Intelligent crop thinning method based on machine vision

Abstract

The invention discloses an intelligent crop thinning method based on machine vision, and belongs to the field of agricultural intelligent equipment. According to the method, RGBD images of crop seedlings are acquired through an image acquisition module, and effective seedling areas are extracted by using four-stage mechanisms of HSV color segmentation, morphological filtering, contour geometric screening and depth information filtering; based on camera calibration, mapping seedling plant centroids into world coordinates, calculating adjacent plant spacing, and counting a local mean value as a dynamic judgment reference; a weighted scoring function is constructed by integrating the seedling height, the canopy area and the color uniformity, and a competition elimination type thinning decision is automatically executed in an over-dense area; the two-dimensional sliding table is controlled to drive the pressure feedback clamping and shearing device to accurately position and remove redundant seedlings. According to the method, a'perception decision execution 'millisecond-level closed loop is realized, the operation efficiency is improved by 5 times, and the thinning accuracy is gt; the seedling thinning method has the advantages that the seedling thinning rate is increased by 96%, the retention rate of robust seedlings is increased by 19.3%, three embodiments of low cost, high performance and cluster cooperation are provided, the seedling thinning method is suitable for automatic seedling thinning of various drill-sown crops such as corn, wheat and cotton, and the agricultural precision management level is promoted to be improved.

Description

Technical Field

[0001] This invention relates to a complete technical solution for precise identification of crop seedlings, intelligent assessment of plant spacing, and automated thinning operations that integrates machine vision and multi-sensor information fusion, belonging to the field of agricultural intelligent equipment technology. Background Technology

