Seed screening method and system based on laser radar technology

Abstract

The invention provides a seed screening method and system based on a laser radar technology, and relates to the technical field of agricultural automation, and the method comprises the steps: extracting geometric features and defect features from optimized point cloud data; extracting mildew and insect pest spectral features from the optimized spectral data; a multi-dimensional feature set is obtained; calculating seed plumpness, roundness and surface roughness based on the optimized point cloud data; calculating a sag index and a crack feature based on point cloud edge detection; fusing the geometric features, the defect features and the pest spectral features to generate a comprehensive score; a machine learning model is utilized to execute final grading, and a first probability predictor based on laser radar features and a second probability predictor based on pest spectral features are trained respectively; and inputting the probability outputs of the two predictors into a second-layer grading model of the meta-learning architecture, and outputting a final quality grade. According to the method, efficient and accurate seed screening is realized through automatic process and algorithm design.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural automation, in particular to a seed screening method and system based on laser radar technology. BACKGROUND

[0002] In the field of seed screening, some traditional screening methods rely on mechanical screening means. Traditional mechanical screening usually relies on some simple physical devices, such as screens with different pore sizes, gravity sorting machines, etc., to screen according to the basic physical characteristics of seeds, such as size and weight. For example, for large-diameter seeds such as corn seeds, the pore size difference of the screen is often used to separate seeds that do not meet the size requirements; through the gravity sorting machine, according to the different weights of the seeds, the lighter impurities or poorly developed seeds are separated out. However, this traditional method may have many limitations. During the screening process, since the mechanical parts are in direct contact with the seeds, it is easy to cause damage to the seeds, which may result in a high breakage rate; in addition, the traditional method has a limited scope of application, and is mostly suitable for large-diameter seeds, and may not be able to effectively screen small-diameter grass seeds, Chinese herbal medicine seeds, etc. At the same time, the traditional mechanical screening equipment may need to replace parts regularly due to frequent wear and tear of mechanical parts, resulting in high maintenance costs. SUMMARY

[0003] The technical problem to be solved by the present application is to provide a seed screening method and system based on laser radar technology, which realizes efficient and accurate seed screening through an automated process and algorithm design.

[0004] To solve the above technical problems, the technical solutions of the present application are as follows: In a first aspect, a seed screening method based on laser radar technology, the method comprising: Step S1: arranging the seeds in a single column on a conveyor belt with indentations through a vibrating feeder, suppressing the rolling of the seeds, generating a single-column ordered seed queue and conveying it to a detection area; Step S2: based on the single-column ordered seed queue, using high-pressure airflow to blow the seeds to remove impurities and unqualified seeds, and obtaining a clean seed queue; Step S3: synchronously scanning the clean seed queue: obtaining three-dimensional point cloud raw data of the seeds through 360° rotation scanning of the laser radar; and collecting spectral image raw data of the seeds through the hyperspectral camera; Step S4: denoising and filtering the three-dimensional point cloud raw data and the spectral image raw data to generate optimized point cloud data and optimized spectral data; Step S5: based on the optimized point cloud data, segmenting and identifying the boundaries of single seeds, marking the adhered seeds, and outputting an independent seed unit dataset; Step S6: Extracting geometric features and defect features from the optimized point cloud data; extracting moldy and insect-damaged spectral features from the optimized spectral data; to obtain a multi-dimensional feature set; Step S7: Calculating seed plumpness, roundness, and surface roughness based on the optimized point cloud data; calculating indentation index and crack features based on point cloud edge detection; fusing geometric features, defect features, and insect-damaged spectral features to generate a comprehensive score; using a machine learning model to perform final grading, training a first probability predictor based on laser radar features and a second probability predictor based on insect-damaged spectral features; inputting the probability outputs of the two predictors into the second layer grading model of the meta-learning architecture to output the final quality grade.

[0005] Further, step S1: arranging the seeds in a single column on the transmission belt with indentations through the vibration feeder, suppressing seed rolling, generating a single-column ordered seed queue, and conveying it to the detection area, including: Based on the disordered seed queue output by the vibration feeder, the seed rolling is suppressed by the indentation structure set on the surface of the transmission belt, a single-column ordered seed queue is generated, and it is conveyed to the detection area at a constant speed.

[0006] Further, step S2: based on the single-column ordered seed queue, using high-pressure airflow to blow the seeds to remove impurities and unqualified seeds, obtaining a clean seed queue, including: Based on the single-column ordered seed queue, set the blowing parameters; Based on the set blowing parameters, remove impurities and unqualified seeds through a pressure-adjustable high-pressure airflow blowing device to obtain a clean seed queue.

[0007] Further, step S3: synchronously scanning the clean seed queue: obtaining seed three-dimensional point cloud raw data through 360° rotation scanning of the laser radar; collecting seed spectral image raw data through the hyperspectral camera, including: Based on the clean seed queue, based on the disordered seed queue output by the vibration feeder, obtain the device parameters; Based on the device parameters, perform synchronous scanning: the laser radar is driven by a rotating motor to realize 360° scanning, obtaining seed three-dimensional point cloud raw data; the hyperspectral camera collects seed spectral image raw data within the complete cycle of the seed passing through the detection area.

[0008] Further, step S4: denoising and filtering the three-dimensional point cloud raw data and spectral image raw data to generate optimized point cloud data and optimized spectral data, including: Based on the three-dimensional point cloud raw data, using radius filtering to remove outliers, generating baseline point cloud data; Based on the spatial coordinate information of the reference point cloud data, the original spectral image data is registered and optimized: the spatial mapping relationship of the spectral image is established according to the spatial coordinate information; the Gaussian filter is used to eliminate the environmental light interference, and the optimized spectral data with a signal-to-noise ratio ≥ 35 dB is generated; and the spatial registration parameters are synchronously output; Based on the spatial registration parameters, the reference point cloud data is subjected to density enhancement processing, and the optimized point cloud data with a resolution ≤ 0.1 mm is generated.

[0009] Further, step S5: based on the optimized point cloud data, the single-seed boundary is segmented and identified, the adhered seeds are marked, and the independent seed unit data set is output, including: Based on the optimized point cloud data, the single-seed boundary is identified by using the Euclidean clustering segmentation algorithm, and the preliminary segmentation result containing the adhered area is generated; Based on the preliminary segmentation result, the adhered area with a point cloud distance less than 80% of the minimum diameter of the seed is marked, and the independent seed unit data set with topological relationship is output.

