing module, the i-th bit naturally corresponds to the i-th adsorption hole, the i-th seed tray hole, and the i-th pneumatic reseeding unit, without the need for software algorithms to perform pixel coordinate to physical coordinate mapping transformation. When reseeding is required, the embedded control board directly sends the reseeding command corresponding to the i-th recognition result to the i-th pneumatic reseeding unit.

[0060] Preferably, the image acquisition module and the pneumatic reseeding execution module are arranged sequentially along the conveyor belt conveying direction, and the distance between them is an integer multiple of the hole spacing of the seed tray, so as to ensure that the correspondence between each seed tray hole and each pneumatic reseeding unit remains unchanged during the process of moving from the image acquisition module to the pneumatic reseeding execution module.

[0061] The image processing module converts the category prediction results of each adsorption site output by the YOLOv8n model into a binary recognition result array and returns it for storage. The length of the binary recognition result array is the same as the total number of adsorption sites, and each bit in the array is used to characterize whether a seed exists at the corresponding adsorption site. Preferably, 1 in the array represents the presence of a seedling, and 0 represents the absence of a seedling.

[0062] The industrial control computer sends the binarized recognition result array to the embedded control board. Preferably, the industrial control computer and the embedded control board are connected via a serial communication interface (UART) with a communication baud rate set to 115200bps. The data frame format sent from the industrial control computer to the embedded control board is as follows: the first byte is the frame header (fixed to 0xAA), the second byte is the data length (equal to the total number of adsorption holes), the third to the second-to-last byte are the bit-by-bit data of the binarized recognition result array, each bit occupies 1 byte, 0x01 indicates seedlings present, 0x00 indicates no seedlings present, 0xFF indicates recognition failure, and the last byte is the check bit, using XOR check or CRC8 check. After receiving the complete data frame, the embedded control board first performs frame header check and CRC check. If the check passes, it returns an acknowledgment byte (0x55) to the industrial control computer. Upon receiving the acknowledgment byte, the industrial control computer confirms successful transmission. If the check fails or the acknowledgment byte is not received within a timeout period, the industrial control computer automatically retransmits the current data frame. After the number of retransmissions reaches a preset threshold (preferably 3 times), it is determined that the communication has timed out.

[0063] After receiving the binary recognition result array, the embedded control board checks bit by bit whether there is a missing seed in each hole of the seed tray. If the data of the i-th bit is 0, it means that no seed was picked up in the i-th adsorption hole, then there is a missing seed in the i-th seed tray hole, and the embedded control board generates a reseeding sub-instruction for the i-th pneumatic reseeding unit.

[0064] The pneumatic reseeding execution module includes multiple independently controllable pneumatic reseeding units, each corresponding to a specific number and position of the adsorption holes on the seeding roller. Each pneumatic reseeding unit is independently controlled by an embedded control board.

[0065] Each pneumatic reseeding unit includes a seed storage chamber, a seed dispensing channel, pneumatic nozzles, and a solenoid valve. The seed storage chamber, located at the top of the unit, is used to pre-store seeds for reseeding, and has a seed inlet at its bottom. The seed dispensing channel is vertically positioned below the seed storage chamber, with its upper end connected to the seed inlet and its lower end connected to the seed inlet of the pneumatic nozzle. The air inlet of the pneumatic nozzle is connected to an air source via an air supply line, which is equipped with a solenoid valve. The control terminal of the solenoid valve is electrically connected to an embedded control board. The outlet of the pneumatic nozzle faces downwards, directly opposite the seed tray holes on the conveyor belt.

[0066] Each pneumatic reseeding unit is isolated from the others by partitions to prevent seeds from splashing into adjacent seed trays and causing overseeding during reseeding. A rotatable seed metering wheel is located at the bottom of the seed storage chamber, with a receiving groove on its outer circumference adapted to the seed size. The seed metering wheel is driven to rotate by a stepper motor. When the solenoid valve receives an opening command, the seed metering wheel rotates one position synchronously, releasing a single seed from the receiving groove into the seed metering channel. Simultaneously, compressed gas is ejected through a pneumatic nozzle, blowing the seed into the corresponding seed tray hole, completing precise single-seed reseeding.

