A weed growth point positioning method and system based on visual analysis

Abstract

The application discloses a weed growth point positioning method and system based on visual analysis, and relates to the technical field of visual analysis, which comprises the following steps: firstly, acquiring a farmland image and segmenting and extracting initial plant feature segments; analyzing non-adjacent segments by using a binary classification determination model, non-destructively copying and merging associated regions based on a virtual canvas, reconstructing a complete associated subgraph, and solving the problems of crop shielding and interweaving; directly inputting the subgraph into a multi-branch deep convolutional neural network, switching independent parameter sets for cyclic reentry, and sequentially completing absolute removal of protected crops and fine species classification of weeds; and finally, extracting a plant topological skeleton for pixel-level determination according to a species matching Bayesian network positioning model, and calculating and outputting weed growth point coordinates by counting spatial distribution density. The application effectively overcomes farmland overlapping interference, greatly reduces edge calculation power consumption, and significantly improves positioning accuracy and robustness.

Description

Technical Field

[0001] This invention relates to the field of visual analysis technology, and in particular to a method and system for locating weed growth points based on visual analysis. Background Technology

[0002] With the deepening development of smart agriculture and precision agriculture, the use of agricultural robots in the field for automated mechanical weeding or precise targeted chemical spraying has become a core development direction for improving agricultural production efficiency, reducing herbicide usage, and protecting the ecological environment. In automated field operation systems, machine vision technology plays a crucial role as the "environmental perception center." A prerequisite for achieving precise weeding is the ability to use a visual analysis system to extract and analyze the three-dimensional physical coordinates of the core "growth points" of weeds with high accuracy in a real natural farmland environment.

[0003] Currently, computer vision-based weed identification and localization technologies in farmland primarily rely on a combination of image processing and deep learning algorithms. The conventional technical approach typically includes: first, acquiring field images using visual sensors; then, using edge detection algorithms or color space conversion techniques for preliminary image segmentation to remove the soil background and extract initial visual fragments of the plants; subsequently, employing deep learning models such as deep convolutional neural networks (CNNs) to perform feature encoding and semantic recognition on the plant pixel matrix. Through these techniques, the vision system can initially achieve the classification and filtering of protected crops and target weeds in farmland scenarios.

[0004] In real natural farmland environments, plant growth exhibits extremely high complexity and diversity. On the one hand, crops and weeds often exhibit highly intertwined and mutually occluded physical forms under natural growth conditions, which inevitably results in the plant leaf images acquired by visual sensors appearing as spatially non-adjacent fragments on a two-dimensional projection. On the other hand, weeds have a wide variety of biological species, and their inherent growth forms (such as the exposed taproot form and the rosette self-occluding form) differ significantly. Furthermore, agricultural field operations require edge computing devices to complete real-time inference with extremely low latency under limited computing power. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for locating weed growth points based on visual analysis, which includes acquiring a visual image of a target area and segmenting the visual image into multiple independent closed regions.

[0008] The contents of each closed region are filtered to extract the initial plant feature fragments of each closed region;

[0009] By analyzing the initial plant feature fragments, correlation analysis is performed between closed regions, and optimization processing such as region addition or region removal is performed.

[0010] The closed region after optimization is used as an associated subgraph;

[0011] For each of the associated subgraphs, feature filtering is performed to filter out protected crops and extract the associated subgraphs of the remaining pure weeds;

[0012] A pre-trained classification network is used to classify species in each feature-filtered associated subgraph. Based on the species classification, the corresponding growth morphology is matched and the localization model is called. The localization model is then used to parse each feature-filtered associated subgraph and output the location of weed growth points.

[0013] As a preferred embodiment of the weed growth point localization method based on visual analysis described in this invention, the image content segmentation includes: calling the Canny edge detection algorithm to perform boundary analysis on the visual image, extracting the set of pixels in the visual image where the pixel grayscale undergoes a step change, and outputting the set of pixels as a two-dimensional image boundary contour.

[0014] Based on the boundary contour of the two-dimensional image, a morphological closing operation is performed to generate multiple completely closed regions in the two-dimensional image space; the closed regions are independent image blocks that are absolutely isolated by the boundary in geometric topology.

[0015] The visual image is converted to the HSV color space, and a color channel threshold range is set in each closed region to filter the semantic features of the closed regions. The set of closed regions obtained after filtering out the closed regions in the color channel threshold range is used as the initial plant feature fragment.

[0016] As a preferred embodiment of the visual analysis-based weed growth point localization method of the present invention, the association analysis includes inputting the initial plant feature fragments in each closed region into a pre-trained feature encoding network to generate corresponding deep content encoding vectors; the deep content encoding vectors are latent space multidimensional tensors that characterize the spatial location, geometric shape and physiological texture features of the initial plant feature fragments in the image.

[0017] Extract any two non-adjacent closed regions in the two-dimensional image space, perform feature concatenation operation on the depth content encoding vectors corresponding to the two non-adjacent closed regions, and generate a fused interaction vector;

[0018] The fused interaction vector is input into a pre-trained binary classification association determination model for analysis; the binary classification association determination model is a multilayer perceptron neural network classifier containing multiple fully connected network layers and terminal sigmoid activation functions.

[0019] Using the binary classification association determination model, we analyze any two non-adjacent deep content encoding vectors to determine whether the occluded part can be reproduced according to the growth pattern of plants. If it can be reproduced, the output is that there is an association; otherwise, the output is that there is no association.

[0020] As a preferred embodiment of the weed growth point localization method based on visual analysis described in this invention, the optimization process includes: if the result output by the binary classification association determination model is that there is an association, then the closed region that intersects with the occluded part in pixel position is taken as a supplementary region of the closed region with association.

[0021] For closed regions that are directly or indirectly related, create an independent virtual canvas;

[0022] When the closed region participates in the construction of the associated subgraph, the pixels corresponding to the original closed region at the bottom layer are not merged or modified. Instead, the pixel features of the associated closed region and all the corresponding supplementary regions are copied and extracted on the corresponding virtual canvas.