[0010] Further, step S6: geometric features and defect features are extracted from the optimized point cloud data; moldy and insect-damaged spectral features are extracted from the optimized spectral data; and a multi-dimensional feature set is obtained, including: Based on the independent seed unit data set, the volume, aspect ratio and surface curvature distribution are extracted from the optimized point cloud data to generate a geometric feature set; the recess depth and crack direction are extracted by point cloud normal mutation detection to generate a defect feature set; Based on the spatial positioning information of the independent seed unit data set, the reflectivity of the 420-510 nm wave band is extracted from the optimized spectral data as the moldy feature, and the absorbance of the 680-780 nm wave band is extracted as the insect-damaged feature, and the spectral feature set composed of the moldy feature and the insect-damaged feature is output; Based on the geometric feature set, the defect feature set and the insect-damaged spectral feature set, a multi-dimensional feature set is obtained.

[0011] Further, step S7: based on the optimized point cloud data, the seed fullness, roundness and surface roughness are calculated; based on the point cloud edge detection, the recess index and crack feature are calculated; the geometric feature, the defect feature and the insect-damaged spectral feature are fused to generate a comprehensive score; the machine learning model is used to perform final grading, and a first probability predictor based on the laser radar feature and a second probability predictor based on the insect-damaged spectral feature are trained; the probability outputs of the two predictors are input into the second layer grading model of the meta-learning architecture, and the final quality grade is output, including: Based on the multi-dimensional feature set, the grading is performed: based on the optimized point cloud data, the seed fullness, roundness and surface roughness are calculated, the fullness is calculated based on the volume deviation of the point cloud and the standard volume; The concave index and crack feature are calculated based on point cloud edge detection. The crack feature is calculated by an edge detection algorithm to calculate the percentage of the total length of the crack to the surface area. The geometric feature weight is set to 0.4, the defect feature weight is 0.3, and the pest spectrum feature weight is 0.3 to generate a comprehensive score. The final grading is performed by a machine learning model: a first probability predictor based on the laser radar feature and a second probability predictor based on the pest spectrum feature are trained respectively; the meta-learning architecture integrates the probability outputs of the two predictors by a Stacking integrator, and the final grading model is XGBoost.

[0012] In a second aspect, a seed screening system based on laser radar technology comprises: A conveying module is configured to arrange the seeds in a single column on a conveying belt with concaves by a vibrating feeder, suppress the rolling of the seeds, generate a single-column ordered seed queue, and transport the seed queue to a detection area; An obtaining module is configured to obtain a clean seed queue by blowing the seeds with high-pressure airflow based on the single-column ordered seed queue to remove impurities and unqualified seeds in terms of shape; A scanning module is configured to synchronously scan the clean seed queue: obtain three-dimensional point cloud raw data of the seeds by 360° rotation scanning of a laser radar; and obtain spectral image raw data of the seeds by a hyperspectral camera; A preprocessing module is configured to perform denoising and filtering processing on the three-dimensional point cloud raw data and the spectral image raw data to generate optimized point cloud data and optimized spectral data; An identification module is configured to segment and identify the boundary of a single seed based on the optimized point cloud data, mark the adhered seeds, and output an independent seed unit dataset; An extraction module is configured to extract geometric features and defect features from the optimized point cloud data, and extract moldy and pest spectrum features from the optimized spectral data to obtain a multi-dimensional feature set; A processing module is configured to calculate seed fullness, roundness, and surface roughness based on the optimized point cloud data, calculate a concave index and a crack feature based on point cloud edge detection, fuse the geometric features, the defect features, and the pest spectrum features to generate a comprehensive score, and perform final grading by a machine learning model: a first probability predictor based on the laser radar feature and a second probability predictor based on the pest spectrum feature are trained respectively; the probability outputs of the two predictors are input into a second-level grading model of a meta-learning architecture to output a final quality grade.

[0013] In a third aspect, a computing device comprises: One or more processors; A storage device configured to store one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0014] In a fourth aspect, a computer readable storage medium stores a program which, when executed by a processor, implements the method.

[0015] The above scheme of the present application at least has the following beneficial effects: By fusing vibration feeding, laser radar, hyperspectral imaging and machine learning technologies, the seed screening efficiency is improved. In the pretreatment stage, the vibration feeder cooperates with the notched conveyor belt to realize single-column ordered transmission of seeds, high-pressure airflow is used to remove impurities and unqualified seeds, and invalid detection is reduced. In data acquisition and processing, laser radar and hyperspectral camera are used in multi-modal cooperation, combined with denoising, registration and other operations, to comprehensively capture the shape and internal characteristics of the seeds and construct a multi-dimensional evaluation system. In the intelligent grading link, multiple models are used for cooperative decision-making to output quantitative indexes, the whole process is automated, the system parameters are adjustable, and the system can be compatible with multiple crop seed detection. In addition, the non-contact design reduces maintenance costs, and the system can be continuously optimized through machine learning, and can detect subtle defects with high precision and early warning of quality problems. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of a seed screening method based on laser radar technology provided by an embodiment of the present application.

[0017] Figure 2 is a schematic diagram of a seed screening system based on laser radar technology provided by an embodiment of the present application. DETAILED DESCRIPTION

[0018] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0019] As Figure 1 shown, an embodiment of the present application proposes a seed screening method based on laser radar technology, which comprises the following steps: Step S1: arranging the seeds in a single column on a notched conveyor belt through a vibration feeder to suppress the rolling of the seeds, generating a single-column ordered seed queue and conveying it to a detection area; Step S2: based on the single-column ordered seed queue, using high-pressure airflow to blow the seeds to remove impurities and unqualified seeds, and obtaining a clean seed queue; Step S3: synchronized scanning of the clean seed queue: obtain the original data of the three-dimensional point cloud of the seed through 360° rotation scanning of the laser radar; obtain the original data of the spectral image of the seed through the hyperspectral camera; Step S4: denoising and filtering processing of the three-dimensional point cloud original data and the spectral image original data to generate optimized point cloud data and optimized spectral data; Step S5: based on the optimized point cloud data, segment and identify the single seed boundary, mark the adhered seeds, and output the independent seed unit data set; Step S6: extract geometric features and defect features from the optimized point cloud data; extract moldy and insect-damaged spectral features from the optimized spectral data; and obtain a multi-dimensional feature set; Step S7: based on the optimized point cloud data, calculate the seed fullness, roundness and surface roughness; based on the edge detection of the point cloud, calculate the indentation index and crack feature; fuse the geometric features, defect features and insect-damaged spectral features to generate a comprehensive score; use a machine learning model to perform final grading, train a first probability predictor based on the laser radar features and a second probability predictor based on the insect-damaged spectral features; input the probability outputs of the two predictors into the second layer grading model of the meta-learning architecture, and output the final quality grade.

[0020] In the embodiment of the present application, the vibration feeder and the notched conveyor belt are used to realize the ordered arrangement of the single seed, and the high-pressure airflow is used to remove impurities and unqualified seeds, laying a foundation for subsequent detection; the laser radar and the hyperspectral camera are used for multi-modal data acquisition, and combined with denoising, segmentation, feature extraction and other processing, a multi-dimensional seed quality evaluation system is constructed; the dual probability predictor and the meta-learning architecture are used to fuse the physical form and the inherent quality features, realize high-precision grading, and output quantitative indicators for easy management; through the automatic process and algorithm design, efficient and accurate seed screening is realized.

