Weed recognition method and system based on convolutional neural network and its application in spraying weed control

Abstract

The invention discloses a weed recognition method and system based on a convolutional neural network and an application in spraying weeding. The weed recognition method based on a convolutional neural network comprises a weed recognition model based on YOLOv8‑seg. The weed recognition model has the characteristics of lightweight model backbone and accurate detection. Firstly, a new lightweight backbone network is constructed by improving ShuffleNetV2 and StarNet networks. Secondly, a C2f_Star module is innovatively proposed to improve the model accuracy while reducing the number of parameters, and an ECA attention mechanism module is added to the model network node. The precise spraying method and system involved in the weeding application comprise algorithms for weed positioning, classification and path planning, and a precise variable spraying algorithm. The invention adopts a precise variable spraying method and system, which can effectively reduce the use of herbicides and the pollution to the environment while achieving precise and efficient weeding.

Description

Technical Field

[0001] The present invention relates to the field of intelligent plant protection agricultural equipment and agricultural robots, and in particular to a weed recognition method and system based on a convolutional neural network and its application in spraying and weeding (precision spraying method and system for field weeds). Background Art

[0002] Weeds are a major challenge in agricultural production. They are highly adaptable, highly fertile, and grow quickly. Weeds compete with crops for limited nutrients and sunlight during crop growth, severely restricting the early growth and development of crops and causing losses and waste of limited resources in agricultural production. In addition, the presence of weeds also provides a medium for the growth of pests and diseases, which can easily lead to outbreaks of pests and diseases and reduce crop yields and quality. Therefore, weed control is a major task in the early management of field crops. Accurately removing weeds in the early stages of crop growth is of great significance to the growth of crops.

[0003] Field weeding technology is mainly divided into manual weeding, mechanical weeding and chemical weeding. Manual weeding refers to the removal of weeds by manual labor instead of using chemical agents or mechanical equipment. However, manual weeding has high labor intensity and low labor efficiency. In addition, with the increasing aging of the population and labor shortage, the cost of manual weeding is also increasing. Mechanical weeding is a green and sustainable way of weeding and the most ideal way of field weeding. However, there are many problems in mechanical weeding. For example, the weeding knife needs to go deep into the soil layer to weed, which has large operation resistance. Mechanical weeding disturbs the crop root system in the soil layer, affecting the early growth of crops. It is difficult to remove weeds near crops and weeds between plants. It is suitable for large-scale farmland weeding operations. Chemical weeding is a method of controlling or removing weeds using chemical agents. At present, this method is widely used. It mostly uses sprayers to spray herbicides indiscriminately over a large area for field weeding. This method is low in cost and high in efficiency. However, large-scale spraying of pesticides not only causes a large amount of pesticide waste and pollutes the environment, but also causes large-scale pesticide residues in farmland, which is not conducive to the sustainable development of modern agriculture. Therefore, precise variable spraying can not only effectively control weeds, but also reduce the use of herbicides and pollution to the environment.

[0004] In response to the current problems of pesticide waste and environmental pollution caused by large-scale spraying of pesticides and the problem of inability to efficiently kill field weeds due to single precision spraying, the present invention adopts computer vision technology and rectangular coordinate robotic arms to jointly realize precise variable spraying for graded spraying treatment of weeds, effectively improving the problems of pesticide waste and environmental pollution. Summary of the invention

[0005] In order to overcome the above shortcomings, the present invention provides a weed identification method and system based on convolutional neural network and its application in spraying weed control. The weed identification method and system of the present invention adopt a weed identification segmentation model based on YOLOv8-seg, and optimize and improve the detection ability, recognition accuracy, model parameter quantity and other issues of small targets in the field. Its application in spraying weed control involves a precise spraying method and system, including a weed positioning and grading algorithm for weed positioning, grading and planning, and a precise variable spraying method for achieving precise spraying, aiming to reduce the use of herbicides and environmental pollution while achieving precise and efficient weed control.

[0006] The technical solution adopted by the present invention to solve the technical problem is:

[0007] Weed recognition methods based on convolutional neural networks include

[0008] S1 Create a dataset

[0009] Use a camera to collect at least 1,000 field images, and then perform data enhancement on the collected images;

[0010] The data enhancement method includes any one of rotation, cropping, mirroring, scaling, and translation, or a combination thereof, to obtain a data-enhanced image set;

[0011] Label the crops and weeds in the enhanced image set and save the corresponding label files;

[0012] Randomly select images and corresponding label files according to the ratio of 7:2:1 to divide the image set into a training set, a validation set, and a test set;

[0013] S2 builds a weed recognition model

[0014] Establish a weed recognition model based on convolutional neural network;

[0015] The weed recognition model is a weed recognition model based on YOLOv8-seg;

[0016] The backbone network of the weed recognition model starts from the input layer, firstly processes the input data through the convolution and maximum pooling layer (Conv_maxpool|0), then extracts the backbone features through the backbone feature extraction module (Triple ShuffleBlock|1, 3, 5, 6), and then further forward propagates through the advanced feature extraction module (StarBlock|2, 4), and then splices the features output by the backbone network with the features of the upsampling layer through the branch processing layer, and performs feature fusion through the feature fusion module (SPPF|7) to obtain multiple feature maps of different sizes, and finally outputs the recognized weed image after processing by the output layer;

[0017] The weed recognition model uses a lightweight backbone network ShuffleNetV2 and StarNet to jointly establish a lightweight backbone network: Triple ShuffleBlock and StarBlock modules are used to perform feature extraction and backbone network simplification;

[0018] The backbone feature extraction module (Triple ShuffleBlock) is improved by ShuffleBlock in the lightweight backbone network ShuffleNetV2 module. The Triple ShuffleBlock module includes three parallel branches and defines the division ratio at the same time. By reasonably allocating the computing resources of different types of convolution operations, each branch performs different convolution operations respectively, and then connects the outputs of each branch and performs channel fusion.

[0019] The advanced feature extraction module (StarBlock) adopts a structure of deep separable convolution, parallel fully connected layers and element-level multiplication, combined with residual connection design, to achieve more efficient feature extraction and fusion; the StarBlock module is improved by StarNet;

[0020] The branch processing process of the weed recognition model based on YOLOv8-seg includes:

[0021] In the first branch, the feature map first passes through the upsampling layer (Upsample|8), then is concatenated with the output of a previous layer (Concat|9), and is enhanced through the ECA layer (ECA|10). It is then further processed through the C2f_Star layer (C2f_Star|11), and then upsampled (Upsample|12) and concatenated (Concat|13). The features are enhanced through the ECA layer (ECA|14), and finally the first segmentation result (Segment) is obtained through the C2f_Star layer (C2f_Star|15) and the convolution layer (Conv|16).

