A method for determining the location of plant specimen picking based on visual detection
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-16
- Publication Date
- 2026-08-14
AI Technical Summary
然而，植物标本具有结构脆弱、形态不规则、材质多样以及易碎易损等特点，对抓取操作提出了极高要求，传统自动化抓取使用通用抓取策略极易导致标本破损或抓取失败；
[0012]与现有技术相比，本发明的有益效果是：采集植物标本的RGB图像及深度图像，并进行预处理，有效抑制环境的噪声与光照干扰，显著提升后续最优抓取点分析的输入数据质量；基于预处理后的RGB图像及深度图像，利用深度卷积模型确定台纸区域，生成候选抓取点位，结合空间位置特征、语义类别特征与几何属性特征构建植物标本抓取特征集，利用多维特征融合，避免单一特征导致的误判，提高了候选抓取点位筛选的科学性；将所述植物标本抓取特征集输入至构建的抓取可行性评估模型，输出各候选抓取点位的抓取可行性评分及抓取参数配置，依据抓取可行性评分对候选抓取点位进行排序，实现通过数据驱动的可行性评估模型的量化决策，替代人工经验判断，提高抓取策略的一致性；结合抓取安全约束条件，确定最优抓取点位，筛选出最佳可执行抓取点位，有效防止对植物标本造成二次伤害；本发明能够精准定位植物标本中适合机械臂抓取的区域，保证植物标本的完整性，适用于标本馆或数字化处理中心的批量作业场景。
Smart Images
Figure CN122574100A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision detection technology, specifically a method for determining the location of plant specimens based on visual detection. Background Technology
[0002] Plant specimens are important physical data for plant taxonomy, systematics, and biodiversity research. They are usually made from collected plant materials through processes such as pressing, drying, sterilization, and binding, and then fixed on standard-sized mounting paper with labels. In herbaria, research institutions, and university laboratories, plant specimens often require digital scanning, information entry, classification, retrieval, or automated storage. With the advancement of specimen digitization and intelligent warehousing systems, the use of automated equipment such as robotic arms to non-destructively grasp and transfer specimens has become a key link in improving management efficiency. However, plant specimens are characterized by their fragile structure, irregular shape, diverse materials, and fragility, which places extremely high demands on the grasping operation. Traditional automated grasping methods using general grasping strategies are prone to specimen damage or grasping failure. Therefore, there is an urgent need for an intelligent method that integrates multimodal visual information, adaptively identifies key structural features of plant specimens, and combines grasping feasibility analysis. To this end, we now provide a method for determining the grasping point of plant specimens based on visual detection. Summary of the Invention
[0003] The purpose of this invention is to provide a method for determining the location of plant specimens based on visual detection.
[0004] The objective of this invention can be achieved through the following technical solution: a method for determining the location of plant specimen picking based on visual detection, comprising the following steps: Step S1: Acquire RGB and depth images of the plant specimens and preprocess them; Step S2: Extract features from the preprocessed RGB and depth images to obtain candidate capture points and construct a plant specimen capture feature set; Step S3: Input the plant specimen capture feature set into the constructed capture feasibility assessment model to obtain the capture feasibility score and capture parameter configuration for each candidate capture point; Step S4: Based on the output of the grasping feasibility assessment model and combined with the grasping safety constraints, obtain the optimal grasping point.
[0005] Furthermore, the process of collecting RGB and depth images of plant specimens includes: The plant specimens include a mounting paper and plant materials and labels fixed thereon; An RGB-D camera is mounted on a rigid, three-dimensionally adjustable bracket. The RGB-D camera includes a color image sensor and a depth sensor, which are used to simultaneously output RGB images and depth images of plant specimens.
[0006] Furthermore, the preprocessing of the acquired RGB images includes: A Gaussian filter is used to reduce noise in the acquired RGB images; After noise reduction, white balance correction is performed on the noise-reduced RGB image. An adaptive histogram equalization algorithm is used to first convert the white balance corrected RGB image into a grayscale image to obtain a brightness representation. Then, the grayscale image is divided into multiple local regions, and histogram equalization is performed independently on each region to enhance the local contrast of the plant specimen. Finally, bilinear interpolation is used to smooth the grayscale discontinuity between adjacent regions to obtain an enhanced grayscale image. An enhanced RGB image is obtained based on the enhanced grayscale image and the original RGB image.