[0023] On the virtual canvas, pixel-level spatial merging is performed based on the replicated and extracted pixel features to complete the construction of one of the associated subgraphs;

[0024] Meanwhile, based on the data isolation characteristics of the replication and extraction operation, each closed region can be replicated in parallel and exist in multiple different associated subgraphs, thereby achieving region supplementation;

[0025] If the correlation analysis results of a closed region with all other closed regions in the image are not found, and it cannot be used as a supplementary region to participate in the construction of the correlation sub-image, then the current closed region is determined to be background noise that cannot be effectively classified. All pixels corresponding to the current closed region are deleted from the underlying storage to achieve the region removal.

[0026] As a preferred embodiment of the visual analysis-based weed growth point localization method of the present invention, the feature filtering includes: constructing the pre-trained classification network, wherein the pre-trained classification network is based on a deep convolutional neural network with spatial pixel mapping in its model architecture, and extracting spatial features of the local receptive field of the input image by sliding through multi-layer cascaded convolutional kernels;

[0027] The pre-trained classification network is internally configured with two independent outputs, each bound to two independently trained sets of network parameters, forming a multi-branch visual recognition network.

[0028] Each of the aforementioned association subgraphs is directly input into the pre-trained classification network;

[0029] The pre-trained classification network is invoked to include the first set of independent training parameters specifically trained for identifying farmland crops, and the first output terminal configured in binary output mode.

[0030] The classification network configured with the first set of independent training parameters is used to identify the associated subgraph; binary logic values ​​are output through the first output terminal: 0 indicates that the current associated subgraph is weeds, and 1 indicates that the current associated subgraph is not weeds;

[0031] When the binary logic value indicates that the current associated subgraph is not weeds, the current associated subgraph is removed to complete the filtering; and the associated subgraphs that are not removed after the filtering is completed are used as the associated subgraphs after feature filtering.

[0032] As a preferred embodiment of the visual analysis-based weed growth point localization method of the present invention, the species classification includes re-inputting the feature-filtered association subgraph into the classification network;

[0033] The pre-trained classification network calls the second set of independently trained network parameters obtained from general training for multiple weeds, and the second output terminal configured as a multi-classification species output mode;

[0034] The classification network with the second set of independently trained network parameters is used to re-identify the feature-filtered association subgraph; the corresponding weed biological species are directly output through the second output terminal as the result of the species classification.

[0035] As a preferred embodiment of the visual analysis-based weed growth point localization method of the present invention, the method of calling the localization model according to the corresponding growth morphology of the species classification includes matching the localization model for each species classification result through a preset growth feature matching table; wherein each localization model is a Bayesian network obtained by targeted training based on different types of root and stem growth features.

[0036] A morphological skeleton extraction algorithm is invoked to refine the associated sub-images after feature filtering, retaining only the topological skeleton reflecting the main stem and branch structure of the plant. The topological skeleton is then input into a matching Bayesian network, which determines whether each pixel position in the topological skeleton is a growth point. The pixel positions of the topological skeleton are converted into binary representations of 1 and 0, where 1 indicates a growth point and 0 indicates a non-growth point.

[0037] The distribution density of the growth points in space is statistically analyzed; the portion with a distribution density greater than a preset density threshold is taken as the final positioning result of the positioning model, and the regional coordinates of the final positioning result are output.

[0038] Secondly, the present invention provides a weed growth point location system based on visual analysis, including a data acquisition unit, which acquires a visual image of a target area and segments the visual image into multiple independent closed regions.

[0039] The filtering unit filters the contents of each closed region and extracts the initial plant feature fragments of each closed region;

[0040] The processing unit analyzes the initial plant feature fragments, performs correlation analysis between closed regions, and performs optimization processing such as region supplementation or region removal; the closed regions after optimization processing are used as correlation subgraphs.

[0041] The extraction unit performs feature filtering on each of the associated subgraphs, filters out protected crops, and extracts the associated subgraphs of the remaining pure weeds;

[0042] The localization unit uses a pre-trained classification network to classify species in each feature-filtered associated sub-graph. Based on the species classification, it matches the corresponding growth morphology and calls the localization model. The localization model is then used to parse each feature-filtered associated sub-graph and outputs the location of the weed growth point.

[0043] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the weed growth point location method based on visual analysis as described in the first aspect of the present invention.

[0044] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the weed growth point location method based on visual analysis as described in the first aspect of the present invention.

[0045] The beneficial effects of this invention are as follows: By performing edge detection, closed region segmentation, and color semantic filtering on the visual image of the target area, plant-related image fragments can be effectively extracted, reducing background interference from soil, shadows, and debris, and improving the accuracy of subsequent identification. Through feature encoding networks and binary classification association judgment models, association analysis is performed on non-adjacent closed regions. Even with leaf occlusion, plant overlap, or partial missing parts, related regions can be supplemented and reconstructed, improving the integrity of the weed image structure. A virtual canvas is used to construct association subgraphs, avoiding direct modification of original pixel data, allowing the same closed region to participate in multiple association subgraphs in parallel, improving the flexibility and robustness of data processing. Furthermore, by first removing protected crops through a multi-branch classification network and then classifying weeds, the risk of crop misidentification can be reduced. By combining different weed growth forms with corresponding localization models and using skeleton extraction and Bayesian networks to determine the location of growth points, the coordinates of weed growth points can be accurately output, providing a reliable basis for precision spraying, mechanical weeding, and intelligent agricultural robot operations. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of a method for locating weed growth points based on visual analysis.

[0048] Figure 2 A computer device diagram illustrating a method for locating weed growth points based on visual analysis. Detailed Implementation

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0050] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0051] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0052] Reference Figure 1 and Figure 2 This is one embodiment of the present invention, which provides a method for locating weed growth points based on visual analysis, including the following steps:

[0053] S1: Obtain a visual image of the target region, and segment the visual image into multiple independent closed regions.

[0054] In a specific embodiment of the present invention, a visual acquisition device mounted on agricultural equipment (such as a weeding robot, an unmanned field operation vehicle, or a tractor frame) is used for image acquisition. The visual acquisition device includes, but is not limited to, a high-resolution industrial RGB color camera, a binocular stereo camera, or an RGB-D depth camera. Preferably, the visual acquisition device is fixedly mounted vertically downwards or at a predetermined tilt angle above the chassis or frame of the agricultural equipment, ensuring that the field of view (FOV) of its lens completely covers the target area of ​​the farmland surface or between crop rows to be treated.