[0022] In the second branch, the feature map is directly concatenated with the output of a previous layer (Concat|17), and feature enhancement is performed through the ECA layer (ECA|18). After being processed by the C2f_Star layer (C2f_Star|19) and the convolution layer (Conv|20), the second segmentation result (Segment) is obtained.

[0023] In the third branch, the feature map is also concatenated with the output of a previous layer (Concat|21), enhanced by the ECA layer (ECA|22), and processed by the C2f_Star layer (C2f_Star|23) to finally obtain the third segmentation result (Segment);

[0024] The entire processing process makes full use of multiple feature extraction, enhancement and aggregation operations to achieve accurate segmentation of input image data;

[0025] The C2f_Star layer is obtained by replacing the Bottleneck module in the C2f module of the model network with the StarBlock module;

[0026] The C2f_Star module first adjusts the number of channels through the convolution layer, then divides the adjusted feature map into several blocks, extracts features through convolution kernels of different sizes, and further processes features through the StarBlock module, then concatenates the output feature maps, and finally adjusts the output feature maps through the convolution layer;

[0027] The C2f_Star module uses the StarBlock module to perform element-by-element multiplication operations (StarOperation) to significantly improve the nonlinear expression ability and performance of the model;

[0028] An ECA layer is provided at the model fusion node, and the ECA layer is a channel attention module;

[0029] The ECA layer first performs global average pooling on the input feature map, then generates channel attention through one-dimensional convolution, passes the convolved features through the Sigmoid activation function to generate channel attention weights, and multiplies the attention weights and the original input feature map element by element between channels to obtain the output feature map;

[0030] S3 model training and validation

[0031] The training set generated by S1 is input into the weed recognition model for training. After the model training is completed, the optimal weight is selected as the weight of the model to generate the weed recognition model;

[0032] Verify and test the weed recognition model through the validation set and test set;

[0033] Application of S4 model and weed identification

[0034] Capture field pictures in real time and input them into the weed recognition model to identify and segment weeds, and obtain weed target frames and mask images.

[0035] The present invention also adopts the following technical solutions to solve the technical problems:

[0036] Weed recognition system based on convolutional neural network, including

[0037] A data processing unit is used to obtain field images, process and annotate them, and then divide them into a training set, a validation set, and a test set;

[0038] A weed recognition model building unit, used to establish a weed recognition model based on a convolutional neural network; the weed recognition model includes an input layer, a backbone network, a branch processing layer, an upsampling layer and an output layer; the weed recognition model is a weed recognition model based on YOLOv8-seg;

[0039] Model training and verification unit, used for training, verification and testing of weed recognition models;

[0040] The weed recognition unit is used to capture field images in real time and input the weed recognition model to identify and segment the weeds, thereby obtaining a weed target frame and a mask image.

[0041] The present invention also adopts the following technical solutions to solve the technical problems:

[0042] Precision spraying methods for weeds in the field include:

[0043] S1 Weed Image Collection

[0044] Real-time image acquisition of crop areas;

[0045] S2 weed recognition segmentation

[0046] Input the image into the weed recognition model to obtain the weed target frame and mask image;

[0047] S3 Weed Location and Classification

[0048] Weed location and classification based on S2 acquired images;

[0049] S31 Weed Location

[0050] The geometric center of the model output target box is located and the mask area is calculated, where the center of the target box is set as the target center point, and the target coordinates are output, and the mask area of ​​a single object in the image is calculated using a summation function;

[0051] Get the coordinates and area of ​​the center point of the weeds, and output the coordinates and area in a list;

[0052] Among them, the positioning algorithm uses the target bounding box to utilize Calculate the mask area using cv2.contourArea function;

[0053]

[0054] It is known that the image captured by the camera is 1280*960, which corresponds to the real ground 256*192mm, that is, 5 pixels correspond to the real distance 1mm. Convert the image coordinates to the real coordinates. Assume that the image coordinates are (x img ,y img ), the real coordinate is (x real ,y real ), the target pixel area is Area img , the target real area is Area real ;

[0055]

[0056] S32 Weed Classification

[0057] When the weed target frame mask area occupies 40% or more of the total image area, it is defined as the full spraying area, and continuous spraying operations will be carried out on this area;

[0058] When the weed target frame mask area accounts for less than 40% of the total image area, it is defined as a fixed-point precision spraying area, and the weeds in this area are classified into weed grades according to the mask area:

[0059] If the weed target frame mask area is less than 225mm 2 When , the judgment is jumped out and the weed level is assigned to level "1";

[0060] If the weed area is larger than 225mm 2 If the weed area is less than 900mm, the judgment will continue. 2 , then the judgment will be jumped out and the weed level will be assigned to level "2". If the weed area is greater than 900mm 2 , then the weed level is assigned to level "3";

[0061] Return all the weed coordinates and grade values ​​in the image to a list and output them one by one;

[0062] S4 Path Planning

[0063] Execute path planning for precise weed spraying based on the output coordinate list;

[0064] If it is the entire spraying area, the planning algorithm will not be used for planning, and the spraying command will be executed directly;

[0065] If it is a fixed-point precision spraying area, it will be planned through the planning algorithm:

[0066] Ant colony algorithm is used for path planning. First, the data information output by S32 weed classification is received, and several ants are placed at each target coordinate point. Then, each ant searches for the next target coordinate point until all coordinate points are traversed, and the pheromone between each coordinate point is updated. The above process is repeated until the number of iterations is zero, and then the planned coordinate list is output.