[0021] In a preferred embodiment of the present application, the above step S1: the vibration feeder is used to arrange the seeds in a single row on the notched conveyor belt, to suppress the rolling of the seeds, to generate a single-row ordered seed queue and to transport it to the detection area, which can include: Step S1-1, step S1-2, based on the disordered seed queue output by the vibration feeder, the notched structure arranged on the surface of the conveyor belt is used to suppress the rolling of the seeds, to generate a single-row ordered seed queue and to transport it to the detection area at a constant speed.

[0022] In the embodiment of the present application, the preliminary control of seed output is realized by adjusting the parameters of the vibrating feeder, which lays the foundation for subsequent orderly arrangement and can be flexibly adjusted according to the characteristics of the seeds to ensure stable and controllable output; the rolling of the seeds is inhibited by the concave marks of the conveyor belt to ensure the stable posture during detection and improve the accuracy of data acquisition; at the same time, the seeds are orderly arranged in a single column to avoid missed detection and improve the detection efficiency; and the seeds are conveyed at a constant speed to create stable conditions for the detection equipment.

[0023] In the embodiment of the present application, the specific steps include: Step S1-1, when the vibrating feeder is working, the relationship between the vibration frequency, amplitude and other parameters and the output flow and speed of the seeds needs to be calculated, a corresponding model of the vibration parameters and the output state of the seeds is established through preliminary experiments or empirical data, and the vibration frequency and amplitude are adjusted according to the characteristics of the seeds such as the type, size and weight to make the seeds output at a suitable flow and speed, and an unordered seed queue is initially formed.

[0024] Step S1-2, the shape (such as circular, square, etc.) and size (depth, diameter or side length, etc.) of the concave marks and the spacing between the concave marks are calculated according to the shape and size of the seeds; the size of the concave marks is slightly larger than the size of the seeds to ensure that the seeds can be embedded in the concave marks, and the spacing needs to be determined according to the speed of the conveyor belt and the working frequency of the detection equipment to ensure that the seeds have a suitable distribution density on the conveyor belt and at the same time ensure that the detection equipment has enough time to scan and analyze a single seed.

[0025] By monitoring the distribution of the seeds on the conveyor belt, the spacing between the seeds and the arrangement uniformity are calculated; the position of the seeds is detected in real time by using a sensor (such as a photoelectric sensor), and if it is found that the spacing between the seeds does not meet the requirements or the arrangement is not uniform, the output parameters of the vibrating feeder are adjusted or the speed of the conveyor belt is fine-tuned to gradually form a single-column and orderly queue of the seeds.

[0026] According to the working speed and processing capacity of the detection equipment, the constant speed of the conveyor belt is calculated; the speed of the conveyor belt is matched with the scanning speed and data processing speed of the detection equipment to ensure that the time for the seeds to stay in the detection area meets the detection requirements; at the same time, the speed of the conveyor belt is monitored in real time by using a speed sensor, and if the speed fluctuates, the speed of the driving motor of the conveyor belt is adjusted by using a closed-loop control system to maintain a constant speed.

[0027] In a preferred embodiment of the present application, the above step S2: based on the single-column and orderly seed queue, the seeds are swept by high-pressure airflow to remove impurities and unqualified seeds, and a clean seed queue is obtained, which can include: Step S2-1, based on the single-column and orderly seed queue, the sweeping parameters are set; Step S2-2, based on the set blowing parameters, remove impurities and unqualified seeds by the pressure-adjustable high-pressure airflow blowing device to obtain a clean seed queue.

[0028] In the embodiments of the present application, by understanding the characteristics of seeds and impurities, the blowing parameters are set specifically, which can avoid the misblowing of qualified seeds and the residue of impurities, and improve the accuracy of pretreatment; by the pressure-adjustable high-pressure airflow blowing device, the airflow pressure is accurately set, which can efficiently remove impurities and reduce the impact on qualified seeds, ensuring the pretreatment quality and efficiency, maintaining the blowing stability through real-time monitoring and adjustment, and continuously outputting high-quality seed queue.

[0029] In the embodiments of the present application, the specific steps include: Step S2-1, before pretreatment, the seed sample needs to be measured for basic parameters, including the weight, size, shape, density of seeds, and the type (such as debris, dust, shriveled seeds), weight ratio and size distribution of impurities; the physical characteristic data of seeds and impurities are collected through sampling detection and manual observation.

[0030] Step S2-2, according to the weight and density difference of seeds and impurities, combined with the principle of fluid mechanics, the airflow pressure range that can blow away impurities and unqualified seeds but retain qualified seeds is estimated; through preliminary experiments, a corresponding relationship model of different seed types and airflow pressure is established, the measured seed characteristic data is substituted into the model to preliminarily determine the pressure value; at the same time, considering the factors such as the speed of the conveying belt and the distance between seeds, the pressure value is fine-tuned to ensure that impurities and unqualified seeds can be effectively blown away when the seeds pass through the blowing area at a uniform speed.

[0031] According to the running speed of the conveying belt and the coverage length of the blowing device, the residence time of the seeds in the blowing area is calculated, which is used as the basis for the blowing time; at the same time, combined with the required blowing effect (such as completely removing impurities of a certain size and weight), the blowing time is adjusted through experimental verification; if the impurities are not removed completely, the blowing time is appropriately extended; if qualified seeds are overblown, the blowing time is shortened until the best blowing effect is achieved.

[0032] During the blowing process, the pressure value of the high-pressure airflow is monitored in real time by the pressure sensor, and compared with the set pressure value; if the pressure fluctuates, the output power of the airflow generating device (such as an air compressor) is fed back and adjusted to maintain the pressure stable; at the same time, according to the running state of the conveying belt and the conveying condition of the seeds, the working state of the blowing device is dynamically adjusted to ensure that each seed can be effectively blown.

[0033] In a preferred embodiment of the present application, the above step S3: synchronously scanning the clean seed queue: obtaining seed three-dimensional point cloud raw data through 360° rotary scanning of the laser radar; and obtaining seed spectral image raw data through the hyperspectral camera, can include: Step S3-1, based on the clean seed queue and the disordered seed queue output by the vibrating feeder, obtaining device parameters; Step S3-2, based on the device parameters, performing synchronous scanning: the laser radar is driven to rotate by a motor to realize 360° scanning, and obtain seed three-dimensional point cloud raw data; and the hyperspectral camera collects seed spectral image raw data within a complete cycle of the seed passing through the detection area.

