A seedling disease monitoring and seedling removal system and method for a semi-automatic transplanter

Abstract

The present invention provides a seedling disease monitoring-seedling removal system and method for a semi-automatic transplanter, which belongs to the field of agricultural robots. The seedling disease monitoring-seedling removal system of the present invention includes a seedling panoramic information head-up acquisition device and a light and simple retractable bad seedling removal device. The seedling panoramic information head-up acquisition device dynamically and reliably captures the panoramic information of the falling seedlings. In the process of compound detection of seedling diseases, the seedling RGB image is self-input and layer-by-layer compared to the multi-layer classification linear image solidification template, so as to realize the accurate and rapid diagnosis of diseased seedlings in the free fall of seedlings, thereby effectively improving the intelligent operation level of the semi-automatic transplanter in the field.

Description

Technical Field

[0001] The invention relates to the field of agricultural robots, and in particular to a seedling disease monitoring-seedling removal system and method for a semi-automatic transplanter. Background Art

[0002] In recent years, the planting machinery used in my country is mainly semi-automatic. In the field seedling transplanting operation, the diseased seedlings grow poorly and the bacteria spread and infect each other after transplanting, which in turn affects the yield and causes waste. Most of the existing field semi-automatic transplanting machine seedling detection technologies are mainly based on the installation of photoelectric, electromagnetic and other on-off sensor feedback devices in the seedling cup, which trigger the feedback of the presence and counting of seedlings in the seedling cup through the on-off signal, but its detection index is single and it is impossible to diagnose diseased seedlings; the current proposal is to install a high-speed camera above the seedling cup at the seedling drop point to obtain seedling image information from a bird's-eye view, detect the presence and leaf spread of seedlings, but there are prominent problems of spatial interference and field of view occlusion in the seedling operation, and it is even more impossible to achieve complete measurement of seedlings in the seedling cup and disease spot detection from a bird's-eye view. Due to the high speed and high frequency of field semi-automatic transplanting operations, farmers cannot check the quality of seedlings one by one, so there is an urgent need for an efficient and intelligent seedling information detection system to ensure transplanting quality. Summary of the invention

[0003] In view of the shortcomings in the prior art, the present invention provides a seedling disease monitoring and seedling removal system and method for a semi-automatic transplanter, which effectively solves key problems such as effective detection of seedling health information in pots and accidental planting of diseased seedlings in field transplanting operations.

[0004] The present invention achieves the above technical objectives through the following technical means.

[0005] A seedling disease monitoring and seedling removal system for a semi-automatic transplanter, comprising a seedling panorama information head-up acquisition device and a light and simple retractable bad seedling removal device;

[0006] The device for obtaining panoramic information of seedlings in pots comprises a laser beam sensor and an RGB-D camera. The laser beam sensor is symmetrically installed on the inner side of the funnel, and the RGB-D camera is arranged on one side of the seedling guide tube. An opening II is arranged on the seedling guide tube facing the RGB-D camera.

[0007] A light and simple retractable bad seedling rejection device is tiltedly fixed on the other side of the seedling guide tube, and the light and simple retractable bad seedling rejection device includes a driving device, a telescopic rod and double-sided seedling rejection plates connected in sequence, the driving device is also connected to a DC motor, and a pressure sensor is fixed at the bottom of the double-sided seedling rejection plates; the end of the light and simple retractable bad seedling rejection device is directly opposite to the opening I, the opening I is arranged on the seedling guide tube, and the height of the opening I is lower than the opening II;

[0008] The laser beam sensor, RGB-D camera, DC motor and pressure sensor all transmit signals to the controller.

[0009] In the above technical solution, the openings I and II are both square and have the same size. Width of opening Among them, H m is the maximum height of the seedlings in the pot during the planting period, d is the diameter of the seedling guide tube, and S1 is the horizontal distance between the RGB-D camera and the vertical center line of the seedling guide tube.

[0010] In the above technical solution, the device for obtaining panoramic information of seedlings in pots also includes a fill light for transmitting signals with the controller, and the fill light is fixed on the inner wall of the seedling guide tube at an angle downward.

