Plant growing point detection method

Abstract

The embodiment of the invention discloses a plant growing point detection method, relates to the technical field of agricultural monitoring, and is used for solving the problem that a crop growing point cannot be accurately detected and positioned in the prior art. The plant growing point detection method comprises the following steps: acquiring a multi-dimensional image and depth data of a target scene through a dual-camera system; and processing the multi-dimensional image and the depth data by using a pre-trained plant growing point detection model, identifying a plant growing point in the target scene, and carrying out spatial positioning on the plant growing point. According to the invention, the stereoscopic vision technology is introduced, the dual-camera system is used for capturing the image and depth information of the scene, and the depth information data and the conventional image data are creatively combined for use, so that not only can crops be identified, but also the positions of the crops in the three-dimensional space can be accurately positioned, and the detection of plant growing points is realized.

Description

Technical Field

[0001] The invention relates to the technical field of agricultural monitoring, and in particular to a method for detecting plant growth points. Background Art

[0002] In the current agricultural technology context, the development of precision agriculture is increasingly dependent on advanced monitoring and automation technologies. Especially in terms of crop growth monitoring and management, traditional methods rely on manual observation, which is not only inefficient but also unable to achieve high accuracy and real-time feedback.

[0003] With the development of image processing and machine vision technology, there are technologies for crop growth monitoring based on deep learning monocular vision systems in the existing technology. These systems usually rely on a large amount of image data to train models to identify crops and assess their growth conditions. However, the systems for crop detection based on monocular vision systems have the following main problems:

[0004] 1. Existing monocular vision systems in agricultural applications usually rely only on two-dimensional image data and cannot effectively obtain spatial depth information of crops, which limits their performance in spatial positioning accuracy, especially in complex agricultural environments. In this case, a single perspective often cannot provide enough information to support accurate crop growth status assessment and growth point positioning.

[0005] 2. Recognition problems in complex environments: In environments with multi-layered vegetation or where the background and crop colors are similar, it is difficult for a monocular vision system to accurately distinguish between crops and backgrounds. This not only affects the detection accuracy of crop growth points, but also reduces the efficiency of crop health monitoring and pest and disease management.

[0006] 3. Inefficient data processing: Existing agricultural monitoring systems often need to process large amounts of image data, which can lead to slow processing when running on a single CPU or GPU, affecting real-time decision-making. Therefore, a method that can more efficiently process and analyze large-scale data is needed.

[0007] 4. High-precision requirements: With the development of precision agriculture technology, the accuracy requirements for detection systems are getting higher and higher. Traditional methods often cannot meet the needs of efficient agricultural production in terms of accuracy, especially in the fields of fine operations such as fixed-point fertilization and precise spraying. Summary of the invention

[0008] In view of this, an embodiment of the present invention provides a plant growth point detection method to solve the problems that the existing crop growth monitoring technology lacks the ability to obtain depth information, cannot accurately locate plant growth points, is ineffective in crop identification under occlusion and complex backgrounds, and has poor monitoring effect.

[0009] An embodiment of the present invention provides a method for detecting plant growth points, the method comprising:

[0010] Acquire multi-dimensional image and depth data of the target scene through a dual-camera system;

[0011] The multi-dimensional image and depth data are processed using a pre-trained plant growth point detection model to identify plant growth points in the target scene and spatially locate them.

[0012] In a first embodiment, after acquiring the multi-dimensional image and depth data of the target scene by the dual camera system, the method further includes:

[0013] Performing data preprocessing on the multidimensional image and depth data of the target scene to obtain a preprocessed image;

[0014] A deep learning model is used to extract key point features of plant growth points from the preprocessed image, and model training is performed to obtain a plant growth point detection model.

[0015] In a second embodiment, the data preprocessing of the multi-dimensional image and depth data of the target scene includes:

[0016] The images obtained by the two cameras of the dual-camera system are synchronously aligned in time, and the images obtained by the two cameras are geometrically corrected by mathematical transformation to obtain a first image;

[0017] The depth data of the target scene is converted into color code to obtain a visual depth information image.