[0067] Preferably, the pneumatic pressure of the pneumatic reseeding execution module is set to 0.5MPa to 0.6MPa, and the conveyor belt speed is set to 50mm / s. Under this parameter combination, the system can achieve the best reseeding accuracy and efficiency.

[0068] This embodiment sets up differentiated processing logic for various abnormal operating conditions.

[0069] The failure handling process is as follows: When the image processing module runs the YOLOv8n model for inference, if the category prediction result of a certain adsorption pore site cannot reach the preset confidence threshold (e.g., confidence level below 0.25) due to poor image quality, abnormal lighting, or low model confidence, the image processing module will generate a recognition failure flag for that adsorption pore site (e.g., return -1).

[0070] Based on the identification failure flag, the industrial control computer stops generating and issuing reseeding instructions for the adsorption hole. This means it doesn't perform any reseeding action on the corresponding seed tray hole, implementing a "no reseeding for missed seeds" strategy. The core logic of this strategy is: it's better to forgo the opportunity to reseed a missed hole than to mistakenly reseed a seed in a qualified hole that already contains seeds, thus avoiding damage to adjacent qualified seedlings or waste of seed resources due to incorrect reseeding.

[0071] Meanwhile, the industrial control computer marks the location information of all failed hole positions on the human-machine interface of the monitor with a highlighted color or a flashing pattern, and generates a log file containing the hole number, hole number, and failure timestamp. This log file is used by operators at the end of the production line for manual secondary inspection and replanting of the marked holes. The log file is stored locally on the industrial control computer and can be uploaded to the database of the production management system via the network.

[0072] The communication timeout handling process is as follows: When the communication link between the industrial control computer and the embedded control board times out, that is, after the industrial control computer sends out the binary recognition result array, it does not receive a reception confirmation signal from the embedded control board within a preset time threshold, or the embedded control board does not receive data from the industrial control computer within a preset time threshold, then a communication timeout is determined to have occurred.

[0073] At this point, the system abandons the unfinished re-seeding task corresponding to the current hole position on the seed tray, and the embedded control board resets to standby mode, awaiting the next re-seeding control command from the industrial control computer. Simultaneously, the industrial control computer records a communication timeout log to facilitate troubleshooting by maintenance personnel.

[0074] The process for handling cylinder jamming is as follows: When a pneumatic re-seeding unit of the pneumatic re-seeding execution module experiences cylinder jamming, i.e., the pneumatic nozzle fails to complete the jetting action within a preset time after the solenoid valve is opened, or the cylinder piston cannot reciprocate normally, the embedded control board detects the abnormality through the position sensor or pressure sensor on the cylinder.

[0075] Upon detecting cylinder jamming, the system immediately abandons the unfinished reseeding task corresponding to the current seed tray position and resets to begin the reseeding control process for the next seed tray position, preventing the continuous operation of the entire production line from being blocked due to a single point of failure. Simultaneously, the embedded control board uploads the cylinder jamming fault information to the industrial control computer, which then issues an audible and visual alarm to prompt maintenance personnel to handle the issue.