[0011] In the above technical solution, the RGB-D camera is fixed at the end of the L-shaped camera bracket, and the L-shaped camera bracket is fixed on one side of the seedling guide tube.

[0012] A seedling disease monitoring and seedling removal method for a semi-automatic transplanter, comprising a panoramic dynamic capture method of RGB-D images of falling seedlings, a composite detection method for seedling diseases in the seedlings, and a method for removing diseased seedlings while they are falling;

[0013] The method for capturing the panoramic dynamic image of the falling seedlings in a pot RGB-D image is specifically as follows:

[0014] 1) The fully opened cup bottom cover blocks the laser beam of the sensor during the falling process, triggering the field scanning countdown signal of the controller. The controller starts to synchronize the timing intervals T1 and T2, and the seedlings in the pot begin to fall freely;

[0015] 2) After the interval T1, the fill light performs a single fill light operation in the tube;

[0016] 3) After the interval T2, the RGB-D camera performs a single field scan of the seedlings. The RGB-D camera captures the seedlings image at the set vertical viewing angle [0, β] and horizontal viewing angle [-α, +α] to obtain a panoramic three-dimensional depth point cloud set A and a two-dimensional RGB image B of the seedlings.

[0017] The composite detection method for pot seedling diseases includes a disease model training sub-method and an RGB image disease area positioning and identification sub-method;

[0018] 1) Disease model training sub-method

[0019] ① Construct a training library of the most common diseases in the stem-leaf parts of tomato and cucumber seedlings in pots during the planting period;

[0020] ② The image dataset of high-diseases in the stem-leaf area of ​​tomatoes and cucumbers was manually labeled with multi-layer features according to seedling type, disease location and disease type, and a training set was prepared;

[0021] ③ The seedling species in the training set are classified as high-level features, the diseased parts are classified as middle-level features, and the disease types are classified as low-level features. The training set is implanted into the deep learning training network in the order of high-middle-low for layer-by-layer training;

[0022] ④ Perform weighted fusion on the high-medium-low feature parameters obtained by layer-by-layer training of seedling types, disease sites and disease types, and obtain a multi-layer classification linear image solidification template of seedling types-occurrence sites-disease types fusion features;

[0023] 2) RGB image defect area location and recognition sub-method

[0024] ① Preprocessing the two-dimensional RGB image B to obtain a two-dimensional RGB image B2 with enhanced disease-healthy part features;

[0025] ② Automatically input the image B2 into the seedling type image classification template in the multi-layer classification linear image solidification template for comparison one by one to determine the seedling type;

[0026] ③ According to the type of potted seedlings, the image classification template of the diseased part of the potted seedlings is automatically called, and the image B2 is autonomously divided into the stem part image B3 and the leaf part image B4 through the stem-leaf segmentation point; the images B3 and B4 are automatically input into the diseased part classification template in the multi-layer classification linear image solidification template for automatic comparison to locate the diseased part;

[0027] ④ According to the disease location result, the disease type image classification template of the corresponding part of the seedling in this type of pot is automatically called; the disease area is self-processed, the stem disease part is cropped and enlarged to obtain a two-dimensional image B5, and the leaf part is cropped and enlarged to obtain a two-dimensional image B6; the image B5 or B6 is automatically input into the disease type classification template of the multi-layer classification linear image solidification template for one-to-one comparison to determine the disease type;

[0028] The method for removing the diseased seedlings on the way down is specifically as follows:

[0029] Through the stem-leaf segmentation point, the panoramic three-dimensional depth point cloud set A(α, β, D(α, β)) of the seedlings in pot is segmented into a depth point cloud set A1(α1, β1, D1(α1, β1)) containing only the seedlings leaves and a depth point cloud set A2(α2, β2, D2(α2, β2)) containing only the seedlings (16) stems; if the diseased part is detected as a leaf, the depth point cloud set A1 of the leaf is automatically called; if the diseased part is detected as a stem, the depth point cloud set A2 of the stem is automatically called; if the diseased part is detected as a leaf and a stem, the depth point cloud set A1 of the leaf and the depth point cloud set A2 of the stem are all automatically called.