[0018] In a third embodiment, the color coding conversion of the depth data of the target scene includes:

[0019] The depth data of the target scene is converted to color coding according to the following second formula:

[0020] I d = alpha × (1 - (Z 0 -Z min ) / (Z max -Z min ))

[0021] In the second formula above, I d is the color value of the converted depth information image corresponding to the target position in the target scene; alpha is the preset conversion coefficient used to map the depth value to the grayscale range of 0-255; Z 0 is the original depth value of the target position in the target scene; Z min , Z maxare the minimum and maximum depth values ​​in the target scene, respectively.

[0022] In a fourth embodiment, the method of using a pre-trained plant growth point detection model to process the multi-dimensional image and depth data, identifying plant growth points in a target scene and spatially locating the plant growth points includes:

[0023] Inputting the first image and the depth information image into the plant growth point detection model;

[0024] identifying plant growth points in the first image by using the plant growth point detection model, and spatially locating key points of the plant growth points in the first image;

[0025] marking plant growth points in the first image to obtain a second image;

[0026] The second image is output.

[0027] In a fifth implementation manner, the spatially locating key points of plant growth points in the first image includes:

[0028] The spatial coordinates of the key points of the plant growth points in the first image are determined by the following third formula:

[0029]

[0030] Where: X, Y, Z are the spatial coordinates of the key point of the plant growth point in the first image; x_k, y_k are the pixel positions of the key point of the plant growth point identified in the first image; c_x, c_y are the pixel coordinates of the optical center of the dual camera system in the first image; Z 0 _k is the original depth value of the key point of the plant growth point in the first image in the target scene; f is the focal length of the dual camera system;

[0031] The step of marking a plant growth point in the first image comprises:

[0032] The position of the plant is marked in the first image in the form of a bounding box, and the spatial coordinates of key points of the plant growth point within the bounding box are marked.

[0033] In a sixth embodiment, the method of acquiring multi-dimensional images and depth data of a target scene by using a dual camera system includes:

[0034] Through the dual-camera system, the depth data of each position in the target scene is determined based on the following first formula:

[0035] Z = f × B / (x1 - x2 + epsilon)

[0036] Wherein, Z is the depth value of the target position in the target scene; f is the focal length of the dual-camera system; B is the baseline distance between the two cameras in the dual-camera system; x1 and x2 represent the horizontal coordinates of the target position in the two camera images respectively; and epsilon is a preset constant.

[0037] In the seventh implementation, in the process of performing model training to obtain the plant growth point detection model, the loss function expression used is:

[0038] L = lambda_det × L_det + lambda_kp × L_kp + lambda_depth × L_depth

[0039] Among them, L is the total loss function, which is the value that needs to be minimized during model training; lambda_det, lambda_kp, and lambda_depth are preset weight parameters, which respectively control the contribution ratio of detection loss, key point detection loss, and depth estimation loss to the total loss; L_det is the loss of the detection task, which uses cross entropy loss or squared error loss to measure the model's prediction accuracy for the target position; L_kp is the loss of key point detection, which is used to evaluate the model's ability to identify specific points in an image; L_depth is the loss of depth estimation, which is used to evaluate the error between the depth value predicted by the model and the true value.

[0040] In an eighth embodiment, after identifying the plant growth points in the target scene and spatially locating them, the method further includes:

[0041] Acquiring manual detection data of plant growth points in the target scene;

[0042] Evaluating the plant growing point detection model by a cross-validation method;

[0043] The parameters of the plant growth point detection model are adjusted according to the evaluation results.

[0044] A plant growth point detection solution provided by an embodiment of the present invention has the following beneficial effects:

[0045] 1. Enhancement of depth information: By introducing stereoscopic vision technology and using a dual-camera system to capture the image of the scene and its depth information, it is not only possible to identify crops, but also to accurately locate the position of crops in three-dimensional space, greatly improving the accuracy of positioning, solving the problem of lack of depth information in existing technologies, and enabling the detection of plant growth points.

[0046] 2. Adaptability to complex environments: Innovatively combining depth information with conventional image data, using depth information to assist image correction and synchronization ensures the accuracy and consistency of data input, which not only improves the accuracy of subsequent image processing, but also optimizes the effect of feature extraction, significantly improving the adaptability and robustness of the model to complex environments. Depth information helps the system distinguish between foreground and background, and can work effectively even when color and texture information are insufficient to distinguish.

