Rapeseed seedling stage heterosis positioning method and heterosis removing method

Abstract

The application provides a rape seedling stage hetero-plant positioning method and a hetero-plant removing method, wherein the rape seedling stage hetero-plant positioning method comprises the following steps: obtaining an image containing a plurality of seedling stage plants in a plot; identifying all the seedling stage plants in the image to obtain plant pictures, and configuring index information for each plant picture; extracting features of each plant picture based on high-dimensional features of a preset dimension to obtain a feature vector; calculating the Euclidean distance between any two feature vectors to determine the outlier index of each feature vector; presetting an outlier index threshold to screen plant pictures corresponding to feature vectors whose outlier indexes exceed the preset outlier index threshold; and determining the position information of the screened plant pictures according to the index information to obtain the position information of the hetero-plant. The application is used to solve the problem that in the prior art, a large amount of variety and feature data is required in the database, and it is still difficult to achieve high-precision hetero-plant identification.

Description

Technical Field

[0001] This invention relates to the application of image recognition technology in the field of seed production agriculture, and particularly to a method for locating and removing weeds during the rapeseed seedling stage. Background Technology

[0002] Hybridization refers to the first-generation hybrids produced by crossing two different crop varieties or lines. Hybrid crops exhibit significant heterosis and are an important pathway for developing crop production. Taking rapeseed as an example, the seedling stage is a crucial period for seed production. Removing major weeds during this stage provides a better growing environment for breeding and effectively reduces the workload of field inspections and weed removal during the bolting and flowering stages. Therefore, weed removal during the seedling stage is an important task in hybrid crop seed production. Typically, only one or two specific varieties can be planted in hybrid seed production fields. Any other varieties present are considered weeds and need to be removed. Current traditional weed removal methods primarily rely on manual field inspections and visual comparison to identify and remove weeds from the seedlings.

[0003] Existing technologies that rely entirely on manual labor to distinguish weeds are highly dependent on the agricultural experience of the workers. Workers need long-term experience in agricultural work to be able to conduct field inspections, which has high labor training costs. In addition, relying entirely on manual weed identification is inefficient and makes it difficult to completely remove weeds in large seed production fields, thus affecting the purity assessment of seed production fields. Therefore, image recognition technology is applied to the field of crop weed identification.

[0004] A Chinese patent with publication number CN114140702A discloses a method for removing weeds and improving seed purity in hybrid rice seed production fields. During the rice plant development stage, a drone equipped with a CCD camera is used to photograph rice plants in the field, generating real-time video stream data. Rice images are collected from the real-time video stream data, and the images are then used to identify and detect the presence of weeds using a weed identification model. If a weed is detected, its position coordinates in the image are output, including the row and column positions of the plant. These coordinates are then transmitted to a weed removal device, which identifies and removes the weed at its location.

[0005] The aforementioned existing technologies directly identify hybrid plants using hybrid plant identification models. Each identification requires pre-determining the hybrid plant variety and confirming the model's ability to identify that variety. Since different hybrid plant varieties may exist in the same plot, pre-confirmation of each hybrid plant's variety is necessary for identification, resulting in low applicability. With over 140,000 rice varieties, establishing a database encompassing all varieties is impractical; hybrid plants not existing in the database cannot be identified. Furthermore, the characteristic differences between hybrid plants and the target plant are small, leading to low accuracy in direct hybrid plant identification and poor hybrid plant removal. This makes it difficult to meet the seed production field's requirement of over 99.7% seed purity. High-precision hybrid plant models require massive amounts of data, and the database, in addition to a vast number of varieties, needs extensive characteristic data for each variety, presenting significant technical challenges.

[0006] In view of this, it is necessary to improve the existing methods for detecting hybrid plants in the seedling stage in order to solve the above problems. Summary of the Invention

[0007] The purpose of this invention is to disclose a method for locating and removing hybrid plants during the rapeseed seedling stage, in order to solve the problem in the prior art that the hybrid plant varieties need to be known in advance and the hybrid plants need to be uploaded to the database in advance. Therefore, the database needs to upload pictures of almost all varieties of the same crop and a large amount of feature data for each variety, which still makes it difficult to achieve high-precision hybrid plant identification.

[0008] To achieve the above objectives, the present invention provides a method for locating hybrid plants during the rapeseed seedling stage, comprising the following steps:

[0009] Obtain an image containing multiple seedlings in the plot;

[0010] Identify all seedlings in the image to obtain plant images, and configure index information for each plant image;

[0011] Features of each plant image are extracted based on high-dimensional features of a preset dimension to obtain a feature vector.