[0007] Furthermore, the preprocessing of the acquired depth images includes: The acquired depth image is subjected to geometric restoration and smoothing, and each pixel in the depth image is traversed to identify invalid depth values; For regions identified as containing invalid depth values, a multi-stage iterative filling strategy is used for repair. After the filling is completed, a bilateral filter is applied to the depth image to make the overall depth image smoother and the geometric structure more accurate.
[0008] Furthermore, the process of extracting features from the preprocessed RGB and depth images to construct a feature set for plant specimen extraction includes: A deep convolutional model for image pixel classification is constructed. The preprocessed RGB image is imported into the deep convolutional model, and each pixel in the RGB image is classified into a category. Each pixel is assigned to one of the predefined functional region categories, including: table paper region, plant body region, label region and abnormal state region, thereby determining the position of all pixels in the image that belong to the table paper region. Based on the determined area of the paper tray, multiple candidate gripping points are generated within this area. These candidate gripping points are generated using a uniform grid sampling method to ensure that the effective gripping area of the paper tray is covered. For each generated candidate capture point, its spatial location features, semantic category features, and geometric attribute features are extracted by combining the pixel classification results with the repaired depth image, and the three are correlated to construct a plant specimen capture feature set.
[0009] Furthermore, the process of inputting the plant specimen retrieval feature set into the constructed retrieval feasibility assessment model to obtain the retrieval feasibility score and retrieval parameter configuration for each candidate retrieval point includes: The plant specimen capture feature set corresponding to each candidate capture point is input into the capture feasibility assessment model that has been trained. The capture feasibility assessment model outputs the capture feasibility score and capture parameter configuration of the candidate capture point. All candidate crawling points are sorted from high to low according to their corresponding crawling feasibility scores, and the sorting results are used as the final output of the model.
[0010] Furthermore, the process of constructing a crawling feasibility assessment model includes: A crawling feasibility assessment model is constructed, which adopts a multilayer perceptron structure, and the parameters of the constructed crawling feasibility assessment model are initialized. Collect sample data, which includes a variety of plant specimen capture feature sets. Each capture feature set corresponds to a specific candidate capture point location and its capture result label. The capture result label includes capture success or failure and the corresponding capture parameter configuration. The sample data is randomly divided into a training set and a test set according to a preset ratio; The feasibility assessment model for crawling is trained under supervision using the training set, and the parameters are optimized using the mean squared error loss function to obtain the training results. The training results are validated using a test set. If the model's accuracy on the test set meets the preset threshold, the training of the crawling feasibility assessment model is completed. If it does not meet the preset threshold, the relevant parameters are adjusted and the training is repeated until the training results meet the preset threshold or the number of iterations reaches the preset upper limit, thus completing the training and obtaining the crawling feasibility assessment model.
[0011] Furthermore, based on the output of the grasping feasibility assessment model and combined with the grasping safety constraints, the process of obtaining the optimal grasping point includes: Based on the candidate grabbing point ranking results output by the grabbing feasibility assessment model, it is determined whether the candidate grabbing point meets the preset grabbing safety constraints. The grabbing safety constraints include: the point is located within the table paper area, the minimum distance from the plant body and label area, the local surface flatness, and the distance from the physical boundary of the table paper. If the aforementioned grabbing safety constraints are met, then the candidate grabbing point is determined as the optimal grabbing point. If the grabbing safety constraints are not met, the next candidate point in the sorted list is checked sequentially until the first candidate grabbing point that meets all safety constraints is found and it is determined as the optimal grabbing point. If none of the candidate capture points meet the safety constraints, the anomaly handling mechanism will be triggered, prompting manual intervention or repositioning of the plant specimen.