[0076] The complete workflow of the system of this invention is described below: Step S101: After the system starts, the industrial computer, embedded control board, image acquisition module and pneumatic re-seeding execution module complete the initialization self-test in sequence. In step S102, the seeding drum rotates and adsorbs seeds from the hopper. The adsorption holes with seeds adsorbed continue to rotate, while the adsorption holes without seeds adsorbed are empty. Step S103: The speed detection module detects the current running speed of the conveyor belt in real time and transmits the speed signal to the industrial control computer. Step S104: The industrial control computer receives the current running speed, matches it with multiple preset speed gear ranges, determines the current speed gear, and calls the pre-stored photo timing parameters corresponding to the current speed gear according to the current speed gear. In step S105, the industrial control computer sends a soft trigger signal to the image acquisition module at a trigger time that is ahead of the time when the hole position of the hole in the image acquisition module reaches the center of the field of view of the image acquisition module, according to the shooting timing parameters. Step S106: The image acquisition module responds to the soft trigger signal to perform image acquisition, obtains the original image data of each adsorption hole on the seeding roller, and uploads the original image data to the industrial control computer. In step S107, the industrial control computer receives the original image data, the image processing module performs cropping preprocessing on the original image data, and uses the Canny edge detection algorithm to extract edge feature maps; In step S108, the image processing module concatenates the original image data with the edge feature map and inputs it into the YOLOv8n model for forward inference. The YOLOv8n model outputs the prediction results of whether there are seedlings or not at each adsorption pore. In step S109, the image processing module converts the category prediction results of each adsorption pore into a binary recognition result array and stores it. In the array, 1 indicates that there are seedlings and 0 indicates that there are no seedlings. For adsorption pores that fail to be recognized, the corresponding position in the array is marked as -1. In step S110, the industrial control computer sends the binarized recognition result array to the embedded control board; Step S111: The embedded control board receives the binary recognition result array, determines bit by bit whether there is a missing seed in each hole of the seed tray, generates a replanting sub-instruction for hole 41 marked as 0 (no seedling), and does not generate a replanting instruction for hole 41 marked as -1 (recognition failed). Step S112: When the missing seed tray hole moves with the conveyor belt to the area directly below the corresponding pneumatic reseeding unit, the embedded control board sends an opening command to the solenoid valve of the pneumatic reseeding unit to execute the reseeding action. Step S113: If a communication timeout or cylinder jamming occurs during the replay process, the system abandons the currently unfinished replay task and resets to enter the next replay cycle, while issuing a fault alarm. Step S114: Repeat steps S103 to S113 until all acupoints have been detected and re-seeded.

[0077] Experimental verification data: To verify the actual performance of the precision seeding and replanting detection system of the present invention, a simulated working condition experimental platform was built. Comparative experiments were conducted using the scheme of the present invention and two types of traditional detection algorithms to verify the comprehensive effect of the system in terms of seedling identification in plug trays, slot matching, and real-time performance.

[0078] The experimental conditions are as follows: The experiment was conducted in a constant temperature and light environment in a laboratory. The core experimental conditions are as follows: Hardware platform: It adopts a Hikvision 5-megapixel industrial area array camera, paired with a ring diffused LED light source to provide stable illumination; the working pressure of the pneumatic system is set to 0.5MPa; the conveyor belt running speed is set to 50mm / s, consistent with the actual production line conditions.

[0079] Sample setup: The model training set contains 3,000 images of seedling holes in plug trays, covering typical scenarios such as different light intensities, seedling sizes, impurities on the tray surface, and edge holes; the test set selects 5 groups of representative single-row plug tray images, each group of images corresponding to 6 consecutive physical holes, covering various typical working conditions such as single hole without seedlings, multiple holes without seedlings, edge holes, and all holes with seedlings. The ground truth values ​​of all test samples were confirmed by manual annotation.

[0080] Comparison Algorithms: Three sets of comparison schemes were set up, namely the scheme of this invention (Canny edge preprocessing + YOLOv8n serial detection scheme), pure morphological detection algorithm, and manual contour feature detection algorithm.

[0081] The experimental output array is defined as follows: the array length corresponds one-to-one with the physical holes. 1 indicates that there are no seedlings (triggering replanting), 0 indicates that there are seedlings (not triggering replanting), and -1 indicates that the hole was not identified (not triggering replanting, following the principle of better to have fewer than more).

