A weed identification method and system based on image recognition

By constructing a multidimensional data field based on polarization images and depth data, and combining polarization difference and finite element analysis, the optical noise and occlusion problems of weed identification under fully biodegradable mulch film were solved, and high-precision weed identification was achieved.

CN123119232APending Publication Date: 2026-11-17YANCHENG AGRI MASCH TECH PROMOTION CENT
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202512037991.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-12-31
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing image recognition technologies face challenges in complex farmland scenarios covered by fully biodegradable mulch films, such as optical noise interference and feature degradation, making it difficult to accurately locate and classify targets under the film, especially when weeds are difficult to identify under strong shading conditions.

Method used

By acquiring the original polarization image and depth point cloud data, a spatially aligned multidimensional data field is constructed using the homography transformation matrix. The reflection component of the mulch film is eliminated by combining Stokes vector and polarization difference model. The stress distribution tensor characteristics are obtained by combining inverse finite element analysis. Differential discrimination is performed using partitioned mask data to identify monocotyledonous and dicotyledonous plants.

Benefits of technology

Under strong light and complex curved surface conditions, the leaf vein texture of plants under the film is reproduced with high fidelity, which improves the accuracy and efficiency of weed identification, reduces the false alarm rate, and improves the precision and efficiency of weeding operations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123119232A_ABST
    Figure CN123119232A_ABST
Patent Text Reader

Abstract

This invention relates to the field of image recognition technology, specifically to a weed identification method and system based on image recognition. The method includes acquiring original polarization images and depth point cloud data; extracting feature point sets from mulch film and matching them with a standard agronomic template; calculating the homography transformation matrix; constructing a spatially aligned multidimensional data field containing optical and depth layers based on the homography transformation matrix; generating partitioned mask data marking the location information of corn and soybean strips; performing parallel processing on the layers within the spatially aligned multidimensional data field to synthesize a composite feature matrix containing internal texture and external force; using the partitioned mask data as a logical gate to perform differential discrimination on the composite feature matrix: in the corn strip mask area, filtering is performed based on single-point concentrated vertical force features and monocot texture features; in the soybean strip mask area, filtering is performed based on double-point dispersed tension features and dicot texture features; and marking raised targets that fail the filtering as weed data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically to a weed identification method and system based on image recognition. Background Technology

[0002] In the field of machine vision perception in smart agriculture, automated detection and classification of field targets are fundamental to achieving precision farming. Existing image recognition technologies are typically based on visible light sensors, using convolutional neural networks to extract surface texture, color, and edge contour features of targets. This method is relatively mature in open-field cultivation scenarios, but in complex scenarios covered with semi-transparent media, such as fully biodegradable mulch, visual recognition systems face significant challenges.

[0003] First, plastic film mulching leads to severe optical noise interference and feature degradation. As a semi-transparent medium, fully biodegradable plastic film exhibits strong and irregular specular reflection under natural light, creating highly saturated areas in images and causing dynamic range overflow. Simultaneously, diffuse reflection signals from targets beneath the film are scattered and attenuated as they penetrate the medium, resulting in significant loss of high-frequency texture details. Traditional feature extraction algorithms based on the RGB color space struggle to effectively segment targets under mixed lighting conditions with low signal-to-noise ratios and high dynamic ranges.

[0004] Secondly, identifying targets under mulch film is a typical fine-grained classification problem under strong occlusion conditions. During the crop emergence stage, targets (soybeans, corn, and weeds) are all located beneath the medium. Visual sensors cannot directly acquire the leaf morphology of the plants; they can only observe the three-dimensional bulges formed by the deformation of the mulch film under stress. Soybeans and corn, as plants from different families and genera, along with the associated noxious weeds, exhibit highly similar geometric features in the bulges formed under the mulch film. Existing image-based morphological processing methods struggle to perceive depth information and cannot construct an inverse mapping relationship from surface deformation to the internal stress source, causing the classifier to fail in the absence of explicit texture features.

[0005] Therefore, for complex farmland scenarios covered by semi-transparent media, this paper proposes an image recognition-based weed identification method and system to overcome the strong polarization optical noise and physical obstruction on the surface of the medium and achieve accurate positioning and fine-grained classification of hidden targets under the film in the case of severe degradation of visual features. Summary of the Invention

[0006] The purpose of this invention is to provide a weed identification method and system based on image recognition. The method involves screening based on single-point concentrated vertical force features and monocot texture features in the corn belt masked area; and screening based on double-point dispersed tension features and dicot texture features in the soybean belt masked area; and marking the raised targets that fail the screening as weed data.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A weed identification method based on image recognition, comprising: The original polarization image and depth point cloud data are acquired, the feature point set of the mulch film is extracted and matched with the standard agronomic template, and the homography transformation matrix is ​​calculated. Based on the homography transformation matrix, a spatially aligned multidimensional data field containing optical and depth layers is constructed, and partitioned mask data with marked positional information of corn and soybean strips is generated. Parallel processing is performed on layers within a spatially aligned multidimensional data field. In the optical layer, Stokes vectors are calculated, and the high-polarization reflection component of the mulch film is removed using a polarization difference model to generate the perspective texture features under the film. In the depth layer, inverse finite element analysis is performed based on the elastic properties of the thin film to transform the deformation data into stress distribution tensor features. Thus, a composite feature matrix containing internal texture and external stress is synthesized. Using partitioned mask data as a logical gate, differential discrimination is performed on the composite feature matrix: in the corn strip mask area, screening is performed based on single-point concentrated vertical force features and monocot texture features; in the soybean strip mask area, screening is performed based on double-point dispersed tension features and dicot texture features; all raised targets that fail the screening are marked as weed data.