[0030] Furthermore, the method for removing diseased seedlings on the way down is divided into single-site disease degree diagnosis and multi-site disease degree diagnosis according to the disease site.

[0031] Furthermore, the single-site disease degree diagnosis is specifically as follows:

[0032] Stem: The coordinates of 2D image B5 are automatically aligned to the depth point cloud set A1, and the proportion of stem disease area in the point cloud of the stem part is calculated. If N1 ≥ 10%, the stem disease is judged to be serious and the seedling removal signal is activated; A11 is the point cloud data volume of the stem disease occurrence area, D A1 The depth point cloud data volume of the whole stem;

[0033] Leaf: The coordinates of 2D image B6 are automatically aligned to the depth point cloud set A2, and the percentage of leaf disease area in the leaf point cloud is calculated. If N2 ≥ 15%, the leaf disease is judged to be serious and the seedling removal signal is activated; A22 is the point cloud data volume of the leaf disease occurrence area, D A2 The depth point cloud data volume of the entire blade;

[0034] The controller controls the light and simple retractable bad seedling removal device by activating the seedling removal signal to remove the diseased seedlings.

[0035] Furthermore, the diagnosis of the degree of disease in multiple parts is specifically as follows: the coordinates of the two-dimensional images B5 and B6 are automatically aligned to the depth point cloud sets A1 and A2, and the proportion of each diseased area in the point cloud of each part is calculated respectively. If the proportion of any point cloud is greater than the set value, the seedling removal signal is activated; the controller controls the light and simple retractable bad seedling removal device by activating the seedling removal signal to achieve the removal of diseased seedlings.

[0036] Furthermore, the interval time The interval time Where L2 is the vertical distance from the bottom of the seedling cup to the entrance of the seedling guide tube, L3 is the vertical distance from the RGB-D camera to the entrance of the seedling guide tube, g is the gravity acceleration of free fall, and Δt is the set time difference between fill light and scanning operations.

[0037] Furthermore, the vertical viewing angle The horizontal viewing angle Among them, H m is the maximum height of the seedlings in the pot during the planting period, d is the diameter of the seedling guide tube, and S1 is the horizontal distance between the RGB-D camera and the vertical center line of the seedling guide tube.

[0038] The beneficial effects of the present invention are as follows: the present invention first reliably captures the dynamic panoramic information of the falling seedlings in pots through an RGB-D camera, and in the process of compound detection of seedling diseases in pots, the multi-site-multi-disease complex features are trained layer by layer through deep learning, and then the RGB image of the seedlings in pots is self-input and layer-by-layer comparison is performed to accurately locate and identify the diseased area, so as to achieve accurate and rapid diagnosis and elimination of diseased seedlings in the free fall of seedlings in pots. The present invention effectively solves the outstanding problems of disease transmission and infection and low transplant quality caused by the planting of diseased seedlings in the semi-automatic transplanting operation of seedlings in pots in the field, and effectively improves the intelligent operation level of the semi-automatic field transplanter. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a schematic diagram of the structure of the device for obtaining the panoramic information of seedlings in pots from a head-up perspective according to the present invention;

[0040] Figure 2 It is a structural schematic diagram of a light and simple retractable bad seedling removal device of the present invention;

[0041] Figure 3 This is a schematic diagram of the double-opening structure of the tube wall of the seedling guide tube of the present invention;

[0042] Figure 4 This is a schematic diagram of the vertical field of view compression of the RGB-D camera of the present invention;

[0043] Figure 5 This is a schematic diagram of the horizontal field of view compression of the RGB-D camera of the present invention;

[0044] In the figure, 1. Seedling cup, 2. Cup bottom cover, 3. Laser beam sensor, 4. Funnel, 5. Seedling guide tube, 6. Opening I, 7. L-shaped camera bracket, 8. RGB-D camera, 9. Opening II, 10. Fill light, 11. DC motor, 12. Drive device, 13. Telescopic rod, 14. Pressure sensor, 15. Double-sided seedling picking plate, 16. Pot seedlings. DETAILED DESCRIPTION

[0045] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited thereto.