[0055] During the movement of the agricultural equipment, the onboard control system controls the visual acquisition device to automatically and continuously capture images of the target area based on a set time interval or trigger pulses output by the wheel displacement encoder. Furthermore, to overcome sudden changes in light intensity, backlighting, and plant shading in natural farmland environments, a uniform supplementary lighting source (such as an LED strobe array) linked to the camera is simultaneously triggered during image capture to obtain high-quality images with uniform illumination and no light spot interference.

[0056] Furthermore, the image content segmentation includes calling the Canny edge detection algorithm to perform boundary analysis on the visual image, extracting the set of pixels within the visual image whose grayscale values ​​undergo a step change, and outputting the pixel set as a two-dimensional image boundary contour. Based on the two-dimensional image boundary contour, a morphological closing operation is performed to generate multiple completely closed regions in the two-dimensional image space; the closed regions are independent image blocks that are geometrically and topologically isolated.

[0057] In a specific embodiment of the present invention, the specific implementation process of calling the Canny edge detection algorithm for boundary analysis is as follows: First, the acquired color visual image is converted into a grayscale matrix, and a Gaussian filter is applied to it for spatial smoothing to effectively filter out high-frequency noise points such as fine soil particles and light spots in the natural farmland environment; then, a differential operator is used to perform two-dimensional convolution calculation on the smoothed image along the horizontal and vertical directions to obtain the gradient magnitude and gradient direction of each pixel. The "step change in pixel grayscale" mentioned here refers to a sharp increase or decrease in the grayscale value (i.e., brightness or intensity) of adjacent pixels within a local spatial range of the image (for example, a pixel instantly transitions from a dark soil background to the bright green outer edge of a plant leaf). This phenomenon, which appears as a sudden change in the brightness of the target's physical contour in spatial vision, corresponds to the local first derivative maximum region of the image grayscale function in terms of underlying mathematical features. Finally, the algorithm performs non-maximum suppression operation on these pixels with step changes in grayscale along the calculated gradient direction, eliminating false or overly wide edge responses caused by light and shadow diffusion, refining and locking the true edge responses into precise lines at the single-pixel level, thereby accurately extracting the set of pixels representing the true physical boundary of the plant, and outputting the entire set of pixels as a continuous and clear two-dimensional image boundary contour.

[0058] In a specific embodiment of the present invention, the morphological closing operation performed based on the boundary contour of the two-dimensional image is specifically implemented by two cascaded basic mathematical morphological steps: "dilation" and "erosion". The system first presets a structural element with a specific local spatial geometry and size. The structural element is used to perform a dilation operation on the extracted boundary contour of the two-dimensional image, driving the boundary pixels to extend outward, thereby forcibly bridging and stitching the microscopic broken edges and adjacent isolated endpoints caused by farmland light and shadow interference or insufficient contrast in space. Subsequently, the same structural element is used to immediately perform an erosion operation on the dilated image, driving the image boundary to shrink inward again, so as to remove the redundant amplification of the boundary outer contour caused by the previous dilation operation, so that the macroscopic boundary of the plant is restored to the true physical size and position. This "expansion followed by erosion" cascade operation can perfectly heal the topological breakpoints on the boundary line without changing the overall area and macroscopic shape of the target. It forces the scattered edges with tiny cracks to connect into continuous closed loops, thereby generating multiple independent image blocks (i.e., the closed regions) that are absolutely isolated and completely closed in geometric topology in the two-dimensional image space. This provides a physical mask carrier with tight boundaries and no data leakage for subsequent pixel-level semantic feature filtering and virtual canvas separation.

[0059] S2: Filter the contents of each closed region and extract the initial plant feature fragments of each closed region.

[0060] The visual image is converted to the HSV color space (Hue, Saturation, Value) to eliminate interference from variations in natural light intensity and shadows on color recognition, as the HSV space is far less sensitive to brightness than the RGB space. Within each closed region, color channel threshold intervals are defined (predefined upper and lower limits for hue (H), saturation (S), and value (V) representing the green phenotype of a plant. For example, a specific angular interval is set on the H channel to lock in the "green" spectral band, and corresponding lower limits are set on the S and V channels to exclude non-plant interference pixels that are too dark or too pale. These three numerical intervals together constitute a three-dimensional color envelope, i.e., the color channel threshold intervals). Semantic feature filtering is then applied to these closed regions. The set of closed regions obtained after filtering out the closed regions within the color channel threshold intervals serves as the initial plant feature fragment. Specifically, the system iterates through all independent closed regions generated by the preceding morphological closing operation and extracts the HSV values ​​of all pixels within each region. If the HSV values ​​of most pixels (or pixels meeting a certain percentage threshold) within a closed region fall within a preset color channel threshold range, the system semantically classifies it as a "plant entity" (such as a leaf or stem). Conversely, if pixels are distributed outside the threshold range (e.g., exhibiting the brownish color of soil or the grayish-white color of stones), the system semantically classifies them as "non-plant background." Through this logical judgment, the system performs a filtering operation, removing all closed regions judged as non-plant backgrounds from the memory set. The final set of closed regions that conform to the semantic characteristics of plant color is extracted and defined as the aforementioned "initial plant feature fragment."

[0061] It is important to note that S2's filtering is not applied to individual pixels within a region, but rather to the entire closed region.

[0062] S3: By analyzing the initial plant feature fragments, correlation analysis is performed between closed regions, and optimization processing such as region addition or region removal is executed. The closed regions after optimization are used as correlation subgraphs.

[0063] Specifically, the association analysis includes inputting the initial plant feature fragments in each of the closed regions into a pre-trained feature encoding network to generate corresponding deep content encoding vectors; the deep content encoding vectors are latent space multidimensional tensors that characterize the spatial location, geometric shape, and physiological texture features of the initial plant feature fragments in the image.

[0064] In a specific embodiment of the present invention, the "feature encoding network" is essentially a deep learning model that serves as the system's underlying visual feature extractor (Backbone). Its network architecture is preferably a multi-layered stacked deep convolutional neural network (CNN, such as the ResNet series or MobileNet series) or a vision transformer.