[0012] Compared with existing technologies, the beneficial effects of this invention are: RGB and depth images of plant specimens are acquired and preprocessed to effectively suppress environmental noise and light interference, significantly improving the input data quality for subsequent optimal grasping point analysis; based on the preprocessed RGB and depth images, a depth convolution model is used to determine the paper tray area, generating candidate grasping points; a plant specimen grasping feature set is constructed by combining spatial location features, semantic category features, and geometric attribute features; multi-dimensional feature fusion is used to avoid misjudgments caused by single features, improving the scientific rigor of candidate grasping point selection; the plant specimen grasping feature set is input into... The constructed grasping feasibility assessment model outputs grasping feasibility scores and grasping parameter configurations for each candidate grasping point. Based on the grasping feasibility scores, the candidate grasping points are ranked, enabling quantitative decision-making through a data-driven feasibility assessment model, replacing manual experience-based judgment and improving the consistency of grasping strategies. Combined with grasping safety constraints, the optimal grasping point is determined, and the best executable grasping point is selected, effectively preventing secondary damage to plant specimens. This invention can accurately locate areas within plant specimens suitable for robotic arm grasping, ensuring the integrity of plant specimens, and is suitable for batch operation scenarios in herbaria or digitization centers. Attached Figure Description
[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0014] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0015] like Figure 1 As shown, the method for determining the location of plant specimen picking based on visual detection includes the following steps: Step S1: Acquire RGB and depth images of the plant specimens and preprocess them; Step S2: Extract features from the preprocessed RGB and depth images to obtain candidate capture points and construct a plant specimen capture feature set; Step S3: Input the plant specimen capture feature set into the constructed capture feasibility assessment model to obtain the capture feasibility score and capture parameter configuration for each candidate capture point; Step S4: Based on the output of the grasping feasibility assessment model and combined with the grasping safety constraints, obtain the optimal grasping point.
[0016] It should be further explained that, in the specific implementation process, the process of collecting RGB images and depth images of plant specimens includes: In practical applications at herbariums or digitization centers, the objects to be processed are plant specimens, which include mounting paper and plant materials and labels fixed thereon. The mounting paper refers to the standard paper base used to support and fix plant specimens. It is usually white or light-colored and uniform in size. After the plant specimens are pressed and dried, they are fixed to the mounting paper by adhesive or sewing. It is often attached with a label containing information about the collection of the plant specimens. It is the main force-bearing area for grasping the plant specimens. Based on the size and placement of the paper tray and the layout of the operating platform, the imaging system is installed and debugged. The high-resolution RGB-D camera is mounted on a rigid bracket that can be adjusted in three dimensions. The RGB-D camera includes a color image sensor and a depth sensor, and can output RGB images and depth images simultaneously. Adjust the distance between the camera lens and the surface of the paper, and the shooting angle, so that the center of the lens is at a preset vertical distance from the plane of the paper. The distance between the camera lens and the paper is set between 600mm and 1000mm to ensure that the lens fully covers the entire paper and obtains the image resolution required for subsequent analysis. The system features a multi-source lighting environment, including a ring light surrounding the camera, a diffused top light above, and side soft lights. The ring light illuminates the plant specimen surface at a 45° angle of incidence, providing uniform frontal illumination. The diffused top light projects diffused light downwards through a diffuser, reducing localized contrast in dark plant materials or label areas. The side soft lights enhance the surface texture details and three-dimensionality of dried inflorescences, curled leaves, and other features of the plant specimens. Image acquisition is performed using an external trigger mode, which triggers the camera to capture images via sensor signals. This ensures that the plant specimen remains stationary during the acquisition process and that the position is fixed during the capture, thereby improving the consistency of image acquisition. It also ensures that all light sources maintain a constant output during camera capture, avoiding uneven lighting, overexposure, or underexposure, thus obtaining clear, high-contrast RGB images and depth images with minimal geometric distortion. After completing the installation of the RGB-D camera and configuring the multi-light source lighting environment, debugging operations are performed. Several representative standard plant specimens are selected as test samples. The plant specimens contain various typical state characteristics, such as curled leaves, broken stems, curled edges of the mounting paper, label obstruction, or partial damage. The RGB and depth images of the test samples are obtained through the aforementioned image acquisition process. The clarity, brightness uniformity, and specimen structure contrast of the obtained images are observed to see if they meet the requirements for subsequent analysis. If the images have problems such as blurriness, uneven lighting, shadow obstruction, or loss of surface details, the camera's focal length and the position of the light source are adjusted accordingly.
[0017] For example, if the reflection on the mounting paper causes the label area to be overexposed, appropriately reduce the brightness of the ring light or diffused top light, or fine-tune its illumination angle to avoid the direction of specular reflection; if the weak parts of the plant specimen (such as dried inflorescences) lack detail due to insufficient lighting, moderately increase the brightness of the side soft light or optimize its incident angle to enhance the visibility of the texture; if the image is blurry overall or in parts, fine-tune the camera focus until the edge of the mounting paper and the texture of the specimen are clearly distinguishable.