[0046] A seedling disease monitoring and seedling removal system for a semi-automatic transplanter comprises a seedling panorama information head-on acquisition device, a light and simple retractable bad seedling removal device and a seedling guide tube wall double-opening structure.

[0047] like Figure 1As shown, the device for obtaining panoramic information of seedlings in pots comprises a laser beam sensor 3, an L-shaped camera bracket 7, an RGB-D camera 8 and a fill light 10; the laser beam sensor 3 is symmetrically mounted on the inner side of the funnel 4, and the shortest vertical distance between the laser beam sensor 3 and the seedling cup 1 is L1; the L-shaped camera bracket 7 is fixed on the right side of the seedling guide tube 5, and the distance between the L-shaped camera bracket 7 and the vertical center line of the seedling guide tube 5 is S3; the RGB-D camera 8 is fixed at the end of the L-shaped camera bracket 7, the horizontal viewing angle of the RGB-D camera 8 is symmetrical along the vertical center line of the seedling guide tube 5, and the horizontal distance between the RGB-D camera 8 and the vertical center line of the seedling guide tube 5 is S1; the opening II 9 is located at the lower right of the entrance of the seedling guide tube 5; the fill light 10 is fixed to the right inner wall of the seedling guide tube 5 at an angle θ1 obliquely to the lower left, and the fill light 10 is located at the upper part of the opening II 9; in this embodiment, L1=10cm, S1=20cm, and S3=25cm. The RGB-D camera 8 captures the depth-color information of the whole view of the pot seedling 16 through the opening II 9. The funnel 4 is arranged at the upper end of the seedling guide tube 5.

[0048] like Figure 1 , 2 As shown in Figure 3, the light and simple retractable bad seedling removal device comprises a DC motor 11, a driving device 12, a telescopic rod 13, a pressure sensor 14 and a double-sided seedling removal plate 15; one end of the driving device 12 is connected to one end of the telescopic rod 13, and the other end is connected to the DC motor 11. The double-sided seedling removal plate 15 is horizontally installed above the other end of the telescopic rod 13, and the pressure sensor 14 is fixed to the bottom of the double-sided seedling removal plate 15. The opening I6 is located on the left side of the tube wall below the seedling guide tube 5, and the height of the opening I6 is lower than the opening II9; the light and simple retractable bad seedling removal device is fixed to the bottom plate on the left side of the seedling guide tube 5 obliquely to the upper right, the distance between the light and simple retractable bad seedling removal device and the vertical center line of the seedling guide tube 5 is S2, and the end of the light and simple retractable bad seedling removal device is opposite to the opening I6; in this embodiment, S2=13cm.

[0049] like Figure 3 As shown, the opening I6 and the opening II9 form a double opening structure of the seedling guide tube wall, and the opening I6 and the opening II9 are both square and have the same size, and the height (H) and width (W) of the opening are respectively:

[0050]

[0051] Where: H m is the maximum height of the pot seedlings in the planting period, d is the diameter of the seedling guide tube, and H in this embodiment is m =12cm, d=15cm.

[0052] The laser beam sensor 3, the RGB-D camera 8, the fill light 10, the DC motor 11 and the pressure sensor 14 all transmit signals to the controller.

[0053] A seedling disease monitoring and seedling removal method for a semi-automatic transplanter comprises a method for capturing a panoramic dynamic RGB-D image of falling seedlings, a composite detection method for seedling diseases in the seedlings, and a method for removing diseased seedlings while they are falling.

[0054] (1) Figure 4 , 5 As shown in the figure, the method for capturing the panoramic dynamic image of the falling seedlings in a pot RGB-D image is as follows:

[0055] Step 1: The seedling cup 1 rotates to the top of the center of the funnel 4, and the bottom cover 2 of the seedling cup 1 automatically opens due to gravity. The fully opened bottom cover 2 blocks the light beam of the laser beam sensor 3 during the falling process, thereby triggering the field scanning countdown signal of the controller. The controller starts to synchronize the timing T1 and T2, and the seedling 16 in the pot starts to fall freely.