[0008] The process of solving the homography transformation matrix includes: The raw data of the field operation scene were collected using a multimodal camera. The raw polarization image included intensity maps of four polarization directions: 0°, 45°, 90° and 135°, and an image pixel coordinate system was established. The depth point cloud data included three-dimensional coordinate values ​​based on the camera coordinate system. In the raw polarization image, the crop emergence position on the surface of the mulch film was identified, and its centroid coordinates were extracted as the mulch film feature point set. The standard agronomic template is a pre-constructed virtual two-dimensional dot matrix, generated based on preset row spacing and plant spacing parameters, representing an ideal planting distribution without distortion, and establishing a standard agronomic coordinate system. The random sampling consensus algorithm is used to establish a correspondence between the feature point set of the mulch film and the standard agronomic template, eliminating mismatched points caused by mulch film damage and irregular holes, and solving the homography transformation matrix that maps the image pixel coordinate system to the standard agronomic coordinate system using the least squares method.

[0009] The process of acquiring the spatially aligned multidimensional data field and partitioned mask data includes: The original polarized image is transformed by perspective projection using a homography transformation matrix, and the pixel gray values ​​are reconstructed by bilinear interpolation to generate an optical layer under orthophoto projection. The depth point cloud data is mapped to the standard agronomic coordinate system through the same transformation matrix and rasterized to generate a depth layer, thereby constructing a pixel-level aligned spatially aligned multidimensional data field. The partition mask data is generated based on the geometric boundaries of the standard agronomic template. A binary logic diagram is drawn in the standard agronomic coordinate system. The pixels in the region corresponding to the corn planting row are marked as the first logic value, and the pixels in the region corresponding to the soybean planting row are marked as the second logic value, thus obtaining partition mask data that marks the position information of the corn strip and the soybean strip.

[0010] The process of obtaining the under-membrane perspective texture features includes: For each pixel in the optical layer, the Stokes vector and linear polarization degree, representing the polarization state of the light wave, are calculated based on the radiation intensity in the four polarization directions. The polarization difference model is constructed based on the physical principle that the reflected light from the surface of the fully biodegradable mulch film has high polarization characteristics, while the diffuse reflected light from the biological tissue under the film has depolarization characteristics. The specific calculation logic is as follows: the total light intensity value in the Stokes vector is multiplied by the linear polarization degree value, and then multiplied by the suppression coefficient dynamically adjusted based on the ambient light intensity to calculate the high polarization component value representing the specular reflection of the mulch film. Then, the high polarization component value is subtracted from the total light intensity value to obtain the perspective texture intensity value containing only the diffuse reflection information under the film, generating the perspective texture feature under the film that can penetrate the semi-transparent mulch film.

[0011] The process of obtaining the force distribution tensor features includes: The raised region in the depth layer is modeled as an isotropic elastic film surface, and the Young's modulus and Poisson's ratio parameters of the mulch film are set. The edge contour of the raised region is used as the relative displacement boundary condition, and the height value in the depth layer is used as the known displacement field input. The nodal force vectors that cause the deformation are solved by the inverse operation of the finite element method. For each mesh node, the direction and magnitude of the force are calculated, and a force distribution tensor describing the stress concentration state is constructed. The force distribution tensor features include three-dimensional feature components: principal stress amplitude, stress gradient divergence, and the angle between the principal stress direction and the normal of the mulch film, which are used to characterize the contact characteristics when an object under the film lifts the mulch film at the physical and mechanical level.

[0012] The steps for generating the composite feature matrix include: The underlying perspectral texture features and stress distribution tensor features are numerically normalized to map the texture grayscale values ​​and tensor feature values ​​to a unified dimensionless range. While maintaining consistent spatial resolution, the normalized texture feature channels and stress tensor feature channels are concatenated in the depth dimension to construct a multi-channel composite feature matrix. Each feature vector of the composite feature matrix simultaneously encodes the spectral texture attributes and top membrane mechanical properties of the object under the membrane at that location.

[0013] When the partition mask data indicates that the current region is a corn belt, the monocotyledon discrimination logic is activated. The single-point concentrated vertical force characteristic is defined as follows: the force distribution tensor shows the existence of a unique stress maximum center, and the principal stress direction is perpendicular to the surface of the mulch film, characterizing the mechanical behavior of the sharp top membrane of the coleoptile of monocotyledonous plants. The monocotyledon texture feature is defined as: in the membrane perspective texture feature, it shows significant parallel texture directionality, characterizing parallel leaf vein features.

[0014] When the partition mask data indicates that the current region is a soybean belt, the dicotyledonous discrimination logic is activated; The dual-point dispersed tension characteristic is defined as follows: the stress distribution tensor shows two separate stress concentration points, exhibiting a linear arched stress distribution, and there is a significant horizontal tension component, characterizing the mechanical behavior of dicotyledonous plants when they emerge from the soil. The dicotyledonous texture feature is defined as follows: in the under-membrane perspective texture feature, the texture gradient histogram exhibits an isotropic distribution feature without a clear dominant direction, characterizing the reticulate leaf vein feature.