[0047] 3. Efficient data processing and analysis: This invention adopts advanced deep learning algorithms, especially in the feature extraction and model training stage, and trains an effective plant growth point detection model by designing a targeted loss function, which improves the speed and efficiency of processing large-scale data. At the same time, the model can effectively handle image variations caused by environmental changes, enhancing the robustness of the model.

[0048] 4. Cross-validation and model optimization: The cross-validation method is used to systematically evaluate the performance of the plant growth point detection model, ensuring the stability and reliability of the model under different conditions. Through this method, the model parameters are continuously adjusted and optimized to ensure that the best detection effect can be achieved in practical applications. It can achieve accurate positioning of crop growth points and real-time monitoring of health status, greatly improving the automation and intelligence level of agricultural production.

[0049] The comprehensive application of the above-mentioned innovative points of the present invention not only significantly improves the detection accuracy of plant growth points, but also optimizes the efficiency and stability of the entire detection process, providing strong technical support for precision agriculture and making crop management more scientific and automated. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 A flow chart of a first embodiment of a plant growth point detection method provided by the present invention;

[0052] Figure 2 A flow chart of a second embodiment of the plant growth point detection method provided by the present invention;

[0053] Figure 3 This is a flow chart of Example 3 of the plant growth point detection method provided by the present invention. DETAILED DESCRIPTION

[0054] The present invention relates to a plant growth point detection method based on stereoscopic vision, which uses a dual-camera system to capture the image and depth data of the scene to achieve accurate positioning and monitoring of the growth points of crops. The technology is implemented through five main steps: data acquisition, data preprocessing, feature extraction and model training, plant growth point detection, and result analysis. In the data acquisition stage, two cameras are used to obtain images of crops and their corresponding depth information from different perspectives to ensure that sufficient space and depth data are obtained. Data preprocessing includes synchronization and correction of images and color coding conversion of depth information, aiming to improve the quality and availability of data. The feature extraction and model training link uses a deep learning algorithm to process image data through a multi-layer convolutional network, learn the complex features of the image, and train a detection model that can accurately identify and locate plant growth points. The plant growth point detection step uses the trained detection model to analyze new image data to accurately identify the boundaries and growth point positions of crops. Finally, in the result analysis stage, the performance of the model is evaluated by statistical methods such as cross-validation, and the parameters of the detection model are adjusted according to the evaluation results to optimize the detection effect. This invention can accurately locate the plant growth point in space, significantly improving the automation and accuracy of agricultural production, and has important application value in the fields of precision agriculture, crop monitoring and management. By accurately monitoring the growth of crops, it helps to optimize agricultural operations such as fertilization, irrigation and pest control, thereby improving crop yield and quality.

[0055] A plant growth point detection method according to an embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0056] It should be clear that the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] Figure 1 Flow chart of the first embodiment of the plant growth point detection method provided by the present invention, as shown in FIG. Figure 1 As shown, the method of this embodiment may include:

[0058] Step 101: Acquire multi-dimensional images and depth data of a target scene through a dual-camera system.

[0059] In this embodiment, the acquisition of image data and depth data is mainly achieved by configuring a dual camera system (C1 and C2). The two cameras shoot the same target scene from different perspectives, thereby acquiring rich spatial and depth information.

[0060] Preferably, this embodiment uses a dual-camera system to calculate the depth data of each position in the target scene based on the following first formula (1):

[0061] Z = f × B / (x1 - x2 + epsilon) (1)

[0062] Wherein, Z is the depth value of the target position in the target scene, representing the depth distance from the camera (e.g., a binocular camera in a dual-camera system) to the object; f is the focal length of the dual-camera system, that is, the distance from the dual-camera lens to the imaging sensor during imaging; B is the baseline distance between the two cameras in the dual-camera system, which refers to the horizontal distance between the centers of the two camera lenses, and determines the basic scale of depth estimation; x1 and x2 represent the horizontal coordinates of the target position in the two camera images, respectively; and epsilon is a preset constant.

[0063] In practical applications, the stereoscopic vision system of the embodiment of the present invention can be widely used in agricultural robots, drone monitoring and field automation equipment to monitor the growth status of crops, the occurrence of pests and diseases, and the maturity of crops in real time. Through precise depth information, these devices can more accurately locate crops, optimize operations such as spraying and pruning, and ultimately improve the efficiency of agricultural production and the quality of crops. For example, by analyzing the spatial distribution and height differences between different crops, planting planning and field management can be effectively carried out. In addition, depth information can also help determine the growth stage of crops, which is of great guiding significance for timed fertilization and irrigation. In general, the stereoscopic vision system not only provides richer and more accurate information than the traditional single camera system, but also can greatly improve the level of agricultural automation and intelligence through reasonable technical configuration and depth algorithm optimization, providing strong technical support for the development of modern precision agriculture.

