Tea garden intelligent precision spraying system and spraying method based on machine vision

Abstract

The present invention provides a machine vision-based intelligent precision spraying system and method for tea gardens, belonging to the technical field of tea garden machinery. The method comprises: step S1: constructing a tea disease detection model; step S2: collecting on-site images of the tea garden; and step S3: transmitting the on-site images collected in step S2 to an NVIDIA Jetson Orin Nano main control system, and detecting whether pests and diseases are present using the detection model obtained in step S1. If so, the location of the pests and diseases is calculated and spraying is performed. This method achieves precise identification of pests and diseases in tea gardens and precise spraying.

Description

Technical Field

[0001] The present invention relates to the technical field of tea garden machinery, and in particular to a machine vision-based intelligent precision spraying system and a spraying method for a tea garden. Background Art

[0002] Currently, tea is susceptible to various diseases during its growth process, leading to a decline in tea quality and yield. Accurately detecting tea diseases is a crucial technical prerequisite for tea farmers to implement early prevention and control measures and reduce losses. However, traditional tea disease identification methods require manual operation and are easily affected by complex background interference, resulting in low recognition accuracy and missed or false detections. Currently, common tea garden spraying methods mainly include manual targeted spraying and uniform large-scale spraying using agricultural machinery and drones. These methods are time-consuming and labor-intensive, with high labor costs, pesticide waste, and environmental pollution, which are not conducive to the sustainable development of agricultural ecology.

[0003] Furthermore, the rise of ecological tea garden cultivation practices in hilly and mountainous areas, which intercrop green manure crops, has led to challenges in trenching in hilly and mountainous areas, where soil compaction is already severe. This has led to complex root systems and weed entanglement in cutting tools. With ridge spacing of only about 60 centimeters, most existing tillage and fertilization machines are unable to meet the operational requirements of ecological tea gardens. Existing micro-tillage machines for tea gardens cannot reach the required trenching depth in hilly and mountainous ecological tea gardens, and manual fertilization and covering are required after trenching. Due to the higher moisture content and adhesiveness of the soil in southern China, the trenches created by micro-tillage machines are uneven, making fertilization uneven and reducing the efficiency of tillage operations in tea gardens. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a tea garden intelligent precision spraying system and spraying method based on machine vision, which can realize the precise identification of pests and diseases and the precise spraying of pesticides during mechanical operations in the tea garden.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A machine vision-based intelligent precision spraying system for a tea garden comprises an electric engine 1, a frame 2, a travel wheel 3, a gearbox 4, an armrest device 11, and a precision spraying device 12;

[0007] The electric engine 1 is installed behind the frame 2, the frame 2 is welded to the gearbox 4, and the armrest device 11 is installed above the gearbox 4;

[0008] The precision spraying device 12 includes an identification camera 12-1, a camera bracket 12-2, a medicine box bracket 12-3, a medicine box 12-4, an electric pump 12-5, a solenoid valve 12-6, a spraying bracket 12-7, a spraying hose 12-8, a nozzle 12-9, an outer transverse sliding rail 12-10, an inner transverse sliding rail 12-11, a longitudinal sliding rail 12-12 and an NVIDIA Jetson Orin Nano main control system 12-13, the camera bracket 12-2 is welded to the front end of the frame 2, and identification cameras 12-1 are installed at both ends of the camera bracket 12-2 for identifying tea plant diseases and pests. The liquid medicine is transported to the solenoid valve 12-6 through the electric pump 12-5. The two ends of the spray hose 12-8 are respectively connected to the solenoid valve 12-6 and the nozzle 12-9. An outer horizontal sliding rail 12-10, an inner horizontal sliding rail 12-11 and two longitudinal sliding rails 12-12 are installed on both sides of the spray bracket 12-7, wherein the end of the extension section of the outer horizontal sliding rail 12-10 is connected to a longitudinal sliding rail 12-12, and the sliding block of the inner horizontal sliding rail 12-11 is connected to a longitudinal sliding rail 12-12. The NVIDIA Jetson Orin Nano main control system 12-13 is used to receive and analyze the tea plant disease area information detected by the identification camera 12-1, and control the nozzle 12-9 to spray the diseased area.

[0009] Furthermore, it also includes an inclined ditching device 6 and a belt transmission device 10;

[0010] The electric engine 1 transmits power to the gearbox 4 through the belt transmission device 10. The gearbox 4 has an output shaft 1 and an output shaft 2. The output shaft 1 of the gearbox 4 is connected to the running wheel 3 to drive the running wheel 3 to rotate.

[0011] The inclined furrowing device 6 includes a chain transmission box 6-1, a transmission shaft 6-2, a cross coupling 6-3, a rotary tillage device frame 6-4, a rotary tillage transmission mechanism 6-5, and a rotary tillage mechanism 6-6. The bottom plate below the chain transmission box 6-1 is fixed to the front end of the frame 2 by bolts. The left and right sides of the chain transmission box 6-1 are connected to a rotary tillage transmission mechanism 6-5 through a cross coupling 6-3. A rotary tillage mechanism 6-6 is connected to each of the two rotary tillage transmission mechanisms 6-5. The output shaft 2 of the gearbox 4 transmits power to the chain transmission box 6-1 through a chain drive, and drives the cross coupling 6-3, the rotary tillage transmission mechanism 6-5, and the rotary tillage mechanism 6-6 to operate;

[0012] The angle between the axial direction of the transmission shaft 6-2 and the axis of the rotary tillage transmission mechanism 6-5 is 70°, and the axis of the rotary tillage transmission mechanism 6-5 forms an angle greater than 70° and less than 90° with the forward direction. The rotary tillage device frame 6-4 is covered on the outside of the cross coupling 6-3, and the rotary tillage device frame 6-4 is respectively fixed to the chain drive box 6-1 and the rotary tillage transmission mechanism 6-5 by bolts.