[0012] Calculate the Euclidean distance between any two eigenvectors and determine the outlier index for each eigenvector;

[0013] A preset outlier index threshold is used to filter plant images corresponding to feature vectors whose outlier index exceeds the preset outlier index threshold.

[0014] The location information of the hybrid plants is obtained by determining their location information based on the index information of the screened plant images.

[0015] As a further improvement of the present invention, the step of identifying all seedlings in the image to obtain a plant image includes: identifying the seedlings in the image using a preset seedling plant model, and determining a plant image of appropriate size based on the identified seedlings.

[0016] As a further improvement of the present invention, before the step of extracting features from each plant image based on high-dimensional features of a preset dimension to obtain feature vectors, the method further includes: standardizing the size of each plant image so that the size of each plant image is consistent.

[0017] As a further improvement of the present invention, before the step of extracting features of each plant image based on high-dimensional features of a preset dimension to obtain feature vectors, the method further includes: distinguishing between the foreground plant part and the background plant part of each plant image in order to retain the foreground plant part.

[0018] As a further improvement of the present invention, the step of extracting features from each plant image based on high-dimensional features of a preset dimension to obtain a feature vector includes the following sub-steps:

[0019] Pre-deploy multiple models of different orders of magnitude;

[0020] Based on actual computing power, one or more orders of magnitude models are dynamically invoked to simultaneously extract features from multiple plant images to obtain feature vectors.

[0021] As a further improvement of the present invention, the step of calculating the Euclidean distance between any two feature vectors and determining the outlier index of each feature vector includes the following sub-steps:

[0022] Calculate the Euclidean distance between each eigenvector and other eigenvectors to form a feature matrix, where each row of the feature matrix corresponds to an eigenvector.

[0023] The outlier index of the eigenvector is obtained by summing each row of the feature matrix.

[0024] As a further improvement of the present invention, the step of screening plant images corresponding to feature vectors whose outlier index exceeds a preset outlier index threshold includes the following sub-steps:

[0025] Obtain the planting ratio of each variety of seedlings in the plot;

[0026] The outlier threshold for the variety to be identified is determined based on the planting proportion of the variety to be identified.

[0027] Plant images corresponding to feature vectors whose outlier index exceeds the outlier index threshold of the variety to be identified are selected as images of the variety to be identified.

[0028] As a further improvement of the present invention, the step of configuring index information for each plant image includes the following sub-steps: configuring index information for each plant image of each image, wherein the index information includes the image image ID and the plant image number corresponding to the plant image.

[0029] As a further improvement of the present invention, the high-dimensional features need to be extracted by training the model through an optimized loss function in order to minimize the Euclidean distance between similar features and maximize the Euclidean distance between dissimilar features.

[0030] This invention also discloses a method for removing weeds during the rapeseed seedling stage, based on the above-mentioned method for locating weeds during the rapeseed seedling stage, comprising the following steps:

[0031] Set a land parcel area threshold and compare the current land parcel area with the threshold.

[0032] If the current plot area is smaller than the plot area threshold, an unmanned vehicle is used to collect images in real time to obtain the location information of weeds and remove them in real time.

[0033] If the current plot area is larger than the plot area threshold, multiple images covering the entire current plot are collected by drone to obtain the location information of weeds. A path is planned based on the location information of the weeds to achieve remote removal.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] First, images containing multiple plants in the plot are acquired. Then, plants in the seedling stage are identified in the images to obtain plant images containing only one plant. High-dimensional features of the plant images are extracted to obtain feature vectors. The Euclidean distance between any two feature vectors is calculated, and the outlier index of each plant is obtained through the Euclidean distance. Each calculated outlier index is compared with a preset outlier index threshold. Plant images corresponding to outlier indices with values ​​greater than the outlier index threshold are selected. The plant corresponding to the plant image is determined to be a hybrid plant. The location information of the selected plant images is determined based on the index information of the selected plant images. This location information is the location of the hybrid plant. Compared to existing technologies that require prior knowledge of the hybrid variety before identification, and that the variety must exist in a database, which necessitates uploading massive amounts of variety images and providing extensive feature data for each variety, still falls short of the 99.7% purity requirement for hybrid breeding, this invention extracts high-dimensional features from each plant in the image, calculates the Euclidean distance between feature vectors to obtain an outlier index, and compares the outlier index with a preset outlier index threshold. When the outlier index exceeds the threshold, the plant corresponding to that outlier index is identified as a hybrid, without prior knowledge of the hybrid variety. Furthermore, this invention does not require the hybrid to exist in a database for identification, and the outlier index-based method improves efficiency and identification accuracy compared to uploading numerous features to a database, effectively meeting the 99.7% purity requirement for hybrid breeding. Due to the index information configuration, when a hybrid plant image is identified, its location is directly pinpointed based on the index information for subsequent removal. Attached Figure Description

[0036] Figure 1 This is a flowchart of the hybrid identification and location process in the seedling stage hybrid location method of the present invention;

[0037] Figure 2 This is a flowchart illustrating the specific steps in step S2 of the present invention.