[0067] S5 precision variable spray

[0068] According to the planned coordinate list and the weed grade value corresponding to each coordinate point, corresponding actions are executed to achieve precise variable spraying;

[0069] S51 dual stepper motor movement control

[0070] When the Y-axis deceleration stepper motor is working, the stepper motor shaft can rotate, the rotation of the shaft drives the guide rail synchronous belt to rotate, the rotation of the synchronous belt drives the movement of the slider, and then drives the third guide rail connected to the slider; when the X-axis deceleration stepper motor is working, it drives the synchronous belt on the third guide rail to rotate, and then drives the nozzle connected to the slider to achieve precise movement between coordinate points;

[0071] The dual stepper motor control algorithm includes:

[0072] First, after receiving the first position coordinates (x1, y1) and the second position coordinates (x2, y2), the straight line length L is calculated, the unit step length n is defined, and the total number of steps N is calculated;

[0073]

[0074]

[0075] Then calculate the movement increments Δx and Δy in each cycle in the x and y directions, that is, in N cycles, every time the x direction moves Δx, the y direction moves Δy;

[0076]

[0077] Since floating point operations may introduce errors, the last step is to move directly to the end point, that is, after N-1 cycles of Δx and Δy driving pulses, position correction movement is performed in the Nth cycle to compensate for the calculation error, that is, the Nth cycle moves x error ,y error ;

[0078] x error =(x2-x1)-Δx*(N-1);

[0079] y error =(y2-y1)-Δy*(N-1);

[0080] S52 spray quantity control

[0081] Variable spray control algorithm:

[0082] By assigning duty cycles to weeds of different grade values;

[0083]

[0084] The frequency of the PWM signal is set to 100 Hz, and the switch state of the nozzle solenoid valve (i.e., the spraying amount) is controlled by adjusting the duty cycle of the PWM signal. In this way, a PWM control signal is generated and output to control the switch of the nozzle solenoid valve to ensure that the target spraying amount is accurately achieved to respond to the spraying amount requirements of different weeds.

[0085] The present invention also adopts the following technical solutions to solve the technical problems:

[0086] Field weed precision spraying system, including

[0087] A weed positioning and grading system includes an image acquisition camera and a microcomputer; the image acquisition camera acquires images of the crop area, transmits the images to the microcomputer for processing by the weed recognition model, outputs the recognition results into a weed coordinate list and corresponding weed grade information through positioning, grading and path planning algorithms, and sends the data to a variable spraying system;

[0088] The variable spraying system includes a rectangular coordinate mechanical arm component and a spraying component;

[0089] The rectangular coordinate robot arm assembly includes an X-axis deceleration stepper motor, a Y-axis deceleration stepper motor, a plurality of guide rails, a plurality of sliders, a synchronous belt, a slider base, and a nozzle module;

[0090] The guide rails are fixed on the frame of the robot arm assembly through the guide rail base, wherein the two guide rails are arranged in parallel and parallel to the travel direction of the system, the Y-axis reduction stepper motor is fixed on one end of the two guide rails, and the synchronous belts on the two guide rails are connected through a coupling, so that the two guide rails are consistent during operation, and the two sliders are respectively matched with the two guide rails and fixed at the same position of the synchronous belt, so that the sliders can move freely along the guide rail direction;

[0091] The third guide rail in the guide rail is fixed on the two sliders through the slider base so that it is perpendicular to the travel direction and moves synchronously with the two sliders. The third slider is matched on the third guide rail and connected to the synchronous belt on the guide rail so that it can move freely along the guide rail direction. The X-axis reduction stepper motor is installed at the end of the third guide rail through the reducer to drive the synchronous belt on the guide rail, so as to accurately determine the position of the slider to achieve precise positioning;

[0092] The rectangular coordinate robot arm assembly controls the coordinated movement of two stepper motors through a dual-motor control algorithm to achieve rapid and accurate movement of the nozzle module in a certain area;

[0093] The spraying assembly includes a diaphragm pump, a medicine box, an overflow valve, a high-pressure medicine box, a medicine discharge pipe and a solenoid valve;

[0094] The diaphragm pump is used to extract medicine from the medicine box when receiving a control instruction, and transport it to the high-pressure medicine box through the medicine inlet pipe. A pressure sensor is installed in the high-pressure medicine box to identify the pressure of the medicine liquid in the high-pressure medicine box. When the internal pressure reaches a certain value, the high-pressure medicine flows back to the medicine box through the overflow valve, and the diaphragm pump stops working; the solenoid valve is arranged on the medicine discharge pipe and connected between the high-pressure medicine box and the nozzle module, that is, one end of the medicine discharge pipe is connected to the high-pressure medicine box, and the other end is connected to the nozzle module. When the rectangular coordinate mechanical arm assembly is controlled to reach the position above the weeds and a spraying instruction is received, the solenoid valve is controlled to perform variable precision spraying on the weeds;

[0095] The walking navigation system includes a frame platform, a driving wheel assembly and a navigation assembly; wherein

[0096] The frame platform is used to place various hardware facilities; the driving wheel assembly includes wheels, wheel connecting frames, and steering stepper motors. The frame platform is fixedly connected to the driving wheel assembly, and all four wheels have the ability to steer and drive. A medicine box, a hardware box, a control cabinet, a battery, a front antenna, and a rear antenna are placed on the frame platform. The control cabinet is equipped with a microcomputer, a single-chip microcomputer, a motor driver, etc., and a cabinet door is provided on the side of the frame platform; a diaphragm pump, an overflow valve, a high-pressure medicine box, an electromagnetic valve, etc. are provided in the hardware box, and a cabinet door is provided on the side of the frame platform; the frame platform is provided with a plurality of control panels, ... The chassis separates various hardware and reduces exposure to the natural environment; the rectangular coordinate mechanical arm assembly is arranged below the rear end of the frame platform in the walking direction, and the image acquisition camera is arranged at the bottom of the front end; the navigation assembly includes a satellite navigation device and a laser radar module, and global positioning is performed by combining the laser radar with the satellite navigation system. The front antenna and the rear antenna are arranged at the top of the frame platform to provide location information. The laser radar module is arranged at the front end of the frame platform and fixed by a laser radar bracket to provide real-time road condition information for the variable spraying robot to autonomously walk in the field;

[0097] The power supply system includes a battery and a voltage converter. The power output by the battery is passed through the voltage converter to provide different power requirements for each system and component. The battery also provides a charging interface.

[0098] The beneficial effects brought by the present invention are:

[0099] 1. The present invention proposes a weed recognition model based on YOLOv8-seg. The weed recognition model improves the model C2f module and adds an attention mechanism module ECA, which significantly improves the model detection accuracy; the model backbone network is lightweighted, thereby improving the model detection speed and parameter quantity.

[0100] 2. The present invention proposes an algorithm for weed positioning, grading and planning, and combines it with the weed recognition model to form a complete field weed precision spraying system and method application. When the real-time collected image is processed by model recognition, positioning, grading and planning algorithms, the weed control signal can be directly output to the single-chip microcomputer for precision spraying operation.