[0013] Furthermore, the rotary tillage mechanism 6-6 includes a furrowing and soil-throwing knife 6-61, a mounting cutter disc 6-62, a rotary tillage knife shaft 6-63, and a soil-throwing iron sheet 6-64. The rotary tillage mechanism 6-6 is installed in a left-right opposite relationship. When viewed from the front, the angle formed by the disc surfaces of the mounting cutter discs 6-62 on the left and right sides is an acute angle greater than 20° and less than 40°. Each mounting cutter disc 6-62 is provided with 5 arc-shaped furrowing and soil-throwing knives 6-61 distributed circumferentially. The back of the furrowing and soil-throwing knife 6-61 is welded with a circular soil-throwing iron sheet 6-64. A rotary tillage knife shaft 6-63 is coaxially arranged with a mounting cutter disc 6-62 and fixedly connected. A rotary tillage knife shaft 6-63 is fixed on the output shaft of a rotary tillage transmission mechanism 6-5.

[0014] The above device can realize precise spraying of tea gardens through the precise spraying device 12.

[0015] A spraying method of a tea garden intelligent precision spraying system based on machine vision, the method comprising:

[0016] Step S1: Construct a tea disease detection model, specifically:

[0017] Step S11: collecting tea disease datasets containing different environments and types;

[0018] Step S12: annotating the tea disease dataset collected in step S11 and dividing it into a training set, a validation set, and a test set in proportion;

[0019] Step S13: performing data enhancement and expansion on the training set in step S12;

[0020] Step S14: Construct a tea disease detection model based on GDE-YOLOv8n: introduce a global attention mechanism (GAM) at the end of the Neck of the original YOLOv8n model, use a diverse branch block (DBB) in the Neck of the original YOLOv8n model combined with the C2f module in the Neck network, and replace the original CIoU loss function with the EIoU loss function;

[0021] Step S15: training and experimenting with the model constructed in step S14 to obtain a detection model;

[0022] Step S16: Deploy the detection model obtained in step S15 on the NVIDIA Jetson Orin Nano main control system for operation;

[0023] Step S2: Collecting on-site pictures of the tea garden;

[0024] Step S3: The tea garden site images collected in step S2 are transmitted to the NVIDIA Jetson Orin Nano main control system, and the detection model obtained in step S1 is used to detect whether there are any pests and diseases. If so, the location of the pests and diseases is calculated and spraying is carried out.

[0025] Furthermore, in step S11, a tea disease dataset containing different environments and types is collected, specifically: pictures of tea diseases and pests in different regions, different types of tea diseases and pests, and different environmental conditions are collected.

[0026] Furthermore, in step S12, the dataset is labeled and divided into training set, validation set and test set in proportion. Specifically, the disease images collected in step S11 are labeled in TXT format using the LabelImg image labeling tool to form a label file, and finally the labeled dataset is divided into training set, validation set and test set in a ratio of 7:2:1.

[0027] Furthermore, in step S13, data enhancement is performed on the training set divided and labeled in step S12. Specifically, the training set in step S12 is transformed, including brightness conversion, noise addition, and random angle rotation, and the original data set is expanded four times.

[0028] Furthermore, step S15: the model constructed in step S14 is trained and experimented to obtain a detection model, specifically: the Backbone network is used to extract features of the input image to obtain input features of three scales; the feature pyramid (FPN) and path aggregation network (PAN) are used in the Neck network to fuse the input features of the three scales to obtain feature output maps of three scales; the Detection Head performs final regression and classification operations on the feature output maps of the three scales, predicts the bounding box regression value of each anchor box and the confidence of the target existence, and uses the regression prediction value to adjust the anchor box to generate the final prediction box.

[0029] The beneficial effects of the present invention are that it can realize accurate identification of pests and diseases and accurate spraying of pesticides during mechanical operations in tea gardens. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a structural diagram of a machine vision-based intelligent precision spraying system for tea gardens according to an embodiment of the present invention;

[0031] Figure 2 This is a structural diagram of a machine vision-based intelligent precision spraying system for tea gardens according to an embodiment of the present invention;

[0032] Figure 3 It is a structural schematic diagram of the inclined ditching device;

[0033] Figure 4 Schematic diagram of the rotary tillage mechanism;

[0034] Figure 5 It is a structural diagram of a double-disc covering device;

[0035] Figure 6 This is a structural diagram of a precision spraying device;

[0036] Figure 7 This is a structural diagram of a precision spraying device;

[0037] Figure 8 This is a schematic diagram of the device working in the field;

[0038] Figure 9 This is the GDE-YOLOv8n model structure diagram;

[0039] Figure 10 for Figure 9 A magnified view of the backbone network.

[0040] Figure 11 for Figure 9 A magnified view of the neck network (Neck) and detection head (Detection Head);

[0041] Figure 12 This is the GAM model structure diagram;

[0042] Figure 13 This is the DBB model structure diagram;

[0043] Figure 14 This is the principle diagram of the EIoU loss function;

[0044] Figure 15 This is a flow chart of the spraying method of the tea garden intelligent precision spraying system based on machine vision;

[0045] Figure 16 This is a sub-flowchart of step S1 in the spraying method.

[0046] Description of labels:

[0047] 1. Electric engine; 2. Frame; 3. Travel wheels; 4. Gearbox; 5. Double-disc soil covering device; 6. Inclined trenching device; 7. Depth limiter; 8. Soil retaining mechanism; 9. Fertilizer discharging device; 10. Belt drive; 11. Handrail; 12. Precision spraying device; 13. Positioning block; 14. Fertilizer discharging hose; 15. Front bracket; 16. Battery; 17. Soil retaining plate; 18. Tea tree;

[0048] 5-1. Pole; 5-2. Notched disc blade; 5-3. V-shaped double disc shaft; 5-4. Bearing seat; 5-5. Shield; 5-6. Cover plate; 5-7. Cover plate connector; 5-8. Scraper plate; 5-9. Cover mounting plate; 5-10. Fertilizer pipe;

[0049] 6-1, chain drive box; 6-2, drive shaft; 6-3, cross coupling; 6-4, rotary tillage device frame; 6-5, rotary tillage transmission mechanism; 6-6 rotary tillage mechanism;