[0038] Figure 3 This is a flowchart illustrating the specific steps in step S3 of the present invention.

[0039] Figure 4 This is a flowchart illustrating the specific steps in step S4 of the present invention.

[0040] Figure 5 This is a flowchart illustrating the specific steps in step S5 of the present invention.

[0041] Figure 6 This is a flowchart of the method for removing weeds during the rapeseed seedling stage in this invention;

[0042] Figure 7 This is a flowchart illustrating steps S0 and S7 in this invention.

[0043] Figure 8 This is a schematic diagram of the impurity removal operation instructions in the invention. Detailed Implementation

[0044] The present invention will now be described in detail with reference to the embodiments shown in the accompanying drawings. However, it should be noted that these embodiments are not intended to limit the present invention. Equivalent changes or substitutions in function, method, or structure made by those skilled in the art based on these embodiments are all within the scope of protection of the present invention.

[0045] In summary, the rapeseed seedling hybrid location method disclosed in the various embodiments of this application can be used to identify and locate hybrid plants in a field during the seedling stage of hybrid crops. Taking rapeseed as an example, in the same hybridization field, the variety used for hybridization is usually one variety or two designated varieties as the male and female parents. Other varieties appearing in the field besides the one or two varieties used for hybridization are considered hybrids. To ensure the purity of hybrid seeds, it is necessary to remove hybrids from the field. Rapeseed hybridization includes seedling stage hybridization and bud stage hybridization. Seedling stage hybridization uses plant characteristics such as seedling appearance, leaf morphology, leaf color, and leaf margin as standards to distinguish hybrids for removal. The difference between bud stage hybridization and seedling stage hybridization is that seedling appearance characteristics are replaced by flower bud morphology characteristics, and leaf morphology characteristics are replaced by plant type characteristics. Compared with seedling appearance characteristics and leaf morphology characteristics, flower bud morphology characteristics and plant type characteristics are more difficult to distinguish. Therefore, seedling stage hybridization is a more commonly used hybridization method in hybrid seed production. However, even though identifying hybrid plants during the seedling stage is less difficult than identifying them during the bud stage, existing methods for identifying hybrid plants still require prior knowledge of the hybrid plant varieties in the plot and ensuring that all hybrid plant varieties exist in the database. Therefore, the database needs to upload a massive number of crop plant variety photos, and as new crop varieties are added, the database needs to be maintained in real time, uploading new variety plant photos. During seedling identification, features such as seedling appearance, leaf morphology, leaf color, and leaf margins need to be identified. To improve accuracy, the database needs to upload a massive number of photos containing these features for each variety to support feature identification. Creating such a database is extremely difficult, and even if it were implemented, it would still be difficult to cover all plant characteristics. When hybrid plant characteristics are not present in the database, hybrid plant identification becomes difficult, leading to incomplete identification and removal of hybrid plants, thus failing to achieve the 99.7% seed purity required for hybrid seed production.

[0046] The rapeseed seedling hybrid location method disclosed in this application identifies seedlings in the plot beforehand and captures images of each plant. Index information is configured for each plant image. Feature vectors of high-dimensional features are extracted from each plant image, and the Euclidean distance between any two feature vectors is calculated. The outlier index corresponding to each plant image is calculated using the obtained Euclidean distance. The outlier index is compared with a preset outlier index threshold to determine whether the corresponding plant image is a hybrid. Compared to existing seedling hybrid removal methods, this hybrid location method, which extracts high-dimensional features and calculates outliers using the extracted feature vectors, eliminates the need for prior knowledge of the hybrid variety or for the hybrid variety to exist in a database, effectively reducing the types of variety images that need to be uploaded to the database. The method of extracting feature vectors to calculate the outlier index effectively improves the accuracy of hybrid identification. Simultaneously, it further reduces the number of feature images required for each variety in the database, reducing the model training workload and making it suitable for practical applications. Furthermore, once hybrid plants are identified, the index information configured in the plant image facilitates the fastest possible location of the identified hybrid plants for subsequent removal, thereby improving the efficiency of hybrid plant removal and further enhancing the purity of seed production.