[0034] In the embodiment of the present application, by reasonably setting the device parameters of the laser radar and the hyperspectral camera, the data is ensured to be complete and accurate to reflect the seed morphology and spectral characteristics, laying a solid data foundation for subsequent analysis; by using the 360° scanning of the laser radar, even if the seed rolls slightly, the morphology information can be captured in all directions, improving the completeness of the three-dimensional point cloud data; the hyperspectral camera completely collects the spectral image of the detection area, providing rich data for the analysis of the intrinsic quality of the seed; and the coaxial synchronous scanning of the two ensures the spatiotemporal consistency of the data, which facilitates the fusion of the morphology and spectral characteristics, improves the data utilization value and the accuracy of the seed quality evaluation.

[0035] In the embodiment of the present application, the specific steps include: Step S3-1, calculating the required sampling frequency of the laser radar and the hyperspectral camera according to the moving speed of the seed; in order to ensure that the seed at each position in the detection area can be captured, the sampling frequency needs to ensure that a sufficient number of data samples are obtained within the time of the seed passing through the detection area; for example, if the time of the seed passing through the detection area is t, and n data samples are expected to be obtained, then the sampling frequency is at least n / t.

[0036] According to the size and complexity of the seed morphology, the scanning resolution of the laser radar and the image resolution of the hyperspectral camera are determined; for small-volume seeds with rich surface details, the resolution needs to be improved to accurately capture their characteristics; otherwise, the resolution can be appropriately reduced to improve the data processing efficiency. Combined with the spectral reflection characteristics of the seed, the spectral acquisition range and the number of wavebands of the hyperspectral camera are set to ensure that the spectral information related to the quality of the seed can be covered.

[0037] Step S3-2, according to the speed of the seed passing through the detection area and the length of the detection area, the rotating speed of the rotating motor is calculated; it is ensured that the laser radar can complete at least one 360° rotating scanning within the time of the seed passing through the detection area, and the scanning starting and ending positions can completely cover the seed; during the scanning process, the rotating angle of the laser radar is monitored in real time through the angle sensor, and the actual angle is compared with the preset scanning path; if there is an angle deviation, the rotating speed and direction of the rotating motor are adjusted and fed back to ensure the accuracy of the scanning path. In combination with the moving track of the seed, the point cloud data obtained at different rotating angles is subjected to coordinate conversion and splicing calculation, so that the three-dimensional point cloud data generated finally can completely and accurately present the three-dimensional morphology of the seed.

[0038] The time points of the seed entering and leaving the detection area are detected by using the photoelectric sensor, the residence time of the seed in the detection area is determined, the time interval and displacement distance between adjacent two frames of images are calculated according to the moving speed of the seed and the frame acquisition speed of the hyperspectral camera, the acquisition parameters are adjusted to ensure that there is neither too much overlap between adjacent frames of images to cause data redundancy, nor too large interval to cause information loss, and continuous and complete spectral image acquisition is realized. The time sequence marking and integration calculation are performed on the collected multiple frames of spectral images, and an original data set reflecting the complete spectral change of the seed in the detection area is formed.

[0039] In the equipment installation stage, the optical axis centers of the laser radar and the hyperspectral camera are overlapped through accurate measurement and calibration, so that the data collected by the two can be accurately corresponded to the same position of the seed; a synchronous triggering mechanism of the laser radar and the hyperspectral camera is established, the two are triggered to start working at the same time when the seed enters the detection area, the time consistency of the data collected by the two is ensured through hardware clock synchronization or software time stamp calibration, and the error is controlled within a very small range; during the data acquisition process, the working states and acquisition progress of the two are monitored in real time, if the time is not synchronized or the data acquisition is abnormal, timely adjustment and correction are performed, and the stability and accuracy of the synchronous scanning are ensured.

[0040] In a preferred embodiment of the present application, the above step S4: denoising and filtering processing is performed on the three-dimensional point cloud original data and the spectral image original data to generate optimized point cloud data and optimized spectral data, which can include: Step S4-1, based on the three-dimensional point cloud original data, radius filtering is adopted to remove outlier noise points to generate reference point cloud data; Step S4-2, based on the space coordinate information of the reference point cloud data, the spectral image original data is registered and optimized: the space mapping relationship of the spectral image is established according to the space coordinate information; Gaussian filtering is adopted to eliminate environmental light interference to generate optimized spectral data with a signal-to-noise ratio of ≥35 dB; the space registration parameters are output synchronously; Step S4-3, based on the spatial registration parameters, the reference point cloud data is subjected to density enhancement processing to generate optimized point cloud data with a resolution of ≤0.1 mm.

[0041] In the embodiment of the present application, the outlying noise points in the three-dimensional point cloud data are effectively removed through radius filtering, the interference of noise on the seed morphology characteristics is reduced, and the reference point cloud data is ensured to truly reflect the seed morphology, laying a solid foundation for feature extraction; the spectral image and the point cloud data are accurately registered with the help of spatial mapping, the environmental light interference is eliminated with the cooperation of Gaussian filtering, the signal-to-noise ratio of the spectral image is improved, and the misjudgment of spectral features is reduced; then, the point cloud data resolution is improved through density enhancement processing, and the data spatial consistency is ensured in combination with spatial registration, thereby improving the data quality and enhancing the accuracy and reliability of the seed morphology and quality feature analysis.

[0042] In the embodiment of the present application, the specific steps include: Step S4-1, according to the average size of the seed and the scanning accuracy, the radius threshold of radius filtering is set. For each point in the point cloud, the number of neighborhood points within the set radius range is calculated with the point as the center; if the number of neighborhood points is lower than a certain threshold, it indicates that the point is an outlying point, which is removed from the point cloud data; otherwise, the point is retained. By traversing the entire three-dimensional point cloud original data and performing the above judgment and processing point by point, the outlying noise points generated due to dust, sensor errors, etc. are removed, and relatively pure reference point cloud data is generated.

[0043] Step S4-2, according to the spatial coordinate information of each point in the reference point cloud data, the corresponding positions of these points in the spectral image are determined, the spatial mapping relationship between the spectral image pixel points and the point cloud data points is established through coordinate conversion and matching algorithm, and the consistency of the image and the point cloud data in space is ensured.

[0044] The noise distribution characteristics of the spectral image original data are analyzed, and the parameters (such as window size, standard deviation) of Gaussian filtering are determined; for each pixel point in the image, the weighted average calculation is performed according to the neighborhood pixel values of the pixel point and according to the weight of Gaussian distribution, the image is smoothed and the interference caused by the change of environmental light is removed, the signal-to-noise ratio of the image is improved to ≥35 dB, and the optimized spectral data is generated. At the same time, the coordinate conversion parameters and the matching relationship involved in the registration process are recorded as spatial registration parameters and output.