[0002] In traditional crop cultivation, to ensure reasonable plant density and nutrient distribution, thinning must be carried out after emergence, which involves manually removing overly dense, weak, or deformed seedlings. This process relies entirely on the operator's experience and judgment, resulting in significant problems such as high labor intensity, low efficiency (approximately 0.1 hectares / hour per person), strong subjectivity, high rate of missed seedlings, and a high rate of mistaken removal of healthy seedlings. These issues directly affect the quality of the crop population and the final yield in the later stages. Existing mechanical thinning equipment mostly uses fixed-spaced punching or random removal methods, such as disc-type thinners and chain-type seedling pullers. These devices lack the ability to sense the actual distribution of seedlings and cannot achieve variable operations of "removing on demand." This easily leads to a mechanical damage rate as high as 15% to 20%, and they are prone to accidentally damaging healthy seedlings or missing densely populated areas, making it difficult to meet the development needs of precision agriculture. In recent years, some studies have attempted to apply deep learning technology to crop identification, but most have focused on pest and disease diagnosis or yield prediction at maturity, with few applications for real-time control in thinning scenarios. The main technical bottlenecks are: firstly, the high computational complexity of deep learning models makes them difficult to deploy at the edge, hindering their ability to meet the real-time requirements of field operations; secondly, the "visual detection, decision-making, planning, and mechanical execution" closed-loop system faces engineering challenges in dynamic farmland environments, including response lag, inaccurate positioning, and poor anti-interference capabilities; and thirdly, existing solutions are complex in structure and expensive, making them unsuitable for the dispersed planting patterns of smallholder farmers in my country. Therefore, there is an urgent need for a visually guided intelligent thinning method with a clear technical route, rapid system response, moderate economic cost, and the ability to be deployed in batches in an engineered manner. This method can achieve integrated closed-loop control of "seeing seedlings, judging distance, determining action", break through the technical bottleneck of traditional thinning mode, and promote the intelligent upgrading of field management. Summary of the Invention To address the aforementioned technical problems, this invention aims to develop an intelligent thinning method and system that integrates machine vision, multi-sensor information fusion, and electromechanical precision collaborative control. This system is suitable for density control and refined field management of row-sown crops such as corn, wheat, cotton, and sorghum during the 2-4 leaf stage after emergence. The technical solution of this invention integrates image acquisition, intelligent analysis, real-time decision-making, and precise execution mechanisms on a mobile operation platform to construct a closed-loop system of "perception, decision-making, and execution," thereby achieving intelligent control across the entire chain and improving thinning accuracy and operational efficiency. The technical solution comprises two levels: software algorithm processes and hardware system architecture. I. The software algorithm process includes the following steps: S1: Synchronous Image Acquisition and Transmission The image acquisition module, mounted on the mobile operation platform, simultaneously acquires color and depth images of crop seedlings in the field. The image acquisition adopts a vertical downward shooting method, and LED supplementary lighting ensures uniform illumination. The acquired raw image data is transmitted in real time to the image processing and decision module via USB 3.0 or Ethernet interface, with a transmission delay of less than 50 milliseconds. S2: Image Preprocessing and Seedling Area Extraction The received images are processed sequentially as follows: First, white balance correction, histogram equalization, and Gaussian filtering are performed to eliminate illumination color cast and improve image contrast. The preprocessed color image is converted from RGB color space to HSV color space, and threshold segmentation is performed based on the vegetation color sensitive range (H:3585, S>30, V>50) to generate an initial seedling area mask. Morphological opening and closing operations are performed on the mask to remove noise points with an area of ​​less than 50 pixels and fill in the internal cavities of leaves. Contour geometric features are extracted using Canny edge detection to screen connected regions that meet the criteria of "height > 3cm, aspect ratio 0.8-1.5". Depth image information is used to exclude non-seedling interference targets with a ground height < 2cm or > 15cm, and finally, the effective seedling area is determined. S3: Plant spacing calculation and dynamic benchmark assessment Based on the effective seedling area: calculate the geometric moments of the outline of each seedling and extract the centroid pixel coordinates (u,v); map the pixel coordinates to the actual spatial position (X,Y,Z) in the world coordinate system based on the camera calibration parameters (intrinsic and extrinsic matrices); after sorting the seedlings by row direction, calculate the Euclidean distance d_i between adjacent seedlings; calculate the average plant distance \bar{d} within a local continuous range of three seedlings and use it as a dynamic judgment benchmark to adaptively adjust the subsequent decision threshold. S4: Comprehensive score of seedling physiological status For each identified seedling, the following three physiological indicators are calculated: seedling height h (the vertical height of the plant obtained based on depth information), canopy area A (the actual projected area of ​​the seedling area outline), and color uniformity C (the color consistency score of the leaf area in HSV space). Then, a comprehensive physiological state score S for the seedling is generated according to the weighted formula S = 0.4·h + 0.35·A + 0.25·C. The score is used to quantify the robustness of the seedling. S5: Competitive Elimination Thinning Decision Generation Based on the plant spacing data and physiological status scores, the following decision logic is constructed: if the adjacent plant spacing d_i is greater than the preset upper limit threshold d_max, it is marked as a missing seedling point; if d_i is within the allowable range [d_min, d_max], the status quo is maintained; if d_i is less than the lower limit threshold d_min, a competition mechanism is triggered: the physiological status scores S of two adjacent plants are compared, and the one with the lower score is marked as the target to be removed; if the spacing of three or more consecutive plants is less than d_min, an adaptive removal strategy is activated: the single plant with the highest score S in the area is retained, and all others are marked as targets to be removed; finally, a thinning execution instruction containing the world coordinates (X,Y,Z) of the target seedling is generated. S6: Precise Execution and Status Feedback After receiving the thinning execution command, the control and communication module: parses it into target position coordinates and converts it into motor drive pulse signals; drives the XZ two-dimensional electric slide to move the end clamping shear device to the target position with a positioning accuracy of ±2mm; vertically descends to perform stem base cutting, with the shearing force controlled within a 24mm diameter range; the pressure sensor built into the clamping shear head monitors the contact pressure in real time, and stops immediately when the pressure >3 Newtons to prevent excessive soil disturbance; and feeds back the execution completion status and