[0015] A weed identification system based on image recognition, comprising: The raw data acquisition module acquires raw polarization images and depth point cloud data, extracts the feature point set of the mulch film and matches it with a standard agronomic template, and calculates the homography transformation matrix. The data transformation and recognition module, based on the homography transformation matrix, constructs a spatially aligned multidimensional data field containing optical and depth layers, and generates partitioned mask data that marks the location information of the corn and soybean strips. The multidimensional data processing module performs parallel processing on layers within a spatially aligned multidimensional data field. In the optical layer, it calculates the Stokes vector and uses a polarization difference model to remove the high-polarization reflection component of the mulch film, generating the perspective texture features under the film. In the depth layer, it performs inverse finite element analysis based on the elastic properties of the thin film, transforming the deformation data into a stress distribution tensor feature. This allows for the synthesis of a composite feature matrix that includes internal texture and external stress. The weed data judgment module uses partitioned mask data as a logical gate to perform differential discrimination on the composite feature matrix: in the corn strip mask area, it filters based on single-point concentrated vertical force features and monocot texture features; in the soybean strip mask area, it filters based on double-point dispersed tension features and dicot texture features; all raised targets that fail the screening are marked as weed data.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention utilizes Stokes vector calculation and polarization difference model, based on the physical characteristics of highly polarized light reflected from the mulch film and depolarized diffuse reflection from organisms beneath the film, to mathematically separate high-light noise from biological texture at the data level. Addressing the challenge of large anisotropic reflectance differences in mulch film elevation areas, a dynamic adjustment mechanism for the suppression coefficient based on the coupling of local curvature and incident angle is introduced. By calculating the cosine of the angle between the line of sight and the mulch film normal, the Fresnel approximation principle is used to automatically enhance the de-reflection intensity at the elevation edge and reduce the suppression coefficient in the vertical observation area. This enables the system to penetrate semi-transparent media and faithfully reproduce the leaf vein texture of plants beneath the film under strong light and complex curved surface conditions, providing pure optical feature input for subsequent classification.

[0017] 2. This invention overcomes the limitations of traditional visual technologies that rely solely on two-dimensional geometric shapes for identification, solving the problem of confusion caused by similar shapes of targets under the film. By introducing inverse finite element analysis, the visual morphology problem is transformed into a contact mechanics problem. The deformation data in the depth layer is used to inversely calculate the internal nodal force vectors that cause the deformation, constructing a force distribution tensor that includes principal stress amplitude, stress gradient divergence, and orientation angle. Addressing the pre-tensioning effect generated along the direction of travel during film laying, this invention employs an anisotropic stiffness matrix to correct the displacement-force mapping relationship, eliminating morphological misjudgments caused by the film's own tension, achieving a deep perception of the biological essence of targets under the film, and improving identification accuracy.

[0018] 3. This invention overcomes the problems of positioning deviation and classification logic conflict in unstructured farmland environments by constructing a spatially aligned multidimensional data field and partitioned masking logic. Addressing perspective distortion caused by machinery vibration and mulch film undulation, it utilizes the matching of mulch film feature point sets with standard agronomic templates to calculate the homography transformation matrix, forcibly correcting the chaotic original image into an orthophoto conforming to agronomic standards, thus achieving pixel-level spatial alignment of the data. Based on this, it employs partitioned masking data as a logic gating mechanism, implementing a divide-and-conquer differentiated discrimination strategy. This not only significantly reduces the false alarm rate across regions but also optimizes the allocation of computing resources through decoupled computation of a dual-branch network, significantly improving the accuracy and efficiency of weeding operations under complex intercropping patterns. Attached Figure Description

[0019] Figure 1This is a flowchart illustrating a weed identification method based on image recognition according to the present invention. Figure 2 This is a logical schematic diagram of an image recognition-based weed identification method according to the present invention; Figure 3 This is a schematic diagram of the structure of a weed identification system based on image recognition according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: This invention proposes a weed identification method based on image recognition, the process of which is as follows: Figure 1 As shown, the logic of the method is as follows: Figure 2 As shown, it includes: The original polarization image and depth point cloud data are acquired, the feature point set of the mulch film is extracted and matched with the standard agronomic template, and the homography transformation matrix is ​​calculated. Based on the homography transformation matrix, a spatially aligned multidimensional data field containing optical and depth layers is constructed, and partitioned mask data with marked positional information of corn and soybean strips is generated. Parallel processing is performed on layers within a spatially aligned multidimensional data field. In the optical layer, Stokes vectors are calculated, and the high-polarization reflection component of the mulch film is removed using a polarization difference model to generate the perspective texture features under the film. In the depth layer, inverse finite element analysis is performed based on the elastic properties of the thin film to transform the deformation data into stress distribution tensor features. Thus, a composite feature matrix containing internal texture and external stress is synthesized. Using partitioned mask data as a logical gate, differential discrimination is performed on the composite feature matrix: in the corn strip mask area, screening is performed based on single-point concentrated vertical force features and monocot texture features; in the soybean strip mask area, screening is performed based on double-point dispersed tension features and dicot texture features; all raised targets that fail the screening are marked as weed data.

[0022] The process of solving the homography transformation matrix includes: The raw data of the field operation scene were collected using a multimodal camera. The raw polarization image included intensity maps of four polarization directions: 0°, 45°, 90° and 135°, and an image pixel coordinate system was established. The depth point cloud data included three-dimensional coordinate values ​​based on the camera coordinate system. In the raw polarization image, the crop emergence position on the surface of the mulch film was identified, and its centroid coordinates were extracted as the mulch film feature point set. The standard agronomic template is a pre-constructed virtual two-dimensional dot matrix, generated based on preset row spacing and plant spacing parameters, representing an ideal planting distribution without distortion, and establishing a standard agronomic coordinate system. The random sampling consensus algorithm is used to establish a correspondence between the feature point set of the mulch film and the standard agronomic template, eliminating mismatched points caused by mulch film damage and irregular holes, and solving the homography transformation matrix that maps the image pixel coordinate system to the standard agronomic coordinate system using the least squares method.

[0023] Data source: Acquire raw polarization images (resolution such as 2448×2048) and depth point clouds; use polarization images as the main source of feature extraction because their edge sharpness is better than that of point clouds.