[0056] Step 2: After the interval T1, the fill light 10 performs a single fill light operation in the tube under the control of the controller. After the fill light operation is completed, the fill light 10 is set to a dormant state and waits for the next trigger; wherein the time T1 is:

[0057]

[0058] Wherein L2 is the vertical distance from the bottom of the seedling cup 1 to the entrance of the seedling guide tube 5, L3 is the vertical distance from the RGB-D camera 8 to the entrance of the seedling guide tube 5, g is the gravity acceleration of free fall, Δt is the set time difference between the fill light and the scanning operation, in this embodiment, L2 = 20 cm, L3 = 30 cm, g = 9.8 m / s 2 , Δt=0.05;

[0059] Step 3: After the interval T2, the RGB-D camera 8 performs a single seedling field scanning operation under the control of the controller. After the scanning operation is completed, the RGB-D camera 8 is put into a dormant state and waits for the next trigger; wherein the time T2 is:

[0060]

[0061] At the same time, during the field of view scanning operation, the RGB-D camera 8 captures the image of the seedling 16 at the vertical viewing angle [0, β] and the horizontal viewing angle [-α, +α] set by the controller; wherein the vertical viewing angle β and the horizontal viewing angle α are respectively:

[0062]

[0063] Step 4: The RGB-D camera 8 obtains a panoramic three-dimensional depth point cloud set A and a two-dimensional RGB image B of the seedling 16.

[0064] (2) A composite detection method for seedling diseases in potted plants, including a disease model training sub-method and an RGB image disease area location and recognition sub-method;

[0065] Disease model training sub-method, specifically:

[0066] Step 1: construct a training library of high-disease diseases in the stem-leaf parts of tomato and cucumber seedlings in pots during the appropriate planting period, as shown in Table 1:

[0067] Table 1 Training database of high-risk diseases in the stem-leaf parts of tomato and cucumber seedlings in pots during the appropriate planting period

[0068]

[0069] Step 2: manually label the tomato and cucumber stem-leaf disease-prone image dataset (obtained in advance) with multi-layer features according to seedling type, disease location, and disease type, and make a training set;

[0070] Step 3: Classify the seedling species in the training set as high-level features, the diseased parts as middle-level features, and the disease types as low-level features; and use the open source Resnet50 network and Pytorch open source model algorithm to implant the training set into the deep learning training network in the order of high-medium-low for layer-by-layer training;

[0071] Step 4, the high-medium-low feature parameters obtained by layer-by-layer training of seedling types, diseased parts and disease types are weightedly fused through SVM classification to obtain a multi-layer classification linear image solidification template of seedling type-occurrence part-disease type fusion features. The multi-layer classification linear image solidification template includes a seedling type image classification template, an occurrence part image classification template and a disease type image classification template.

[0072] The RGB image defect area location and recognition sub-method is as follows:

[0073] Step 1: pre-process the panoramic two-dimensional RGB image B of the seedlings in the pot in the controller to obtain a two-dimensional RGB image B2 with enhanced features of diseased and healthy parts;

[0074] Step 2: automatically input the image B2 into the seedling type (high-level feature) image classification template in the multi-layer classification linear image solidification template for comparison one by one. The template category with the highest similarity between the image B2 and the layer is the seedling type;

[0075] Step 3: According to the seedling type result, the controller automatically calls the image classification template of the diseased part (middle-layer features) of the seedlings of this type, and autonomously divides the image B2 into the stem part image B3 and the leaf part image B4 through the stem-leaf segmentation point; the images B3 and B4 are automatically input into the diseased part (middle-layer features) classification template in the multi-layer classification linear image solidification template for automatic comparison to locate the diseased part;

[0076] Step 4. According to the disease location result of step 3, the disease type (low-level feature) image classification template of the corresponding part of the seedling in the pot is automatically called; the diseased area is self-processed, the stem diseased part is cropped and enlarged to obtain image B5, and the leaf part is cropped and enlarged to obtain image B6; image B5 or B6 is automatically input into the disease type (low-level feature) classification template of the multi-layer classification linear image solidification template for one-to-one comparison, and the category with the highest similarity is the disease type.