[0065] The specific implementation mechanism is as follows: when the isolated initial plant feature fragment (i.e., the closed region image block that has completed semantic filtering) is input into the feature encoding network, the multi-level cascaded convolutional kernels inside the network will perform spatial sliding scanning of the local receptive field and layer-by-layer feature extraction. Among them, the shallow network structure is mainly responsible for capturing and responding to the basic geometric shape and physical contour (e.g., the serrated shape of the edge of the plant leaf, the linear orientation and curvature of the main stem); the deep network structure is responsible for extracting more abstract physiological texture features (e.g., the mesh-like distribution texture of leaf veins, the local color gradient distribution of the leaf surface) and the absolute or relative spatial position information of the fragment in the original field of view.

[0066] Through a series of forward propagation operations within the network (such as convolution, nonlinear activation, and pooling dimensionality reduction), the original RGB / HSV pixel data, which was originally highly redundant, high-dimensional, and difficult to directly perform logical operations, is extremely compressed and highly condensed. Finally, the network's final layer (such as a global average pooling layer) outputs a fixed-dimensional mathematical array (i.e., the deep content encoding vector). This network not only completely removes the remaining invalid and redundant information from feature fragments but also perfectly encapsulates the macroscopic morphology, microscopic texture, and spatial orientation attributes of the local plant fragment in the form of a purely numerical multidimensional tensor. This encoding conversion process standardizes the chaotic image fragments into latent space vectors that the model can directly compute and interact with, thus laying an indispensable mathematical foundation for subsequent steps such as cross-regional feature stitching operations and "occlusion reproduction and correlation inference" based on plant growth patterns.

[0067] Extract any two non-adjacent closed regions in the two-dimensional image space, and perform feature concatenation operation on the depth content encoding vectors corresponding to the two non-adjacent closed regions to generate a fused interaction vector.

[0068] The fused interaction vector is input into a pre-trained binary classification association determination model for analysis; the binary classification association determination model is a multilayer perceptron neural network classifier containing multiple fully connected network layers and terminal sigmoid activation functions.

[0069] Using the binary classification association determination model, we analyze any two non-adjacent deep content encoding vectors to determine whether the occluded part can be reproduced according to the growth pattern of plants. If it can be reproduced, the output is that there is an association; otherwise, the output is that there is no association.

[0070] It should be noted that, in a specific embodiment of the present invention, the multilayer perceptron neural network classifier (i.e., a binary classification association determination model) achieves "reproduction of the occluded portion according to the plant's growth pattern," and its specific implementation mechanism is as follows:

[0071] First, when the fused interaction vector containing features of two non-adjacent closed regions is input to the classifier, the front fully connected network layer of the multilayer perceptron essentially acts as a "feature decoder and spatial mapper." The network utilizes plant biological structural patterns (such as the linear extension trajectory of the main stem, the natural curvature of branches and leaves, and connection orientation) embedded in the weight parameters during pre-training to decode the high-dimensional fused interaction vector and roughly project or generate a two-dimensional probability mask matrix in the original two-dimensional image space. This mask region generated at the pixel level represents the "occluded portion" predicted by the model for physically connecting these two isolated segments.

[0072] It is important to note that this pixel-level occlusion reproduction does not require generating photorealistic images with high-precision physiological textures or lighting details. Instead, it aims to quickly generate a rough footprint that roughly represents the location and geometric extent of plant remnants (such as connecting stems covered by soil, straw, or crops) in two-dimensional space.

[0073] Subsequently, the classifier's post-network layer evaluates this roughly generated pixel-level occlusion. If the reproduced part can reasonably bridge the two non-adjacent initial segments in physical coordinates, and its span and direction conform to the spatial limits of plant growth, the sigmoid activation function at the end outputs a probability value close to 1, which is to say, it is determined that "there is a correlation".

[0074] The multilayer perceptron not only outputs logically related binary classification results on a macroscopic level, but also generates a "masking region" with clearly defined pixel coordinate boundaries in the underlying memory. This roughly reproduced pixel region will directly serve as the geometric target area when the system performs "region supplementation" operations. This allows the system to accurately filter out other fragmented closed regions scattered within the masking range through simple two-dimensional spatial intersection and collision operations, and capture and extract them as supplementary regions for related segments.

[0075] Furthermore, the optimization process includes, if the binary classification association determination model outputs a correlation, then the closed region that intersects with the occluded part in pixel position is taken as a supplementary region for the correlated closed region. Specifically, the system uses the occluded part (i.e., the predicted virtual connection mask) roughly generated by the aforementioned model in a two-dimensional plane as a spatial target area and maps it back to the original image coordinate system. Subsequently, the system traverses the remaining isolated closed regions in the image and uses a two-dimensional collision detection method to determine whether the pixel coordinates of these closed regions fall into or touch the boundary range of the predicted mask. Whenever spatial coordinate overlap (i.e., pixel position intersection) occurs, it is captured and defined as a "supplementary region".

[0076] Suppose that in actual farmland operations, the main stem of a target weed is severely obscured by a horizontally fallen, withered crop stalk. In the initial image segmentation and filtering stage, the system extracts two physically non-adjacent, clear plant segments: segment A (the crown leaves of the weed exposed above the stalk) and segment B (the root tissue of the weed located below the stalk). Simultaneously, within the gaps in the stalk, the system also segments an extremely small closed region C (a small patch of green pixels darkened by shadow). Due to the small size and isolation of region C, without contextual information, the system is highly likely to misclassify it as background noise. However, in this optimization step, the model analyzes the latent space features of segments A and B, determines a correlation between them based on plant growth patterns, and roughly generates a two-dimensional "predicted occlusion region" (i.e., a predicted main stem connectivity mask) between them. At this point, the system performs spatial coordinate comparison and finds that the pixel coordinates of the isolated closed region C fall exactly within or partially intersect with the coverage area of ​​this "predicted occlusion region." Therefore, the system is spatially convinced that region C is not soil or noise, but a real plant fragment that has survived in the gaps of the weed due to being obscured by straw. The system then formally recalls the closed region C and marks it as a "supplementary region," enabling it to participate in the complete physical reconstruction of the subsequent associated subgraphs along with fragments A and B.