[0047] For a specific implementation of the method for locating hybrid plants in rapeseed seedlings (hereinafter referred to as the locating method) disclosed in this invention, please refer to... Figures 1 to 5 The illustrated process for identifying and locating hybrid plants includes steps S1 to S6. This location method aims to identify each seedling in an image containing multiple seedlings, obtaining a plant image for each identified plant, and configuring index information for each plant image for subsequent location. High-dimensional features of each plant image are extracted, and Euclidean distance is calculated using the obtained feature vectors to obtain the outlier index corresponding to each plant image. The outlier index is compared one by one with a preset outlier index threshold. When the outlier index exceeds the threshold, it indicates a significant difference between the plant image corresponding to that outlier index and other plant images. Therefore, the plant in the image corresponding to that outlier index is considered a hybrid. The location of the filtered impurities is determined using the pre-configured index information for subsequent removal. Compared to traditional hybrid identification methods, this application obtains the outlier index by calculating Euclidean distance. By comparing the outlier index with a preset outlier index threshold, it can be determined whether the plant image is a hybrid. It can identify and judge hybrid plants in the seedling stage with high accuracy without the need to establish a full crop variety database, so as to meet the accuracy requirement of more than 99% for hybrid breeding.

[0048] It should be noted that the rapeseed seedling hybrid identification method of this application is used for removing hybrids from seedlings in plots used for hybrid breeding, including but not limited to plots planted with rapeseed and other crops. The crops planted in the hybrid breeding plots are usually one or two specific varieties. Any plants appearing in the plot other than the specific variety are considered hybrids. Hybrids are usually seeds of other varieties or weeds mixed with the seeds of the specific variety during planting (weeds do not affect the purity of the breeding). Therefore, the number of hybrids in the plot will be significantly less than the number of specific variety plants used for breeding. Based on this situation, the method of this application includes the following steps:

[0049] S1. Obtain images containing multiple seedlings in the plot. In step S1, mobile equipment with shooting capabilities (including drones, mobile vehicles, etc.) is used to photograph the seedlings in the plot, resulting in multiple images containing multiple rows of seedlings of the main planting type in the plot. This step prepares the conditions for subsequent identification and location of weeds.

[0050] S2. Identify all seedlings in the image to obtain plant images, and configure index information for each plant image. In this embodiment, step S2 is implemented based on a deep learning target detection model trained on the YOLOv5 architecture. It uses low-dimensional feature recognition and labeling to perform a first-stage identification and screening of all identified targets (i.e., plants) in the plot without distinguishing between varieties. The resulting images are sets containing specific varieties or hybrid plants. This variety-indiscriminate seedling identification and screening achieves an accuracy of over 99.7% with a relatively small training set, and effectively reduces the learning pressure and the number of identifications required for subsequent accurate identification of hybrid plants, thereby improving the accuracy of hybrid plant identification. A first model is established to identify all seedlings in the image obtained in step S1. Then, each identified plant is cropped into a plant image. Index information is established between the plant image and the image obtained in S1, so that after identifying the location of hybrid plants, it can be associated with the image and directly located in the plot for removal.

[0051] S3. Extract features from each plant image based on high-dimensional features of a preset dimension to obtain a feature vector. High-dimensional features are extracted from all seedling images captured in step S2 using a convolutional neural network, with one high-dimensional feature extracted from each image. As an example, in this implementation, the high-dimensional feature can be a 1*2048 feature vector. The dimension value can be freely set according to the accuracy and speed requirements of the computing platform during actual use. Using a smaller dimension value can improve computing speed, but at the cost of some accuracy. Conversely, using a larger dimension value will sacrifice computing speed to improve recognition accuracy. A dimension of 2048 is a value obtained after multiple experiments that yields high computational accuracy and appropriate computational speed.

[0052] S4. Calculate the Euclidean distance between any two feature vectors and determine the outlier index of each feature vector. Extract high-dimensional features from each plant image in step S3 to obtain feature vectors. Calculate the Euclidean distance between any two feature vectors and form a similarity matrix. The sum of the Euclidean distances in each row of the similarity matrix is ​​the outlier index of the corresponding plant image. Use the obtained outlier index value to determine whether the corresponding plant image is a hybrid.