[0077] (3) The method for removing diseased seedlings during falling is as follows: by using the stem-leaf segmentation point, the panoramic depth point cloud set A(α, β, D(α, β)) of the 16 seedlings in the pot is divided into the depth point cloud set A1(α1, β1, D1(α1, β1)) containing only the leaf depth point cloud set A1(α1, β1, D1(α1, β1)) of the 16 seedlings in the pot and the depth point cloud set A2(α2, β2, D2(α2, β2)) containing only the stem depth point cloud set A1(α1, β1, D1(α1, β1)) of the 16 seedlings in the pot;

[0078] If the diseased part is detected as a leaf, the depth point cloud set A1 (α1, β1, D1 (α1, β1)) of the leaf is automatically called; if the diseased part is detected as a stem, the depth point cloud set A2 (α2, β2, D2 (α2, β2)) of the stem is automatically called; if diseases are detected in both the leaf and the stem, the depth point cloud set A1 (α1, β1, D1 (α1, β1)) of the leaf and the depth point cloud set A2 (α2, β2, D2 (α2, β2)) of the stem are all automatically called;

[0079] 1) Diagnosis of disease severity at a single site

[0080] Stem: The coordinates of the 2D image B5 are automatically aligned to the depth point cloud set A1(α1, β1, D1(α1, β1)), and the calculation is:

[0081]

[0082] Where N1 is the percentage of stem disease area in the point cloud of the stem part, D A11 is the point cloud data volume of the stem disease occurrence area, D A1 is the depth point cloud data volume of the entire stem. If N1≥10%, it is determined that the stem disease is serious and the seedling removal signal is activated.

[0083] Leaves: The coordinates of the 2D image B6 are automatically aligned to the depth point cloud set A2(α2, β2, D2(α2, β2)), and the calculation is:

[0084]

[0085] Where N2 is the percentage of leaf disease area in the leaf point cloud, D A22 is the point cloud data volume of the leaf disease occurrence area, D A2 is the depth point cloud data volume of the entire leaf. If N2≥15%, it is determined that the leaf disease is serious and the seedling removal signal is activated.

[0086] 2) Diagnosis of disease severity in multiple locations

[0087] The coordinates of the two-dimensional images B5 and B6 are automatically aligned to the depth point cloud sets A1 and A2, and the point cloud proportions of each disease area and each part area are calculated respectively. If the proportion value of any place is greater than the set value, the seedling removal signal is activated.

[0088] The controller activates the seedling removal signal and immediately drives the DC motor 11 to rotate forward, so that the telescopic rod 13 enters the interior of the seedling guide tube 5 through the opening Ⅰ6. When the controller detects that the weight data of the pressure sensor 14 changes, it immediately sends a reverse signal to the DC motor 11 to retract the telescopic rod 13, thereby effectively taking the diseased seedlings out of the seedling guide tube 5 and completing the removal.

[0089] In the description of the present invention, it is necessary to understand that the terms "left", "right", "vertical", "horizontal", etc., indicating directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention.

[0090] The embodiments are preferred implementations of the present invention, but the present invention is not limited to the above-mentioned implementations. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essential content of the present invention belong to the protection scope of the present invention.