[0077] For closed regions that are directly or indirectly related, an independent virtual canvas is created. Specifically, firstly, the system physically categorizes several closed regions that have been determined to be directly or indirectly related through preliminary correlation analysis, grouping these closed regions belonging to the same real plant into an independent "set of related closed regions." Subsequently, for each generated "set of related closed regions," the system dynamically allocates and distributes a dedicated blank two-dimensional matrix space in the memory of the edge computing device, perfectly aligned with the original image coordinate system. This space is defined as the aforementioned "independent virtual canvas." The correspondence here follows a strict mapping logic: that is, "one set of related closed regions absolutely and uniquely corresponds to one dedicated virtual canvas."

[0078] When the closed regions participate in the construction of the associated subgraph, the pixels corresponding to the original closed regions at the bottom layer are not merged or modified. Instead, on the corresponding virtual canvas, the pixel features of the associated closed regions and all the corresponding supplementary regions are replicated and extracted. On the virtual canvas, pixel-level spatial merging is performed based on the replicated and extracted pixel features to complete the construction of one associated subgraph.

[0079] In other words, the region supplementation uses the dedicated virtual canvas as the operating platform and the corresponding "associated closed region set" as the index. All the original closed regions contained in the set, as well as all the "supplementary regions (i.e. real physical fragments)" successfully captured based on the spatial occlusion prediction in the set, are uniformly used as the replication targets. The original pixel features (including their original spatial coordinates and color semantics) of these replication targets are jointly extracted and "copied and mapped" onto the virtual canvas corresponding to the set in a lossless manner.

[0080] Meanwhile, based on the data isolation characteristics of the replication and extraction operation, each closed region can be replicated in parallel and exist in multiple different associated subgraphs, thereby achieving region supplementation.

[0081] It's important to understand that in real-world field conditions, weeds and protective crops often have intricate interweaving and overlapping branches. This physical overlap in three-dimensional space, projected onto a two-dimensional visual image, easily causes the tissues of different plants (such as two intersecting leaves) to visually merge into a single, connected "closed region." Traditional image instance segmentation or feature clustering algorithms typically employ a "mutually exclusive" resource allocation mechanism. This means that once a closed region is assigned to plant A, plant B completely loses that overlapping feature, resulting in the forced "hollowing out" of plant B's local features and severe discontinuities in its geometric topology. This directly leads to branch breakage and failure to locate growth points during subsequent skeleton extraction.

[0082] The "parallel replication" mechanism designed in this invention breaks the exclusive resource contention of traditional algorithms. When the system performs spatial intersection determination, if it finds that a closed region (i.e., a physically overlapping region) falls within the "predicted occlusion mask" range of multiple different plants, the system utilizes the data isolation characteristics of the virtual canvas to treat this shared closed region as a legitimate supplementary region and replicates it losslessly, independently, and in parallel to the respective dedicated virtual canvases of these multiple plants. This non-destructive "one-to-many" mapping design ensures that even in severely intertwined complex scenes, each intertwined plant (regardless of how its underlying pixels overlap) can reconstruct the most complete, topologically intact, and data-free dedicated relational subgraph on its respective virtual canvas. This mechanism safeguards the absolute integrity of the target plant's macroscopic structure from the data source, laying an irreplaceable data foundation for subsequent high-precision morphological topological skeleton extraction.

[0083] If the correlation analysis results of a closed region with all other closed regions in the image are not found, and it cannot be used as a supplementary region to participate in the construction of the correlation sub-image, then the current closed region is determined to be background noise that cannot be effectively classified. All pixels corresponding to the current closed region are deleted from the underlying storage to achieve the region removal.

[0084] S4: Perform feature filtering on each of the associated subgraphs to filter out protected crops and extract the associated subgraphs of the remaining pure weeds.

[0085] The feature filtering includes constructing the pre-trained classification network, which is a deep convolutional neural network based on spatial pixel mapping in terms of model architecture. The network extracts spatial features of the local receptive field of the input image by sliding through multiple cascaded convolutional kernels.

[0086] Specifically, the convolutional kernel slides continuously pixel-by-pixel across the width and height of the two-dimensional image matrix with a preset stride. At each hovering position, the weight parameters inside the convolutional kernel perform element-wise multiplication and accumulation operations with the corresponding element-wise input pixel blocks covered by the current local receptive field. This local receptive field mechanism ensures that the network processes only a limited physical spatial neighborhood at any given time. While abstracting it into a feature map, it perfectly preserves the relative positional relationships and absolute topological connectivity of plant micro-features (such as leaf edge serrations and local lesions) on the original plane, and endows the model with "translation invariance" in dealing with the random growth coordinates of weeds in nature. Based on this, through a multi-layered cascaded network architecture design, the above-mentioned local sliding convolutional operation is sequentially chained along the depth direction of the network. As the network layers deepen, the equivalent local receptive field mapped by the post-convolutional kernel on the original input image expands exponentially. The shallow network is mainly responsible for capturing the basic color gradient and primary geometric boundaries at the micro level, while the deep network uses its expanded receptive field to perform spatial aggregation and non-linear splicing of these shallow discrete features in the macro dimension. This reconstructs the fragmented bottom lines into high-order abstract biological semantic features representing the structure of specific plant organs (such as the overall shape of cotyledons, the spatial arrangement of the main stem and branches). In this way, a complete feature decoding path is established from pixel-level local contours to plant-level macro morphology.

[0087] The pre-trained classification network is configured with two independent outputs, each bound to two independently trained sets of network parameters, forming a multi-branch visual recognition network.