abnormal information to the control module to form a closed-loop monitoring system. S7: Cyclic Operations and Continuous Execution After a single clearing action is completed, the actuator returns to the initial standby position; the system then starts the next inspection cycle and repeats steps S1 to S6; the total response time of the cycle is less than 800 milliseconds, supporting continuous uninterrupted operation at a travel speed of 0.8 m / s, with a seedling loss rate of <3%. II. The hardware system architecture is mainly divided into the following five core modules: 1. Image acquisition module A high-frame-rate industrial camera (such as the Basler acA1920155um) or an RGBD camera (such as the Intel RealSense D435i) is mounted on the front of the mobile platform at a height of 5070 cm above the ground, shooting vertically downwards. The camera operates in the visible light band and is equipped with a ring LED fill light (color temperature 6500K, power 12W), which adapts to different lighting conditions such as sunny days, cloudy days, and dusk through PWM dimming. The acquired images have a resolution of no less than 1280×960 pixels and are transmitted in real time to the image processing and decision-making module via USB 3.0 or Gigabit Ethernet, with a transmission latency of less than 50 milliseconds. 2. Image Processing and Decision Module This module uses an embedded industrial computer (such as NVIDIA Jetson Orin NX with a computing power of 100 TOPS) as the core and runs customized vision algorithms. The specific processing flow is as follows: (1) Image preprocessing: First, perform white balance correction on the original image to eliminate illumination color bias; then perform histogram equalization to enhance the image contrast; finally, use a 5×5 Gaussian filter to smooth the image and suppress sensor noise. (2) Seedling area extraction: Adopt a four-level filtering mechanism - color filtering (convert the RGB image to the HSV color space, set the vegetation sensitive interval H: 3585, S>30, V>50 according to the color characteristics of crop seedlings at the 24-leaf stage, and generate a binary mask), morphological filtering (perform an opening operation on the mask with a 3×3 structuring element to remove noise points with an area less than 50 pixels; then perform a closing operation to fill the holes inside the leaves), geometric feature filtering (extract the contours through Canny edge detection and filter out the connected regions that meet the conditions of "height>3cm, aspect ratio 0.81.5, contour fullness>0.85"), depth information filtering (for the depth map obtained by the RGBD camera, filter out non-seedling objects (such as weeds, soil clods, stones) with a height from the ground less than 2cm or greater than 15cm, and finally determine the effective seedling area). (3) Row spacing calculation: Through centroid extraction (calculate the geometric moments of the contour of each seedling area to obtain the centroid pixel coordinates (u, v)), coordinate mapping (use the pre-calibrated camera internal parameters (focal lengths fx, fy, principal point cx, cy) and external parameters (rotation matrix R, translation vector T), and convert the pixel coordinates to the actual position (X, Y, Z) in the world coordinate system through the perspective transformation formula), distance calculation (sort the seedlings in the row direction (Y coordinate) and calculate the Euclidean distance di between adjacent seedlings. At the same time, calculate the average row spacing \bar{d} of three consecutive plants as the dynamic benchmark. When \bar{d}<18cm, automatically lower the d_min threshold to avoid excessive clearing in areas with dense sowing). (4) Thinning decision: Set the optimal row spacing range [d_min, d_max] (usually 1825cm). The decision logic is as follows: If di<d_min, trigger the competition mechanism, compare the physiological state scores S = 0.4·h + 0.35·A + 0.25·C (h is the seedling height, A is the canopy area, C is the color uniformity score) of the two plants, and mark the one with the lower score as the clearing target; if d{i1}, di, d{i+1} are all less than d_min continuously, start the skip retention strategy: calculate the S values of all the seedlings in this section, retain the seedling corresponding to max(S), and mark all the others for clearing; if di>d_max, it is judged as a missing seedling, and record the position for subsequent replanting reference. 3. Control and Communication Module A hierarchical collaborative architecture of "host computer + slave computer" is adopted: the host computer (industrial control computer) completes complex image processing, outputs the world coordinates (X,Y,Z) of the target seedling to be cleared and the execution parameters, and sends them down via TCP / IP protocol; the slave computer (STM32F407 microcontroller) receives instructions via CAN bus (500kbps baud rate) or RS485 and parses them into motor control pulse signals. The X-axis servo motor (400W, with 17-bit encoder) drives the slide table to move horizontally, and the Z-axis stepper motor (Type 57, 1.8° step angle) controls vertical lifting. The slave computer monitors the motor position, pressure sensor values ​​and solenoid valve status in real time and feeds back the execution results to the host computer. In terms of safety mechanisms, a hardware emergency stop button and software timeout protection are set up. If the actuator fails to reach the target position within 2 seconds or the pressure value is abnormal, the machine will stop immediately and an alarm will sound. 4. Actuator Module Adopting a modular design, it consists of an XZ two-dimensional electric slide table and an end-effector shearing device. The slide table has an X-axis travel of 600mm and a Z-axis travel of 300mm, with a repeatability of ±2mm. It uses a ball screw drive (5mm lead) and linear guide rails to ensure high-speed and stable operation. The shearing head of the shearing device is driven by a miniature cylinder (16mm diameter, 30mm stroke). The blade is made of SKD11 tool steel with a 30° cutting angle, which can easily cut the base of crop stems up to 24mm in diameter. The shearing head has a built-in thin-film pressure sensor (range 0-10N). When the contact pressure is >3N, it is determined to be in contact with the soil and the pressure is immediately stopped. In terms of response time, the total time from receiving the command to completing the shearing action is <800ms. It supports operation at a travel speed of 0.8m / s with a seedling loss rate of <3%. 5. Mobile carrier platform It can be equipped with an electric tracked chassis (80cm wide, ground pressure <20kPa) or suspended on an existing tractor (5080 horsepower). The platform travel speed is adjustable from 0.3 to 0.8m / s. It is equipped with an RTK GPS navigation system and the working path deviation is <5cm. The seedling extraction steps specifically include: Convert the original image from the RGB color space to the HSV color space; Set a sensitive range for green vegetation (H:3585, S>30, V>50) and perform color threshold segmentation; Morphological opening operations are performed on the segmentation results to remove minor noise, and closing operations are performed to fill in holes. Canny edge detection and contour analysis were used to screen connected regions with a height greater than 3 cm and an aspect ratio between 0.8 and 1.5 as candidate seedling areas; By combining depth images, non-seedling targets with a height of less than 2 cm or more than 15 cm above the ground are excluded, thus completing the identification of effective seedling areas. This invention is applicable to field management operations during the 2-4 leaf stage of corn after sowing. By adjusting the HSV threshold parameter and plant spacing judgment standard, it can be extended to automated thinning scenarios for row-sown crops such as wheat, cotton, sorghum, and soybeans. The core innovation of this invention: 1. Multimodal fusion seedling area extraction algorithm: It innovatively combines HSV color space segmentation, morphological denoising, contour geometry screening and depth information filtering to construct a four-level progressive filtering mechanism, which improves the seedling identification accuracy to over 96% under complex field lighting and background interference conditions. 