[0050] 6-61, trenching and soil throwing knife; 6-62, installing the cutterhead; 6-63, rotary tillage blade shaft; 6-64, soil throwing iron sheet;

[0051] 7-1. Depth-control wheel adjustment lever; 7-2. Depth-control wheel;

[0052] 9-1, fertilizer discharging box; 9-2, outer groove wheel fertilizer discharging device; 9-3, fertilizer discharging device bracket;

[0053] 12-1, Identification camera; 12-2, Camera bracket; 12-3, Medicine box bracket; 12-4, Medicine box; 12-5, Electric pump; 12-6, Solenoid valve; 12-7, Sprayer bracket; 12-8, Sprayer hose; 12-9, Spray nozzle; 12-10, Outer horizontal sliding rail; 12-11, Inner horizontal sliding rail; 12-12, Longitudinal sliding rail; 12-13, NVIDIA Jetson Orin Nano main control system;

[0054] 18-1, distal end of the tea tree; 18-2, proximal end of the tea tree. DETAILED DESCRIPTION

[0055] To illustrate the technical content, achieved objectives and effects of the present invention in detail, the following description is given in conjunction with the embodiments and accompanying drawings.

[0056] Please refer to Figures 1 to 8 , the embodiments provided by the present invention are:

[0057] A machine vision-based intelligent precision spraying system for a tea garden comprises an electric engine 1, a frame 2, a walking wheel 3, a gearbox 4, a double-disc covering device 5, an inclined ditching device 6, a depth limiting device 7, a retaining mechanism 8, a fertilizer discharge device 9, a belt drive device 10, an armrest device 11, and a precision spraying device 12. The electric engine 1 is mounted behind the frame 2, the frame 2 is welded to the gearbox 4, the armrest device 11 is mounted above the gearbox 4, the inclined ditching device 6 is mounted in front of the frame 2, the fertilizer discharge device 9 is mounted above the front end of the frame, the double-disc covering device 5 is mounted below the fertilizer discharge device 9, the depth limiting device 7 is mounted in front of the inclined ditching device 6 through a front bracket 15, and the retaining mechanism 8 is symmetrically mounted on both sides of the front bracket 15.

[0058] In the embodiment of the present invention, the electric engine 1 transmits power to the gearbox 4 through the belt transmission device 10. The gearbox 4 is provided with an output shaft 1 and an output shaft 2. The output shaft 1 of the gearbox 4 is connected to the running wheel 3 to drive the running wheel 3 to rotate.

[0059] In the embodiment of the present invention, the inclined furrowing device 6 includes a chain transmission box 6-1, a transmission shaft 6-2, a cross coupling 6-3, a rotary tillage device frame 6-4, a rotary tillage transmission mechanism 6-5, and a rotary tillage mechanism 6-6. The bottom plate of the chain transmission box 6-1 is fixed to the front end of the frame 2 by bolts. The left and right sides of the chain transmission box 6-1 are connected to a rotary tillage transmission mechanism 6-5 through a cross coupling 6-3. Each of the two rotary tillage transmission mechanisms 6-5 is connected to a rotary tillage mechanism 6-6. The output shaft 2 of the gearbox 4 transmits power to the chain transmission box 6-1 through a chain drive, and drives the cross coupling 6-3, the rotary tillage transmission mechanism 6-5, and the rotary tillage mechanism 6-6 to operate. The axial direction of the transmission shaft 6-2 and the axis of the rotary tillage transmission mechanism 6-5 are at an angle of 70 degrees, and the axis of the rotary tillage transmission mechanism 6-5 forms an angle greater than 70 degrees and less than 90 degrees with the forward direction. The rotary tillage device frame 6-4 is covered on the outside of the cross coupling 6-3, and the rotary tillage device frame 6-4 is fixed to the chain drive box 6-1 and the rotary tillage transmission mechanism 6-5 respectively by bolts.

[0060] In the embodiment of the present invention, the rotary tillage mechanism 6-6 includes a trenching and throwing blade 6-61, a mounting cutter disc 6-62, a rotary tillage blade shaft 6-63, and an iron sheet 6-64 for throwing soil. The rotary tillage mechanism 6-6 is installed opposite to each other on the left and right sides. When viewed from the front, Figure 3As shown, the angle α formed by the disc surfaces of the left and right mounted cutter discs 6-62 is greater than 20° and less than 40°, and each mounted cutter disc 6-62 is circumferentially evenly distributed with five arc-shaped trenching and soil-throwing knives 6-61. The backs of the trenching and soil-throwing knives 6-61 are welded with circular soil-throwing iron plates 6-64. A rotary tillage blade shaft 6-63 is coaxially arranged and fixedly connected to a mounted cutter disc 6-62, and a rotary tillage blade shaft 6-63 is fixed to the output shaft of a rotary tillage transmission mechanism 6-5. While breaking the soil, the soil can be thrown away. The mounted cutter discs 6-62 on the left and right sides are staggered at a certain angle, so that the trenching and soil-throwing knives 6-61 enter the soil first each time the soil is cut, and the rotary tillage mechanism 6-6 completes the initial trenching and simultaneously throws away the soil at the bottom of the trench. The tilted and staggered rotary tillage method reduces the cutting contact area between the trenching blades and the soil, thereby reducing the resistance during the trenching process.

[0061] In an embodiment of the present invention, the fertilizer discharge device 9 includes a fertilizer discharge box 9-1, an outer groove wheel fertilizer discharger 9-2, a battery 16, and a fertilizer discharge device bracket 9-3. The outer groove wheel fertilizer discharger 9-2 is fixed to the bottom of the fertilizer discharge box 9-1 by bolts, the bottom plate of the fertilizer discharge box 9-1 is fixed to the fertilizer discharge device bracket 9-3 by bolts, and the bottom plate of the fertilizer discharge bracket 9-3 is fixed to the frame 2 by bolts. The battery 16 is used to control the rotation speed of the outer groove wheel fertilizer discharger 9-2 to adjust the amount of fertilizer discharge. The fertilizer discharge port of the outer groove wheel fertilizer discharger 9-2 is connected to the fertilizer guide pipe 5-10 of the double disc covering device 5 through a fertilizer discharge hose 14. The double disc covering device 5 applies fertilizer to the ditch while performing secondary ditching. The groove wheel rotation speed and the size of the fertilizer discharge outlet of the fertilizer discharge box 9-1 are controlled by the battery 16 to control the amount of fertilizer discharge.