[0018] It should be further explained that, in the specific implementation process, the preprocessing of the acquired RGB images includes: A Gaussian filter was used to denoise the acquired RGB image. The filter kernel size was set to 5×5 and the standard deviation was 1.0. The Gaussian filter effectively suppressed high-frequency noise in the RGB image by weighted averaging of each pixel and its neighboring pixels, making the RGB image smoother overall while preserving the main edge structure of the plant specimen. After noise reduction, white balance correction is performed on the noise-reduced RGB image. The grayscale world algorithm is used to adjust the gain of each color channel to eliminate color shift caused by differences in light source color temperature. An adaptive histogram equalization algorithm is used to convert the white balance-corrected RGB image into a grayscale image to obtain a brightness representation. Then, adaptive histogram equalization is performed on the grayscale image to divide the RGB image into multiple local regions. The grayscale histogram of each region is calculated independently and equalization is performed to enhance the local contrast of the plant specimen. Bilinear interpolation is used to smooth the grayscale discontinuity between adjacent regions to obtain the enhanced grayscale image. Based on the enhanced grayscale image and the original RGB image, the pixel values of each color channel are adjusted proportionally to generate an RGB image with enhanced details.
[0019] It should be further explained that, in the specific implementation process, the preprocessing of the acquired depth images includes: The acquired depth image is subjected to geometric restoration and smoothing. Each pixel in the depth image is traversed to identify invalid depth values, which include pixels with a value of 0, invalid identification values, or pixels that exceed the effective range of the depth sensor. For regions identified as containing invalid depth values, a multi-stage iterative filling strategy is used for repair. In the first stage, all invalid pixels in the depth image are traversed. The invalid pixel refers to the pixel position corresponding to the invalid depth value. A 3×3 neighborhood is constructed with each invalid pixel as the center. Eight neighboring pixels other than the center pixel are selected in the neighborhood. If there is a valid depth value in the neighborhood, the pixel is filled with the average of the valid depth values in the neighborhood and marked as valid. In the second stage, for pixels that are still invalid pixels after the first stage of filling, that is, pixels that do not have a valid depth value in their 3×3 neighborhood, the global reference depth average is calculated using the valid depth values of the identified paper area, and the remaining invalid pixels are filled with this reference depth, thereby completing the filling and repair of the depth image. After the filling is completed, a bilateral filter is applied to the depth image. The standard deviation of the spatial domain is set to 3 pixels and the standard deviation of the depth domain is set to 0.5 mm. The bilateral filter performs a weighted average of each pixel and its neighboring pixels in two dimensions: spatial distance and depth difference. While preserving the significant depth jump boundary between the paper plate area and the main body area of the plant specimen, it effectively suppresses surface noise in the depth image, making the overall depth image smoother and the geometric structure more accurate.
[0020] It should be further explained that, in the specific implementation process, the process of extracting features from the preprocessed RGB image and depth image to obtain candidate grasping points and constructing a plant specimen grasping feature set includes: A deep convolutional model for image pixel classification is constructed. The preprocessed RGB image is imported into the deep convolutional model, and each pixel in the RGB image is classified into a category. Each pixel is assigned to one of the predefined functional region categories, including: table paper region, plant body region, label region and abnormal state region, thereby determining the position of all pixels in the image that belong to the table paper region. Based on the determined area of the paper tray, multiple candidate gripping points are generated within this area. These candidate gripping points are generated using a uniform grid sampling method to ensure that the effective gripping area of the paper tray is covered. For each generated candidate capture point, its spatial location features, semantic category features, and geometric attribute features are extracted by combining the pixel classification results with the repaired depth image, and the three are correlated to construct a plant specimen capture feature set.
[0021] in: Spatial location features: These represent the absolute and relative spatial location information of each candidate grab point within the tabletop area in the image coordinate system. They include the two-dimensional coordinates of the candidate point, the distance from the four boundaries of the tabletop, and the minimum Euclidean distance to the nearest non-tabletop area (such as the plant body or label). For example, for a candidate point located in the middle of the tabletop, its spatial location features include: horizontal coordinate, vertical coordinate, distance from the left boundary, distance from the top boundary, and distance from the safety boundary. Semantic category features: These represent the functional region type information of each pixel in the image, describing the distribution of the mounting paper, the main body of the plant specimen, the label, and abnormal conditions. For example, semantic category features include classification labels such as "this area belongs to the mounting paper area", "the neighboring area contains a withered inflorescence", and "there is a paper label below". Geometric attribute features: These are numerical data used to quantify the physical feasibility of grasping a surface. They consist of a combination of multiple geometric indices, each representing a graspability attribute. For example, geometric attribute features include local surface flatness, relative height offset, normal vector tilt angle, and safety boundary distance. For a candidate region, its geometric attribute features can be specifically expressed as: flatness 0.35 mm, height offset 1.2 mm, tilt angle 8°, and safety distance 28 mm.