[0053] S5. Set a preset outlier index threshold and filter plant images corresponding to feature vectors whose outlier index exceeds the preset outlier index threshold. The outlier threshold is set in two ways: when the plot contains only one variety and when it contains two varieties for hybridization. When the plot contains only seedlings of one variety, the outlier threshold value is dynamically adjusted according to actual needs. When the plot contains hybrids of two varieties, assuming the two varieties are A and B, the planting ratio of A and B in the plot needs to be known in advance in the step of setting the outlier threshold. When the planting ratio of A is 80% and the planting ratio of B is 20%, the outlier threshold for B is identified as 0.2±0.05 (error value). Since the planting ratio of B at 20% is lower than that of A at 80%, B has a larger outlier value than A. Therefore, the outlier threshold is set based on the outlier value of B. When the outlier calculated in step S6 is greater than the preset outlier threshold of 0.2±0.05, the plant image corresponding to the outlier index can be determined to be a hybrid.

[0054] S6. Determine the location information of the selected plant images based on their index information to obtain the location information of the hybrid plants. By using the plant images obtained in step S5, and combining them with the index information established between each plant image and the image in step S2, the location of the hybrid plants in the plot can be accurately located, so that the hybrid plants can be accurately removed. This achieves the goal of removing hybrids from the hybrid breeding plot and improving the breeding purity of the hybrid breeding plot, in order to meet the requirement of a purity of over 99% for hybrid breeding.

[0055] Combination Figure 2 As shown, in this embodiment, step S2 includes sub-steps S21 to S25:

[0056] S21. Identify seedling plants in the image using a pre-defined seedling plant model. First, preliminary dataset preparation is required. In this embodiment, field images of nearly 1000 different rapeseed varieties are used, and the rapeseed plants are labeled to establish a dataset for detecting rapeseed seedlings. The pre-defined seedling plant detection model employs a deep learning algorithm. Using the aforementioned dataset for detecting rapeseed seedlings, and trained with the YOLOv5 object detection pipeline, this model can detect the location coordinates of all seedling plants in the image obtained in step S1.

[0057] S22. Determine a plant image of appropriate size based on the identified seedlings. The image obtained in step S1 contains images of multiple rapeseed seedlings. After identifying and detecting the location of all seedlings using a preset seedling model in step S21, step S22 extracts all identified seedlings into a single plant image.

[0058] S23. Standardize the size of each plant image to ensure that all plant images are the same size. Resize all identified plant images to the same size for subsequent high-dimensional feature extraction.

[0059] S24. Distinguish between the foreground and background plant parts in each plant image to retain the foreground plant part. Taking rapeseed as an example, the background of its seedlings is always the soil / ground, so the background is fixed. Rapeseed seedlings are small and not enough to cover the ground. Therefore, in step S25, the image processing method distinguishes between the foreground rapeseed plant part and the background ground part, and removes the background part, retaining only the foreground part, which effectively improves the accuracy of subsequent feature extraction.

[0060] S25. Configure index information for each plant image in each image. The index information includes the image ID and plant image number corresponding to the plant image. In step S22, all plant images are captured, and each plant image is numbered according to its detection ID. An index is established between the number and the image ID of the plant image. When a plant image is subsequently detected as a hybrid plant, the position of the hybrid plant in the image can be found through the index information between the number and the image ID, so that the hybrid plant can be removed in subsequent steps.

[0061] Combination Figure 3 As shown, in this embodiment, step S3 includes sub-steps S31 to S32:

[0062] S31. Pre-deploy multiple models of different orders of magnitude. Estimate the machine's computing power and deploy multiple models of different orders of magnitude simultaneously, including: 32 models, 16 models, and 1 model. One or more models can be dynamically called according to the actual computing power. Since all plant images are resized to the same size in step S23, the multiple models deployed in this step can extract / process the features of a batch of plant images at once. If step S23 is omitted, the plant images will be of different sizes, and only one plant image can be processed at a time, resulting in wasted computing power.

[0063] S32. Dynamically call one or more orders of magnitude models based on actual computing power to simultaneously extract high-dimensional features from multiple plant images to obtain feature vectors. Train the high-dimensional feature extraction model through an optimized loss function, minimizing the Euclidean distance between similar feature vectors and maximizing the Euclidean distance between dissimilar feature vectors. Taking a total of 32 plant images as an example, the plant image size is fixed in step S23, and the 32-image model can be directly called to process 32 plant images in batches, effectively improving feature extraction efficiency and extracting feature vectors from 32 plant images at once. If 33 plant images need to be processed, the following two methods can be used: call the 32-image model to process 32 plant images at once, and fill the remaining image with 31 blank images and continue calling the 32-image model; or call the 32-image model and the 1-image model simultaneously, with the 32-image model batch processing 32 plant images and the 1-image model processing the remaining image. In step S32, the number of plant images processed is dynamically adjusted according to the computing power of the deployment environment, usually occupying 80%-90% of the device's computing power, reserving computing power for other programs, so as to fully utilize the device's computing power. Through the above feature extraction, the feature vector of each plant image is obtained, which is used to calculate the outlier index to identify hybrid plants.