[0082] The manually labeled truth values ​​for the 5 test samples are as follows: like Figure 2 As shown: Sample 1: [1,0,0,0,0,0], that is, no seedlings in the first hole, and seedlings in the other 5 holes; like Figure 3 As shown: Sample 2: [0,0,0,1,1,0], that is, no seedlings in holes 4 and 5, and seedlings in the other 4 holes; like Figure 4 As shown: Sample 3: [0,0,1,0,0,0], that is, no seedlings in the 3rd hole, and seedlings in the other 5 holes; like Figure 5 As shown: Sample 4: [0,0,0,1,0,1], that is, no seedlings in holes 4 and 6, and seedlings in the other 4 holes; like Figure 6 As shown: Sample 5: [0,0,0,0,0,0], meaning that all 6 holes had seedlings, and there were no holes without seedlings. This invention uses Canny edge extraction as image preprocessing, and inputs the edge map into the YOLOv8n model to complete the final judgment. The recognition results are as follows: Sample 1 output: [1,0,0,0,0,-1], Error condition: The 6th acupoint (image edge area) failed to be identified, with no false positives; Sample 2 output: [0,0,0,0,0,-1], error: no seedlings were missed in holes 4 and 5, and identification failed in hole 6; Sample 3 output: [0,0,0,0,0,0], Error: No seedlings were missed in the 3rd hole; Sample 4 output: [0,0,0,0,0,0], error: missed detection of no seedlings in holes 4 and 6; Sample 5 output: [0,0,0,0,0,-1], Error: The 6th acupoint failed to be identified, with no false positives.

[0083] The pure morphological algorithm determines the presence or absence of seedlings through threshold segmentation, opening and closing operations, and area filtering. The recognition results are as follows: Sample 1 output: [0,0,0,1,0,0], error: missed detection of no seedlings in hole 1, false positive of seedlings in hole 4 (false positive). Sample 2 output: [0,0,0,0,0,0], Error: No seedlings were detected in holes 4 and 5; Sample 3 output: [0,1,0,0,0,0], error condition: no seedlings in the slot misaligned, the true value is no seedlings in the 3rd hole, but it is misidentified as no seedlings in the 2nd hole; Sample 4 output: [0,0,0,0,0,0], Error: No seedlings were detected in holes 4 and 6; Sample 5 output: [0,0,0,1,0,0], error: no real seedling-free holes, seedling-free judgment was generated for the 4th hole out of thin air (false positive).

[0084] The manual contour feature algorithm makes a determination by extracting contour shape, roundness, and size features. The recognition results are as follows: Sample 1 output: [0,0,0,1,0,0], error: missed detection of no seedlings in hole 1, false positive of seedlings in hole 4 (false positive). Sample 2 output: [0,0,0,0,1,1], error: missed detection of no seedlings in hole 4, false positive of seedlings in hole 6 (false positive). Sample 3 output: [0,1,0,0,0,0], error condition: no seedlings in the slot misaligned, the true value is no seedlings in the 3rd hole, but it is misidentified as no seedlings in the 2nd hole; Sample 4 output: [0,0,0,0,0,0], Error: No seedlings were detected in holes 4 and 6; Sample 5 output: [0,0,0,0,0,1], error: no real seedling-free holes, seedling-free judgment was generated for the 6th hole out of thin air (false positive).

[0085] Based on the test results of the above 5 groups of 30 holes, the results are compared from the 5 dimensions that the replanting system focuses on, as shown in the table below:

[0086] Based on the business logic of the replanting system, further analysis is as follows: The solution of this invention did not produce a single false positive. All acupoints that were determined to be without seedlings corresponded to real scenarios without seedlings, which perfectly matched the control logic of the replanting system, which is to avoid the problem of seed waste and seedling overlap caused by erroneous replanting. Both traditional algorithms have obvious false positives and are prone to performing incorrect replanting on normal seedlings.