[0064] Step 102: Use a pre-trained plant growth point detection model to process the multi-dimensional image and depth data, identify plant growth points in the target scene, and spatially locate them.

[0065] In this embodiment, a pre-trained plant growth point detection model is used in combination with depth data to analyze the image of the target scene to identify the plant growth points in the image. The output results of the model can be directly used to guide specific operations in agricultural production, such as precise fertilization and disease control.

[0066] Figure 2 This is a flow chart of Embodiment 2 of the plant growth point detection method provided by the present invention, as shown in FIG. Figure 2 As shown, the method of this embodiment may include:

[0067] Step 201: Acquire multi-dimensional images and depth data of a target scene through a dual-camera system.

[0068] In this embodiment, the process of step 201 is similar to step 101 of the above method embodiment and will not be repeated here.

[0069] Step 202: preprocess the multi-dimensional image and depth data of the target scene to obtain a preprocessed image.

[0070] In the embodiment of the present invention, after the original image and depth data are collected, necessary preprocessing operations are performed, including synchronization and geometric correction of the image and color coding conversion of the depth data. The main purpose of these operations is to improve the data quality by adjusting the image so that the images taken by the two cameras can be aligned, and converting the depth information into a format more suitable for computer processing. In this way, the processed data is more convenient for effective feature extraction and analysis in subsequent steps, providing support for more complex image analysis tasks.

[0071] Specifically, step 202 may include the following steps A10 to A20:

[0072] Step A10: synchronously align the images obtained by the two cameras of the dual-camera system in time, and perform geometric correction on the images obtained by the two cameras through mathematical transformation to obtain a first image.

[0073] In this embodiment, image synchronization and correction are key steps to ensure image data quality and subsequent processing accuracy. This process involves two main technologies: time synchronization and geometric correction. Time synchronization ensures that the image data obtained from the two cameras are aligned in time, which is particularly important for capturing objects in dynamic environments. Geometric correction adjusts the images in the two views through mathematical transformation (applying affine transformation or perspective transformation) to eliminate image distortion caused by camera position deviation and make the positions of the same object in the two camera views correspond, which usually involves rotation and translation correction.

[0074] Step A20: Perform color coding conversion on the depth data of the target scene to obtain a visualized depth information image.

[0075] In this embodiment, color coding conversion is the process of converting depth data into a visual 2D image, which is of great significance for both manual inspection and training of machine learning models. Depth image data is usually not directly applicable to traditional image processing algorithms. In this embodiment, by color coding it, the depth data can be converted into grayscale or color image data so that it can be processed by standard image processing technology or visual recognition algorithms.

[0076] Preferably, in the implementation of the present invention, the depth data of the target scene is converted into color code according to the following second formula (2):

[0077] I d= alpha × (1 - (Z 0 -Z min ) / (Z max -Z min )) (2)

[0078] In the second formula (2) above, I d is the color value of the converted depth information image corresponding to the target position in the target scene; alpha is the preset conversion coefficient used to map the depth value to the grayscale range of 0-255; Z 0 is the original depth value of the target position in the target scene; Z min , Z max are the minimum and maximum depth values ​​in the target scene, respectively.

[0079] In the embodiment of the present invention, by converting the depth data into a color-coded image, the growth of crops and terrain changes can be observed intuitively, thereby performing more accurate growth analysis and yield prediction. In addition, by using this conversion in machine vision, different objects and regions in the target scene image can be identified and segmented by color differences.

[0080] In this embodiment, the image synchronization, geometric correction and color coding conversion in the data preprocessing of step 202 are the basis for realizing high-quality visual data analysis, which can improve the availability of data and play an indispensable role in improving the level of agricultural automation, enhancing the perception ability of robots and optimizing data visualization. In addition, the accuracy of decision-making and the efficiency of operation can be improved through precise data processing.

[0081] Step 203: Use a deep learning model to extract key point features of plant growth points from the preprocessed image, and perform model training to obtain a plant growth point detection model.