Claims

1. A seedling disease monitoring and seedling removal system for a semi-automatic transplanter, characterized in that: It includes a device for obtaining panoramic information of seedlings in pots and a light and simple retractable bad seedling removal device; The device for obtaining panoramic information of seedlings in pots comprises a laser beam sensor (3) and an RGB-D camera (8), wherein the laser beam sensor (3) is symmetrically mounted on the inner side of the funnel (4), the RGB-D camera (8) is arranged on one side of the seedling guide tube (5), and an opening II (9) is arranged on the seedling guide tube (5) facing the RGB-D camera (8); A light and simple retractable bad seedling rejection device is tiltedly fixed on the other side of the seedling guide tube (5), the light and simple retractable bad seedling rejection device comprising a driving device (12), a telescopic rod (13) and double-sided seedling rejection plates (15) connected in sequence, the driving device (12) is also connected to a DC motor (11), and a pressure sensor (14) is fixed at the bottom of the double-sided seedling rejection plates (15); the end of the light and simple retractable bad seedling rejection device is directly opposite to the opening I (6), the opening I (6) is arranged on the seedling guide tube (5), and the height of the opening I (6) is lower than that of the opening II (9); The laser beam sensor (3), the RGB-D camera (8), the DC motor (11) and the pressure sensor (14) all transmit signals to the controller.

2. The seedling disease monitoring and seedling removal system according to claim 1 is characterized in that: The openings I (6) and II (9) are both square and have the same size. Width of opening Among them, H m is the maximum height of the seedlings in the pot during the planting period, d is the diameter of the seedling guide tube, and S1 is the horizontal distance between the RGB-D camera (8) and the vertical center line of the seedling guide tube (5).

3. The seedling disease monitoring and seedling removal system according to claim 1, characterized in that: The device for obtaining panoramic information of seedlings in pots from a head-up perspective also includes a fill light (10) for performing signal transmission with the controller, and the fill light (10) is fixed to the inner wall of the seedling guide tube (5) in a manner tilted downward.

4. The seedling disease monitoring and seedling removal system according to claim 1, characterized in that: The RGB-D camera (8) is fixed to the end of the L-shaped camera bracket (7), and the L-shaped camera bracket (7) is fixed to one side of the seedling guide tube (5).

5. A seedling disease monitoring and seedling removal method for a semi-automatic transplanter, characterized in that: It includes a method for capturing the panoramic dynamic RGB-D images of falling seedlings in pots, a method for detecting the diseases of seedlings in pots, and a method for removing the diseased seedlings while they are falling. The method for capturing the panoramic dynamic image of the falling seedlings in a pot RGB-D image is specifically as follows: 1) The fully opened cup bottom cover (2) blocks the laser beam of the laser beam sensor (3) during the falling process, triggering the field scanning countdown signal of the controller, and the controller starts to synchronize the timing intervals T1 and T2, and at the same time, the seedling in the pot (16) starts to fall freely; 2) After the interval T1, the fill light (10) performs a single fill light operation in the tube; 3) After the interval T2, the RGB-D camera (8) performs a single field scanning operation of the seedlings in the pot, and the RGB-D camera (8) captures the image of the seedlings in the pot at a set vertical viewing angle [0, β] and horizontal viewing angle [-α, +α] to obtain a panoramic three-dimensional depth point cloud set A and a two-dimensional RGB image B of the seedlings in the pot; The composite detection method for pot seedling diseases includes a disease model training sub-method and an RGB image disease area positioning and identification sub-method; 1) Disease model training sub-method ① Construct a training library of the most common diseases in the stem-leaf parts of tomato and cucumber seedlings in pots during the planting period; ② The image dataset of high-diseases in the stem-leaf area of ​​tomatoes and cucumbers was manually labeled with multi-layer features according to seedling type, disease location and disease type, and a training set was prepared; ③ The seedling species in the training set are classified as high-level features, the diseased parts are classified as middle-level features, and the disease types are classified as low-level features. The training set is implanted into the deep learning training network in the order of high-middle-low for layer-by-layer training; ④ Perform weighted fusion on the high-medium-low feature parameters obtained by layer-by-layer training of seedling types, disease sites and disease types, and obtain a multi-layer classification linear image solidification template of seedling types-occurrence sites-disease types fusion features; 2) RGB image defect area location and recognition sub-method ① Preprocessing the two-dimensional RGB image B to obtain a two-dimensional RGB image B2 with enhanced disease-healthy part features; ② Automatically input the image B2 into the seedling type image classification template in the multi-layer classification linear image solidification template for comparison one by one to determine the seedling type; ③ According to the type of potted seedlings, the image classification template of the diseased part of the potted seedlings is automatically called, and the image B2 is autonomously divided into the stem part image B3 and the leaf part image B4 through the stem-leaf segmentation point; the images B3 and B4 are automatically input into the diseased part classification template in the multi-layer classification linear image solidification template for automatic comparison to locate the diseased part; ④ According to the disease location result, the disease type image classification template of the corresponding part of the seedling in this type of pot is automatically called; the disease area is self-processed, the stem disease part is cropped and enlarged to obtain a two-dimensional image B5, and the leaf part is cropped and enlarged to obtain a two-dimensional image B6; the image B5 or B6 is automatically input into the disease type classification template of the multi-layer classification linear image solidification template for one-to-one comparison to determine the disease type; The method for removing the diseased seedlings on the way down is specifically as follows: Through the stem-leaf segmentation point, the panoramic three-dimensional depth point cloud set A(α, β, D(α, β)) of the seedlings (16) is segmented into a depth point cloud set A1(α1, β1, D1(α1, β1)) containing only the seedlings (16) leaves and a depth point cloud set A2(α2, β2, D2(α2, β2)) containing only the seedlings (16) stems; if the diseased part is detected as a leaf, the depth point cloud set A1 of the leaf is automatically called; if the diseased part is detected as a stem, the depth point cloud set A2 of the stem is automatically called; if diseases are detected in both the leaves and the stems, the depth point cloud set A1 of the leaves and the depth point cloud set A2 of the stems are all automatically called.