[0045] Step S4-3, the spatial coordinate calibration and adjustment of the reference point cloud data are performed by using the spatial registration parameters, so as to ensure that the point cloud data and the spectral image are completely aligned in space; then, according to the target resolution (≤0.1 mm), the interpolation algorithm (such as nearest neighbor interpolation, linear interpolation, etc.) is used to insert new points in the sparse area of the point cloud data, and the density of the point cloud is increased. Through point-by-point calculation and insertion, the resolution of the point cloud data meets the requirements, and finally the optimized point cloud data with a resolution of ≤0.1 mm is generated.

[0046] In a preferred embodiment of the present application, the step S5 of segmenting and identifying the single seed boundary based on the optimized point cloud data, marking the adhered seeds, and outputting the independent seed unit data set can include: Step S5-1, based on the optimized point cloud data, using the Euclidean clustering segmentation algorithm to identify the single seed boundary, and generating a preliminary segmentation result containing the adhered region. Step S5-2, based on the preliminary segmentation result, marking the adhered region with a point cloud distance less than 80% of the minimum diameter of the seed, and outputting an independent seed unit data set with topological relationship.

[0047] In the embodiment of the present application, the Euclidean clustering segmentation algorithm is used to efficiently identify the single seed boundary according to the spatial distribution of the point cloud data, quickly segment the dense point cloud data, reduce manual intervention and complex calculation, and improve the efficiency and automation of seed separation; by reasonably setting the adhered judgment threshold, the adhered seed region is accurately identified, and the adhered region is marked and the topological relationship is established, providing clear data identification and structural information for separation processing; finally, an independent seed unit data set with topological relationship is output, which clearly identifies the seed identity and spatial relationship, and provides a reliable and orderly data foundation for seed feature extraction and quality classification.

[0048] In the embodiment of the present application, the specific steps include: Step S5-1, according to the average size of the seeds and the density of the point cloud data, set the distance threshold of the Euclidean clustering segmentation algorithm; the threshold is used to measure the Euclidean distance between two points in the point cloud space, to determine whether the points belong to the same cluster (i.e. the same seed). Traverse each point in the optimized point cloud data, search for neighborhood points with a distance less than the set threshold from the current point, and classify these neighborhood points into a cluster; repeat this process until all points are classified into the corresponding cluster; in the clustering process, the algorithm can automatically identify the boundaries of different seeds according to the spatial distribution of the point cloud, even if the seeds are partially overlapped, the approximate area of the single seed can be preliminarily distinguished, thereby generating a preliminary segmentation result containing the adhered region.

[0049] Step S5-2, after obtaining the preliminary segmentation result, the minimum diameter of the seeds in different seed categories is measured to determine the reference size for adhesion judgment; then, the boundary of each seed region in the preliminary segmentation is analyzed, and the minimum distance of the point cloud between adjacent seed regions is calculated. If the distance of the point cloud between certain adjacent regions is less than 80% of the minimum diameter of the corresponding seed, the region is determined as an adhesion region, and is marked; after marking, the topological relationship (such as the adjacent relationship, the containing relationship, etc.) between the seed regions is established to clarify the position and connection of each seed unit in the whole data, and finally an independent seed unit data set containing complete topological information is output; when establishing the topological relationship, the spatial position, morphological characteristics of the seeds and the distribution of the adhesion region need to be considered comprehensively to ensure that the topological relationship can accurately reflect the actual relationship between the seeds.

[0050] In a preferred embodiment of the present application, the step S6 of extracting geometric features and defect features from the optimized point cloud data and extracting moldy and insect-damaged spectral features from the optimized spectral data to obtain a multi-dimensional feature set can include: Step S6-1, based on the independent seed unit data set, volume, aspect ratio and surface curvature distribution are extracted from the optimized point cloud data to generate a geometric feature set; and defect features are extracted by point cloud normal mutation detection to generate a defect feature set; Step S6-2, based on the spatial positioning information of the independent seed unit data set, the reflectivity of the 420-510 nm wave band is extracted from the optimized spectral data as a moldy feature, and the absorbance of the 680-780 nm wave band is extracted as an insect-damaged feature, and a spectral feature set composed of the moldy feature and the insect-damaged feature is output; Step S6-3, based on the geometric feature set, the defect feature set and the insect-damaged spectral feature set, a multi-dimensional feature set is obtained.

[0051] In the embodiment of the present application, the geometric features quantify the fullness of the seeds by volume and aspect ratio, and identify deformities by surface curvature distribution to provide an intuitive basis for grading; the defect features accurately identify the physical damage on the seed surface to evaluate the integrity by the depth of the recess and the direction of the crack to avoid the influence of surface defects on the performance of the seeds; the spectral features use the 420-510 nm and 680-780 nm wave bands to respectively realize non-destructive detection of early mold and internal insect damage, and make up for the blind area of surface detection; the multi-dimensional feature set formed by the fusion of geometric, defect and spectral features evaluates the quality of the seeds and overcomes the limitations of single features.

[0052] In the embodiment of the present application, the specific steps include: Step S6-1: The point cloud data of the independent seed unit is discretized in three-dimensional space, dividing it into small cubic voxels. The number of voxels containing the point cloud is counted, multiplied by the volume of a single voxel, and summed to obtain the approximate volume of the seed. Principal component analysis (PCA) is used to determine the principal axis direction of the seed, and the maximum length along the principal axis and the maximum width perpendicular to the principal axis are calculated. The ratio of these two is the aspect ratio, which is used to evaluate the fullness and deformity of the seed.

[0053] Calculate the normal vector of each point in the point cloud data, and determine the curvature by analyzing the rate of change of the normal vectors of adjacent points; statistically classify the curvature values ​​(such as flat, convex, concave) to generate a curvature distribution histogram, which reflects the complexity of the seed surface.

[0054] Search for regions in the point cloud data where the normal vector direction changes abruptly. These regions typically correspond to the concave edges of the seed surface. The depth of the concavity is determined by calculating the vertical distance from the center point of the concave region to the surrounding normal surface. Identify continuous linear low-density regions in the point cloud data, which may represent cracks. The main direction of the crack is determined by fitting a linear model, and the length and width of the crack are analyzed to generate a crack feature description.

[0055] Step S6-2: Calculate the average reflectance of the 420-510nm band within the optimized spectral data. Due to the influence of surface microbial metabolites, moldy seeds typically exhibit characteristic reflectance changes in this band. The degree of mold is determined by comparing the reflectance with that of healthy seeds.

[0056] Within the 680-780nm wavelength range, the average absorption rate of this band is calculated. Seeds affected by pests have changed internal tissue structure, resulting in a significantly higher absorption rate in this band compared to healthy seeds. By setting an absorption rate threshold, it can be determined whether seeds have been attacked by pests.