[0062] In an embodiment of the present invention, the double-disc covering device 5 includes a vertical pole 5-1, a notched disc knife 5-2, a V-shaped double-disc shaft 5-3, and a bearing seat 5-4. A positioning block 13 is provided on the frame 2. The positioning block 13 is provided with a slot for inserting the pole. The pole 5-1 is provided with evenly distributed positioning holes in the vertical direction. The pole 5-1 inserted into the slot is locked and positioned by passing through the positioning holes thereon and the bolts and nuts on the positioning block 13. The height of the double-disc covering device 5 can be adjusted by locking and positioning through the positioning holes of different heights on the pole 5-1. A V-shaped double-disc shaft 5-3 is provided at the bottom of the pole 5-1, and the two notched disc knives 5-2 arranged in a V shape are installed on the V-shaped double-disc shaft 5-3 through the bearing seat 5-4.

[0063] In an embodiment of the present invention, the double-disc covering device 5 also includes a shield 5-5, a covering plate 5-6, a covering plate connector 5-7, a scraper plate 5-8 and a cover mounting plate 5-9. A cover mounting plate 5-9 is welded to the vertical pole 3. The shield 5-5 is fixed to the cover mounting plate 5-9 on the vertical pole 5-1 by bolts. A fertilizer guide pipe 5-10 is provided on the shield 5-5, and the fertilizer eventually passes through the fertilizer guide pipe 5-10 and falls into the opened ditch. The covering plate connector 5-7 is welded to the vertical pole 5-1, and a connecting hole is provided on the connector 5-7. The covering plate 5-6 and the covering plate connector 5-7 are installed and fixed by pins. The scraper plate 5-8 is located on the outside of the two notched disc cutters and is fixed to the shield 5-5.

[0064] In an embodiment of the present invention, the depth limiting device 7 includes a front bracket 15, a depth limiting wheel adjusting rod 7-1, and a depth limiting wheel 7-2. A fixed tube is welded on the front bracket 15, and the depth limiting wheel adjusting rod 7-1 is inserted into the fixed tube. The fixed tube and the depth limiting wheel adjusting rod 7-1 are matched using a positioning pin to achieve the adjustment of the trenching depth.

[0065] It also includes a soil retaining plate 17. Hinge holes are provided on both sides of the front bracket 15. The soil retaining plate 17 is connected to the front bracket 15 through a hinge.

[0066] In the embodiment of the present invention, the precision spraying device 12 includes an identification camera 12-1, a camera bracket 12-2, a medicine box bracket 12-3, a medicine box 12-4, an electric pump 12-5, a solenoid valve 12-6, a spraying bracket 12-7, a spraying hose 12-8, a nozzle 12-9, an outer transverse sliding rail 12-10, an inner transverse sliding rail 12-11, a longitudinal sliding rail 12-12 and an NVIDIA Jetson OrinNano main control system 12-13, the camera bracket 12-2 is welded to the front end of the frame 2, and identification cameras 12-1 are installed at both ends of the camera bracket 12-2 for identifying tea plant diseases and insect pests. The liquid medicine is transported to the solenoid valve 12-6 through the electric pump 12-5, and the two ends of the spray hose 12-8 are respectively connected to the solenoid valve 12-6 and the nozzle 12-9. An outer horizontal sliding guide 12-10, an inner horizontal sliding guide 12-11 and two longitudinal sliding guides 12-12 are installed on both sides of the spray bracket 12-7, wherein the end of the extension section of the outer horizontal sliding guide 12-10 is connected to a longitudinal sliding guide 12-12, which is responsible for spraying the far end 18-1 of the tea tree, and the sliding block of the inner horizontal sliding guide 12-11 is connected to a longitudinal sliding guide 12-12, which is responsible for spraying the proximal end 18-2 of the tea tree, so that the nozzle 12-9 can be aimed at the diseased area of ​​the tea ridge to realize point spraying. The NVIDIA Jetson The Orin Nano main control system 12-13 is used to receive and analyze information about tea disease areas detected by the camera 12-1, and then control the corresponding nozzles to precisely spray the diseased areas, which can reduce the use of pesticides and reduce pollution to the environment.

[0067] In an embodiment of the present invention, a method for using a tool-tilted trenching, fertilizing and covering soil device is provided, and the method is performed in the following steps: (1) before working, the depth of the trenching is adjusted by the depth-limiting device 7 according to the agronomic requirements of the ecological tea garden fertilization and the actual situation; (2) when trenching, the trenching and throwing knives 6-61 of the two rotary tillage mechanisms 6-6 installed symmetrically and tilted in front are used to cut the soil, and the throwing iron sheet 6-64 throws the soil at the bottom of the trench to complete the preliminary trenching of the soil between the tea ridges; (3) when the double-disc covering device 5 is working, the outer circumferential edge of the notched disc knife 5-2 is used to cut the soil. Since the double discs are installed in a V-shape, with a narrow front and wide back arrangement, when the whole machine moves forward, the soil is affected by the side The pushing action produces a push to achieve secondary trenching and complete the leveling of the trench shape; (4) When applying fertilizer, the amount of fertilizer discharged is controlled by the outer groove wheel fertilizer discharger 9-2, and the two ends of the fertilizer discharge hose 14 are respectively connected to the fertilizer discharge port at the lower end of the outer groove wheel fertilizer discharger 9-2 and the fixed fertilizer guide pipe 5-10 on the double disc covering device 5, so that the fertilizer falls into the shaped trench bottom; (5) The covering plate 5-6 is connected to the double disc covering device 5 through the connector welded on the covering plate 5-6, pushing the soil on both sides into the trench bottom to achieve soil covering; (6) When the device moves in the tea garden, the recognition camera 12-1 collects tea disease information in front in real time. When the recognition camera 12-1 detects a disease, it immediately sends it to the NVIDIA The Jetson Orin Nano main control system 12-13 analyzes and processes the center coordinate data of the diseased area, and then controls the solenoid valve 12-6 of the corresponding nozzle 12-9 to open, accurately spraying the diseased area, reducing the use of pesticides, protecting the environment, and promoting the sustainable development of agricultural ecology.