[0064] The loss function mentioned above is:

[0065] Where L is the loss value. Let A1 be the Euclidean distance between A0 and A0. Let α be the Euclidean distance between B and A0, and let α be the minimum Euclidean distance correction coefficient between B and A0. This value α is typically set around 0.3 and dynamically adjusted according to actual conditions. When training the high-dimensional feature extraction model, three seedling images are read in each time. Two of these are images of the same variety, A0 and A1, and the other is an image of a seedling of variety B. To distinguish between different varieties, the model needs to be trained using the aforementioned loss function and the rapeseed seedling dataset from step S21, differentiating between two different varieties each time. In this embodiment, we take seedlings of variety A and variety B as examples. During one training process, images of A0 and A1 are read from the variety A dataset (hereinafter referred to as IDA), and one image of variety B (hereinafter referred to as B image) is read from the variety B dataset (hereinafter referred to as IDB). IDA and IDB each contain approximately 100 images of rapeseed seedlings of the corresponding variety for feature extraction. This implementation uses a ReID-type feature extraction network with ResNet50 and triple loss. First, images A0, A1, and B are read, and the ResNet50 features of these three images are extracted using the model. Then, subsequent steps calculate the Euclidean distance between A0 and A1, and between A0 and B, respectively. The triple loss is calculated and backpropagated to the training model to optimize the parameters in the ResNet50 feature extraction network. After multiple rounds of training and iteration, the model is trained until the Euclidean distance between the feature vectors of seedlings of the same variety approaches 0, and the Euclidean distance between the feature vectors of seedlings of different varieties approaches 1. The seedling rapeseed plant dataset in step S21 also includes 10 datasets such as IDC, IDE, and IDF, containing field images of 1000 different rapeseed varieties. Each dataset is trained sequentially using the above steps.

[0066] Combination Figure 4 As shown, in this embodiment, step S4 includes sub-steps S41 to S42:

[0067] S41. Calculate the Euclidean distance between each feature vector and other feature vectors to form a feature matrix, where each row of the feature matrix corresponds to a feature vector. Suppose the image contains m seedlings, so m plant images are obtained through identification. After calculating the Euclidean distance between any two feature vectors, a feature matrix consisting of m*m Euclidean distances can be obtained. Each high-dimensional feature needs to be calculated with itself once, and the resulting value is 0; the remaining Euclidean distance values ​​are floating-point numbers that are close to 0 or close to 1.

[0068] S42. Summing each row of the feature matrix yields the outlier index of the eigenvector. Summing each row of the feature matrix from step S41 yields an outlier index table consisting of m values.

[0069] Combination Figure 5 As shown, in this embodiment, step S5 includes sub-steps S51 to S53:

[0070] S51. Obtain the planting ratio of each variety in the seedling stage of the plot. In the case of two rapeseed varieties, A and B, where the planting ratios of varieties A and B are not the same, the rapeseed variety with a smaller planting ratio has a larger outlier index than the rapeseed variety with a larger planting ratio. Therefore, it is necessary to know the planting ratio of the two varieties in the plot in advance.

[0071] S52. Determine the outlier threshold for the variety to be identified based on its planting proportion. When the entire plot is planted with the same variety of rapeseed, the outlier threshold is usually set at around 0.3, and dynamically adjusted according to the actual situation. When the plot includes both variety A and variety B rapeseed, the planting proportions of variety A and variety B in the plot need to be known in advance when setting the outlier threshold. When the planting proportion of variety A is 80% and the planting proportion of variety B is 20%, the outlier threshold for identifying variety B is 0.2 ± 0.05 (error value). Since the planting proportion of variety B at 20% is lower than that of variety A at 80%, variety B will have a larger outlier value than variety A after calculation. Therefore, the outlier threshold is set based on the outlier value of variety B.