[0087] The detection results of this invention correspond one-to-one with the physical holes, eliminating the problem of misalignment across slots and ensuring accurate cylinder replanting position. Traditional algorithms are susceptible to image offset, disk surface impurities, and contour interference, resulting in hole matching errors and directly causing the replanting action to fail.

[0088] The single-frame inference time of this invention is 140ms. Combined with the conveyor belt speed of 50mm / s and the spacing between holes, image acquisition, recognition and command issuance can be completed before the hole reaches the cylinder replanting station, thus meeting the timing requirements for real-time replanting.

[0089] The current solution lacks stability in recognizing holes (such as the 6th hole) in the edge region of the image, and the recall rate in small, seedless scenes needs to be improved. In the future, the problem of missed detection can be further improved by expanding the training samples in the edge region and optimizing the image cropping range.

[0090] Experimental results show that the precision seeding and replanting detection scheme proposed in this invention, which integrates Canny edge preprocessing and YOLOv8n serial processing, significantly outperforms traditional pure morphological algorithms and manual contour feature algorithms in three core engineering indicators: slot matching accuracy, false positive control, and seedling identification accuracy. Furthermore, its inference speed can match the real-time replanting requirements at a conveyor belt speed of 50 mm / s. It can stably and reliably complete the identification of seedling presence in plug trays and the triggering control for replanting, demonstrating significant engineering application value.

[0091] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A visual recognition-based intelligent system for precise pneumatic reseeding of missing seeds in seed trays, characterized in that: include: The image acquisition module is used to acquire real-time images of each adsorption hole on the seeding drum and obtain raw image data. The image processing module is used to classify and identify whether each adsorption pore in the original image data has a seed based on a deep learning target detection model, and outputs a binary recognition result array corresponding to each adsorption pore. The control module is used to receive the binarized recognition result array and generate replay control instructions; The pneumatic reseeding execution module includes multiple sets of independently controllable pneumatic reseeding units that correspond one-to-one with each adsorption hole position. Each pneumatic reseeding unit responds to the corresponding sub-instruction in the reseeding control command and performs a reseeding action for the hole position in the seed tray corresponding to the missing adsorption hole position. The image acquisition module performs image acquisition at a trigger time determined by pre-stored shooting timing parameters corresponding to multiple discrete speed levels. The trigger time has an advance relative to the time when the acupoint reaches the center of the field of view of the image acquisition module, and different speed levels correspond to different shooting timing parameters.

2. The intelligent system for precise pneumatic reseeding of seed trays based on visual recognition as described in claim 1, characterized in that, The deep learning object detection model is a lightweight YOLOv8n model. The YOLOv8n model adopts a three-segment network structure, specifically including: The Backbone network, based on an improved CSPDarknet structure, includes a C2f module and an SPPF fast spatial pyramid pooling layer, which are used to downsample the input image by 8x, 16x and 32x respectively, and output feature maps of different scales in three layers: P3, P4 and P5. The Neck multi-scale fusion network adopts a PAN-FPN bidirectional feature pyramid structure, and all deep and shallow features in the PAN-FPN structure are fused and interacted using C2f modules. The decoupled detection head adopts an anchor-free design, completely separating the classification branch from the coordinate regression branch, and independently predicting the class confidence and bounding box coordinate distribution respectively; The input image size of the YOLOv8n model is a preset small size, which is smaller than the conventional input size of a general object detection model. The YOLOv8n model also enables a rect rectangular training strategy to adapt to the non-square, slender image features of plug seedlings.

3. The intelligent system for precise pneumatic reseeding of seed trays based on visual recognition according to claim 2, characterized in that, The image processing module is further configured to extract edge features from the original image data using the Canny edge detection algorithm before inputting the original image data into the YOLOv8n model, thereby generating an edge feature map; The edge feature map and the original image data are input together into the YOLOv8n model for recognition; In this system, the recognition result of the YOLOv8n model serves as the decision-making entity, while the output of the Canny edge detection algorithm serves as an auxiliary reference feature.