[0022] It should be further explained that, in the specific implementation process, the process of inputting the plant specimen retrieval feature set into the constructed retrieval feasibility assessment model to obtain the retrieval feasibility score of each candidate retrieval point and its retrieval parameter configuration includes: The plant specimen capture feature set corresponding to each candidate capture point is input into the capture feasibility assessment model that has been trained. The capture feasibility assessment model outputs the capture feasibility score and capture parameter configuration of the candidate capture point. All candidate crawling points are sorted from high to low according to their corresponding crawling feasibility scores, and the sorting results are used as the final output of the model.
[0023] It should be further explained that, in the specific implementation process, the process of constructing the feasibility assessment model includes: A crawling feasibility assessment model is constructed, which adopts a multilayer perceptron structure, and the parameters of the constructed crawling feasibility assessment model are initialized. Collect sample data, which includes a variety of plant specimen capture feature sets. Each capture feature set corresponds to a specific candidate capture point location and its capture result label. The capture result label includes capture success or failure and the corresponding capture parameter configuration. The sample data is randomly divided into a training set and a test set according to a preset ratio; The feasibility assessment model for crawling is trained under supervision using the training set, and the parameters are optimized using the mean squared error loss function to obtain the training results. The training results are validated using a test set. If the model's accuracy on the test set meets the preset threshold, the training of the crawling feasibility assessment model is completed. If it does not meet the preset threshold, the relevant parameters are adjusted and the training is repeated until the training results meet the preset threshold or the number of iterations reaches the preset upper limit, thus completing the training and obtaining the crawling feasibility assessment model.
[0024] It should be further explained that, in the specific implementation process, the process of obtaining the optimal grab point location based on the output of the grab feasibility assessment model and combined with the grab security constraints includes: Based on the candidate grabbing point ranking results output by the grabbing feasibility assessment model, it is determined whether the candidate grabbing point meets the preset grabbing safety constraints. The grabbing safety constraints include: the point is located within the table paper area, the minimum distance from the plant body and label area, the local surface flatness, and the distance from the physical boundary of the table paper. If the aforementioned grabbing safety constraints are met, then the candidate grabbing point is determined as the optimal grabbing point. If the grabbing safety constraints are not met, the next candidate point in the sorted list is checked sequentially until the first candidate grabbing point that meets all safety constraints is found and it is determined as the optimal grabbing point. If none of the candidate capture points meet the safety constraints, the anomaly handling mechanism will be triggered, prompting manual intervention or repositioning of the plant specimen.
[0025] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications or equivalent substitutions made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A method for determining the location of plant specimen grasping points based on visual detection, characterized in that, Includes the following steps: Step S1: Acquire RGB and depth images of the plant specimens and preprocess them; Step S2: Extract features from the preprocessed RGB and depth images to obtain candidate capture points and construct a plant specimen capture feature set; Step S3: Input the plant specimen capture feature set into the constructed capture feasibility assessment model to obtain the capture feasibility score and capture parameter configuration for each candidate capture point; Step S4: Based on the output of the grasping feasibility assessment model and combined with the grasping safety constraints, obtain the optimal grasping point.
2. The method for determining the location of plant specimen grasping points based on visual detection according to claim 1, characterized in that, The process of collecting RGB and depth images of plant specimens includes: The plant specimens include a mounting paper and plant materials and labels fixed thereon; An RGB-D camera is mounted on a rigid, three-dimensionally adjustable bracket. The RGB-D camera includes a color image sensor and a depth sensor, which are used to simultaneously output RGB images and depth images of plant specimens.