[0088] In a specific embodiment of the present invention, the underlying training mechanism and implementation process of the classification network’s “two independent sets of network parameter sets bound to two independent output ends respectively” are as follows: For the first set of independent training parameter sets bound to the first output end configured in binary output mode, the core training objective is to achieve the absolute and accurate extraction and removal of farmland protection crops; In the offline training stage, the system constructs a large-scale binary classification sample set with the target economic crop as the positive category and various weeds and complex backgrounds as the negative category, and uses the binary cross-entropy loss function to iteratively optimize the network, forcing the model to focus its attention weights on the unique, highly recognizable macroscopic morphology and texture features of the crop, and solidifies them after gradient descent backpropagation to generate the first set of weight tensors for “black and white” coarse screening. For the second set of independent training parameters bound to the second output terminal configured as a multi-classification species output mode, the core training objective is to achieve high-precision, fine-grained species identification of the remaining pure weed images. During this training phase, crop data is completely excluded from the sample set, and a high-quality multi-classification sample set covering all known weed species in the target farmland scene is specifically constructed. A categorical cross-entropy loss function is used to constrain the network. Under this mechanism, the model is forced to learn and extract subtle biological differences between different weed species in dimensions such as local phyllotaxis, micro-edge serrations, and specific color gradients. Finally, a second set of weight tensors is solidified to output the distribution probability of specific weed species. Through this independent training mechanism based on different task orientations and different data distribution boundaries, the same deep convolutional neural network architecture can successfully generate two independent high-dimensional feature parameter spaces that do not interfere with each other, respectively, to handle "coarse screening" and "fine classification". During real-time inference in the field, only dynamic loading and switching of the parameter set are required to complete the re-entry identification of multiple branches of a single model.

[0089] Each of the associated subgraphs is directly input into the pre-trained classification network.

[0090] The pre-trained classification network is invoked, which includes a first set of independent training parameters specifically trained for identifying farmland crops, and a first output terminal configured in binary output mode.

[0091] The classification network configured with the first set of independent training parameters is used to identify the associated subgraph; binary logic values ​​are output through the first output terminal: 0 indicates that the current associated subgraph is weeds, and 1 indicates that the current associated subgraph is not weeds.

[0092] When the binary logic value indicates that the current associated subgraph is not a weed (i.e., it is determined to be the protected crop), the current associated subgraph is removed, completing the filtering. The protected crop is an economic plant entity in farmland that needs to be preserved from weeding. In the underlying algorithm cache, all pixel features contained in the target associated subgraph are synchronously reset to zero, and an absolute erasure operation is performed at the data level to implement the feature filtering, thereby completely filtering out the protected crop. The associated subgraph that has not been removed after filtering is then used as the associated subgraph after feature filtering.

[0093] In real-world automated weeding operations, protecting crops from accidental damage is a far higher priority than accurately identifying every single weed. Therefore, instead of using a conventional single-pass multi-classification network to simultaneously identify crops and weeds, the system employs a dedicated binary classification pre-inspection channel for "friend or foe" identification. This design grants the algorithm the highest level of "crop protection privilege"—meaning that if the system detects even the faintest crop features, it immediately resets all corresponding pixels to zero in the underlying physical memory. This completely eliminates any possibility of residual crop features leaking to the subsequent robotic arm or laser targeting module, physically isolating crops from the risk of misjudgment and accidental damage by weeding equipment at the algorithm's underlying level, thus ensuring the safety baseline of automated agricultural operations.

[0094] S5: Use a pre-trained classification network to classify species in each feature-filtered associated sub-graph, match the corresponding growth morphology according to the species classification, call the localization model, and use the called localization model to parse each feature-filtered associated sub-graph to output the location of weed growth points.

[0095] Specifically, the species classification includes re-inputting the feature-filtered association subgraph into the classification network. It also involves calling the second independently trained set of network parameters obtained from general training for multiple weeds in the pre-trained classification network, and the second output terminal configured for a multi-classification species output mode.

[0096] In a specific embodiment of the present invention, the species classification stage does not involve re-calling or cascading a completely new deep neural network. Instead, it is achieved by dynamically reloading the parameter set and physically switching the output of the pre-trained classification network. From the perspective of the distribution patterns of low-level features in computer vision, whether it is the "coarse screening for binary classification of protected crops" performed in the pre-stage or the "fine-grained multi-classification of various weeds" performed in this stage, the basic biological visual features (such as the geometric contours of leaf edges, color gradient distribution, and primary reticulated leaf vein texture) that need to be extracted in the first half of the network have extremely high homology and overlap. If two completely independent network models are deployed for these two tasks, it will not only lead to a huge amount of redundant computation in the low-level feature extraction operators (such as shallow and mid-level convolutional kernels), but also consume a lot of the limited on-board video memory (VRAM) and memory resources of agricultural equipment. Therefore, the present invention utilizes this high similarity of features and uses the backbone architecture of the classification network as a general feature extraction container. In actual operation, the system does not need to repeatedly destroy and instantiate massive network models in GPU memory. It only needs to quickly switch between different task-oriented network weight tensors at the memory level (i.e., call the second set of independently trained network parameter sets) and activate the corresponding multi-class terminal neurons (i.e., the second output). This lightweight re-entry mechanism of "changing parameters and output heads without changing the network backbone" not only maximizes the generalization ability of the intermediate layer feature extraction of the convolutional neural network, but also completely eliminates the model loading overhead and peak GPU memory pressure brought about by calling a new network. Thus, while ensuring high-precision recognition across tasks, it provides highly reliable high-concurrency and low-latency underlying algorithm execution efficiency for complex real-time automated weeding operations in farmland.

[0097] The classification network with the second set of independently trained network parameters is used to re-identify the feature-filtered association subgraph; the corresponding weed biological species are directly output through the second output terminal as the result of the species classification.

[0098] Furthermore, the method of calling the localization model according to the corresponding growth morphology of the species classification includes matching the localization model for each species classification result through a preset growth feature matching table; wherein each localization model is a Bayesian network obtained by targeted training based on different types of root and stem growth features.

[0099] In a specific embodiment of the present invention, the "growth feature matching table" is essentially a priori biological knowledge mapping dictionary embedded in the underlying system. Because the growth morphologies of weeds in nature vary greatly (e.g., broadleaf weeds like dandelions grow in a rosette-like basal form, grasses like foxtail grass grow upright in clumps, while bindweed grows as vines), the distribution patterns of their "growth points" (i.e., the root-stem junction or meristem) in geometric topology are drastically different. Using a single model can easily lead to offsets in the positioning coordinates. Therefore, this matching table accurately maps the specific weed species category output by the classification network to its corresponding "biological morphological category (e.g., rosette type, upright branching type, creeping type, etc.)"; the system then dynamically routes and calls the corresponding specific positioning model based on this morphological category. The parameter configuration within each positioning model is strictly tailored to the characteristics of that morphology (e.g., stem convergence angle characteristics, primary and secondary branch ratio characteristics).