[0068] Please refer to Figures 9 to 16 , a spraying method of a tea garden intelligent precision spraying system based on machine vision, the method is:

[0069] Step S1: Construct a tea disease detection model, specifically:

[0070] Step S11: collecting tea disease datasets containing different environments and types;

[0071] Step S12: annotating the tea disease dataset collected in step S11 and dividing it into a training set, a validation set, and a test set in proportion;

[0072] Step S13: performing data enhancement and expansion on the training set in step S12;

[0073] Step S14: constructing a tea disease detection model based on GDE-YOLOv8n;

[0074] Step S15: training and experimenting with the model constructed in step S14 to obtain a detection model;

[0075] Step S16: Deploy the detection model obtained in step S15 on the NVIDIA Jetson Orin Nano main control system for operation;

[0076] Step S2: Collecting on-site pictures of the tea garden;

[0077] Step S3: The tea garden site images collected in step S2 are transmitted to the NVIDIA Jetson Orin Nano main control system, and the detection model obtained in step S1 is used to detect whether there are any pests and diseases. If so, the location of the pests and diseases is calculated and spraying is carried out.

[0078] This method can achieve precise spraying of pests and diseases in tea gardens, reduce the use of pesticides, protect the environment, and is conducive to the sustainable development of agricultural ecology.

[0079] In step S11, the disease dataset was collected from the tea garden, and 1,811 pictures with different types of tea diseases and different environmental conditions were obtained.

[0080] In step S12, the disease images collected in step S11 are annotated in TXT format using the LabelImg image annotation tool to create a label file. The annotated dataset is then divided into a training set, a validation set, and a test set in a ratio of 7:2:1. The training set is used to learn model parameters and optimize the model through algorithms such as backpropagation. The validation set is used during training to evaluate model performance, adjust hyperparameters (such as the learning rate and number of network layers), or select different model architectures to prevent overfitting of the model to the training data. The test set serves as an objective evaluation standard for the model's final performance, simulating the model's ability to generalize to unknown data in real-world scenarios. It is only used to evaluate the final model after training is complete.

[0081] In step S13, data augmentation operations are performed on the training set divided and labeled in step S12, including brightness conversion, noise addition, and random angle rotation, ultimately expanding the original data set by four times.

[0082] In step S14, Figure 9 As shown in the figure, the GDE-YOLOv8n tea disease detection model includes a backbone network (Backbone), a neck network (Neck), and a detection head (Detection Head) connected in sequence. The backbone network is used to extract feature information from the image for use by the subsequent network; the neck network is used to better utilize the feature information extracted by the backbone network and plays a role in feature fusion; the detection head uses the features extracted by the backbone network and the neck network to make predictions on the input. Compared with the original YOLOv8n model, the GDE-YOLOv8n tea disease detection model introduces a global attention mechanism (GAM) at the end of the neck ( Figure 11 23rd layer in the network), using the Diversity Branch Block (DBB) combined with the C2f module in the Neck network ( Figure 11 The 13th, 16th, 19th and 22nd layers in the

[13] network are replaced by the EIoU loss function to improve the detection accuracy and speed of the model.

[0083] 1) Introducing the Global Attention Mechanism (GAM) at the Neck end of the original YOLOv8n model: Traditional attention mechanisms ignore the importance of retaining information in both channel and spatial dimensions. GAM combines the advantages of the Channel Attention mechanism (CA) and the Spatial Attention mechanism (SA), strengthening the connection between channels and spaces, reducing target information loss in complex environments, and amplifying cross-dimensional feature information. Figure 12 As shown in Figure 1, CA first uses a 3D arrangement to preserve the three-dimensional information in the input feature map; then uses a two-layer Multilayer Perceptron (MLP) with a compression ratio of r to amplify the cross-dimensional channel-space dependency; finally, a Sigmoid activation function is applied to map the variables in the input feature map to between [0,1] to generate a channel attention feature map. The Sigmoid activation function is calculated by formula (1):

[0084] (1)

[0085] Here, x represents the variable in the input feature map.

[0086] SA takes the feature map generated by multiplying the channel attention feature map by the residual of the original input feature map of GAM as input to develop adaptive features. In order to focus on spatial information, SA uses two Convolution performs spatial information fusion. In addition, since the maximum pooling operation reduces information utilization, it is removed to further preserve the feature map. Given the input feature map, the intermediate state and output are defined by formulas (2) and (3), respectively.

[0087] (2)

[0088] (3)

[0089] in represents the input feature map, represents the output feature map, It is an intermediate transition feature. and They are channel attention feature map and spatial attention feature map, Represents element-wise multiplication.

[0090] 2) Using Diverse Branch Blocks (DBB) in combination with the C2f module in the Neck network: Improving model performance often increases computational complexity and inference time. DBB achieves a perfect balance between the two. Figure 13 As shown in Figure 2, DBB uses a complex microstructure during training, which is equivalent to a single convolution during inference, improving model performance while maintaining lightweight. By output channel , input channel and kernel size It is essentially a fourth-order tensor and an optional offset It is based on Channel feature map As input, output Channel feature map ,in and Depend on , padding, and stride configurations. Use To represent the convolution operator, and the offset is expressed as The convolution form is shown in formula (4):

[0091] (4)

[0092] No. On output channels The value at is given by formula (5):

[0093] (5)

[0094] in, Indicates in output channels, and the input channels are , the kernel size is The convolution kernel. yes No. The corresponding channels superior The corresponding relationship is determined by the padding and stride. The linear properties of convolution, including homogeneity and additivity, can be easily derived from formula (5). The specific formulas are shown in formulas (6) and (7).