[0057] Step S6-3 involves standardizing the geometric feature set (volume, aspect ratio, surface curvature), defect feature set (dent depth, crack orientation), and pest spectral feature set (mold reflectance, pest absorption rate) to eliminate dimensional differences between different features. Based on the needs of seed quality assessment, weights are assigned to each feature; for example, the weight of mold features is increased for edible seeds, and the weight of crack features is increased for seeds intended for sowing. Finally, the weighted features are combined into a complete multi-dimensional feature vector.

[0058] In a preferred embodiment of the present application, the above step S7: calculating seed fullness, roundness and surface roughness based on optimized point cloud data; calculating indentation index and crack feature based on point cloud edge detection; fusing geometric features, defect features and pest spectrum features to generate a comprehensive score; using a machine learning model to perform final grading, respectively training a first probability predictor based on laser radar features and a second probability predictor based on pest spectrum features; inputting the probability outputs of the two predictors into the second layer grading model of the meta-learning architecture, and outputting the final quality level, which can include: Step S7-1, grading based on multi-dimensional feature set: calculating seed fullness, roundness and surface roughness based on optimized point cloud data, fullness calculation based on point cloud volume and standard volume deviation; Step S7-2, calculating indentation index and crack feature based on point cloud edge detection, crack feature calculating the percentage of total crack length to surface area by edge detection algorithm; Step S7-3, setting the geometric feature weight to 0.4, the defect feature weight to 0.3, and the pest spectrum feature weight to 0.3 for fusion to generate a comprehensive score; Step S7-4, using a machine learning model to perform final grading: respectively training a first probability predictor based on laser radar features and a second probability predictor based on pest spectrum features; the meta-learning architecture fuses the probability outputs of the two predictors through a Stacking integrator, and the final grading model is XGBoost.

[0059] In the embodiment of the present application, by quantifying fullness, roundness and roughness, the seed geometric morphology can be intuitively evaluated from multiple dimensions, providing a quantitative basis for preliminary screening and improving grading efficiency and accuracy; accurately quantifying indentation index and crack feature can quickly locate physically damaged seeds and ensure overall seed quality; multi-dimensional feature fusion and reasonable weight allocation overcome the limitations of single feature evaluation, making the score meet actual needs; the combination of double probability predictors and meta-learning architecture fully utilizes the advantages of laser radar and hyperspectral data, excavates the implicit relationship between features, accurately responds to complex grading scenarios, adapts to different crop seed quality differences, and improves grading accuracy and stability.

[0060] In the embodiment of the present application, the specific steps include: Step S7-1, calculating the fullness of the seed by point cloud integration, the formula is wherein, is the fullness, is the height, is the projected area; by formula wherein, is the roundness, is the projected area, The roundness of the seed is calculated to reflect the regularity of the shape of the seed, and the height variation of each point in the seed point cloud data is analyzed to count the surface fluctuation degree, and the formula is wherein, is the fluctuation degree, is the total number of data points, is the value of the i-th data point, is the average value of all data points, is the data point.

[0061] The fullness is an important indicator for evaluating whether the seed is healthy and mature, and the roundness can reflect the regularity of the shape of the seed, and then judge whether it is deformed, and the surface fluctuation degree is used to evaluate the roughness of the seed, which is of great significance for identifying mechanical damage or insect damage on the surface of the seed.

[0062] In step S7-2, the edge of the concave area on the surface of the seed is identified by using the point cloud edge detection algorithm, and the formula is wherein, is the percentage of the concave area to the total surface area, which is used as the concave index, is the concave area, is the total surface area, and the higher the percentage, the more serious the concave of the seed surface.

[0063] Similarly, the edge of the crack on the surface of the seed is identified by using the point cloud edge detection algorithm, the total length of the crack is calculated by connecting the edge points of the crack, and the percentage of the total length to the surface area of the seed is calculated; in addition, the direction and width of the crack can also be analyzed to more comprehensively describe the characteristics of the crack.

[0064] For example, for the seed with a large concave index, the system may consider it as a poor quality seed; and for the seed with a long or deep crack, the system may directly determine it as an unqualified seed.

[0065] In step S7-3, the weights of the geometric features, the defect features and the insect spectrum features are determined as 0.4, 0.3 and 0.3 respectively, the fullness, the roundness, the roughness, the concave index, the crack characteristics, and the mold spectrum features are standardized to eliminate the dimensional differences; then according to the weights, the features are weighted and summed to obtain the comprehensive score of each seed; the higher the comprehensive score, the better the quality of the seed.

[0066] ​Step S7-4, respectively, the geometry of the laser radar acquisition and defect characteristics, the spectral characteristics of the hyperspectral camera acquisition as input, with seed actual quality level as label, using support vector machine, random forest algorithm, training the first probability predictor based on laser radar characteristics and the second probability predictor based on pest spectral characteristics, output the probability of seed belonging to different quality levels.

[0067] The output probability of the two probability predictors is received by the Stacking integrator as a new feature input into the XGBoost model for secondary training. Based on these features, the XGBoost model learns the complex relationship between the features and the grade in combination with the seed quality grade label, and finally outputs the quality grade of the seed.

[0068] The step S7-4 further comprises a final quality grading based on machine learning, specifically as follows: Step S7-4-1, data preprocessing: before machine learning modeling, the collected seed feature data needs to be preprocessed, this step includes data cleaning, missing value processing, feature scaling, etc.; data cleaning mainly removes invalid data and outliers to ensure the accuracy and reliability of the data; missing value processing uses interpolation method or other appropriate methods to fill the missing part of the data; feature scaling is to convert different dimensional feature values to the same scale for subsequent model training.

[0069] Step S7-4-2, feature selection: select the most influential features for seed quality grading from numerous features; this can be achieved through correlation analysis, feature importance evaluation, etc.; for example, the Pearson correlation coefficient can be used to calculate the correlation between features and target variables, or random forest model can be used to evaluate the importance of features; through feature selection, the model complexity can be simplified, and the model training efficiency and prediction performance can be improved.

[0070] Step S7-4-3, model training, further comprising: Step S7-4-3-1, after feature selection is completed, a suitable machine learning algorithm is selected, and model training is performed using the pre-processed and feature-selected data. In the preferred scheme of the present application, a suitable machine learning algorithm is selected according to the characteristics of the seed quality grading task. In the embodiments of the present application, a LiDAR and hyperspectral independent probability predictor is used, and a preliminary classification model is constructed based on these two data sources respectively. At the same time, considering that advanced algorithms such as meta-learning (e.g. Stacking) can integrate the prediction results of multiple models and improve the prediction performance, XGBoost is selected as the second layer model for final grading. The pre-processed and feature-selected data set is divided into a training set, a validation set and a test set. The training set is used for model training, the validation set is used for model selection and parameter tuning, and the test set is used for final evaluation of model performance. By reasonably dividing the data set, the stability and reliability of the model training process can be ensured. For the LiDAR and hyperspectral independent probability predictor, the corresponding training set data is used for model construction and training. Support vector machine (SVM), random forest (RandomForest) and other algorithms are used for training of the preliminary classification model. During the training process, the model parameters (such as learning rate, iteration number, regularization parameter, etc.) are adjusted to optimize the model performance. At the same time, the validation set data is used for model validation, and the model with the best performance is selected as the final probability predictor. For the second layer model (such as XGBoost), the output of the LiDAR and hyperspectral probability predictor is used as the input feature, and the corresponding training set label is used for model training. The parameters of XGBoost (such as the number of trees, the depth of the tree, the learning rate, etc.) are adjusted to optimize the model performance.