[0082] In the embodiments of the present invention, the key points of plant growth points generally refer to those specific parts that are crucial to plant growth and development. These key points may include but are not limited to the following:

[0083] Bud: This is the point where new branch or leaf growth begins on a plant, usually located inside a bud scale at the tip or side of a branch. In perennials and woody plants, the bud is the starting point for a new season's growth.

[0084] Nodes: Nodes on plant stems are the points at which leaves, branches, and inflorescences grow. These locations are key areas for new plant tissue to grow from the stem, so they are important for monitoring plant growth.

[0085] Root Apex: The root apex is the growth area of ​​the root and contains the root hairs and apical meristem. These are the key points for roots to grow and expand downward into the soil.

[0086] Flowering point: In flowering plants, the location where the flower buds form is also an important growth point. These are usually located at the top of the stem or at the tip of a branch.

[0087] The specific definition and identification of these key points usually depends on the type and growth stage of the plant. When implementing plant growth point detection, these key point features can be identified through high-resolution image capture and deep data analysis combined with machine learning models.

[0088] In this embodiment, the key point feature extraction of plant growth points is performed based on the architecture of a deep convolutional neural network, and the useful features in the image are extracted by stacking multiple convolutional layers and activation layers. Among them, the convolution layer works by sliding multiple small windows (convolution kernels) on the input image, and the area corresponding to each window is multiplied by the weight of the convolution kernel, and then summed to obtain a pixel point of the output feature map. Each convolution layer usually contains tens to thousands of trainable filters that can capture features ranging from simple edges and color changes to complex objects and shapes. The activation layer usually uses the ReLU (Rectified Linear Unit) activation function to increase the nonlinear ability of the network and help the network learn more complex patterns. This structural configuration can not only effectively process various visual information, but is also particularly effective in processing agricultural applications such as crop disease identification and growth monitoring, allowing machines to quickly and accurately extract necessary information from complex agricultural scenes.

[0089] Since the loss function defines the difference between the model output and the actual label, it is the target that needs to be minimized during the model training process. In a multi-task learning environment, designing an effective loss function is crucial to optimizing model performance. In this embodiment, a composite loss function is designed, which combines the losses of several different tasks to achieve performance optimization in multiple dimensions. Specifically, this embodiment uses the following loss function (3) to train the plant growth point detection model:

[0090] L = lambda_det × L_det + lambda_kp × L_kp + lambda_depth × L_depth (3)

[0091] Among them, L is the total loss function, which is the value that needs to be minimized during model training; lambda_det, lambda_kp, and lambda_depth are preset weight parameters, which respectively control the contribution ratio of detection loss, key point detection loss, and depth estimation loss to the total loss; L_det is the loss of the detection task, which uses cross entropy loss or squared error loss to measure the model's prediction accuracy for the target position; L_kp is the loss of key point detection, which is used to evaluate the model's ability to identify specific points in an image; L_depth is the loss of depth estimation, which is used to evaluate the error between the depth value predicted by the model and the true value.

[0092] The plant growth point detection model training method and loss function configuration provided in this embodiment can greatly improve the accuracy of crop disease detection, growth analysis and yield prediction. For example, through accurate growth point detection and disease identification, agricultural robots can automatically adjust fertilization or spraying strategies to achieve precision agricultural management. At the same time, the depth estimation function enables the drone to adjust its altitude during flight to avoid hitting crops while obtaining high-quality crop growth data. In short, through efficient feature extraction and precise model training, it is possible to better understand and utilize complex agricultural image data, provide scientific and technological support for agricultural production, and achieve the goal of automated and intelligent management. This not only improves the efficiency and accuracy of agricultural operations, but also helps protect the environment and improve food safety.

[0093] Step 204: Use a pre-trained plant growth point detection model to process the multi-dimensional image and depth data, identify plant growth points in the target scene, and spatially locate them.