[0072] S53. Select plant images corresponding to feature vectors whose outlier indices exceed the outlier index threshold of the variety to be identified as images of the variety to be identified. Compare the m values ​​in the outlier index table with the outlier index threshold set in S52. When the outlier index is greater than the set outlier index threshold, the plant image corresponding to the outlier index threshold can be determined as an image of the variety to be identified, and it can be identified as a hybrid plant. This allows for subsequent step S6, where the position of the plant image on the image can be located based on the index relationship between the plant image number and the image ID, so that the hybrid plant can be located and removed.

[0073] To assess the model's accuracy in identifying and locating hybrid plants in the seedling stage, a precision test was conducted: the numbers in the test set IDs indicate the number of plant images, and the variety of each plant image is known. Using the model training method described above, the test set contains 10 categories (i.e., 10 varieties), each category has 10 IDs, and each ID contains 10 images. Images of the same plant from the same plot are placed under the same ID, while images of the same plant from different plots are placed under different IDs but within the same category (introducing images of the same plant from other plots increases the testing difficulty and accuracy). Similarity is tested between any two images in the test set, resulting in a 1000*1000 similarity matrix. The detection result is categorized as similar or dissimilar. The detection result is compared with known results; identical results are 1, and dissimilar results are 0. The accuracy is calculated as: Accuracy = Number of identical results / 1000 * 1000.

[0074] Table 1: Accuracy Testing Table

[0075]

[0076] As shown in Table 1, the accuracy of this detection method can reach 99.9% for a single dataset containing only ten different varieties in the same plot. For hybrid breeding plots that need to identify and remove hybrids, which contain only two different varieties of plants used for hybridization, and usually no more than five hybrids, the detection rate of hybrids is 99.7%-99.9%, which is significantly improved compared to the 60%-70% detection rate of existing technologies. It also effectively reduces the training load and can be put into practical application.

[0077] This invention also discloses a method for removing weeds during the rapeseed seedling stage, based on the weed identification and location process used to implement the location method in the above specific embodiments, combined with... Figure 6 As shown, it includes a preliminary step S0 and a subsequent step S7.

[0078] S0. Determine the mobile work equipment that performs image acquisition in step S1.

[0079] S7. Perform the removal of impurities based on the position of the impurities in the image as described in step S6 above.

[0080] Combination Figure 6 As shown, in this embodiment, step S0 includes the following sub-steps:

[0081] S01. Set a plot area threshold and compare the current plot area with the threshold. When the current plot area is small, select mobile equipment for image acquisition and real-time weed removal as needed to improve weed removal efficiency.

[0082] S021. If the current plot area is less than the plot area threshold, an unmanned vehicle (UAV) is used to collect images in real time. When it is determined that the current plot area is less than the set plot area threshold, an UAV is used as a mobile operating device to collect images. The UAV moves a certain distance within the current plot and then stops, capturing images containing some of the plants within the plot. When using the UAV for image collection, the aforementioned weed identification and positioning system is directly mounted on the UAV, enabling real-time identification of the weed location after image acquisition. Since the current plot area is less than the plot area threshold, the number of plants to be identified in the current plot is also relatively small. Therefore, the computing power of the UAV can support the aforementioned weed identification and positioning process to achieve the purpose of identifying and locating the weeds.

[0083] S022. If the current plot area is larger than the plot area threshold, a drone is used to collect multiple images covering the entire plot. When it is determined that the plot area is larger than the set plot area threshold, a drone is used as a mobile operating device to collect images. The drone flies at low altitude over the plot to collect multiple images covering the entire plot, and uploads the images to the device that performs the above-mentioned weed identification and location process, thereby identifying and locating the location of all weeds in the current plot at once. Because the current plot area is larger than the plot area threshold, the number of plants to be identified in the current plot is large, which is difficult for the unmanned vehicle's computing power to support, and real-time image acquisition and removal is too time-consuming. Therefore, a drone is used to collect images at once, and the weeds are identified and located by a device with greater computing power.

[0084] Combination Figure 7 As shown, in this embodiment, step S7 includes the following sub-steps:

[0085] S711. The unmanned vehicle moves in real time to the location of the weeds as defined in step S6 above to remove the weeds. Since the location of the weeds is associated with and located on the image map, and the image map itself contains RGBD information, the location of the weeds in the current plot can be directly located based on the location of the weeds in the image map, and a path can be constructed to allow the unmanned vehicle to move to the location of the weeds to remove them.