[0101] 3. The present invention proposes a variable spraying control method, which adopts a dual-stepping motor control algorithm to achieve efficient and accurate dual-stepping motor control through reasonable and accurate pulse generation and real-time feedback monitoring, so that it can complete the movement in a coordinated manner; the variable spraying control algorithm is used to control the spraying amount of the constant-pressure pulse variable nozzle in the spraying system, and adjusts the duty cycle of the PWM signal according to the level information of the weeds, which can accurately control the switching state of the nozzle, thereby adjusting the spraying amount. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0103] Figure 1 This is a network structure diagram of the weed recognition model based on YOLOv8-seg described in Example 1;

[0104] Figures 2(a), (b), (c), (d), and (e) are schematic diagrams of the main improved modules of the weed recognition model of Example 1;

[0105] Figure 3 This is a flow chart of the weed grading algorithm described in Example 4;

[0106] Figure 4 This is a flow chart of the path planning algorithm described in Example 4;

[0107] Figure 5 This is a schematic diagram of the structure of the precision spraying robot described in Example 5;

[0108] Figure 6 This is a top view of the structure of the precision spraying robot described in Example 5;

[0109] Figure 7 This is a schematic diagram of the structure of the rectangular coordinate robot arm assembly described in Example 5;

[0110] Figure 8 The variable spray system connection diagram of Example 5;

[0111] Fig. 9 This is a schematic diagram of the workflow of the precision spraying robot described in Example 5.

[0112] Numbers in the figure:

[0113] 1-image acquisition camera, 2-adjustable camera bracket, 3-laser bracket, 4-laser, 5-front antenna, 6-control cabinet, 7-hardware box, 8-rear antenna, 9-medicine box port, 10-chassis, 11-steering stepper motor, 12-wheel connecting frame, 13-sprinkler module, 14-rectangular coordinate robotic arm assembly, 15-wheel, 16-medicine box, 17-diaphragm pump, 18-high-pressure medicine box, 19-battery, 20-frame platform, 21-motor driver, 22-single chip microcomputer, 23-microcomputer, 24-driving wheel assembly, 25-guide rail, 26-slider, 27-profile frame, 28-synchronous belt, 29-X-axis reduction stepper motor, 30-slider base, 31-sprinkler fixing plate, 32-Y-axis reduction stepper motor. DETAILED DESCRIPTION

[0114] The following will be combined with the embodiments of the present invention and the accompanying drawings to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0115] Example 1

[0116] The construction of weed recognition model based on convolutional neural network includes

[0117] S1 Create a dataset

[0118] Use a camera to collect at least 1,000 field images, and then perform data enhancement on the collected images;

[0119] The data enhancement method includes rotation, cropping, mirroring, scaling, and translation to obtain a data-enhanced image set;

[0120] Use the labelimg tool to label crops and weeds in the enhanced image set and save the corresponding label files;

[0121] Randomly select images and corresponding label files according to the ratio of 7:2:1 to divide the image set into a training set, a validation set, and a test set;

[0122] S2 builds a weed recognition model

[0123] Establish a weed recognition model based on convolutional neural network;

[0124] Reference Figure 1 , the weed recognition model includes an input layer, a backbone network, a branch processing layer, an upsampling layer and an output layer;

[0125] The weed recognition model starts from the input layer, firstly processes the input data through the convolution and maximum pooling layer (Conv_maxpool|0), then extracts the main features through the main feature extraction module (Triple ShuffleBlock|1, 3, 5, 6), and then further forward propagates through the advanced feature extraction module (StarBlock|2, 4), and then splices the features output by the main network with the features of the upsampling layer through the branch processing layer, and performs feature fusion through the feature fusion module (SPPF|7) to obtain multiple feature maps of different sizes, and finally outputs the recognized weed image after processing by the output layer;

[0126] The weed recognition model is a weed recognition model based on YOLOv8-seg;

[0127] The weed recognition model uses the lightweight backbone network ShuffleNetV2 and StarNet to jointly establish a new lightweight backbone network: Triple ShuffleBlock and StarBlock modules are used to extract features and simplify the backbone network;

[0128] The backbone feature extraction module (Triple ShuffleBlock, shown in FIG2(e)) is improved from the ShuffleBlock in the lightweight backbone network ShuffleNetV2 (shown in FIG2(d)). The Triple ShuffleBlock module includes three parallel branches and defines a division ratio. By reasonably allocating computing resources for different types of convolution operations, each branch performs different convolution operations respectively, and then connects the outputs of each branch and performs channel fusion.

[0129] The advanced feature extraction module (StarBlock, shown in Figure 2(b)) adopts a structure of depthwise separable convolution, parallel fully connected layers and element-wise multiplication, combined with residual connection design, to achieve more efficient feature extraction and fusion; the StarBlock module is improved from StarNet;

[0130] The weed recognition model based on YOLOv8-seg, the branch processing process includes:

[0131] In the first branch, the feature map first passes through the upsampling layer (Upsample|8), then is concatenated with the output of a previous layer (Concat|9), and is enhanced through the ECA layer (ECA|10). It is then further processed through the C2f_Star layer (C2f_Star|11), and then upsampled (Upsample|12) and concatenated (Concat|13). The features are enhanced through the ECA layer (ECA|14), and finally the first segmentation result (Segment) is obtained through the C2f_Star layer (C2f_Star|15) and the convolution layer (Conv|16).

[0132] In the second branch, the feature map is directly concatenated with the output of a previous layer (Concat|17), and feature enhancement is performed through the ECA layer (ECA|18). After being processed by the C2f_Star layer (C2f_Star|19) and the convolution layer (Conv|20), the second segmentation result (Segment) is obtained.