[0024] Image pixel coordinate system: The top left corner of the image is defined as the origin (0, 0), the horizontal axis to the right is the u-axis, and the vertical axis downwards is the v-axis.

[0025] Feature point extraction algorithm: Adaptive threshold segmentation is used to process the polarization intensity map. Since crop emergence points are green or dark, and there is a grayscale difference compared to the white / semi-transparent mulch, the connected components of the emergence points are extracted, and their centroids (u) are calculated. i v i ( ) serves as the observation point, i.e., the set of feature points on the plastic film. If the crop has not yet emerged, the pre-made black perforated dots on the plastic film can be identified as alternative features.

[0026] Furthermore, the process of extracting the feature point set of the mulch film can also employ a topological invariant geometric constraint matching strategy: When identifying features on the surface of the plastic film, not only are the centroids of individual connected domains extracted, but the Euclidean distance and direction vector between adjacent centroids are also calculated. A local topological graph of the observation points is constructed and its subgraph isomorphic matching is performed with the theoretical topological structure in the standard agronomic template. For soil obstruction or damaged holes on the surface of the plastic film, a virtual repair logic is introduced: if a distance break is detected between two collinear feature points and the break length is an integer multiple of the preset plant spacing, virtual feature points are automatically filled at the break and the filled point set is used for solving the homography matrix.

[0027] Before solving the homography matrix, the extraction of feature points of the plastic film is optimized; Graph construction: All suspected holes / seedlings extracted are used as nodes, and a graph structure is constructed based on neighborhood distance.

[0028] Topological verification: Check the linearity of nodes in the graph. For example, given that the agronomic standard is a plant spacing of 12cm, if three points A, B, and C are found to be approximately collinear, and the distance between AB is 12cm and the distance between BC is 24cm, then it is determined that there is a missing node between BC.

[0029] Virtual repair: A virtual feature point is generated at the midpoint of the BC line, and its weight is assigned lower than that of the actual measured point. It is then added to the point set to participate in the random sampling consensus algorithm.

[0030] Anomaly removal: If a node is outside the main topology network, such as being more than 30cm away from its nearest neighbor, it is judged as a mud spot or a random hole and is removed.

[0031] This invention improves the robustness of positioning under harsh working conditions. Field operations often involve mud and water splashing that obstructs the holes in the plastic film, or the film itself may have irregular tears. Traditional identification based on individual features is prone to missed or false detections, leading to failure in homography matrix calculation. This solution utilizes the periodicity and collinearity prior knowledge of agronomic planting to identify missing information and filter noise through topological relationships, ensuring accurate correction of the image coordinate system even when feature points are lost.

[0032] The standard agronomic coordinate system is a virtual, idealized physical coordinate system.

[0033] Define the top left corner of the field as the origin, the direction due east of the origin as the X-axis, and the direction due north of the origin as the Y-axis; Template construction: Based on the parameters of the "4+2" planting pattern; for example, 2 rows of corn with a row spacing of 40cm, 4 rows of soybeans with a row spacing of 30cm, and a strip spacing of 60cm, the corresponding theoretical coordinate points (X) are generated in the virtual coordinate system. i Y i ), which constitute the theoretical point set.

[0034] Matrix solution: The random sampling consensus algorithm is used to match the observed point set with the theoretical point set, eliminating outliers caused by mulch film damage. The homography transformation matrix H is solved using the least squares method, such that: ; This invention effectively solves the problem of precise positioning in unstructured farmland environments by introducing standard agronomic templates and homography transformation technology. In traditional agricultural machine vision, the vibration of machinery, undulations in the ground, and the installation angle of the camera can all cause severe perspective distortion in the image, rendering the preset spatial logic invalid. This solution does not rely on expensive centimeter-level positioning, but cleverly utilizes the inherent regularity of the mulch film as a reference to forcibly correct the distorted image acquired in real time into a standard orthophoto. This not only significantly reduces hardware costs but also greatly improves the robustness of the system under complex working conditions, ensuring the absolute accuracy of subsequent zone mask generation and laying the spatial geometric foundation for achieving strip-based fine weeding.

[0035] The process of acquiring the spatially aligned multidimensional data field and partitioned mask data includes: The original polarized image is transformed by perspective projection using a homography transformation matrix, and the pixel gray values ​​are reconstructed by bilinear interpolation to generate an optical layer under orthophoto projection. The depth point cloud data is mapped to the standard agronomic coordinate system through the same transformation matrix and rasterized to generate a depth layer, thereby constructing a pixel-level aligned spatially aligned multidimensional data field. The partition mask data is generated based on the geometric boundaries of the standard agronomic template. A binary logic diagram is drawn in the standard agronomic coordinate system. The pixels in the region corresponding to the corn planting row are marked as the first logic value, and the pixels in the region corresponding to the soybean planting row are marked as the second logic value, thus obtaining partition mask data that marks the position information of the corn strip and the soybean strip.

[0036] Optical layer reconstruction: The original polarization image is inversely mapped using the calculated homography transformation matrix. For each point coordinate in the standard coordinate system, its floating-point coordinates in the image pixel coordinate system are found, and the gray value of that point is calculated using bilinear interpolation. This process eliminates perspective distortion and generates an orthographic projection.

[0037] Depth layer reconstruction: The coordinates of the original point cloud are transformed to a standard coordinate system using a matrix transformation. Since the point cloud is discrete, holes may appear after the transformation. Therefore, rasterization is used to project the transformed point cloud onto a standard grid. The grid value is taken as the average or median value of the point cloud height within the region, generating a depth layer aligned with the pixels of the optical layer.

[0038] Multidimensional data field: The final spatially aligned multidimensional data field is a multi-channel matrix, where each pixel location simultaneously contains: [RGB intensity, polarization degree, height].