[0094] In some optional embodiments, step 204 may include the following steps B10-B40:

[0095] Step B10, inputting the first image and the depth information image into the plant growth point detection model;

[0096] Step B20: identifying plant growth points in the first image by using the plant growth point detection model, and spatially locating key points of the plant growth points in the first image;

[0097] In this embodiment, preferably, the spatial coordinates of the key points of the plant growth points in the first image can be determined by the following third formula (4):

[0098] (4)

[0099] Where: X, Y, Z are the spatial coordinates of the key point of the plant growth point in the first image; x_k, y_k are the pixel positions of the key point of the plant growth point identified in the first image; c_x, c_y are the pixel coordinates of the optical center of the dual camera system in the first image; Z 0 _k is the original depth value of the key point of the plant growth point in the first image in the target scene; f is the focal length of the dual camera system.

[0100] The core of the third formula is to convert the 2D pixel coordinates on the image plane into 3D world coordinates. This conversion allows us to obtain the actual physical location of an object from a flat image, which is crucial for agricultural applications such as automated planting, disease management, and harvesting.

[0101] Step B30, marking plant growth points in the first image to obtain a second image;

[0102] In this embodiment, the position of the plant is marked in the first image in the form of a bounding box, and the spatial coordinates of the key points of the plant growth point within the bounding box are marked. The labeled image may be used in the model training process, in which the positions of these key points are marked in detail to improve the accuracy of recognition and the generalization ability of the model. In this way, the model can not only recognize the overall morphology of the plant, but also accurately locate the specific position of the growth point, providing data support for further growth analysis and agricultural management.

[0103] Step B40: output the second image.

[0104] In this embodiment, the user can intuitively know the position of plants in the current target scene based on the output second image and the bounding box displayed in the second image, and know the spatial coordinates of the key points of each plant growth point based on the marks in the second image.

[0105] Figure 3 Flow chart of the third embodiment of the plant growth point detection method provided by the present invention, as shown in Figure 3 As shown, the method of this embodiment may include:

[0106] Step 301: Acquire multi-dimensional images and depth data of a target scene through a dual-camera system.

[0107] Step 302: preprocess the multi-dimensional image and depth data of the target scene to obtain a preprocessed image.

[0108] Step 303: Use a deep learning model to extract key point features of plant growth points from the preprocessed image, and perform model training to obtain a plant growth point detection model.

[0109] Step 304: Use a pre-trained plant growth point detection model to process the multi-dimensional image and depth data, identify plant growth points in the target scene, and spatially locate them.

[0110] In this embodiment, the process of steps 301-304 is similar to steps 201-204 of the above method embodiment, and will not be repeated here.

[0111] Step 305: Obtain manual detection data of plant growth points in the target scene.

[0112] Step 306: Evaluate the accuracy and robustness of the plant growth point detection model by a cross-validation method.

[0113] Since performance evaluation is a key step to confirm whether the model can effectively perform the predetermined task in practical applications. In the detection of plant growth points in the agricultural field, it is crucial to ensure the accuracy, reliability and efficiency of the model. To this end, the cross-validation method is used in this embodiment to evaluate the performance of the plant growth point detection model. This method can effectively avoid overfitting while ensuring the generalization ability of the evaluation.

[0114] The cross-validation method is used to evaluate and improve the model's predictive power for independent data sets. It involves splitting the data set into multiple small subsets, training the model on one subset and testing it on the other subsets, and repeating the process multiple times, selecting different training and test sets each time.

[0115] The specific evaluation process includes:

[0116] Data segmentation: All data is randomly divided into K equal-sized subsets. In agricultural applications, this means dividing all images and related depth information into multiple parts to ensure that each part can independently represent the characteristics of the overall data.

[0117] Iterative testing: Select one of the subsets as the test set and the remaining K-1 subsets as the training set. After the model is trained on the training set, the test set is used to evaluate its performance. This process is repeated K times, each time selecting a different subset as the test set.

[0118] Performance indicator calculation: The test results of each round are evaluated. Common performance indicators include accuracy, recall, precision and F1 score. Each indicator provides different aspects of performance information. Among them, accuracy refers to the ratio of the total number of correctly predicted samples to the total number of samples; recall refers to the ratio of the number of correctly predicted positive samples to the number of actual positive samples; precision refers to the ratio of the number of correctly predicted positive samples to the number of all predicted positive samples; F1 score refers to the harmonic mean of precision and recall, which is an important indicator to measure the robustness of the model.

[0119] Optimal parameter selection: Based on the cross-validation results, adjust model parameters such as learning rate, number of layers, number of neurons, etc. to find the optimal model configuration.