[0100] Furthermore, in a real two-dimensional top-down view of farmland, the actual growth points of weeds are often severely obscured by the top leaves, representing "implicit features" that cannot be directly observed. Locating growth points is essentially a mathematical process of "inferring the spatial convergence probability based on the direction of surrounding visible branches under uncertain occlusion conditions." This is precisely the strength of Bayesian Networks. During the offline training phase of the model, the aforementioned single-pixel-width "topological skeleton" is used as the sole input feature for training. This forces the Bayesian Network to shield itself from all visual interference such as leaf patches and light reflections, focusing instead on learning the spatial connectivity of plant branch nodes, the geometric convergence trend of skeleton line segments converging towards the bottom of the trunk, and the conditional probability distribution for inferring the coordinates of implicit growth points based on these explicit topological features. In actual inference, when the Bayesian Network receives the real-time topological skeleton, it calculates the posterior probability that each pixel location is a "growth point" based on the convergence trend of each branch on the skeleton. Subsequently, the system converts this probability into a binary representation of 1 and 0 (1 represents a growth point, 0 represents a non-growth point) using a probability threshold (a pre-set value).

[0101] A morphological skeleton extraction algorithm is invoked to refine the associated sub-images after feature filtering, retaining only the topological skeleton reflecting the main stem and branch structure of the plant. The topological skeleton is then input into a matching Bayesian network, which determines whether each pixel position in the topological skeleton is a growth point. The pixel positions of the topological skeleton are converted into binary representations of 1 and 0, where 1 indicates a growth point and 0 indicates a non-growth point.

[0102] In a specific embodiment of the present invention, the refinement process of the morphological skeleton extraction algorithm is essentially a two-dimensional spatial iterative stripping process based on strict local connectivity constraints. Using the feature-filtered associated subgraph (i.e., a two-dimensional binary mask representing the macroscopic shape of the plant) as the input benchmark, a preset local pixel neighborhood window (such as a 3×3 pixel probe) is used to scan from the outer physical contour of the plant mask layer by layer towards its internal geometric center. In each round of edge traversal, the algorithm strictly evaluates the retention of each edge pixel according to graph theory and topological decision rules: if removing an edge pixel will not disrupt the local connectivity of the current plant tissue (i.e., will not cause the originally connected stems or leaf veins to break), and the pixel is not the extreme endpoint representing the natural terminal boundary of the plant, then it is marked as a shape redundant pixel and erased from physical memory. This layer-by-layer "erosion" operation from the outside in will be continuously iterated multiple times until no pixel in the mask meets the safe removal condition, at which point it automatically terminates. Through this extreme spatial compression, the robust stems and broad leaves, originally possessing actual physical area, are forced to converge inward into a set of centerlines with a single pixel width and absolute centering. This extremely rigorous image refinement process completely eliminates visual metric information such as the size of the plant entity in a two-dimensional plane projection and the thickness of organs, which can easily interfere with geometric calculations. Ultimately, it extracts without loss a pure geometric topological skeleton that absolutely reflects the extension trajectory of the plant's main stem, the spatial distribution of branch nodes, and the intersection logic of multi-level branches.

[0103] The distribution density of the growth points in space is statistically analyzed; the portion with a distribution density greater than a preset density threshold is taken as the final positioning result of the positioning model, and the regional coordinates of the final positioning result are output.

[0104] In complex pixel-level morphological extrapolation, due to microscopic defects or atypical intersections in localized plant areas, the pre-processing network inevitably outputs some scattered and isolated "false positive" prediction points. Directly using all predicted pixels as output would not only lead to a massive amount of ineffective and redundant movements in downstream weeding actuators (such as robotic arms, blades, or laser generators), but would also severely drag down overall operational efficiency. By introducing a spatial distribution density statistical mechanism, a high-dimensional "spatial clustering filter" is implemented; the preset density threshold automatically filters out and discards sparse and scattered interference noise points, ensuring that the final coordinate focus is absolutely converged to the plant's physical core (i.e., the true root crown center) with the strongest probability response and the densest branching intersections. This not only achieves extremely simplified dimensionality reduction of the coordinate data but also provides the downstream physical attack module with a highly fault-tolerant and clearly defined two-dimensional geometric target area.

[0105] This embodiment also provides a weed growth point location system based on visual analysis, including:

[0106] The acquisition unit acquires a visual image of the target area and segments the visual image into multiple independent closed regions.

[0107] The filtering unit filters the contents of each closed region and extracts the initial plant feature fragments of each closed region.

[0108] The processing unit analyzes the initial plant feature fragments, performs correlation analysis between closed regions, and performs optimization processing such as region supplementation or region elimination; the closed regions after optimization processing are used as correlation subgraphs.

[0109] The extraction unit performs feature filtering on each of the associated subgraphs, filtering out protected crops and extracting the associated subgraphs of the remaining pure weeds.

[0110] The localization unit uses a pre-trained classification network to classify species in each feature-filtered associated sub-graph. Based on the species classification, it matches the corresponding growth morphology and calls the localization model. The localization model is then used to parse each feature-filtered associated sub-graph and outputs the location of the weed growth point.

[0111] This embodiment also provides a computer device applicable to the weed growth point location method based on visual analysis, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the weed growth point location method based on visual analysis as proposed in the above embodiment.

[0112] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0113] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the weed growth point localization method based on visual analysis as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0114] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for locating weed growth points based on visual analysis, characterized in that: This includes acquiring a visual image of the target region and segmenting the visual image into multiple independent closed regions; The contents of each closed region are filtered to extract the initial plant feature fragments of each closed region; By analyzing the initial plant feature fragments, correlation analysis is performed between closed regions, and optimization processing such as region addition or region removal is performed. The closed region after optimization is used as an associated subgraph; For each of the associated subgraphs, feature filtering is performed to filter out protected crops and extract the associated subgraphs of the remaining pure weeds; A pre-trained classification network is used to classify species in each feature-filtered associated subgraph. Based on the species classification, the corresponding growth morphology is matched and the localization model is called. The localization model is then used to parse each feature-filtered associated subgraph and output the location of weed growth points.