[0133] In the third branch, the feature map is also concatenated with the output of a previous layer (Concat|21), enhanced by the ECA layer (ECA|22), and then processed by the C2f_Star layer (C2f_Star|23), and finally the third segmentation result (Segment) is obtained;

[0134] The entire processing process makes full use of multiple feature extraction, enhancement and aggregation operations to achieve accurate segmentation of input image data;

[0135] The C2f_Star layer is obtained by replacing the Bottleneck module in the C2f module of the model network with the StarBlock module;

[0136] The C2f_Star module (shown in FIG. 2( a) ) first adjusts the number of channels through a convolutional layer, then divides the adjusted feature map into several blocks, extracts features through convolution kernels of different sizes, and further processes features through StarBlock modules, then concatenates the output feature maps, and finally outputs the adjusted feature maps through a convolutional layer;

[0137] The C2f_Star module uses the StarBlock module to perform element-by-element multiplication operations (StarOperation) to significantly improve the nonlinear expression ability and performance of the model;

[0138] An ECA layer is provided at the model fusion node, and the ECA layer is a channel attention module;

[0139] The ECA layer (shown in Figure 2(c)) first performs global average pooling on the input feature map, then generates channel attention through one-dimensional convolution, passes the convolved features through the Sigmoid activation function to generate channel attention weights, and multiplies the attention weights by the original input feature map element by element between channels to obtain the output feature map;

[0140] S3 model training and validation

[0141] The training set generated by S1 is input into the weed recognition model for training. After the model training is completed, the optimal weight is selected as the weight of the model to generate the weed recognition model;

[0142] Verify and test the weed recognition model through the validation set and test set;

[0143] A weed recognition model is obtained, which can be used for weed target recognition and image segmentation in crop fields.

[0144] Example 2

[0145] Weed recognition methods based on convolutional neural networks include

[0146] Capture field pictures in real time;

[0147] The weed recognition model described in Example 1 is input in real time to perform weed recognition and segmentation, and a weed target frame and a mask image are obtained.

[0148] Example 3

[0149] Weed recognition system based on convolutional neural network, including

[0150] A data processing unit is used to obtain field images, process and annotate them, and then divide them into a training set, a validation set, and a test set;

[0151] A weed recognition model building unit, used to establish the weed recognition model based on the convolutional neural network described in Example 1; the weed recognition model includes an input layer, a backbone network, a branch processing layer, an upsampling layer and an output layer; the weed recognition model is a weed recognition model based on YOLOv8-seg;

[0152] Model training and verification unit, used for training, verification and testing of weed recognition models;

[0153] The weed recognition unit is used to capture field images in real time and input them into the weed recognition model to perform target detection and semantic segmentation of weeds, and obtain weed target frames and mask images.

[0154] Example 4

[0155] Precision spraying methods for weeds in the field include:

[0156] S1 Weed Image Collection

[0157] Real-time image acquisition of crop areas;

[0158] S2 weed recognition segmentation

[0159] Input the image into the weed recognition model constructed in Example 1 to obtain the identified and segmented weed target frame and mask image;

[0160] S3 Weed Location and Classification

[0161] Weed location and classification are performed on the S2 output image;

[0162] S31 Weed Location

[0163] The geometric center of the model output target box is located and the mask area is calculated, where the center of the target box is set as the target center point, and the target coordinates are output, and the mask area of ​​a single object in the image is calculated using a summation function;

[0164] Get the coordinates and area of ​​the center point of the weeds, and output the coordinates and area in a list;

[0165] Among them, the positioning algorithm uses the target bounding box to utilize Calculate the mask area using cv2.contourArea function;

[0166]

[0167] It is known that the image captured by the camera is 1280*960, which corresponds to the real ground 256*192mm, that is, 5 pixels correspond to the real distance 1mm. Convert the image coordinates to the real coordinates. Assume that the image coordinates are (x img ,y img ), the real coordinate is (x real ,y real ), the target pixel area is Area img , the target real area is Area real ;

[0168]

[0169] S32 Weed Classification

[0170] Reference Figure 3 ,When the weed target box mask area occupies 40% or more of the total image area, it is defined as the full spraying area, and continuous spraying operation will be performed in this area;

[0171] When the weed target frame mask area accounts for less than 40% of the total image area, it is defined as a fixed-point precision spraying area, and the weeds in this area are classified into weed grades according to the mask area:

[0172] If the weed target frame mask area is less than 225mm 2 When , the judgment is jumped out and the weed level is assigned to level "1";

[0173] If the weed area is larger than 225mm 2 If the weed area is less than 900mm, the judgment will continue. 2 , then the judgment will be jumped out and the weed level will be assigned to level "2". If the weed area is greater than 900mm 2 , then the weed level is assigned to level "3";

[0174] Return all the weed coordinates and grade values ​​in the image to a list and output them one by one;

[0175] S4 Path Planning

[0176] Execute path planning for precise weed spraying based on the output coordinate list;

[0177] If it is the entire spraying area, the planning algorithm will not be used for planning, and the spraying command will be executed directly;

[0178] If it is a fixed-point precision spraying area, it will be planned through the planning algorithm, refer to Figure 4 :

[0179] The ant colony algorithm is used for position planning. First, the data information output by S32 weed classification is received, and several ants are placed at each target coordinate point. Then, each ant searches for the next target coordinate point until all coordinate points are traversed, and the pheromone between each coordinate point is updated. The above process is repeated until the number of iterations is zero, and then the planned coordinate list is output;

[0180] S5 precision variable spray

[0181] According to the planned coordinate list and the weed grade value corresponding to each coordinate point, corresponding actions are executed to achieve precise variable spraying;

[0182] S51 dual stepper motor movement control

[0183] When the Y-axis deceleration stepper motor 32 is working, the stepper motor shaft can be rotated, and the rotation of the shaft drives the rotation of the synchronous belt 28 of the guide rail 25, and the rotation of the synchronous belt 28 drives the movement of the slider 26, and then drives the third guide rail 25 connected to the slider 26; when the X-axis deceleration stepper motor 29 is working, it drives the synchronous belt 28 on the third guide rail 25 to rotate, and then drives the nozzle connected to the slider 26 to achieve precise movement between coordinate points;

[0184] The dual stepper motor control algorithm includes:

[0185] First, after receiving the first position coordinates (x1, y1) and the second position coordinates (x2, y2), the straight line length L is calculated, the unit step length n is defined, and the total number of steps N is calculated;

[0186]

[0187] Then calculate the movement increments Δx and Δy in each cycle in the x and y directions, that is, in N cycles, every time the x direction moves Δx, the y direction moves Δy;

[0188]

[0189] Since floating point operations may introduce errors, the last step is to move directly to the end point, that is, after N-1 cycles of Δx and Δy driving pulses, position correction movement is performed in the Nth cycle to compensate for the calculation error, that is, the Nth cycle moves x error ,y error ;

[0190] x error =(x2-x1)-Δx*(N-1);

[0191] y error =(y2-y1)-Δy*(N-1);

[0192] S52 spray quantity control

[0193] Variable spray control algorithm:

[0194] By assigning duty cycles to different levels of weeds;

[0195]

[0196] The frequency of the PWM signal is set to 1kHz, and the on / off state of the nozzle (i.e., the spraying amount) is controlled by adjusting the duty cycle of the PWM signal to ensure that the target spraying amount is accurately achieved;

[0197] For example, when the weed level is "1", the control signal duty cycle is 33%, that is, the sprinkler is open for 33% of the time and closed for 64% of the time in one cycle. The same is true for other weed levels.