[0120] Step 307: Adjust the parameters of the plant growth point detection model according to the evaluation result.

[0121] In this embodiment, a step of evaluating the performance of plant growth points is added, and the performance evaluation process ensures the applicability and reliability of the plant growth point detection model under different environments and conditions. For example, by accurately adjusting the detection model parameters, the machine vision system can better adapt to plant growth point detection under different lighting and background conditions, thereby optimizing operations such as crop spraying, fertilization, and harvesting. This not only improves operating efficiency, but also helps to reduce resource waste, increase crop yields, and improve the overall sustainability of agricultural production. Overall, through strict performance evaluation and systematic parameter optimization, the operating accuracy and efficiency of agricultural robots and monitoring systems can be significantly improved, and high-quality agricultural production can be achieved. The implementation of this method will help promote the development of modern agriculture in the direction of higher technology and more automation.

[0122] The specific implementation method of the third embodiment of the present invention is described below with reference to a specific example:

[0123] Assume that a plant growth point detection system needs to be deployed in a large agricultural planting base to improve the accuracy and efficiency of crop management. The base mainly grows tomatoes, which are crops that are sensitive to the growth environment and whose growth points are difficult to accurately detect. Using the method provided by the present invention, the following method flow is performed:

[0124] S1. Technical equipment configuration

[0125] Camera configuration: Two high-resolution stereo vision cameras are selected and installed on an automatically moving monitoring vehicle. The cameras are aligned and set parallel to a fixed baseline distance B to ensure that a clear binocular view can be captured.

[0126] S2. Data Collection

[0127] The stereo vision camera continuously captures images and corresponding depth data of tomatoes in the base field, which are transmitted to the edge computing device in real time for preliminary processing.

[0128] S3, data processing and model training

[0129] Pre-processing: The captured images are processed using image synchronization and geometric correction techniques to ensure that the two camera views are fully aligned; the depth data is converted into a 2D image format through color encoding for subsequent processing.

[0130] Feature extraction: Use deep learning models, such as modified convolutional neural networks, to extract features from preprocessed images. The network structure design should take into account the need to efficiently process large amounts of data while ensuring that the growth points of tomatoes can be accurately identified and located.

[0131] Model training: Based on the extracted key image features, model training is performed to obtain a tomato growth point detection model.

[0132] S4. Growth point detection and analysis

[0133] Detection algorithm: Use the trained tomato growth point detection model to analyze the image, detect the growth points of the tomato, and mark their precise positions in three-dimensional space.

[0134] Decision support: The test results can be directly used in decision support systems to guide agricultural operations such as precision spraying and targeted fertilization.

[0135] S5. Optimize feedback

[0136] Results analysis: The accuracy and robustness of the tomato growth point detection model were evaluated by using a cross-validation method through testing under actual field conditions. The parameters of the tomato growth point detection model were adjusted according to the test results to optimize the detection algorithm.

[0137] Continuous optimization: The system regularly collects operational feedback and crop growth data for model retraining and optimization, ensuring that the system adapts to changing field environments and crop growth conditions.

[0138] This example demonstrates how to use stereo vision technology and deep learning algorithms to improve the automation and intelligence of crop management in actual agricultural production, especially in accurately monitoring and processing crop growth points in complex growing environments. This technology can significantly improve crop production efficiency and quality.

[0139] In summary, through the embodiments of the present invention, plant growth points can be accurately detected. Applying this method to agricultural robots can enable agricultural robots to perform more complex tasks in a targeted manner, such as precise fertilization, irrigation or harvesting based on the detected accurate spatial position information. For example, in a vineyard, by accurately identifying the growth points of grape branches, the robot can perform fine pruning, which helps control the growth pattern of the grapes and optimize the sunlight exposure and maturity of the fruit. In addition, this technology can also be used in plant pathology. By identifying affected plant growth points early, agricultural workers can quickly take measures to deal with plant diseases or pest problems and reduce crop losses. By continuously monitoring the health of plant growth points, the health and yield of plants can be optimized throughout the growth cycle, improving the sustainability and efficiency of agriculture.

[0140] It should be noted that, in this document, relational terms such as first and second, etc. are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0141] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0142] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for detecting plant growth points, characterized in that: The method comprises: Acquire multi-dimensional image and depth data of the target scene through a dual-camera system; The multi-dimensional image and depth data are processed using a pre-trained plant growth point detection model to identify plant growth points in the target scene and spatially locate them.