[0039] Partition mask generation logic: Since the data has been corrected to the standard agronomic coordinate system, the crop row positions in this coordinate system are fixed. A binary logical graph is generated directly in memory; for example, the corn planting area is marked as 1, the soybean planting area is marked as 2, and other areas are non-planting areas by default.

[0040] This invention uses geometric correction to strictly bind the physical properties of data to their spatial location. The partitioned mask acts as a highly efficient logical filter, eliminating the need for image recognition to determine whether a pixel belongs to the corn or soybean belt; the mask value is sufficient. This not only concentrates the core algorithm's computational power on identifying difficult targets under the mask, significantly improving processing speed, but also avoids cross-regional misjudgments caused by similar crop seedling morphologies, thus significantly enhancing logical reliability.

[0041] The process of obtaining the under-membrane perspective texture features includes: For each pixel in the optical layer, the Stokes vector and linear polarization degree, representing the polarization state of the light wave, are calculated based on the radiation intensity in the four polarization directions. The polarization difference model is constructed based on the physical principle that the reflected light from the surface of the fully biodegradable mulch film has high polarization characteristics, while the diffuse reflected light from the biological tissue under the film has depolarization characteristics. The specific calculation logic is as follows: the total light intensity value in the Stokes vector is multiplied by the linear polarization degree value, and then multiplied by the suppression coefficient dynamically adjusted based on the ambient light intensity to calculate the high polarization component value representing the specular reflection of the mulch film. Then, the high polarization component value is subtracted from the total light intensity value to obtain the perspective texture intensity value containing only the diffuse reflection information under the film, generating the perspective texture feature under the film that can penetrate the semi-transparent mulch film.

[0042] The surface of fully biodegradable mulch film is relatively smooth, producing strong specular reflection under sunlight. According to Fresnel's law of reflection, this reflected light is primarily linearly polarized, exhibiting high polarization characteristics. Conversely, when light penetrates the mulch film and reaches the interior of plant leaves, it undergoes diffuse reflection. The light experiences multiple scatterings within the plant tissue, leading to a degradation of its polarization state, resulting in depolarized light. Based on this physical difference, the reflective surface of the mulch film and the underlying biological texture can be distinguished by analyzing the polarization state of the light.

[0043] Stokes vector calculation: For each pixel in the image, based on the collected radiation intensity values ​​in the four polarization directions of 0°, 45°, 90°, and 135°, it is denoted as... , , , The Stokes vector and the degree of linear polarization characterizing the polarization state of the light wave are calculated.

[0044] The polarization parameters are calculated based on the definition of the Stokes vector. Among them, the first Stokes component... Characterized by total light intensity; second Stokes component Characterized by the difference between the horizontal 0° and vertical 90° polarization components; the third Stokes component. Characterizing the difference in polarization components between the 45° and 135° directions; where: ; ; ; Furthermore, based on the first Stokes component Second Stokes component Third Stokes component Calculate the degree of linear polarization The calculation formula is: ; To ensure the numerical stability of the system, a non-zero check is performed before calculating the division: if the current pixel's... If the value is less than the preset dark current threshold (e.g., 10^{-6}), the linear polarization degree of the pixel is directly forced to be 0 to prevent calculation errors caused by sensor dark noise.

[0045] Polarization difference dereflection is performed to calculate the grayscale reference value used for display; based on the polarization difference model, the high-polarization reflection component of the mulch film is calculated; then, the high-polarization reflection component is subtracted from the grayscale reference value to obtain the feature value of the under-film transparent texture. ; ; ; in, This is the grayscale reference value; This refers to the high-polarization reflection component of the mulch film. These are the feature values ​​of the perspectral texture under the membrane; This is a dynamic suppression coefficient based on ambient light and the angle of incidence. Through this step, the polarized light signal caused by reflection from the mulch film surface is mathematically extracted from the total light intensity, while retaining the unpolarized light signal diffusely reflected beneath the film.

[0046] Dynamic adjustment of the suppression coefficient: The suppression coefficient is a parameter used to adjust the reflectivity removal effect. In this embodiment, an ambient light sensor is connected to monitor the light intensity of the working environment in real time. When the ambient light is strong and the film reflection is severe, the system automatically increases the value of the suppression coefficient to enhance the reflectivity removal effect; when the ambient light is weak, the system automatically decreases the value of the suppression coefficient to retain more image details, thereby achieving dynamic adaptation to different lighting conditions.

[0047] Furthermore, the dynamic adjustment process of the suppression coefficient includes constructing local curvature-incident angle coupling compensation: The normal vector of each pixel on the surface of the plastic film is calculated using a depth layer. The solar azimuth and elevation angles are calculated by combining the time of field operations with GPS coordinates, thereby obtaining the local incident angle of light at each pixel. The inhibition coefficient is not only adjusted with the total ambient light intensity, but also dynamically changes with the local incident angle: when the local incident angle is close to the Brewster angle, the maximum inhibition coefficient weight is assigned; when the local incident angle is perpendicular to the surface of the mulch film, the minimum inhibition coefficient weight is assigned; thereby eliminating the anisotropic reflectivity differences caused by the mulch film bulge.

[0048] First, using depth point cloud data, the unit normal vector of each pixel on the surface of the mulch film is obtained through gradient calculation. Simultaneously, the unit line-of-sight vector of the camera is known. Angle calculation: The dot product of the line-of-sight and the normal is calculated to directly obtain the cosine value of the angle, which is between [0, 1]. When the cosine value of the angle is close to 1, it indicates a vertical downward view, with the weakest reflection; when the cosine value of the angle is close to 0, it indicates a grazing / level view, with the strongest reflection, such as a raised edge.