[0198] This generates and outputs a PWM control signal to control the switch of the nozzle solenoid valve, ensuring that the target spray volume is accurately achieved to respond to the spray volume requirements of different weeds.

[0199] Example 5

[0200] Reference Figure 5 , Figure 6, field weed precision spraying system / precision spraying robot, mainly includes

[0201] A weed positioning and grading system comprises an image acquisition camera 1 and a microcomputer 23; the image acquisition camera 1 is fixed in an adjustable camera bracket 2 and connected to the microcomputer 23, the image acquisition camera 1 acquires images of a crop area, transmits the images to the microcomputer 23, processes the images through the weed recognition model described in Example 1, outputs the recognition results into a weed coordinate list and corresponding weed grade information through positioning, grading and path planning algorithms, and sends the data to a variable spraying system;

[0202] The variable spraying system includes a rectangular coordinate mechanical arm assembly 14 and a spraying assembly; wherein

[0203] Reference Figure 7 The rectangular coordinate robot arm assembly 14 includes an X-axis deceleration stepper motor 29, a Y-axis deceleration stepper motor 32, a plurality of guide rails 25, a plurality of sliders 26, a profile frame 27, a synchronous belt 28, a slider base 30, a nozzle fixing plate 31, and a nozzle module 13. The guide rails 25 are fixed to the robot arm assembly frame through the guide rail base, wherein the two guide rails 25 are arranged in parallel and parallel to the robot's travel direction, so that the slider 26 moves along the corresponding guide rail 25. The Y-axis deceleration stepper motor 32 is fixed to one end of the two guide rails 25, and is connected to the synchronous belts 28 on the two guide rails 25 through a coupling, so that the two guide rails 25 maintain synchronization during operation. The two sliders 26 are respectively matched with the two guide rails 25 and fixed at the same position of the synchronous belt 28, so that the slider 26 can move freely along the direction of the guide rails 25. The third guide rail 25 is fixed on the two sliders 26 through the slider base 30, so that it is perpendicular to the travel direction and moves synchronously with the two sliders 26. The third slider 26 is matched on the third guide rail 25 and connected to the synchronous belt 28 on the guide rail 25, so that it can move freely along the direction of the guide rail 25. The X-axis reduction stepper motor 29 is installed at the end of the third guide rail 25 through a reducer to drive the synchronous belt 28 on the guide rail 25, which is used to accurately determine the position of the slider 26 to achieve precise positioning.

[0204] Specifically, when the Y-axis deceleration stepper motor 32 is working, the rotation of the stepper motor shaft can be realized, and the rotation of the shaft drives the rotation of the synchronous belt 28 of the two guide rails 25, and the rotation of the synchronous belt 28 drives the movement of the corresponding slider 26, and then drives the third guide rail 25 connected to the slider 26; when the X-axis deceleration stepper motor 29 is working, it drives the synchronous belt 28 on the third guide rail 25 to rotate, and then drives the nozzle module 13 connected to its slider 26. The rectangular coordinate robot arm assembly 14 controls the coordinated movement of the two stepper motors through the dual-motor control algorithm to realize the rapid and accurate movement of the nozzle module 13 in a certain area;

[0205] Reference Figure 8 The spraying assembly includes a diaphragm pump 17, a medicine box 16, an overflow valve, a high-pressure medicine box 18, a medicine discharge pipe and a solenoid valve;

[0206] The diaphragm pump 17 is used to extract medicine from the medicine box 16 (medicine box port 9) when receiving a control instruction, and transport it to the high-pressure medicine box 18 through the medicine inlet pipe. A pressure sensor is installed in the high-pressure medicine box 18 to identify the pressure of the medicine liquid in the high-pressure medicine box 18. When the pressure inside the high-pressure medicine box reaches a certain value, the high-pressure medicine flows back to the medicine box 16 through the overflow valve, and the diaphragm pump 17 stops working; the solenoid valve is arranged on the medicine discharge pipe and connected between the high-pressure medicine box 18 and the nozzle module 13, that is, one end of the medicine discharge pipe is connected to the high-pressure medicine box 18, and the other end is connected to the nozzle module 13. When the rectangular coordinate mechanical arm assembly 14 is controlled to reach the position above the weeds and a spraying instruction is received, the solenoid valve is controlled to spray the weeds with variable precision;

[0207] The walking navigation system includes a frame platform 20, a driving wheel assembly 24 and a navigation assembly; wherein

[0208] The frame platform 20 is used to place various hardware facilities; the driving wheel assembly 24 includes a wheel 15, a wheel connecting frame 12, and a steering stepper motor 11. The frame platform 20 is fixedly connected to the driving wheel assembly 24, and the four wheels 15 have the ability to steer and drive. A medicine box 16, a hardware box 7, a control cabinet 6, a battery 19, a front antenna 5, and a rear antenna 8 are placed on the frame platform 20. The control cabinet 6 is equipped with a microcomputer 23, a single-chip microcomputer 22, a motor driver 21, etc., and a cabinet door is set on the side of the frame platform 20; the hardware box 7 is equipped with a diaphragm pump 17, an overflow valve, a high-pressure medicine box 18, an electromagnetic valve, etc., and a cabinet door is set on the side of the frame platform 20. door; the chassis 10 is used to separate various hardware above the frame platform 20 and reduce exposure to the natural environment; the rectangular coordinate mechanical arm assembly 14 is arranged below the rear end of the frame platform 20 in the walking direction, and the image acquisition camera 1 is arranged at the bottom of the front end; the navigation assembly includes a satellite navigation device and a laser radar module, and global positioning is performed by combining the laser radar 4 with the satellite navigation system. The front antenna 5 and the rear antenna 8 are arranged at the top of the frame platform 20 to provide position information. The laser radar module is arranged at the front end of the frame platform 20 and fixed by the laser radar bracket 3 to provide real-time road condition information for the field autonomous walking of the precision spraying robot;

[0209] The power supply system includes a battery 19 and a voltage converter. The power output by the battery 19 is passed through the voltage converter to provide different power requirements for each system and component. The battery 19 also provides a charging interface.