6. The seedling disease monitoring and seedling removal method according to claim 5, characterized in that: The method for removing diseased seedlings on the way down is divided into single-site disease degree diagnosis and multi-site disease degree diagnosis according to the disease site.

7. The seedling disease monitoring and seedling removal method according to claim 6, characterized in that: The single-site disease degree diagnosis is specifically as follows: Stem: The coordinates of 2D image B5 are automatically aligned to the depth point cloud set A1, and the proportion of stem disease area in the point cloud of the stem is calculated. If N1 ≥ 10%, the stem disease is judged to be serious and the seedling removal signal is activated; A11 is the point cloud data volume of the stem disease occurrence area, D A1 The depth point cloud data volume of the whole stem; Leaf: The coordinates of 2D image B6 are automatically aligned to the depth point cloud set A2, and the percentage of leaf disease area in the leaf point cloud is calculated. If N2 ≥ 15%, the leaf disease is judged to be serious and the seedling removal signal is activated; A22 is the point cloud data volume of the leaf disease occurrence area, D A2 The depth point cloud data volume of the entire blade; The controller controls the light and simple retractable bad seedling removal device by activating the seedling removal signal to remove the diseased seedlings.

8. The seedling disease monitoring and seedling removal method according to claim 7, characterized in that: The diagnosis of the degree of disease in multiple parts is specifically as follows: the coordinates of the two-dimensional images B5 and B6 are automatically aligned to the depth point cloud sets A1 and A2, and the proportion of each diseased area in the point cloud of each part is calculated respectively. If the proportion of any point cloud is greater than the set value, the seedling removal signal is activated; the controller controls the light and simple retractable bad seedling removal device by activating the seedling removal signal to achieve the removal of diseased seedlings.

9. The method for monitoring and removing seedlings from potted seedlings according to claim 5, characterized in that: The interval time The interval time Wherein L2 is the vertical distance from the bottom of the seedling cup (1) to the entrance of the seedling guide tube (5), L3 is the vertical distance from the RGB-D camera (8) to the entrance of the seedling guide tube (5), g is the gravity acceleration of free fall, and Δt is the set time difference between the fill light and scanning operations.

10. The seedling disease monitoring and seedling culling method according to claim 5, characterized in that: The vertical viewing angle The horizontal viewing angle Among them, H m is the maximum height of the seedlings in the pot during the planting period, d is the diameter of the seedling guide tube, and S1 is the horizontal distance between the RGB-D camera (8) and the vertical center line of the seedling guide tube (5).

Images