[0049] The suppression coefficient of the current pixel is calculated using a linear interpolation formula. The preset vertical observation suppression coefficient is taken as the minimum value, the preset grazing observation suppression coefficient is taken as the maximum value, and the cosine value of 1 minus the included angle is taken as the weight. The difference between the maximum and minimum values ​​is weighted. Then, the weighted sum of the weight and the difference is added to the minimum value to obtain the suppression coefficient of the current pixel. The coefficient is then substituted into the polarization difference formula for pixel-by-pixel calculation.

[0050] This design solves the problem of residual reflection on complex curved surfaces. When the plastic film is pushed up by crops, it forms bulges. The reflective characteristics of the sun-facing, shadow-facing, and apex surfaces of these bulges are drastically different. The sun-facing surface is close to specular reflection with extremely high polarization. Traditional global coefficients may remove reflections on flat ground, but not completely on bulging slopes. This solution achieves pixel-level adaptive anti-reflection through refined physical optics modeling, greatly improving the under-film visibility in bulging areas (i.e., key target areas).

[0051] This invention utilizes the principle of polarization optics to achieve physical defogging, solving the problem of unclear visual recognition under plastic film mulch. Traditional image enhancement algorithms typically only adjust brightness or contrast, failing to eliminate reflective noise caused by physical obstruction, and may even amplify the noise. This solution, however, starts from the wave nature of light, mathematically separating the reflective layer of the plastic film superimposed on the image from the underlying biological texture layer by analyzing the polarization state of light. This allows the system to penetrate the semi-transparent plastic film, clearly restoring the true texture of the plant leaves beneath, providing high-quality, reflective-free optical feature data for subsequently distinguishing monocots from dicots, significantly improving recognition accuracy under strong light conditions.

[0052] The process of obtaining the force distribution tensor features includes: The raised region in the depth layer is modeled as an isotropic elastic film surface, and the Young's modulus and Poisson's ratio parameters of the mulch film are set. The edge contour of the raised region is used as the relative displacement boundary condition, and the height value in the depth layer is used as the known displacement field input. The nodal force vectors that cause the deformation are solved by the inverse operation of the finite element method. For each mesh node, the direction and magnitude of the force are calculated, and a force distribution tensor describing the stress concentration state is constructed. The force distribution tensor features include three-dimensional feature components: principal stress amplitude, stress gradient divergence, and the angle between the principal stress direction and the normal of the mulch film, which are used to characterize the contact characteristics when an object under the film lifts the mulch film at the physical and mechanical level.

[0053] The reverse finite element analysis process includes: Modeling: The mulch film is regarded as an isotropic linear elastic film, and its Young's modulus (e.g., 200 MPa) and Poisson's ratio (e.g., 0.3) are set. For mulch films that exhibit obvious viscoelasticity and hyperelasticity, the predicted Young's modulus value is obtained by combining field temperature changes, UV aging degree, humidity and other factors.

[0054] Input: Local bulge height data in the depth layer; Boundary conditions: The height of the outer contour of the raised region is 0, and the displacement is 0; Solution logic: Utilize the large deflection equation of the thin film or the force balance equation of the mesh nodes; for each node on the mesh, the vertical resultant force it experiences is related to the relative displacement of the surrounding nodes, and calculate the internal node force distribution required for the thin film to maintain its current local bulge height and shape.

[0055] Force distribution tensor construction: For each node, extract the following physical quantities to construct tensor features: Principal stress amplitude: characterizes the magnitude of the force applied; Stress gradient divergence: characterizes the degree of stress concentration; Principal stress direction angle: The angle between the force vector and the normal to the mulch film.

[0056] Specifically, the process of obtaining the force distribution tensor features is as follows: Mesh discretization and modeling: Local raised areas in the depth layer are discretized into regular quadrilateral mesh cells, with each mesh node corresponding one-to-one with a pixel in the depth layer. For example, the mesh spacing is set to a millimeter-level size (e.g., 2 mm) consistent with the depth camera resolution. The physical model of the mulch film is set as an ideal elastic film, ignoring its bending stiffness and considering only its in-plane tensile stiffness.

[0057] Define boundary conditions and displacement field: including edge constraints and forced displacement input; Edge constraint: Identify the zero contour line profile of the raised area, mark all mesh nodes on the contour line as fixed boundaries, and set their three-dimensional displacement to zero.

[0058] Forced displacement input: For each node inside the raised area, its vertical displacement is forcibly set to the corresponding height value read from the depth layer; its horizontal displacement is temporarily set to zero in the initial state.

[0059] Constructing the stiffness equation and inverse solution: Based on the geometric nonlinearity assumption of small strain and large rotation, a global tangent stiffness matrix is ​​assembled. In conventional forward finite element analysis, the stiffness matrix multiplied by the displacement vector equals the force vector. This embodiment utilizes inverse thinking. Since the displacement vector, i.e., the deformation field of the mulch film, is known, and the stiffness matrix is ​​based on the mulch film material properties (Young's modulus, Poisson's ratio) and geometric dimensions, there is no need for complex matrix inversion or iterative solutions. Sparse matrix multiplication can be directly performed to calculate the nodal force vector required to maintain the current deformation.

[0060] Feature extraction: From the calculated nodal force vectors, the vertical component force at each node is extracted as the principal stress amplitude, and the divergence of the force vector in the neighborhood is extracted as the stress concentration feature, thus completing the construction of the force distribution tensor.