2. Dynamic plant spacing evaluation mechanism: In addition to calculating the absolute distance between adjacent seedlings, the system also introduces the local area adaptive statistical average plant spacing as a dynamic benchmark. The system can automatically adjust the judgment threshold according to the sowing density of the current working block to avoid "one-size-fits-all" misjudgment and improve decision robustness. 3. Competitive Elimination Intelligent Decision Model: Breaking through the traditional single distance judgment mode, it innovatively introduces a multi-dimensional weighted scoring function based on seedling height, canopy area, and color uniformity. In overly dense areas, it automatically executes a leapfrog optimization strategy of "retaining the best seedlings and eliminating the second-best seedlings", which increases the retention rate of healthy seedlings by nearly 20 percentage points. 4. Millisecond-level response electromechanical collaborative system: The system is designed with a hierarchical collaborative control architecture of "industrial computer + microcontroller" and combined with the end pressure feedback safety mechanism to complete the entire set of "detection, positioning and clearing" actions within 800 milliseconds. It supports high-speed continuous operation at 0.8 m / s and the seedling loss rate is less than 3%. The positive and beneficial effects of this invention are as follows: 1. Technological Innovation: For the first time, RGBD multimodal perception, dynamic plant spacing assessment and competitive elimination decision-making mechanism are introduced into the thinning process, breaking through the limitations of traditional mechanical "blind operation" and constructing an integrated closed-loop control system of "perception, decision-making and execution", filling the technological gap in this field; 2. Economic practicality: Operational efficiency is increased by 56 times, labor costs are reduced by 75% to 95%, the retention rate of healthy seedlings is increased by nearly 20 percentage points, the yield per mu is increased by 8% to 12%, the equipment investment payback period is shortened to 12 years, and agricultural production costs are significantly reduced; 3. System Flexibility: The modular design supports three configurations: low-cost popularization, high-performance intelligent, and cluster collaboration, which can cover the needs of all scenarios from small farmers to large farms; by adjusting the HSV threshold and plant spacing parameters, it can be quickly adapted to various row-sown crops such as wheat, cotton, sorghum, and soybeans. 4. Engineering reliability: Millisecond-level response control, pressure feedback safety mechanism, and RTK precision navigation ensure long-term stable operation of the system in complex field environments, with a cumulative fault-free operation time of >200 hours and a mechanical damage rate of <2%, meeting actual production requirements; 5. Environmental adaptability: The four-level filtration mechanism effectively copes with interference from different lighting conditions, soil, and weeds, maintaining an accuracy rate of over 96% under various lighting conditions such as sunny days, cloudy days, and evenings. Attached Figure Description Figure 1 The system's overall architecture diagram fully demonstrates the five core modules of the intelligent seedling thinning system (image acquisition module, image processing and decision-making module, control and communication module, actuator module, and mobile carrier platform) and their collaborative working mechanism. It clearly indicates the data flow interface types (USB3.0 / Ethernet, TCP / IP, CAN bus / RS485) and transmission delay indicators, and uses solid and dashed lines to distinguish the differentiated configuration paths between low-cost and high-performance models. Figure 2 The diagram illustrates the image processing and algorithm flow, detailing the complete algorithm pipeline from raw image input to thinning decision output. It includes a preprocessing unit (white balance correction, histogram equalization, 5×5 Gaussian filtering), a four-level filtering unit (color filtering, morphological filtering, geometric feature filtering, depth information filtering), a coordinate mapping unit, and a decision generation unit. The diagram also marks the computation time and memory usage of each stage. Figure 3 The flowchart for the competitive elimination thinning decision-making logic presents the complete logical branches of the thinning decision-making process in the form of a standard program flowchart. It includes the starting node, the plant spacing judgment module (three-level nested judgment), the competitive scoring sub-process, the skip-style retention strategy, and the coordinate verification output node. The diamond judgment box and the rectangular action box in the figure are marked with the key threshold parameters d_min, d_max and the state transition conditions. Figure 4 The topology diagram of the cluster collaboration system shows the "cloud-edge-device" collaborative architecture consisting of 510 unmanned seedling robots, a central dispatch server, a LoRa wireless communication network, and a high-definition farmland GIS map. It clearly indicates the information broadcasting mechanism between robot nodes, the dynamic task partitioning adjustment logic, and the fault migration path. Detailed Implementation To make the technical solution and beneficial effects of the present invention more specific and clear, a detailed description is provided below with reference to the accompanying drawings and three typical embodiments. Each embodiment features a tiered design in terms of hardware configuration, algorithm complexity, decision-making intelligence, and operational efficiency to meet diverse market demands. The overall system architecture diagram is as follows: Figure 1As shown, the system includes: an image acquisition module for acquiring real-time color and depth images of crop seedlings in the field; an image processing and decision-making module connected to the image acquisition module for executing the image processing, plant spacing calculation, and thinning decision-making algorithm described in claim 1; a control and communication module for receiving decision commands and parsing them into execution signals; an execution mechanism module for responding to execution signals and performing physical removal operations on the target seedlings; and a mobile carrier platform for carrying the above modules and operating in the field. The system fully demonstrates the five core modules of the intelligent thinning system and their collaborative working mechanism. The mobile carrier platform, as the physical base, integrates an electric drive and RTK navigation unit; the image acquisition module uploads the raw image data stream (bandwidth > 300Mbps) to the image processing and decision-making module via a USB 3.0 / Ethernet interface; after running the core algorithm, this module sends target coordinate commands to the control and communication module via the TCP / IP protocol; the control module parses the commands into motor drive pulses based on the CAN bus (500kbps baud rate) or RS485 protocol, driving the execution mechanism to complete the precise thinning action. The diagram specifically marks the differentiated configuration paths for the low-cost model (connected by dashed lines) and the high-performance model (connected by solid lines). The low-cost model uses a Raspberry Pi 4B main controller and a single servo actuator, while the high-performance model is equipped with a Jetson Orin NX edge computing platform and a dual-axis