[0071] Step S7-4-3-2, after the second layer model training is completed, the prediction results of the LiDAR and hyperspectral probability predictor are fused with the prediction results of the second layer model using meta-learning algorithms such as Stacking. The parameters of the fusion algorithm (such as fusion method, weight distribution, etc.) are adjusted to further improve the prediction performance. At the same time, the validation set data is used for validation and optimization of the fusion model, to ensure the stability and reliability of the fusion model in different crops and screening scenarios.

[0072] ‌Step S7-4-4, model evaluation and validation: after model training is completed, model evaluation and validation are needed; this step includes calculating the accuracy, recall rate, F1 score and other indicators of the model, as well as drawing ROC curve and PR curve, etc. Through these evaluation indicators and charts, the performance of the model can be fully understood, and it can be determined whether the model meets the actual application requirements; if the model performance does not meet the requirements, it needs to return to the previous steps for adjustment and optimization.

[0073] Step S7-4-5, model deployment and application: after evaluation and verification, if the model performance meets the requirements, it can be deployed to the actual application scene; during the deployment process, factors such as the running efficiency, stability and scalability of the model need to be considered; at the same time, the model needs to be updated and maintained regularly to adapt to the changing needs of different crops and screening scenarios; in actual application, the trained model can be used to quickly and accurately grade the quality of seeds, providing strong support for agricultural production.

[0074] Step S7-4-6, model optimization and iteration: as the actual application scenario changes and data accumulates, the model needs to be continuously optimized and iterated, which can be achieved by adding new features, improving algorithms, adjusting model parameters, etc.; at the same time, the model can be retrained using new data to improve its accuracy and generalization ability. Through continuous optimization and iteration, the model can better adapt to the needs of different crops and screening scenarios, providing more reliable technical support for agricultural production.

[0075] In summary, the machine learning model construction process of step S7-4 includes data preprocessing, feature selection, model training, model evaluation and verification, model deployment and application, and model optimization and iteration, etc. Through the implementation of these sub-steps, the accuracy and reliability of the machine learning model in the seed quality grading task can be ensured.

[0076] During the training process, the system will use a large amount of sample data to train and verify the model to ensure its generalization ability and stability; at the same time, the system will continuously optimize and adjust the model according to the actual situation to adapt to the changing needs of different crops and screening scenarios; through this step, the system can achieve accurate grading and efficient screening of seed quality, providing strong support for agricultural production; as can be seen from the above embodiments, the seed screening system and method based on laser radar technology provided by the present application have significant technical advantages. First, through the fusion application of laser radar and hyperspectral camera, the system can comprehensively obtain the morphological and insect spectrum characteristics of seeds, providing a more comprehensive information basis for screening; second, through the automatic process and intelligent algorithm design, the system realizes efficient and accurate seed screening, greatly improving the screening efficiency and accuracy; finally, due to the non-contact design, the system effectively reduces the maintenance cost and prolongs the service life. In addition, the embodiments of the present application also fully consider the types and characteristics of seeds, as well as the weight and correlation between different features, ensuring the accuracy and reliability of the screening results.

[0077] As Figure 2 shown, the embodiments of the present application also provide a seed screening system based on laser radar technology, which comprises: The conveying module is used for arranging the seeds in a single row on the notched conveying belt through the vibrating feeder, inhibiting the rolling of the seeds, generating a single-row ordered seed queue, and conveying to a detection area; The obtaining module is used for obtaining a clean seed queue by blowing the seeds with a high-pressure airflow to remove impurities and unqualified seeds based on the single-row ordered seed queue; The scanning module is used for synchronously scanning the clean seed queue: obtaining three-dimensional point cloud original data of the seeds by 360° rotary scanning of a laser radar; and collecting spectral image original data of the seeds by a hyperspectral camera; The preprocessing module is used for denoising and filtering the three-dimensional point cloud original data and the spectral image original data to generate optimized point cloud data and optimized spectral data; The recognition module is used for segmenting and recognizing the boundaries of single seeds based on the optimized point cloud data, marking the adhered seeds, and outputting independent seed unit data sets; The extraction module is used for extracting geometric features and defect features from the optimized point cloud data, extracting moldy and insect-damaged spectral features from the optimized spectral data, and obtaining a multi-dimensional feature set; The processing module is used for calculating seed fullness, roundness and surface roughness based on the optimized point cloud data, calculating a depression index and a crack feature based on edge detection of the point cloud, generating a comprehensive score by fusing the geometric features, the defect features and the insect-damaged spectral features, performing final grading by using a machine learning model, training a first probability predictor based on the laser radar features and a second probability predictor based on the insect-damaged spectral features, inputting the probability outputs of the two predictors into a second-level grading model of a meta-learning architecture, and outputting a final quality grade.

[0078] The above describes the preferred embodiments of the present application. It should be noted that those skilled in the art can make some improvements and refinements without departing from the principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A seed selection method based on lidar technology, characterized in that, The method includes: Step S1: The seeds are arranged in a single column on the grooved conveyor belt by a vibrating feeder to suppress seed rolling, generate a single-column orderly seed queue and transport it to the detection area; Step S2: Based on a single-column ordered seed queue, high-pressure airflow is used to blow away the seeds, removing impurities and seeds with unqualified morphology, to obtain a clean seed queue; Step S3: Simultaneous scanning of the clean seed queue: Obtain raw three-dimensional point cloud data of the seeds by 360° rotating scan with LiDAR; acquire raw spectral image data of the seeds by hyperspectral camera; Step S4: Denoise and filter the original 3D point cloud data and the original spectral image data to generate optimized point cloud data and optimized spectral data; Step S5: Based on the optimized point cloud data, segment and identify the boundaries of individual seeds, mark the adhering seeds, and output the dataset of independent seed units; Step S6: Extract geometric and defect features from optimized point cloud data; extract mold and pest spectral features from optimized spectral data; to obtain a multi-dimensional feature set; Step S7: Calculate seed fullness, roundness, and surface roughness based on optimized point cloud data; calculate indentation index and crack features based on point cloud edge detection; fuse geometric features, defect features, and pest spectral features to generate a comprehensive score; perform final grading using a machine learning model, training a first probability predictor based on lidar features and a second probability predictor based on pest spectral features respectively; input the probability outputs of the two predictors into the second-layer grading model of the meta-learning architecture to output the final quality level.