[0061] This invention introduces inverse finite element analysis, transforming a visual morphological problem into a mechanical and physical problem, thus solving the problem of difficulty in distinguishing similar-shaped targets under plastic film. In two-dimensional images, different plants exert drastically different mechanical forces on the plastic film. Monocotyledonous plants generate extremely high and concentrated stress like needle tips, while dicotyledonous plants generate dispersed tensile stress like arch bridges. By inverting these invisible forces, strong robust features independent of color and texture are obtained; this greatly reduces misjudgments caused by the variability of crop growth forms, enabling the identification of the essential attributes of hidden targets under plastic film.

[0062] Furthermore, based on the modeling of the aforementioned raised region as an isotropic elastic thin film surface, the inverse finite element analysis introduces an anisotropic stiffness matrix: Considering that the mulch film is subjected to mechanical stretching along the longitudinal direction during the laying process, while being relatively relaxed in the transverse direction, the mulch film may have obvious orthotropic mechanical properties; therefore, in the inverse finite element analysis, the longitudinal Young's modulus is set to be greater than the transverse Young's modulus; when solving the nodal force vector, the anisotropic stiffness matrix is ​​used to correct the mapping relationship between displacement and force, so as to distinguish the deformation caused by the tension of the mulch film itself and the deformation caused by the lifting of the object under the film.

[0063] Physical modeling background: When the plastic film laying machine is working in the field, the plastic film is subjected to strong traction and stretching by the mechanical rollers along the direction of travel of the machine (i.e., longitudinal direction), and is in a pre-tensioned state; while it is relatively loose in the direction perpendicular to the direction of travel (i.e., transverse direction). This laying process results in inconsistent mechanical stiffness of the plastic film in the two directions.

[0064] Stiffness matrix construction: The constitutive equation of the mulch film is defined, which describes the relationship between stress and strain. In this equation, the longitudinal Young's modulus is set to be greater than the transverse Young's modulus, for example, 1.5 to 2 times the transverse Young's modulus. The specific ratio is predetermined based on the tensile ratio parameters during mulch film installation. Simultaneously, based on the mechanical properties of orthotropic materials, a reciprocal relationship is established between Poisson's ratio and Young's modulus, meaning the ratio of transverse Poisson's ratio to transverse Young's modulus is equal to the ratio of longitudinal Poisson's ratio to longitudinal Young's modulus. Based on this, the anisotropic stiffness matrix is ​​constructed.

[0065] Inverse solution correction: After finite element mesh generation, the global stiffness equation is assembled using the anisotropic stiffness matrix constructed above. The bulge height data acquired by the depth camera is used as the known displacement field input equation, and the mesh nodal force vectors causing the deformation are solved through inverse calculation.

[0066] Tensor correction: The principal stress directions in the calculated stress distribution tensor are transformed based on the stretching direction of the mulch film to remove the background stress components caused by the pretension of the mulch film itself, and only the disturbance stress caused by the lifting of objects under the film is retained.

[0067] This invention solves the problem of morphological misjudgment caused by the pre-tension of mulch film. In actual fields, because the mulch film is stretched very tightly along the direction of travel, when the same object lifts the mulch film, the deformation along the stretching direction is relatively small, while the deformation in the vertical direction is relatively large. This can cause a round monocotyledonous corn tip to appear as a laterally elongated ellipse under the film, which is easily misjudged as the arch of a dicotyledonous soybean. This solution introduces an anisotropic stiffness matrix, which is equivalent to pre-correcting the stretching within the algorithm. This mathematically offsets the mechanical deviations caused by the mulch film laying process, restores the true stress shape of the object under the film, and significantly improves the classification accuracy in real mechanized operation scenarios, especially for seedling identification where morphological features are easily affected by stretching.

[0068] The steps for generating the composite feature matrix include: The underlying perspectral texture features and stress distribution tensor features are numerically normalized to map the texture grayscale values ​​and tensor feature values ​​to a unified dimensionless range. While maintaining consistent spatial resolution, the normalized texture feature channels and stress tensor feature channels are concatenated in the depth dimension to construct a multi-channel composite feature matrix. Each feature vector of the composite feature matrix simultaneously encodes the spectral texture attributes and top membrane mechanical properties of the object under the membrane at that location.

[0069] Specifically, the steps for generating the composite feature matrix include: Data normalization preprocessing: To eliminate the influence of different physical dimensions on subsequent discrimination logic, range normalization is used to process each feature component. For the under-membrane perspective texture features: the maximum and minimum texture intensity values ​​in the current processing area are statistically analyzed, and the grayscale value of each pixel is mapped to the dimensionless interval [0, 1]. For the stress distribution tensor features: its three feature components are normally normalized independently. Among them, the original unit of the principal stress amplitude is megapascals, the original unit of the stress gradient divergence is megapascals per millimeter, and the original unit of the angle between the principal stress direction and the normal is radians; the values ​​of the above three components are linearly mapped to the interval [0, 1].

[0070] Constructing a multi-channel feature tensor: While maintaining consistent image spatial resolution, the normalized feature components are stacked and stitched along the channel dimension to construct a four-channel composite feature matrix.

[0071] The four channels of the composite feature matrix are physically defined as follows: Channel 1: Normalized under-film perspective texture intensity, characterizing the spectral texture properties of the object; Channel 2: Normalized principal stress amplitude, characterizing the magnitude of the force on the top membrane of the object; Channel 3: Normalized stress gradient divergence, characterizing the degree of concentration or dispersion of stress; Channel 4: Normalized principal stress direction angle, characterizing the degree of deviation of the force direction from the normal of the mulch film.

[0072] This matrix simultaneously encodes texture information and mechanical contact information at each pixel location, providing a standardized multimodal input for subsequent logic gating and discrimination.