2. The plant growth point detection method according to claim 1, characterized in that: After acquiring the multi-dimensional image and depth data of the target scene by the dual camera system, the method further includes: Performing data preprocessing on the multidimensional image and depth data of the target scene to obtain a preprocessed image; A deep learning model is used to extract key point features of plant growth points from the preprocessed image, and model training is performed to obtain a plant growth point detection model.

3. The plant growth point detection method according to claim 2, characterized in that: The data preprocessing of the multi-dimensional image and depth data of the target scene includes: The images obtained by the two cameras of the dual-camera system are synchronously aligned in time, and the images obtained by the two cameras are geometrically corrected by mathematical transformation to obtain a first image; The depth data of the target scene is converted into color code to obtain a visual depth information image.

4. The plant growth point detection method according to claim 3, characterized in that: The color coding conversion of the depth data of the target scene includes: The depth data of the target scene is converted to color coding according to the following second formula: AND d = alpha × (1 - (Z0 - Z min ) / (WITH max - WITH min )); In the second formula above, I d is the color value of the converted depth information image corresponding to the target position in the target scene; alpha is the preset conversion coefficient used to map the depth value to the grayscale range of 0-255; Z0 is the original depth value of the target position in the target scene; Z min , Z max are the minimum and maximum depth values ​​in the target scene, respectively.

5. The plant growth point detection method according to claim 3, characterized in that: The method of using a pre-trained plant growth point detection model to process the multi-dimensional image and depth data, identifying plant growth points in the target scene and spatially locating the plant growth points includes: Inputting the first image and the depth information image into the plant growth point detection model; identifying plant growth points in the first image by using the plant growth point detection model, and spatially locating key points of the plant growth points in the first image; marking plant growth points in the first image to obtain a second image; The second image is output.

6. The method for detecting plant growth points according to claim 5, characterized in that: The spatially locating key points of plant growth points in the first image includes: The spatial coordinates of the key points of the plant growth points in the first image are determined by the following third formula: ； Wherein: X, Y, Z are the spatial coordinates of the key point of the plant growth point in the first image; x_k, y_k are the pixel positions of the key point of the plant growth point identified in the first image; c_x, c_y are the pixel coordinates of the optical center of the dual camera system in the first image; Z0_k is the original depth value of the key point of the plant growth point in the first image in the target scene; f is the focal length of the dual camera system; The step of marking a plant growth point in the first image comprises: The position of the plant is marked in the first image in the form of a bounding box, and the spatial coordinates of key points of the plant growth point within the bounding box are marked.

7. The plant growth point detection method according to claim 1, characterized in that: The method of acquiring multi-dimensional images and depth data of a target scene by using a dual-camera system includes: Through the dual-camera system, the depth data of each position in the target scene is determined based on the following first formula: Z = f × B / (x1 - x2 + epsilon); Wherein, Z is the depth value of the target position in the target scene; f is the focal length of the dual-camera system; B is the baseline distance between the two cameras in the dual-camera system; x1 and x2 represent the horizontal coordinates of the target position in the two camera images respectively; and epsilon is a preset constant.

8. The method for detecting plant growth points according to claim 2, characterized in that: In the process of performing model training to obtain the plant growth point detection model, the loss function expression used is: L = lambda_det × L_det + lambda_kp × L_kp + lambda_depth × L_depth; Among them, L is the total loss function, which is the value that needs to be minimized during model training; lambda_det, lambda_kp, and lambda_depth are preset weight parameters, which respectively control the contribution ratio of detection loss, key point detection loss, and depth estimation loss to the total loss; L_det is the loss of the detection task, which uses cross entropy loss or squared error loss to measure the model's prediction accuracy for the target position; L_kp is the loss of key point detection, which is used to evaluate the model's ability to identify specific points in an image; L_depth is the loss of depth estimation, which is used to evaluate the error between the depth value predicted by the model and the true value.

9. The plant growth point detection method according to claim 2, characterized in that: After identifying the plant growth points in the target scene and spatially locating them, the method further includes: Acquiring manual detection data of plant growth points in the target scene; Evaluating the plant growing point detection model by a cross-validation method; The parameters of the plant growth point detection model are adjusted according to the evaluation results.

Images