[0210] Reference Fig. 9, the working process of this precision spraying robot is:

[0211] When the precision spraying robot walks autonomously in the field, the image acquisition camera 1 reads the crop information in the field in front in real time, and the acquired image is uploaded to the microcomputer 23 in real time and is segmented by crop and weed identification model. The coordinate information and area information of the identified weeds are processed into the coordinate information and grade information of the weeds through the weed positioning and grading algorithm, and sent to the variable spraying system for path planning. The lower computer controls the rotation of the two reduction stepper motors to drive the free movement of the rectangular coordinate mechanical arm, and moves the nozzle coordinates to the top of the target weed position. Combined with the weed grade information, the lower computer sends a spraying instruction to control the solenoid valve to transport the liquid medicine from the discharge pipe of the high-pressure medicine box 18 to the nozzle, and the liquid medicine is sprayed through the nozzle, thereby completing the precise variable spraying of the weeds to kill the weeds.

[0212] It should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A weed recognition method based on a convolutional neural network, characterized in that: include S1 Create a dataset Collect field images and perform data enhancement on the collected images; The data enhancement method includes any one of rotation, cropping, mirroring, scaling, and translation, or a combination thereof, to obtain a data-enhanced image set; Annotating the image set with crops and weeds, and saving corresponding label files; Dividing the image set into a training set, a validation set, and a test set; S2 builds a weed recognition model Establishing a weed recognition model based on a convolutional neural network; the weed recognition model is a weed recognition model based on YOLOv8-seg; The weed recognition model uses a lightweight backbone network ShuffleNetV2 and StarNet to jointly establish a lightweight backbone network: Triple ShuffleBlock and StarBlock modules are used to perform feature extraction and backbone network simplification; The backbone feature extraction module, namely the Triple ShuffleBlock module, is improved from the ShuffleBlock in the lightweight backbone network ShuffleNetV2 module. The Triple ShuffleBlock module includes three parallel branches and defines a division ratio. By reasonably allocating computing resources for different types of convolution operations, each branch performs different convolution operations respectively, and then connects the outputs of each branch and performs channel fusion. The advanced feature extraction module, i.e., the StarBlock module, adopts a structure of deep separable convolution, parallel fully connected layers, and element-wise multiplication, combined with a residual connection design, to achieve efficient feature extraction and fusion; the StarBlock module is improved from StarNet; The backbone network of the weed recognition model starts from the input layer, firstly performs preliminary processing on the input data through the convolution and maximum pooling layers, then performs backbone feature extraction through the backbone feature extraction module, and then performs further forward propagation through the advanced feature extraction module, and then splices the features output by the backbone network with the features of the upsampling layer through the branch processing layer, and performs feature fusion through the feature fusion module to obtain multiple feature maps of different sizes, and finally outputs the identified weed image after processing by the output layer; S3 model training and validation The training set is input into the weed recognition model for training. After the model training is completed, the optimal weight is selected as the weight of the model to generate the weed recognition model; Verify and test the weed recognition model through the validation set and test set; Application of S4 model and weed identification Capture field pictures in real time and input them into the weed recognition model for weed recognition and segmentation to obtain weed target frames and mask images.

2. The weed identification method based on convolutional neural network according to claim 1, characterized in that: The branch processing process of the weed identification model includes: In the first branch, the feature map first passes through the upsampling layer, then is concatenated with the output of a previous layer, and is enhanced through the ECA layer. It is then further processed by the C2f_Star layer, upsampled and concatenated again, and enhanced through the ECA layer. Finally, it passes through the C2f_Star layer and the convolution layer to obtain the first segmentation result. In the second branch, the feature map is directly concatenated with the output of a previous layer, and the feature is enhanced by the ECA layer. After being processed by the C2f_Star layer and the convolution layer, the second segmentation result is obtained. In the third branch, the feature map is also concatenated with the output of a previous layer, enhanced by the ECA layer, and processed by the C2f_Star layer to finally obtain the third segmentation result. Through multiple feature extraction, enhancement and aggregation operations, accurate segmentation of input image data is achieved.

3. The weed identification method based on convolutional neural network according to claim 2, characterized in that: The C2f_Star layer is obtained by replacing the Bottleneck module in the C2f module of the model network with the StarBlock module; The C2f_Star module first adjusts the number of channels through the convolution layer, then divides the adjusted feature map into several blocks, extracts features through convolution kernels of different sizes, and further processes features through the StarBlock modules respectively, then concatenates the output feature maps, and finally adjusts the output feature maps through the convolution layer.

4. The weed identification method based on convolutional neural network according to claim 2, characterized in that: An ECA layer is provided at the model fusion node, and the ECA layer is a channel attention module; The ECA layer first performs global average pooling on the input feature map, then generates channel attention through one-dimensional convolution, passes the convolved features through the Sigmoid activation function to generate channel attention weights, and multiplies the attention weights and the original input feature map element by element between channels to obtain the output feature map.

5. A weed identification system based on a convolutional neural network, characterized by: include A data processing unit is used to obtain field images, process and annotate them, and then divide them into a training set, a validation set, and a test set; The weed recognition model construction unit is used to establish a weed recognition model based on a convolutional neural network; the weed recognition model is a weed recognition model based on YOLOv8-seg; the weed recognition model uses a lightweight backbone network ShuffleNetV2 and StarNet to jointly establish a lightweight backbone network: feature extraction and backbone network simplification are performed through Triple ShuffleBlock and StarBlock modules; the backbone feature extraction module, namely the TripleShuffleBlock module, is improved by ShuffleBlock in the lightweight backbone network ShuffleNetV2 module, and the Triple The ShuffleBlock module includes three parallel branches and defines the division ratio at the same time. By reasonably allocating the computing resources of different types of convolution operations, each branch performs different convolution operations respectively, and then connects and channels the outputs of each branch; the advanced feature extraction module, i.e., the StarBlock module, adopts the structure of deep separable convolution, parallel fully connected layer and element-level multiplication, combined with residual connection design, to achieve efficient feature extraction and fusion; the StarBlock module is improved by StarNet; the weed recognition model includes an input layer, a backbone network, a branch processing layer, an upsampling layer and an output layer; the backbone network of the weed recognition model starts from the input layer, firstly performs preliminary processing on the input data through convolution and maximum pooling layers, then performs backbone feature extraction through the backbone feature extraction module, and then further forward propagates through the advanced feature extraction module, and then splices the features output by the backbone network with the features of the upsampling layer through the branch processing layer, and performs feature fusion through the feature fusion module to obtain multiple feature maps of different sizes, and finally outputs the identified weed image after processing by the output layer; Model training and verification unit, used for training, verification and testing of weed recognition models; The weed recognition unit is used to capture field pictures in real time and input the weed recognition model to perform weed recognition and segmentation to obtain a weed target frame and mask image.