[0095]

[0096] In formula (6), express Channel feature map, ; Represents a feature map matrix of any size, where all elements are real numbers. ,in ; Indicates that the output channel , input channel and kernel size The convolution kernel is composed of In formula (7), Indicates that the output channel , input channel and kernel size The convolution kernel is composed of ; Indicates that the output channel , input channel and kernel size The convolution kernel is composed of .

[0097] Note that additivity is only satisfied when the two convolutions have the same configuration (e.g., number of channels, kernel size, padding, stride, etc.).

[0098] Based on the above two basic properties, six transformations are summarized: batch normalization (BN), branch addition, deep connection, multi-scale operation, average pooling and convolution sequence. The DBB structure in this model uses , and To enhance the original The convolutional layer greatly enriches the feature space by fusing branches of various scales and complexities, improving the performance of the model without additional inference time cost.

[0099] 3) Replace the original CIoU loss function with the EIoU loss function: YOLOv8n uses CIoU as the BBox regression loss, which adds center point distance and aspect ratio on the basis of IoU to provide a more comprehensive metric. However, the second penalty term of CIoU only constrains the aspect ratio similarity and ignores the independent differences in width and height, resulting in a penalty of 0 when matching the aspect ratio, which limits the optimization effect of the model. EIOU, on the other hand, separates the influencing factors of the aspect ratio of the predicted box and the true box based on the penalty term of CIoU, and calculates the width and height of the predicted box and the true box respectively. The loss function consists of three parts: IoU loss, distance loss, and width and height loss. The first two parts continue the method in CIoU, but the height and width loss directly minimizes the difference in width and height between the predicted box and the true box, which enables the model to converge faster and have higher regression accuracy, providing a more accurate fit. As Figure 14 As shown, the EIoU loss function can be calculated by formula (8):

[0100]

[0101] In formula (8) 、 、 They represent IoU loss, distance loss, and width and height loss respectively. is the Euclidean distance between the center point of the real box and the predicted box, and Represent the width and height differences between the true box and the predicted box, respectively. and Represent the width and height of the minimum bounding rectangle of the predicted bounding box and the true bounding box, respectively.

[0102] In step S15, the process of training the constructed GDE-YOLOv8n model using the training set to obtain the detection model is as follows: using the Backbone network to extract features of the input image to obtain input features of three scales; using the feature pyramid (FPN) and path aggregation network (PAN) in the neck network to fuse the input features of the three scales to obtain feature output maps of three scales; the Detection Head performs final regression and classification operations on the feature output maps of the three scales, predicts the bounding box regression value of each anchor box and the confidence of the target existence, and uses the regression prediction value to adjust the anchor box to generate the final prediction box.

[0103] The method of extracting features from input data using the Backbone network is as follows: Figure 10As shown, the input image first passes through the entry convolutional layers of layers 1 and 2 to extract low-level features (edges, textures). This feature information is then passed to the C2f module (layers 3 to 5) to fuse features at different levels, outputting large-scale feature information X5. The C2f module splits the input features into two parts: one part is directly short-circuited to pass shallow-level information; the other part is passed through multiple Bottleneck residual blocks to extract deep-level features. Finally, the shallow and deep features are fused through Concat to enhance gradient flow. The same Conv Convolutional Layer (layer 6) + C2f Module (layer 7) extraction operation is then used to fuse more detailed features, outputting medium-scale feature information X4. Finally, the same Conv Convolutional Layer (layer 8) + C2f Module (layer 9) extracts deeper features, combined with the serial pooling operation of the SPPF layer (layer 10) to fuse multi-scale context information, outputting small-scale feature information X3. The SPPF module applies multiple MaxPool layers to the input feature map to generate multi-scale features. It concatenates the pooling results at different scales with the original features to improve the model's robustness to object size. The backbone network outputs small-scale feature information (X3), medium-scale feature information (X4), and large-scale feature information (X5).

[0104] The method for fusing the input features of the three scales in the Neck network is as follows: Figure 11As shown, the Neck network consists of FPN and PAN. FPN transfers deep feature semantics from top to bottom, while PAN transfers deep object localization from bottom to top. During the PAN bottom-up transfer, the small-scale feature information X3 is input into the first Upsample (11th layer) upsampling module. Simple interpolation achieves a low-cost increase in feature map resolution. This is then combined with the medium-scale feature information X4 in conjunction with the Concat module (12th layer). Finally, the C2f-DBB module (13th layer), which utilizes DBB to improve multi-scale feature fusion, enhances the model's small object detection capabilities while preserving the original information. The output feature P is then passed to the 14th and 18th layers, respectively. Next, through the Upsample (14th layer) upsampling, Concat (15th layer) concatenation, and C2f-DBB module (16th layer), the same operations are performed. Feature P is then aggregated with the large-scale feature information X5 to produce the large-scale feature output map Y5. In the top-down transfer of the FPN, the large-scale feature output image Y5 is halved in size by the Conv module (layer 17), aggregated with feature P by the Concat module (layer 18), and finally passed through the C2f-DBB module (layer 19) to obtain the medium-scale feature output image Y4. Similarly, the medium-scale feature output image Y4 is halved in size by the Conv module (layer 20), aggregated with feature X3 by the Concat module (layer 21), and then passed through the C2f-DBB module (layer 22) to fuse the large, medium, and small-scale feature information. Finally, it passes through the GAM module (layer 23) to improve the model's recognition performance for small-scale detection boxes that are closer in size to tea leaves, resulting in the small-scale feature output image Y3.

[0105] The Detection Head performs the final regression and classification operations on the feature output maps at three scales as follows: The Detection Head uses a decoupled head structure that separates the regression task and the classification task (VFL and DFL). VFL is used as the classification loss function to balance the weights of the target and background in small target detection training and improve the category probability of the predicted target; DFL is combined with EIOU as the regression loss function to quickly gather the neighboring area of ​​the target positioning and obtain the position information of the predicted box. Figure 11 As shown in the figure, the Detection Head receives three input features Y5, Y4, and Y3 of different resolutions from the Neck network (FPN+PAN), which correspond to the detection of large, medium, and small targets respectively. The features of each scale are independently predicted by the decoupling head, and finally the outputs of different scales are weightedly fused to achieve full coverage detection of cross-size targets. At the same time, the Anchor-Free design is used to directly regress the bounding box.