3. The method for determining the location of plant specimen grasping points based on visual detection according to claim 2, characterized in that, The preprocessing of the acquired RGB image includes: A Gaussian filter is used to reduce noise in the acquired RGB images; After noise reduction, white balance correction is performed on the noise-reduced RGB image. An adaptive histogram equalization algorithm is used to first convert the white balance corrected RGB image into a grayscale image to obtain a brightness representation. Then, the grayscale image is divided into multiple local regions, and histogram equalization is performed independently on each region to enhance the local contrast of the plant specimen. Finally, bilinear interpolation is used to smooth the grayscale discontinuity between adjacent regions to obtain an enhanced grayscale image. An enhanced RGB image is obtained based on the enhanced grayscale image and the original RGB image.
4. The method for determining the location of plant specimen grasping points based on visual detection according to claim 3, characterized in that, The preprocessing of the acquired depth images includes: The acquired depth image is subjected to geometric restoration and smoothing, and each pixel in the depth image is traversed to identify invalid depth values; For regions identified as containing invalid depth values, a multi-stage iterative filling strategy is used for repair. After the filling is completed, a bilateral filter is applied to the depth image to make the overall depth image smoother and the geometric structure more accurate.
5. The method for determining the location of plant specimen grasping based on visual detection according to claim 4, characterized in that, The process of extracting features from preprocessed RGB and depth images to construct a feature set for plant specimen capture includes: A deep convolutional model for image pixel classification is constructed. The preprocessed RGB image is imported into the deep convolutional model, and each pixel in the RGB image is classified into a category. Each pixel is assigned to one of the predefined functional region categories, including: table paper region, plant body region, label region and abnormal state region, thereby determining the position of all pixels in the image that belong to the table paper region. Based on the determined area of the paper tray, multiple candidate gripping points are generated within this area. These candidate gripping points are generated using a uniform grid sampling method to ensure that the effective gripping area of the paper tray is covered. For each generated candidate capture point, its spatial location features, semantic category features, and geometric attribute features are extracted by combining the pixel classification results with the repaired depth image, and the three are correlated to construct a plant specimen capture feature set.
6. The method for determining the location of plant specimen grasping points based on visual detection according to claim 5, characterized in that, The process of inputting the plant specimen collection feature set into the constructed collection feasibility assessment model to obtain the collection feasibility score and collection parameter configuration for each candidate collection point includes: The plant specimen capture feature set corresponding to each candidate capture point is input into the capture feasibility assessment model that has been trained. The capture feasibility assessment model outputs the capture feasibility score and capture parameter configuration of the candidate capture point. All candidate crawling points are sorted from high to low according to their corresponding crawling feasibility scores, and the sorting results are used as the final output of the model.
7. The method for determining the location of plant specimen grasping points based on visual detection according to claim 6, characterized in that, The process of building a crawling feasibility assessment model includes: A crawling feasibility assessment model is constructed, which adopts a multilayer perceptron structure, and the parameters of the constructed crawling feasibility assessment model are initialized. Collect sample data, which includes a variety of plant specimen capture feature sets. Each capture feature set corresponds to a specific candidate capture point location and its capture result label. The capture result label includes capture success or failure and the corresponding capture parameter configuration. The sample data is randomly divided into a training set and a test set according to a preset ratio; The feasibility assessment model for crawling is trained under supervision using the training set, and the parameters are optimized using the mean squared error loss function to obtain the training results. The training results are validated using a test set. If the model's accuracy on the test set meets the preset threshold, the training of the crawling feasibility assessment model is completed. If it does not meet the preset threshold, the relevant parameters are adjusted and the training is repeated until the training results meet the preset threshold or the number of iterations reaches the preset upper limit, thus completing the training and obtaining the crawling feasibility assessment model.
8. The method for determining the location of plant specimen grasping points based on visual detection according to claim 7, characterized in that, The process of obtaining the optimal grab point based on the output of the grab feasibility assessment model and the grab safety constraints includes: Based on the candidate grabbing point ranking results output by the grabbing feasibility assessment model, it is determined whether the candidate grabbing point meets the preset grabbing safety constraints. The grabbing safety constraints include: the point is located within the table paper area, the minimum distance from the plant body and label area, the local surface flatness, and the distance from the physical boundary of the table paper. If the aforementioned grabbing safety constraints are met, then the candidate grabbing point is determined as the optimal grabbing point. If the grabbing safety constraints are not met, the next candidate point in the sorted list is checked sequentially until the first candidate grabbing point that meets all safety constraints is found and it is determined as the optimal grabbing point. If none of the candidate capture points meet the safety constraints, the anomaly handling mechanism will be triggered, prompting manual intervention or repositioning of the plant specimen.