4. The intelligent system for precise pneumatic reseeding of seed trays based on visual recognition according to claim 1, characterized in that, The deep learning target detection model identifies seedling-free adsorption pores based on the following criteria: When no seeds are adsorbed at the adsorption pore, the pore has a regular circular outline; when seeds are adsorbed at the pore, the pore has an irregular shape.

5. The intelligent system for precise pneumatic reseeding of seed trays based on visual recognition according to claim 4, characterized in that, The image acquisition module and the pneumatic seeding execution module are arranged sequentially along the conveyor belt conveying direction. The camera coordinate system of the image acquisition module, the physical coordinates of each hole in the seed tray, and the cylinder coordinates of the pneumatic seeding execution module are aligned through an equally spaced physical structure design. The pneumatic seeding units of the pneumatic seeding execution module are physically mapped one-to-one with each adsorption hole of the seeding roller and each hole in the seed tray at equal intervals.

6. The intelligent system for precise pneumatic reseeding of seed trays based on visual recognition according to claim 1, characterized in that, The control module includes an industrial computer and an embedded control board; The industrial control computer is communicatively connected to the image acquisition module and is used to receive the raw image data and run the image processing module; The embedded control board is communicatively connected to the industrial computer and is used to receive the binarized recognition result array and generate the replay control command. The embedded control board is also electrically connected to the control terminal of each pneumatic re-seeding unit of the pneumatic re-seeding execution module, and is used to send the re-seeding control command to each pneumatic re-seeding unit.

7. The intelligent system for precise pneumatic reseeding of seed trays based on visual recognition according to claim 6, characterized in that, The industrial control computer is also used for: The corresponding photo-taking sequence is matched according to the current running speed of the conveyor belt, and the image acquisition module is controlled to acquire images according to the photo-taking sequence. The image acquisition module uses an industrial camera, which responds to the photo-taking command of the industrial control computer via a soft trigger to acquire images and uploads the acquired BGR format image data to the industrial control computer.

8. The intelligent system for precise pneumatic reseeding of seed trays based on visual recognition according to claim 1, characterized in that, The image processing module performs the recognition process of the deep learning object detection model, which includes: The original image data is cropped and then input into the deep learning target detection model for forward inference. The deep learning target detection model outputs the category prediction results for each adsorption pore site, including seedling category and no seedling category. The category prediction results of each adsorption pore site are converted into the binary recognition result array and returned for storage, so as to be called in subsequent replay.

9. The intelligent system for precise pneumatic reseeding of seed trays based on visual recognition according to claim 7, characterized in that, The image processing module generates a recognition failure flag for adsorption wells where the recognition result is failed. The industrial control computer stops generating and issuing replanting instructions for the adsorption wells based on the recognition failure flag, so as to implement the strategy of not replanting missed seeds. When a communication timeout occurs between the industrial control computer and the embedded control board, or when the pneumatic re-seeding execution module experiences cylinder jamming, the unfinished re-seeding task corresponding to the current hole position is abandoned, and the re-seeding control process for the next hole position is reset.

10. The intelligent system for precise pneumatic reseeding of seed trays based on visual recognition according to claim 2, characterized in that, The loss function of the YOLOv8n model is: Loss(total)=λ(box)×Loss(CIoU)+λ(cls)×Loss(BCE)+λ(dfl)×Loss(DFL); Wherein, Loss(CIoU) is the CIoU bounding box regression loss, used to constrain the overlap area, center distance and aspect ratio of the predicted box and the ground truth box; Loss(BCE) is the binary cross-entropy classification loss, used for multi-class seedling identification; Loss(DFL) is the distribution focus loss, used to predict the probability distribution of discrete coordinate intervals, and λ(box), λ(cls), and λ(dfl) are the preset weight coefficients of the corresponding loss functions.

Images