[0106] In step S15, the GDE-YOLOv8n model obtained after training is compared with other target detection algorithms. The experimental results are shown in Table 1:

[0107] Table 1 Comparative study of GDE-YOLOv8n model and other target detection algorithms

[0108]

[0109] In Table 1, the accuracy rate represents the ratio of the number of correctly predicted samples in the test set to the number of samples in the predicted value, and also represents the possibility that the sample belongs to a certain category. The recall rate represents the ratio of the number of correctly predicted samples in the test set to the number of samples in the true value, and represents the probability of correctly identifying a certain category. The mean average precision refers to the average precision of all classes during the training process of the neural network model, which represents the overall performance of the model on all classes. To ensure accuracy, the IoU threshold of the mean average precision evaluation index is set to 0.5 to comprehensively evaluate the performance of the model. In addition, the parameter quantity index reflects the performance of the model in terms of resource occupation, the floating-point operation number index reflects the performance of the model in terms of computational efficiency, the model size index reflects the performance of the model in terms of deployment feasibility, and the frames per second index reflects the performance of the model in terms of detection speed. The calculation formula for the accuracy rate is shown in formula (9), the calculation formula for the recall rate is shown in formula (10), the calculation formula for the mean average precision is shown in formula (11), and the calculation formula for the average precision is shown in formula (12):

[0110]

[0111] In equations (9) to (12), true positives are the number of samples correctly predicted by the model as tea disease, false positives are the number of samples incorrectly predicted by the model as tea disease, false negatives are the number of samples incorrectly predicted by the model as background, and true negatives are the number of samples correctly predicted by the model as background. The number of representative categories in this study The value is set to 3. Where p(r) is the precision value corresponding to the recall rate r.

[0112] The experimental results in Table 1 show that the GDE-YOLOv8n model achieved a recognition accuracy of 91.7%, the best performance among all models. This represents a 3.1% improvement over the traditional YOLOv8n model. In terms of model parameter count, floating-point operations, and model size, GDE-YOLOv8n only saw negligible improvements over YOLOv8n, with the difference being negligible. Notably, the detection speed reached 80.0 FPS, far exceeding the 4.8 FPS required for real-time detection scenarios, demonstrating real-time performance. In summary, GDE-YOLOv8n offers superior overall performance.

[0113] In step S16, the NVIDIA Jetson Orin Nano is a high-performance edge computing platform designed for AI and machine learning applications. It integrates an NVIDIA Ampere architecture GPU with 1024 CUDA cores and 32 Tensor cores; a six-core ARM Cortex-A78AE CPU with a maximum clock speed of 1.5GHz; and 8GB of 128-bit LPDDR5 memory. The development board includes four USB ports, one HDMI port, one RJ45 network port, and 40 I / O ports for connecting various peripherals. Given its compact size and excellent energy efficiency, the GDE-YOLOv8n model was deployed on this device for real-time detection and location of tea plantation diseases, as well as disease information processing and analysis, providing critical technical support for the subsequent control of the spraying execution system.

[0114] In the spraying execution system, there are 4 solenoid valves and 4 nozzles. The solenoid valve is a 12V DC pressurized normally closed copper valve with a working pressure range of 0.02-0.8Mpa. The rear end of the nozzle pipeline is equipped with a fine-tuning switch, and the spraying amount can be adjusted by rotating the switch. The spraying pipeline uses a PU pneumatic transparent hose, so that the working condition of the spraying pipeline can be directly observed during operation.

[0115] In the monitoring system, the liquid level sensor measures the amount of medicine in the medicine box through exposed parallel wires and converts it into an analog signal that can be collected and read by the main control system to achieve medicine quantity monitoring. Its operating voltage is DC3-5V and the operating current is less than 20mA.

[0116] The method includes the following intelligent precision spraying system for tea gardens based on machine vision:

[0117] It includes an electric engine 1, a frame 2, a travel wheel 3, a gearbox 4, an armrest device 11 and a precision spraying device 12;

[0118] The electric engine 1 is installed behind the frame 2, the frame 2 is welded to the gearbox 4, and the armrest device 11 is installed above the gearbox 4;

[0119] The precision spraying device 12 includes an identification camera 12-1, a camera bracket 12-2, a medicine box bracket 12-3, a medicine box 12-4, an electric pump 12-5, a solenoid valve 12-6, a spraying bracket 12-7, a spraying hose 12-8, a nozzle 12-9, an outer transverse sliding rail 12-10, an inner transverse sliding rail 12-11, a longitudinal sliding rail 12-12 and an NVIDIA Jetson Orin Nano main control system 12-13, the camera bracket 12-2 is welded to the front end of the frame 2, and identification cameras 12-1 are installed at both ends of the camera bracket 12-2 for identifying tea plant diseases and pests. The liquid medicine is transported to the solenoid valve 12-6 through the electric pump 12-5. The two ends of the spray hose 12-8 are respectively connected to the solenoid valve 12-6 and the nozzle 12-9. An outer horizontal sliding rail 12-10, an inner horizontal sliding rail 12-11 and two longitudinal sliding rails 12-12 are installed on both sides of the spray bracket 12-7, wherein the end of the extension section of the outer horizontal sliding rail 12-10 is connected to a longitudinal sliding rail 12-12, and the sliding block of the inner horizontal sliding rail 12-11 is connected to a longitudinal sliding rail 12-12. The NVIDIA Jetson Orin Nano main control system 12-13 is used to receive and analyze the tea plant disease area information detected by the identification camera 12-1, and control the nozzle 12-9 to spray the diseased area.