2. The seed selection method based on lidar technology according to claim 1, characterized in that, Step S1: The seeds are arranged in a single column on the grooved conveyor belt using a vibrating feeder to suppress seed rolling, creating a single-column ordered seed queue and conveying it to the detection area, including: Based on the disordered seed queue output by the vibrating feeder, the seed rolling is suppressed by the indentation structure set on the surface of the conveyor belt, and a single-column ordered seed queue is generated and conveyed to the detection area at a constant speed.

3. The seed selection method based on lidar technology according to claim 2, characterized in that, Step S2: Based on a single-column ordered seed queue, high-pressure airflow is used to purge the seeds, removing impurities and morphologically defective seeds to obtain a clean seed queue, including: Based on a single-column ordered seed queue, set the purge parameters; Based on the set purging parameters, impurities and seeds with unqualified morphology are removed by a high-pressure airflow purging device with adjustable pressure, resulting in a clean seed queue.

4. The seed screening method based on lidar technology according to claim 3, characterized in that, Step S3: Perform synchronous scanning on the clean seed queue: acquire the original three-dimensional point cloud data of the seeds by 360° rotating scanning with LiDAR; Raw data of seed spectral images were acquired using a hyperspectral camera, including: Equipment parameters are obtained based on the clean seed queue and the disordered seed queue output by the vibrating feeder. Based on the equipment parameters, synchronous scanning is performed: the lidar achieves 360° scanning through a rotating motor to acquire the original three-dimensional point cloud data of the seed; the hyperspectral camera acquires the original spectral image data of the seed during the complete cycle of the seed passing through the detection area.

5. The seed selection method based on lidar technology according to claim 4, characterized in that, Step S4: Denoise and filter the original 3D point cloud data and the original spectral image data to generate optimized point cloud data and optimized spectral data, including: Based on the original 3D point cloud data, radius filtering is used to remove outlier noise points and generate baseline point cloud data. Based on the spatial coordinate information of the reference point cloud data, the original spectral image data is registered and optimized: the spatial mapping relationship of the spectral image is established according to the spatial coordinate information; Gaussian filtering is used to eliminate ambient light interference and generate optimized spectral data with a signal-to-noise ratio ≥35dB; and spatial registration parameters are output simultaneously. Based on spatial registration parameters, density enhancement processing is performed on the reference point cloud data to generate optimized point cloud data with a resolution ≤0.1mm.

6. The seed screening method based on lidar technology according to claim 5, characterized in that, Step S5: Based on the optimized point cloud data, segment and identify the boundaries of individual seeds, label the adhering seeds, and output an independent seed unit dataset, including: Based on optimized point cloud data, Euclidean clustering segmentation algorithm is used to identify the boundaries of individual seeds and generate preliminary segmentation results including adhered regions. Based on the preliminary segmentation results, the sticky regions with point cloud spacing less than 80% of the minimum seed diameter are labeled, and the independent seed unit dataset with topological relationships is output.

7. The seed screening method based on lidar technology according to claim 6, characterized in that, Step S6: Extract geometric features and defect features from the optimized point cloud data; Extracting spectral features of mold and pests from optimized spectral data; To obtain a multi-dimensional feature set, including: Based on the independent seed unit dataset, volume, aspect ratio and surface curvature distribution are extracted from optimized point cloud data to generate a geometric feature set; the depth of depression and crack direction are extracted through point cloud normal mutation detection to generate a defect feature set. Based on the spatial positioning information of the independent seed unit dataset, the reflectance of the 420-510nm band is extracted from the optimized spectral data as a mold feature and the absorptivity of the 680-780nm band is extracted as a pest feature, and the spectral feature set composed of the mold feature and the pest feature is output. A multi-dimensional feature set is obtained based on geometric feature set, defect feature set, and pest spectral feature set.

8. The seed screening method based on lidar technology according to claim 7, characterized in that, Step S7: Calculate seed fullness, roundness, and surface roughness based on optimized point cloud data; calculate indentation index and crack features based on point cloud edge detection; fuse... A comprehensive score is generated based on geometric features, defect features, and insect pest spectral features. A machine learning model is used to perform the final grading, training a first probability predictor based on LiDAR features and a second probability predictor based on insect pest spectral features. The probability outputs of the two predictors are input into the second layer of the meta-learning architecture's grading model to output the final quality level, including: Hierarchical grading is performed based on multi-dimensional feature sets: seed fullness, roundness and surface roughness are calculated based on optimized point cloud data, and fullness is calculated based on the deviation value between point cloud volume and standard volume. The indentation index and crack features are calculated based on point cloud edge detection. The crack features are calculated as the percentage of the total crack length to the surface area using the edge detection algorithm. The geometric feature weight is set to 0.4, the defect feature weight is 0.3, and the pest spectral feature weight is 0.

3. These are then fused together to generate a comprehensive score. The final classification is performed using a machine learning model: a first probability predictor based on lidar features and a second probability predictor based on insect spectral features are trained separately; the meta-learning architecture fuses the probability outputs of the two predictors through a stacking ensemble, and the final classification model is XGBoost.

9. A seed screening system based on lidar technology, wherein the system implements the seed screening method based on lidar technology as described in any one of claims 1 to 8, characterized in that, include: The conveying module is used to arrange the seeds in a single row on the grooved conveyor belt by a vibrating feeder, suppress seed rolling, generate a single-row ordered seed queue and transport it to the detection area; The module is used to remove impurities and morphologically unqualified seeds by blowing the seeds with high-pressure airflow based on a single-column ordered seed queue, thereby obtaining a clean seed queue. The scanning module is used to simultaneously scan the clean seed queue: it acquires the original three-dimensional point cloud data of the seeds through a 360° rotating scan using a lidar; and it acquires the original spectral image data of the seeds through a hyperspectral camera. The preprocessing module is used to denoise and filter the raw 3D point cloud data and the raw spectral image data to generate optimized point cloud data and optimized spectral data. The recognition module is used to segment and identify the boundaries of individual seeds based on optimized point cloud data, mark adhering seeds, and output an independent seed unit dataset. The extraction module is used to extract geometric features and defect features from optimized point cloud data; Extract spectral features of mold and pests from optimized spectral data to obtain a multi-dimensional feature set. The processing module calculates seed fullness, roundness, and surface roughness based on optimized point cloud data; calculates indentation index and crack features based on point cloud edge detection; and fuses... A comprehensive score is generated from geometric features, defect features, and insect spectral features; the final grading is performed using a machine learning model, by training a first probability predictor based on lidar features and a second probability predictor based on insect spectral features; the probability outputs of the two predictors are input into the second layer of the meta-learning architecture to output the final quality level.

10. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the seed screening method based on lidar technology as described in any one of claims 1 to 8.

Images