servo slide, fully demonstrating the flexible adaptability of the architecture. Image processing algorithm flowchart as follows Figure 2 As shown, the complete algorithm pipeline from raw image input to thinning decision output is described in detail. The preprocessing unit sequentially performs white balance correction, histogram equalization, and 5×5 Gaussian filtering to optimize the image signal-to-noise ratio. A four-level filtering unit constitutes the core processing chain: the color filtering layer generates an initial vegetation mask based on the HSV threshold; the morphological filtering layer refines the mask through 3×3 structuring element opening and closing operations; the geometric feature filtering layer selects candidate seedling areas based on contour fullness and aspect ratio; and the depth information filtering layer finally removes ground interference. This four-level mechanism works together to improve the recognition accuracy to >96%. The coordinate mapping unit uses pre-calibrated camera intrinsic and extrinsic parameters to achieve real-time conversion from pixels to the world coordinate system. The decision generation unit integrates plant spacing calculation and physiological state scoring, outputting a priority queue of target plants to be removed. The figure shows the computation time (in milliseconds) and memory usage of each step, providing a basis for engineering optimization. The flowchart of the competitive elimination thinning decision-making logic is as follows: Figure 3As shown, the thinning decision logic branch is fully presented in the form of a standard program flow chart. After the start node is triggered, the system first reads the world coordinates and scoring data of the current seedlings. The plant spacing judgment module performs three-layer nested judgment: if d_i > d_max (25 cm), it is marked as a missing seedling and the replanting coordinates are recorded; if d_i ∈ [d_min, d_max], it is skipped; if d_i < d_min (18 cm), it enters the competition scoring sub-process. This sub-process further judges whether it is a continuous overcrowded mode (more than 3 plants continuously): if so, the jumping retention strategy is started (retain the max(S) seedlings and clear the rest); if not, the single-group elimination strategy is executed (clear the min(S) seedlings). All clearing instructions are output to the execution queue after coordinate verification. The key threshold parameters and state transition conditions are marked in the diamond judgment box and rectangular action box in the figure, with strict logic and can be directly converted into embedded control code. Example 1: A low-cost popular technology solution for small farmers This example is designed for small farmers (operating scale < 10 hectares) in hilly and mountainous areas with scattered plots and limited budgets. The core goal is to achieve the basic automated thinning function while controlling costs. Hardware configuration: The image acquisition module selects an OV2640 camera (2 million pixels, frame rate 30fps), uses an industrial camera or RGBD camera, operates in the visible light band, and is equipped with an LED light supplement device to adapt to different lighting conditions; the camera is installed at the front end of the mobile carrier, 50 - 70 cm from the ground, and the collected images are transmitted to the image processing and decision-making module in real time through the USB3.0 or Ethernet interface. The image processing and decision-making module uses a Raspberry Pi 4B main control (4GB memory, 1.5GHz quad-core CPU), the actuator is a simple scissor structure driven by a single servo (shearing torque 15 kg·cm), and the mobile carrier is adapted to a hand-held electric chassis (travel speed 0.3 m / s). The total cost of the system is controlled within 8000 yuan, with outstanding cost performance. As Figure 1As shown by the dashed line connecting the paths, this embodiment employs a lightweight single-modal processing flow, utilizing only RGB image information. It extracts seedling areas through fixed HSV threshold segmentation (H:3585, S>30, V>50) and basic contour filtering (area>50 pixels). Plant spacing calculation uses an offline calibrated pixel distance mapping table, avoiding the computational overhead of real-time 3D reconstruction. The algorithm complexity is reduced by 60%, achieving a processing speed of 2 frames / second on the Raspberry Pi platform, meeting the needs of low-speed operations. The decision logic uses a simplified rule engine, with a preset fixed minimum plant spacing threshold d_min=18cm. When a plant spacing less than the threshold is detected, seedlings on the right are uniformly removed, without introducing a physiological state scoring mechanism. The actuator module includes an XZ two-dimensional electric slide structure and an end clamping and shearing device. The two-dimensional slide is driven by a servo motor or a stepper motor and achieves precise planar positioning through ball screw transmission, with a positioning accuracy of ±2mm. The clamping and shearing device is a pneumatically or electrically controlled shearing head installed at the end of the slide, which can vertically descend to complete the stem base cutting action. The clamping and shearing head is equipped with a miniature pressure sensor feedback unit, which automatically stops pressing down when the contact pressure exceeds a preset value to prevent excessive insertion into the soil and mechanical damage. While sacrificing some level of intelligence, this design significantly reduces the difficulty of software implementation, making it suitable for users with limited technical maintenance capabilities. Operational performance indicators: In field trials at the 24-leaf stage of maize, this embodiment achieved a thinning accuracy rate of 86%, a 12% increase in the retention rate of robust seedlings compared to manual labor, and an operating speed of 0.3 hectares / hour, representing a 3-fold increase in efficiency compared to manual labor. Mechanical damage rate was <5%, and battery life allowed for 4 hours of continuous operation. It is particularly suitable for hilly and terraced areas with complex terrain where large machinery cannot access. Example 2: High-performance intelligent technology solution for large farms. This embodiment is customized for large-scale farms with contiguous planting areas (operating scale > 100 hectares), emphasizing high speed, high precision, all-weather operation capabilities and intelligent decision-making level. Hardware configuration: such as Figure 1 As shown by the solid line connecting the components, the image acquisition module uses an Intel RealSense D435i RGBD camera (1280×720 resolution, 30fps depth streaming), and the image processing core is an NVIDIA Jetson OrinNX edge computing platform (100 TOPS computing power, supporting CUDA acceleration). The actuator is equipped with a dual-axis servo slide (X-axis travel 800mm, Z-axis travel 400mm, repeatability ±2mm) and a hydraulic shear head (shearing force 200N, response time <100ms). The mobile carrier is suspended from an 80-horsepower tractor and equipped with an RTK GPS navigation system. The overall system performance meets industrial-grade standards. like Figure 2As shown, this embodiment implements multimodal deep fusion processing, simultaneously utilizing RGB images and 3D point cloud data. Based on HSV color segmentation, a lightweight YOLOv5s model (7.5M parameters) is introduced for seedling and weed binary classification, effectively distinguishing between crop seedlings and associated weeds (recognition accuracy >95%). A GPU-accelerated real-time stereo matching algorithm is employed to achieve millisecond-level 3D coordinate reconstruction with a processing frame rate of 15fps, providing computational power support for high-speed operations. The method for calculating plant spacing includes: extracting the centroid coordinates (u, v) for each identified seedling; mapping the pixel coordinates