[0120] The spraying method based on the above system is:

[0121] Use identification camera 12-1 to collect on-site pictures of the tea garden;

[0122] The on-site pictures of the tea garden collected by the recognition camera 12-1 are transmitted to the NVIDIA Jetson Orin Nano main control system, and are detected by the detection model obtained in step S15 to determine whether there are pests and diseases. If so, the location of the pests and diseases is calculated, and the nozzle 12-9 is controlled by the outer horizontal sliding guide 12-10, the inner horizontal sliding guide 12-11, and the longitudinal sliding guide 12-12 to move to the location above the pests and diseases, and the nozzle 12-9 is controlled to spray by controlling the opening of the solenoid valve 12-6; if not, there is no need to open the nozzle 12-9 for spraying.

[0123] In summary, the machine vision-based intelligent precision spraying system and spraying method for tea gardens provided by the present invention can realize the accurate identification and precise spraying of diseases and insect pests in tea gardens, and can identify and spray diseases and insect pests during the process of trenching, fertilizing, and covering operations in tea gardens. The intelligent precision spraying system for tea gardens is a system that combines the functions of trenching, fertilizing, covering, and precise spraying. It can not only reduce the use of pesticides and reduce pollution to the environment, but also is energy-saving and environmentally friendly, simple to operate, lightweight and convenient.

[0124] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent transformations made using the contents of the present invention's description and drawings, or directly or indirectly applied in related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A spraying method of a tea garden intelligent precision spraying system based on machine vision, characterized in that: A machine vision-based intelligent precision spraying system for tea gardens is provided, wherein the system comprises an electric engine (1), a frame (2), a travel wheel (3), a gearbox (4), an armrest device (11) and a precision spraying device (12); The electric motor (1) is installed behind the frame (2), the frame (2) is welded to the gearbox (4), and the handrail device (11) is installed above the gearbox (4); The precision spraying device (12) includes an identification camera (12-1), a camera bracket (12-2), a medicine box bracket (12-3), a medicine box (12-4), an electric pump (12-5), a solenoid valve (12-6), a spraying bracket (12-7), a spraying hose (12-8), a nozzle (12-9), an outer transverse sliding guide rail (12-10), an inner transverse sliding guide rail (12-11), a longitudinal sliding guide rail (12-12) and an NVIDIA Jetson Orin Nano main control system (12-13), the camera bracket (12-2) is welded to the front end of the frame (2), and identification cameras (12-1) are installed at both ends of the camera bracket (12-2) for identifying tea plant diseases and insect pests. The liquid medicine is transported to the solenoid valve (12-6) through the electric pump (12-5), and the two ends of the spray hose (12-8) are respectively connected to the solenoid valve (12-6) and the nozzle (12-9). An outer transverse sliding guide rail (12-10), an inner transverse sliding guide rail (12-11) and two longitudinal sliding guide rails (12-12) are installed on both sides of the spray bracket (12-7), wherein the end of the extension section of the outer transverse sliding guide rail (12-10) is connected to a longitudinal sliding guide rail (12-12), and the sliding block of the inner transverse sliding guide rail (12-11) is connected to a longitudinal sliding guide rail (12-12). The NVIDIA Jetson Orin The Nano main control system (12-13) is used to receive and analyze the information of the tea disease area detected by the recognition camera (12-1), and control the spray head (12-9) to spray the diseased area; The method is: Step S1: Construct a tea disease detection model, specifically: Step S11: collecting tea disease datasets containing different environments and types; Step S12: annotating the tea disease dataset collected in step S11 and dividing it into a training set, a validation set, and a test set in proportion; Step S13: performing data enhancement and expansion on the training set in step S12; Step S14: Construct a tea disease detection model based on GDE-YOLOv8n: introduce a global attention mechanism (GAM) at the end of the Neck of the original YOLOv8n model, use a diverse branch block (DBB) in the Neck of the original YOLOv8n model combined with the C2f module in the Neck network, and replace the original CIoU loss function with the EIoU loss function; Step S15: training and experimenting with the model constructed in step S14 to obtain a detection model; Step S16: Deploy the detection model obtained in step S15 on the NVIDIA Jetson Orin Nano main control system for operation; Step S2: Collecting on-site pictures of the tea garden; Step S3: The tea garden site images collected in step S2 are transmitted to the NVIDIA Jetson Orin Nano main control system, and the detection model obtained in step S1 is used to detect whether there are any pests and diseases. If so, the location of the pests and diseases is calculated and spraying is carried out.

2. The spraying method of the tea garden intelligent precision spraying system based on machine vision according to claim 1 is characterized in that: In step S11, a tea disease dataset containing different environments and types is collected, specifically: pictures of tea diseases and pests in different regions, different types of tea diseases and pests, and different environmental conditions are collected.

3. The spraying method of the tea garden intelligent precision spraying system based on machine vision according to claim 1 is characterized in that: In step S12, the training set, validation set, and test set are divided and annotated. Specifically, the images collected in step S11 are divided into training set, validation set, and test set according to the ratio of 7:2:1, and the training set images are annotated in TXT format using the LabelImg image annotation tool to form a label file.

4. The spraying method of the tea garden intelligent precision spraying system based on machine vision according to claim 1 is characterized in that: In step S13, data enhancement is performed on the training set divided and labeled in step S12. Specifically, the training set in step S12 is transformed, including brightness conversion, noise addition, and random angle rotation, and the original data set is expanded four times.

5. The spraying method of the tea garden intelligent precision spraying system based on machine vision according to claim 1 is characterized in that: Step S15: Train and experiment the model constructed in step S14 to obtain a detection model, specifically: use the Backbone network to extract features from the input image to obtain input features at three scales; use the Feature Pyramid (FPN) and Path Aggregation Network (PAN) in the Neck network to fuse the input features at three scales to obtain feature output maps at three scales; the Detection Head performs final regression and classification operations on the feature output maps at three scales, predicts the bounding box regression value of each anchor box and the confidence of the target existence, and uses the regression prediction value to adjust the anchor box to generate the final prediction box.

6. The spraying method of the tea garden intelligent precision spraying system based on machine vision according to claim 1 is characterized in that: The method further comprises: in step S2, using the machine vision-based intelligent precision spraying system for tea gardens to collect on-site pictures of the tea gardens; In step S3, after calculating the location where the pests and diseases occur, the tea garden intelligent precision spraying system based on machine vision is used to carry out spraying treatment.

Images