6. A method for precisely spraying weeds in a field, characterized by: include S1 Weed Image Collection Real-time image acquisition of crop areas; S2 weed recognition segmentation Inputting the image into the weed recognition model according to any one of claims 1 to 5 to obtain the identified and segmented weed target frame and mask image; S3 Weed Location and Classification Weed location and classification based on S2 acquired images; S31 Weed Location The geometric center of the model output target box is located and the mask area is calculated, where the center of the target box is set as the target center point, and the target coordinates are output, and the mask area of ​​a single object in the image is calculated using a summation function; Get the coordinates and area of ​​the center point of the weeds, and output the coordinates and area in a list; S32 Weed Classification When the weed target frame mask area occupies more than the set proportion of the total image area, it is defined as the full spraying area, and continuous spraying operation will be carried out on this area; When the weed target frame mask area is less than the set ratio of the total image area, it is defined as a fixed-point precision spraying area, and the weeds in the area are classified into weed grades according to the mask area: Return all the weed coordinates and grade values ​​in the image to a list and output them one by one; S4 Path Planning Weed spraying path planning based on output coordinate list; If it is the entire spraying area, no path planning is performed and the spraying command is executed directly; If it is a fixed-point precision spraying area, path planning is performed: Ant colony algorithm is used for path planning. First, the data information output by S32 weed classification is received, and several ants are placed at each target coordinate point. Then, each ant searches for the next target coordinate point until all coordinate points are traversed, and the pheromone between each coordinate point is updated. The above process is repeated until the number of iterations is zero, and then the coordinate list after path planning is output. S5 precision variable spray According to the coordinate list after path planning and the weed level value corresponding to each coordinate point, perform corresponding actions: Through the stepper motor movement control, the print head is driven to achieve precise movement between coordinate points; By assigning duty cycles to weeds of different levels, a PWM control signal is generated and output to control the on / off state of the nozzle, thereby achieving precise variable spraying of the target spraying amount.

7. The field weed precision spraying system is characterized by: include A weed positioning and grading system, comprising an image acquisition camera and a microcomputer; the image acquisition camera acquires images of a crop area, transmits the images to the microcomputer, processes the images through the weed recognition model according to any one of claims 1 to 5, outputs the recognition results into a weed coordinate list and corresponding weed grade information through positioning, grading and path planning algorithms, and sends the data to a variable spraying system; A variable spraying system, comprising a rectangular coordinate mechanical arm assembly and a spraying assembly; wherein the rectangular coordinate mechanical arm assembly comprises an X-axis deceleration stepper motor, a Y-axis deceleration stepper motor, and a nozzle module, and the dual-motor control algorithm is used to control the coordinated movement of two stepper motors to achieve rapid and precise movement of the nozzle module; the spraying assembly comprises a high-pressure medicine box, a medicine discharge pipe, and a solenoid valve, wherein the solenoid valve is arranged on the medicine discharge pipe and connected between the high-pressure medicine box and the nozzle module; when the nozzle module reaches the position above the weeds by controlling the rectangular coordinate mechanical arm assembly, the weeds are precisely sprayed with variable amount of medicine by controlling the switch state of the solenoid valve; A walking navigation system includes a frame platform, a driving wheel assembly and a navigation assembly; the rectangular coordinate mechanical arm assembly is arranged at the rear end of the frame platform in the walking direction, and the image acquisition camera is arranged at the front end; the driving wheel assembly includes a wheel and a steering stepper motor; the navigation assembly includes a satellite navigation device and a laser radar module; the laser radar module is arranged at the front end of the frame platform to provide real-time road condition information for the autonomous walking of the precision spraying system in the field; The power supply system, including batteries, is used to provide power requirements for various systems and components.

8. The field weed precision spraying system according to claim 7, characterized in that: The rectangular coordinate robot arm assembly includes an X-axis deceleration stepper motor, a Y-axis deceleration stepper motor, a plurality of guide rails, a plurality of sliders, a synchronous belt, a slider base, and a nozzle module; The guide rails are fixed on the frame of the robot arm assembly, wherein the two guide rails are arranged in parallel and parallel to the direction of travel of the system, the Y-axis reduction stepper motor is fixed at one end of the two guide rails, and the synchronous belts on the two guide rails are connected through a coupling, so that the two guide rails are consistent during operation, and the two sliders are respectively matched with the two guide rails and fixed at the same position of the synchronous belt, so that the sliders can move freely along the direction of the guide rails; The third guide rail in the guide rail is fixed on the two sliders through the slider base, so that it is perpendicular to the travel direction and moves synchronously with the two sliders. The third slider in the slider is matched on the third guide rail and connected to the synchronous belt on the guide rail, so that it can move freely along the direction of the guide rail. The X-axis reduction stepper motor is installed at the end of the third guide rail through the reducer to drive the synchronous belt on the guide rail, which is used to accurately determine the position of the slider to achieve precise positioning.

9. The field weed precision spraying system according to claim 7, characterized in that: The spraying assembly includes a diaphragm pump, a medicine box, an overflow valve, a high-pressure medicine box, a medicine discharge pipe and a solenoid valve; The diaphragm pump is used to extract medicine from the medicine box when receiving a control instruction, and transport it to the high-pressure medicine box through the medicine inlet pipe. A pressure sensor is installed in the high-pressure medicine box to identify the pressure of the medicine liquid in the high-pressure medicine box. When the internal pressure reaches a certain value, the high-pressure medicine flows back to the medicine box through the overflow valve and stops the operation of the diaphragm pump; the solenoid valve is arranged on the medicine discharge pipe and connected between the high-pressure medicine box and the nozzle module.

Images