to the actual position (X, Y, Z) in the world coordinate system based on camera calibration parameters; sorting the seedlings according to the row direction and calculating the Euclidean distance d_i = ||P{i+1} P_i|| between adjacent seedlings; and calculating the average plant spacing \bar{d} within a local range of three consecutive seedlings for dynamically adjusting the judgment benchmark. The thinning decision logic includes: If the distance between two adjacent seedlings, d_i, is less than the preset minimum plant spacing threshold, d_min (1518cm), a removal mechanism is triggered; The system compares the physiological states of the two seedlings and determines the relatively weaker seedling as the target for removal based on a weighted score S = w1·h + w2·A + w3·C, calculated from seedling height h, canopy area A, and color uniformity C; When three or more consecutive seedlings have a spacing lower than the minimum allowable value, an adaptive removal strategy is activated: all seedlings in the segment are traversed, and the seedling with the highest comprehensive score S is selected for retention, while the rest are marked as removal targets. Figure 3 As shown, this embodiment deploys an adaptive intelligent decision engine. The system dynamically optimizes the plant spacing threshold based on the real-time statistics of the average seedling emergence rate of the current row: when the seedling emergence rate is >90%, d_min is automatically lowered to 15cm to avoid excessive sparseness; when the seedling emergence rate is <70%, d_min is raised to 20cm to retain more robust seedlings. It also supports remote OTA upgrades of decision parameters, enabling rapid response to the agronomic requirements of different varieties and regions. The control and communication module adopts an embedded industrial computer and microcontroller collaborative architecture, including: the host computer runs the image processing algorithm and outputs the world coordinates of the target seedling to be cleared; the slave computer receives the coordinate instructions and controls the motor movement and actuator action through pulse signals; bidirectional communication adopts CAN bus or RS485 protocol to ensure reliable transmission of instructions and status feedback, and the system response time is less than 800 milliseconds. The physiological state scoring weighting coefficient supports online learning and optimization, continuously improving the scientific nature of decision-making. Operational performance indicators: Extensive field validation shows that this embodiment achieves a thinning accuracy rate >96.5%, a robust seedling retention rate 19.3% higher than manual thinning, an operation speed of 0.6 hectares / hour, and a daily operation area of ​​8 hectares. The system supports nighttime LED supplemental lighting (illuminance >1000 Lux), positioning accuracy ±2mm, mechanical damage rate <2%, and seedling loss rate <2.8%, with all performance indicators reaching international advanced levels. Example 3: Cluster-based collaborative distributed thinning system This embodiment is designed for ultra-large-scale agricultural operators (tens of thousands of acres of contiguous farmland), and constructs a distributed operation architecture that coordinates "cloud, edge, and terminal" to achieve swarm intelligence and task optimization. Cluster collaboration system topology diagram as follows Figure 4 As shown, the system consists of 510 small unmanned seedling robots (overall dimensions 60×60×80cm, weight <50kg) forming a work formation. Each robot is configured with the same core modules as in Example 2. A central dispatch server is deployed at the field operation management center. Based on a high-definition farmland GIS map (0.1m resolution) and real-time seedling data, a genetic algorithm is used for dynamic task decomposition and path planning to ensure global optimization. Each robot constructs a decentralized information sharing network through a LoRa wireless communication network (operating frequency 470MHz, transmission distance >1.5km, strong penetration). Figure 4 As shown by the communication link arrow, during the operation, the heat map of seedling distance in the processed area, the coordinates of obstacles (such as stones and utility poles), and equipment status information are broadcast in real time to achieve the following collaborative functions: load balancing (the central scheduler dynamically adjusts the task partitioning according to the real-time position and operation progress of each robot to avoid duplicate operations and resource idleness), segment mutual exclusion (a distributed lock mechanism ensures that the operation segments of adjacent robots do not overlap, and the clearing actions do not interfere with each other), fault migration (when a robot malfunctions or runs out of power, the neighboring device automatically takes over its unfinished operation segment, with a task switching time of <30 seconds and system availability >99%), and edge collaboration (robot nodes can share local seedling condition change information (such as early warning of dense sections) point-to-point, realizing rapid response at the edge and reducing cloud communication latency). Large-scale operation efficiency: When 10 robots operate in parallel, the theoretical operation efficiency reaches 6 hectares / hour. The theoretical operation time for a contiguous farmland of 10,000 mu (667 hectares) is shortened to 111 hours (about 4.6 days). In practice, considering task scheduling and equipment maintenance factors, the total operation time can be controlled within 7 days. It improves efficiency by more than 3 times compared to single-machine operation and by more than 60 times compared to manual operation. For example... Figure 4As shown in the topology diagram, the system supports hot-swappable expansion, allowing for flexible increases or decreases in the number of robots based on agricultural needs, demonstrating strong adaptability and investment flexibility. In terms of fault tolerance and safety, the system employs a redundant design, ensuring that a single robot failure does not affect the overall operation. Equipped with an electronic fence and a multi-sensor fusion obstacle avoidance system, it ensures safety in human-robot hybrid operation scenarios. All robot operation data is backed up to the cloud in real time, supporting historical tracking and agronomic analysis. Field testing and verification Verification was conducted for three consecutive years (2022-2024) in experimental fields of the Shijiazhuang Academy of Agricultural and Forestry Sciences in Hebei Province, covering a total area of ​​120 hectares. Soil types included loam and sandy loam, and maize varieties included Zhengdan 958 and Xianyu 335. The test results are as follows: The results showed that the system maintained stable performance under different light conditions (sunny, cloudy, and evening) and different seedling densities (65,110 plants / meter), significantly improved the retention rate of robust seedlings, and did not cause significant disturbance to the soil structure and adjacent root system, resulting in a significant yield increase. This invention systematically solves the automation problem of crop thinning by employing specific image processing algorithms, intelligent decision-making mechanisms, and highly reliable mechatronics design. (Appendix) Figure 4 The system architecture, algorithm flow, decision logic, and cluster topology demonstrated have all been implemented and verified in the three embodiments described above. These three embodiments fully demonstrate the technology's forward-looking nature and industrial value. Extensive field testing has proven their outstanding creativity and practicality, meeting the standards for invention patent authorization. They possess broad prospects for widespread application and can provide solid technical support for the modernization and intelligent transformation of agriculture in my country.