2. The method for locating weed growth points based on visual analysis as described in claim 1, characterized in that: The image content segmentation includes calling the Canny edge detection algorithm to perform boundary analysis on the visual image, extracting the set of pixels in the visual image where the pixel grayscale changes abruptly, and outputting the set of pixels as a two-dimensional image boundary contour. Based on the boundary contour of the two-dimensional image, a morphological closing operation is performed to generate multiple completely closed regions in the two-dimensional image space. The closed region is an independent image block that is absolutely isolated by geometric topology. The visual image is converted to the HSV color space, and a color channel threshold range is set in each closed region to filter the closed regions based on semantic features. The set of closed regions obtained by filtering out the closed regions of the color channel threshold range is used as the initial plant feature fragment.

3. The method for locating weed growth points based on visual analysis as described in claim 2, characterized in that: The association analysis includes inputting the initial plant feature fragments in each of the closed regions into a pre-trained feature encoding network to generate corresponding deep content encoding vectors; the deep content encoding vectors are latent space multidimensional tensors that characterize the spatial location, geometric shape, and physiological texture features of the initial plant feature fragments in the image. Extract any two non-adjacent closed regions in the two-dimensional image space, perform feature concatenation operation on the depth content encoding vectors corresponding to the two non-adjacent closed regions, and generate a fused interaction vector; The fused interaction vector is input into a pre-trained binary classification association determination model for analysis; the binary classification association determination model is a multilayer perceptron neural network classifier containing multiple fully connected network layers and terminal sigmoid activation functions. Using the binary classification association determination model, we analyze any two non-adjacent deep content encoding vectors to determine whether the occluded part can be reproduced according to the growth pattern of plants. If it can be reproduced, the output is that there is an association; otherwise, the output is that there is no association.

4. The method for locating weed growth points based on visual analysis as described in claim 3, characterized in that: The optimization process includes, if the result output by the binary classification association determination model is that there is an association, then the closed region that intersects with the occluded part in pixel position is taken as a supplementary region of the closed region with association. For closed regions that are directly or indirectly related, create an independent virtual canvas; When the closed region participates in the construction of the associated subgraph, the pixels corresponding to the original closed region at the bottom layer are not merged or modified. Instead, the pixel features of the associated closed region and all the corresponding supplementary regions are copied and extracted on the corresponding virtual canvas. On the virtual canvas, pixel-level spatial merging is performed based on the replicated and extracted pixel features to complete the construction of one of the associated subgraphs; Meanwhile, based on the data isolation characteristics of the replication and extraction operation, each closed region can be replicated in parallel and exist in multiple different associated subgraphs, thereby achieving region supplementation; If the correlation analysis results of a closed region with all other closed regions in the image are not found, and it cannot be used as a supplementary region to participate in the construction of the correlation sub-image, then the current closed region is determined to be background noise that cannot be effectively classified. All pixels corresponding to the current closed region are deleted from the underlying storage to achieve the region removal.

5. The method for locating weed growth points based on visual analysis as described in claim 4, characterized in that: The feature filtering includes constructing the pre-trained classification network, which is based on a deep convolutional neural network with spatial pixel mapping in its model architecture. The network extracts spatial features of the local receptive field of the input image by sliding through multiple cascaded convolutional kernels. The pre-trained classification network is internally configured with two independent outputs, each bound to two independently trained sets of network parameters, forming a multi-branch visual recognition network. Each of the aforementioned association subgraphs is directly input into the pre-trained classification network; The pre-trained classification network is invoked to include the first set of independent training parameters specifically trained for identifying farmland crops, and the first output terminal configured in binary output mode. The classification network configured with the first set of independent training parameters is used to identify the associated subgraph; binary logic values ​​are output through the first output terminal: 0 indicates that the current associated subgraph is weeds, and 1 indicates that the current associated subgraph is not weeds; When the binary logic value indicates that the current associated subgraph is not weeds, the current associated subgraph is removed to complete the filtering; and the associated subgraphs that are not removed after the filtering is completed are used as the associated subgraphs after feature filtering.

6. The method for locating weed growth points based on visual analysis as described in claim 5, characterized in that: The species classification includes re-inputting the feature-filtered association subgraph into the classification network; The pre-trained classification network calls the second set of independently trained network parameters obtained from general training for multiple weeds, and the second output terminal configured as a multi-classification species output mode; The classification network with the second set of independently trained network parameters is used to re-identify the feature-filtered association subgraph; the corresponding weed biological species are directly output through the second output terminal as the result of the species classification.

7. The method for locating weed growth points based on visual analysis as described in claim 6, characterized in that: The method of calling the localization model according to the corresponding growth morphology of the species classification includes matching the localization model for each species classification result through a preset growth feature matching table; wherein, each localization model is a Bayesian network obtained by targeted training based on different types of root and stem growth characteristics. A morphological skeleton extraction algorithm is invoked to refine the associated sub-images after feature filtering, retaining only the topological skeleton reflecting the main stem and branch structure of the plant. The topological skeleton is then input into a matching Bayesian network, which determines whether each pixel position in the topological skeleton is a growth point. The pixel positions of the topological skeleton are converted into binary representations of 1 and 0, where 1 indicates a growth point and 0 indicates a non-growth point. The distribution density of the growth points in space is statistically analyzed; the portion with a distribution density greater than a preset density threshold is taken as the final positioning result of the positioning model, and the regional coordinates of the final positioning result are output.

8. A weed growth point localization system based on visual analysis, based on the weed growth point localization method based on visual analysis according to any one of claims 1 to 7, characterized in that: Includes an acquisition unit that acquires a visual image of the target area and segments the visual image into multiple independent closed regions. The filtering unit filters the contents of each closed region and extracts the initial plant feature fragments of each closed region; The processing unit analyzes the initial plant feature fragments, performs correlation analysis between closed regions, and performs optimization processing such as region addition or region removal. The closed region after optimization is used as an associated subgraph; The extraction unit performs feature filtering on each of the associated subgraphs, filters out protected crops, and extracts the associated subgraphs of the remaining pure weeds; The localization unit uses a pre-trained classification network to classify species in each feature-filtered associated sub-graph. Based on the species classification, it matches the corresponding growth morphology and calls the localization model. The localization model is then used to parse each feature-filtered associated sub-graph and outputs the location of the weed growth point.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the weed growth point location method based on visual analysis as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the visual analysis-based weed growth point location method according to any one of claims 1 to 7.

Images