[0086] S712. Mark the locations of all weeds in the image to obtain a weed removal operation instruction map. Remove weeds according to their locations in the weed removal operation instruction map. Reconstruct and stitch together all the image images obtained in step S022 into a complete map. This complete map marks the locations of all weeds, which is the weed removal operation instruction map. Since weed image index information was established in step S25 of the above weed identification and positioning process, taking a total of 10 image images as an example, each image has 20 targets. When the 10th plant in the third image is a weed, its ID in the image image is 3-10, and the image number of the plant it belongs to is 50. When it is determined that the image numbered 50 is a weed, backtracking to IDs 3-10 in the image image, the position of the weed in the image image can be directly located, and it can be marked in the image image by marking it with a red dot or other methods. Since the image image itself contains RGBD information, the position of the weed in the current plot can be directly located based on the position of the weed in the image image. Based on the obtained cleaning operation instruction map, cleaning can be carried out manually or by machine. If machine cleaning is adopted, the machine's movement path in the current plot needs to be planned.

[0087] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

[0088] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0089] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for locating hybrid plants in rapeseed seedlings, characterized in that, Includes the following steps: Obtain an image containing multiple seedlings in the plot; Identify all seedlings in the image to obtain plant images, and configure index information for each plant image; The process of extracting features from each plant image based on high-dimensional features of a preset dimension to obtain a feature vector includes the following sub-steps: Pre-deploy multiple models of different orders of magnitude; Based on actual computing power, one or more orders of magnitude models are dynamically invoked to simultaneously extract features from multiple plant images to obtain feature vectors; Calculate the Euclidean distance between any two eigenvectors and determine the outlier index for each eigenvector; A preset outlier index threshold is used to filter plant images corresponding to feature vectors whose outlier index exceeds the preset outlier index threshold. The location information of the hybrid plants is obtained by determining their location information based on the index information of the screened plant images.

2. The method for locating hybrid plants in rapeseed seedlings according to claim 1, characterized in that, The steps of identifying all seedlings in the image to obtain plant images include: identifying the seedlings in the image using a preset seedling plant model, and determining a plant image of appropriate size based on the identified seedlings.

3. The method for locating hybrid plants in rapeseed seedlings according to claim 2, characterized in that, Before the step of extracting features from each plant image based on high-dimensional features of a preset dimension to obtain feature vectors, the method also includes: standardizing the size of each plant image to make each plant image the same size.

4. The method for locating hybrid plants in rapeseed seedlings according to claim 2, characterized in that, Before the step of extracting features from each plant image based on high-dimensional features of a preset dimension to obtain feature vectors, the method further includes: distinguishing between the foreground and background plant parts of each plant image in order to retain the foreground plant part.

5. The method for locating hybrid plants in rapeseed seedlings according to claim 1, characterized in that, The steps for calculating the Euclidean distance between any two eigenvectors and determining the outlier index for each eigenvector include the following sub-steps: Calculate the Euclidean distance between each eigenvector and other eigenvectors to form a feature matrix, where each row of the feature matrix corresponds to an eigenvector. The outlier index of the eigenvector is obtained by summing each row of the feature matrix.

6. The method for locating hybrid plants in rapeseed seedlings according to claim 1, characterized in that, The steps for selecting plant images corresponding to feature vectors with outlier indices exceeding a preset outlier index threshold include the following sub-steps: Obtain the planting ratio of each variety of seedlings in the plot; The outlier threshold for the variety to be identified is determined based on the planting proportion of the variety to be identified. Plant images corresponding to feature vectors whose outlier index exceeds the outlier index threshold of the variety to be identified are selected as images of the variety to be identified.

7. The method for locating hybrid plants in rapeseed seedlings according to claim 1, characterized in that, The step of configuring index information for each plant image includes the following sub-steps: configuring index information for each plant image of each image, wherein the index information includes the image ID and plant image number corresponding to the plant image.

8. The method for locating hybrid plants in rapeseed seedlings according to claim 1, characterized in that, The high-dimensional features need to be extracted using an optimized loss function to train the model, in order to minimize the Euclidean distance between similar features and maximize the Euclidean distance between dissimilar features.

9. A method for removing unwanted plants during the rapeseed seedling stage, characterized in that... The method for locating hybrid plants in rapeseed seedlings according to any one of claims 1-8 includes the following steps: Set a land parcel area threshold and compare the current land parcel area with the threshold. If the current plot area is smaller than the plot area threshold, an unmanned vehicle is used to collect images in real time to obtain the location information of weeds and remove them in real time. If the current plot area is larger than the plot area threshold, multiple images covering the entire current plot are collected by drone to obtain the location information of weeds. A path is planned based on the location information of the weeds to achieve remote removal.

Images