Claims

1. A machine vision-based intelligent thinning method for crops, comprising software algorithm flow and hardware system architecture, characterized in that, Includes the following steps: S1: Synchronous Image Acquisition and Transmission The image acquisition module, mounted on the mobile operation platform, simultaneously acquires color and depth images of crop seedlings in the field. The image acquisition adopts a vertical downward shooting method, and LED supplementary lighting ensures uniform illumination. The acquired raw image data is transmitted in real time to the image processing and decision module via USB 3.0 or Ethernet interface, with a transmission delay of less than 50 milliseconds. S2: Image Preprocessing and Seedling Area Extraction The received images are processed sequentially as follows: First, white balance correction, histogram equalization, and Gaussian filtering are performed to eliminate illumination color cast and improve image contrast. The preprocessed color image is converted from RGB color space to HSV color space, and threshold segmentation is performed based on the vegetation color sensitive range (H:3585, S>30, V>50) to generate an initial seedling area mask; The mask is filtered by morphological opening and closing operations to remove noise points with an area of ​​less than 50 pixels and fill the voids inside the blade. Combine Canny edge detection to extract contour geometric features and filter connected regions that meet the criteria of "height > 3cm and aspect ratio 0.8-1.5"; By using depth image information to exclude non-seedling interference targets with a ground height of <2cm or >15cm, the effective seedling area is finally determined. S3: Plant spacing calculation and dynamic benchmark assessment Based on the aforementioned effective seedling area: Calculate the geometric moments of the outline of each seedling and extract the centroid pixel coordinates (u,v); Based on camera calibration parameters (intrinsic and extrinsic matrices), pixel coordinates are mapped to actual spatial positions (X, Y, Z) in the world coordinate system; After sorting the seedlings by row direction, calculate the Euclidean distance d_i between adjacent seedlings; The average plant spacing \bar{d} within a range of three consecutive plants in a local area is statistically analyzed and used as a dynamic judgment benchmark to adaptively adjust the subsequent decision threshold. S4: Comprehensive score of seedling physiological status For each identified seedling, the following three physiological indicators were calculated: Seedling height h: Vertical height of the plant obtained based on depth information; Canopy area A: The actual projected area of ​​the seedling area outline; Color uniformity C: Color consistency score of the leaf area in HSV space; Then, a comprehensive physiological state score S for the seedling is generated according to the weighted formula S = 0.4·h + 0.35·A + 0.25·C. The score is used to quantify the robustness of the seedling. S5: Competitive Elimination Thinning Decision Generation Based on the plant spacing data and physiological status scores, the following decision logic is constructed: If the distance between adjacent plants d_i is greater than the preset upper limit threshold d_max, it is marked as a missing seedling point; If d_i is within the allowed range [d_min, d_max], then maintain the status quo; If d_i is less than the lower threshold d_min, a competition mechanism is triggered: the physiological state scores S of two adjacent plants are compared, and the plant with the lower score is marked as the target to be eliminated. If the distance between three or more consecutive plants is less than d_min, then an adaptive removal strategy is activated: retain the single plant with the highest score S in the region, and mark all others as removal targets; Finally, a thinning execution command containing the world coordinates (X,Y,Z) of the target seedling is generated. S6: Precise Execution and Status Feedback After receiving the thinning execution command, the control and communication module then: The target position coordinates are analyzed and converted into motor drive pulse signals. The XZ two-dimensional electric slide table is driven to move the end clamping and shearing device to the target position with a positioning accuracy of ±2mm. The stem base is cut vertically downwards, with the shearing force controlled within a 24mm diameter range; The built-in pressure sensor in the clamping head monitors the contact pressure in real time, and stops the machine immediately when the pressure exceeds 3 Newtons to prevent excessive soil disturbance. The execution completion status and abnormal information are fed back to the control module to form a closed-loop monitoring system. S7: Cyclic Operations and Continuous Execution After a single clearing action is completed, the actuator returns to the initial standby position; the system then starts the next inspection cycle and repeats steps S1 to S6; the total response time of the cycle is less than 800 milliseconds, supporting continuous uninterrupted operation at a travel speed of 0.8 m / s, with a seedling loss rate of <3%.

2. The method as described in claim 1, characterized in that, The image acquisition module uses an industrial camera or an RGBD camera, which operates in the visible light band and is equipped with an LED supplementary lighting device to adapt to different lighting conditions. The camera is installed at the front end of the mobile carrier at a height of 50-70 cm above the ground, and the acquired images are transmitted in real time to the image processing and decision module via a USB 3.0 or Ethernet interface.

3. The method as described in claim 1, characterized in that, The seedling extraction steps specifically include: Convert the original image from the RGB color space to the HSV color space; Set a sensitive range for green vegetation (H:3585, S>30, V>50) and perform color threshold segmentation; Morphological opening operations are performed on the segmentation results to remove minor noise, and closing operations are performed to fill in holes. Canny edge detection and contour analysis were used to screen connected regions with a height greater than 3 cm and an aspect ratio between 0.8 and 1.5 as candidate seedling areas; By combining depth images, non-seedling targets with a height of less than 2 cm or more than 15 cm above the ground are excluded, thus completing the identification of effective seedling areas.

4. The method as described in claim 1, characterized in that, The method for calculating plant spacing includes: Extract the centroid coordinates (u, v) for each identified seedling; Based on camera calibration parameters, pixel coordinates are mapped to actual positions (X, Y, Z) in the world coordinate system. Sort the seedlings according to the row direction, and calculate the Euclidean distance between adjacent seedlings in turn: d_i = ||P{i+1} P_i||. The average plant spacing \bar{d} within a range of three consecutive plants in a local area is statistically analyzed and used to dynamically adjust the judgment benchmark.

5. The method as described in claim 1, characterized in that, The thinning decision logic includes: If the distance between two adjacent seedlings d_i is less than the preset minimum plant spacing threshold d_min (1518cm), the clearing mechanism is triggered. The system compares the physiological states of the two plants and determines the relatively weaker plant as the target for removal based on a weighted score S = w1·h + w2·A + w3·C, which is determined by the weighted score of seedling height h, canopy area A, and color uniformity C. When three or more consecutive seedlings are spaced below the minimum allowable distance, an adaptive removal strategy is activated: all seedlings in the section are traversed, the one with the highest comprehensive score S is selected and retained, and the rest are marked as removal targets.

6. The method as described in claim 1, characterized in that, The actuator module includes an XZ two-dimensional electric slide structure and an end clamping and shearing device; The two-dimensional slide is driven by a servo motor or a stepper motor, and achieves precise planar positioning through ball screw transmission, with a positioning accuracy of ±2mm. The clamping and shearing device is a pneumatically or electrically controlled shearing head, installed at the end of the slide table, which can descend vertically to complete the stem base cutting action; The clamp head is equipped with a miniature pressure sensor feedback unit, which automatically stops pressing down when the contact pressure exceeds the preset value to prevent excessive insertion into the soil and mechanical damage.

7. The method as described in claim 1, characterized in that, The control and communication module adopts a collaborative architecture of embedded industrial computer and microcontroller: The host computer runs the image processing algorithm and outputs the world coordinates of the target seedling to be cleared. The lower-level machine receives coordinate commands and controls the motor movement and actuator actions through pulse signals; Two-way communication uses CAN bus or RS485 protocol to ensure reliable transmission of commands and status feedback, with a system response time of less than 800 milliseconds.

8. The method as described in claim 1, characterized in that, The preprocessing in step S2 includes white balance correction, histogram equalization, and Gaussian filtering; the execution process in step S4 adopts a pressure feedback safety mechanism, which automatically stops the machine when the contact pressure of the clamping head exceeds 3 Newtons.

9. The method according to any one of claims 1 to 8, characterized in that, It is suitable for field management operations during the 2-4 leaf stage after corn sowing. By adjusting the HSV threshold parameters and plant spacing judgment criteria, it can be extended to automated thinning scenarios for row-sown crops such as wheat, cotton, sorghum, and soybeans.

10. A machine vision-based intelligent thinning system for crops, used to implement the method described in claim 19, characterized in that... The system includes: an image acquisition module for acquiring real-time color and depth images of crop seedlings in the field; an image processing and decision-making module connected to the image acquisition module for executing the image processing, plant spacing calculation, and thinning decision-making algorithm as described in claim 1; a control and communication module for receiving decision instructions and parsing them into execution signals; an execution mechanism module for responding to execution signals and performing physical removal operations on the target seedlings; and a mobile carrier platform for carrying the above modules and operating in the field.

Images