[0073] This invention achieves data fusion of biological and physical attributes by constructing a multi-channel composite feature matrix. Single feature sources often have limitations: for example, the texture is unclear under low light or the mechanical features are blurred under soil interference. The composite feature matrix aligns and encodes spectral texture attributes and top membrane mechanical attributes at the pixel level, providing high-dimensional information input for the classifier. This complementarity of multimodal data greatly improves the system's fault tolerance. Even if a certain type of feature fails under specific conditions, such as when the local film is too thick and the texture is not visible, the mechanical bulge still exists, and the correct judgment can still be made based on the other dimension of features, thus ensuring the stability of all-weather operation.

[0074] When the partition mask data indicates that the current region is a corn belt, the monocotyledon discrimination logic is activated. The single-point concentrated vertical force characteristic is defined as follows: the force distribution tensor shows the existence of a unique stress maximum center, and the principal stress direction is perpendicular to the surface of the mulch film, characterizing the mechanical behavior of the sharp top membrane of the coleoptile of monocotyledonous plants. The monocotyledon texture feature is defined as: in the membrane perspective texture feature, it shows significant parallel texture directionality, characterizing parallel leaf vein features.

[0075] When the partition mask data indicates that the current region is a soybean belt, the dicotyledonous discrimination logic is activated; The dual-point dispersed tension characteristic is defined as follows: the stress distribution tensor shows two separate stress concentration points, exhibiting a linear arched stress distribution, and there is a significant horizontal tension component, characterizing the mechanical behavior of dicotyledonous plants when they emerge from the soil. The dicotyledonous texture feature is defined as follows: in the under-membrane perspective texture feature, the texture gradient histogram exhibits an isotropic distribution feature without a clear dominant direction, characterizing the reticulate leaf vein feature.

[0076] A dual-branch coupled network is used to execute the discrimination logic; Network architecture and logic gating: The dual-branch coupled network includes an input adaptation layer, a shared physical encoder, a mask gating module, and two dedicated discriminant branches, including a single-sub-branch and a double-sub-branch.

[0077] Input adaptation layer: The input tensor dimension is 512*512*4, where the first channel is a normalized under-membrane perspective texture feature, and the last three channels are force distribution tensor features, including principal stress amplitude, divergence, and orientation angle. The first layer does not use standard RGB convolution, but instead employs two sets of parallel convolution kernels; the texture convolution group has a size of 7*7, a stride of 1, 1 input channel (texture), and 16 output channels. The initial weights follow a Gabor distribution. The mechanics convolution group has a size of 3*3, a stride of 1, 3 input channels (mechanics), and 16 output channels. The initial weights follow a Laplacian distribution (to enhance edge gradients). The two sets of outputs are concatenated to obtain an initial feature map of 512*512*32.

[0078] Shared physical encoder: The first three stages (Stage 1-3) of a lightweight ResNet-18 after removing the fully connected layers are used as the backbone; Stage 1: Contains 2 residual blocks, with an output dimension of 128*128*64; Stage 2: Contains 2 residual blocks, downsampled, output dimension 64*64*128; Stage 3: Contains 2 residual blocks, downsampled, output dimension 32*32*256; Output a shared physical feature map.

[0079] Mask gating module: Utilizes the generated partitioned mask data as a logic gating signal. When the mask data shows that the current area is a corn belt, the network automatically activates the monocot branch and blocks the dicot branch; when the mask data shows that the current area is a soybean belt, it automatically activates the dicot branch and blocks the monocot branch.

[0080] Discrimination of masked regions in maize: Monocot logic, in the monocot branches, the network focuses on screening targets that conform to the biological characteristics of maize; Single-point concentrated vertical force feature screening: The mechanical feature extraction layer of this branch uses convolution kernel design to specifically detect whether there is a unique stress maximum center in the force distribution tensor, and whether the principal stress direction is perpendicular to the surface of the mulch film, corresponding to the mechanical behavior of the coleoptile of monocotyledonous plants (corn) lifting the mulch film like a needle.

[0081] Monocotyledonous texture feature screening: The texture feature extraction layer of this branch uses a high aspect ratio convolution kernel to detect whether the perspectral texture features under the membrane show significant parallel texture directionality, corresponding to the biological characteristics of parallel leaf veins in maize.

[0082] Judgment result: Only raised targets that simultaneously meet the above-mentioned characteristics of sharp vertical force and parallel texture are judged as corn seedlings; otherwise, they are marked as weeds.

[0083] Discrimination of masked regions in soybeans: Dicotyledonous logic, in which the network focuses on screening targets that conform to the biological characteristics of soybeans in the dicotyledonous branches; Dual-point dispersed tension feature screening: The mechanical feature extraction layer of this branch is designed with cavitary convolution or multi-scale receptive fields to specifically detect whether there are two separate stress concentration points in the stress distribution tensor, corresponding to two cotyledons; or a linear arched stress distribution, corresponding to hypocotyl arching; and detect whether there are significant horizontal tension components, corresponding to the mechanical behavior of cotyledon emergence or hypocotyl arching in dicotyledonous plants (soybeans).

[0084] Dicotyledonous texture feature screening: The texture feature extraction layer of this branch is used to statistically analyze the distribution of texture gradients, focusing on whether the texture gradient histogram shows an isotropic distribution without a clear dominant direction, corresponding to the biological characteristics of soybean reticulate leaf veins.

[0085] Judgment result: Only raised features that simultaneously meet the above-mentioned characteristics of dispersed / arched force and reticulated isotropic texture are judged as soybean seedlings; otherwise, they are marked as weeds.

[0086] All data marked as weeds are aggregated, and the logical coordinates of the weeds in the standard agronomic coordinate system are restored to physical pixel coordinates through the inverse matrix of the homography transformation matrix. Finally, the depth information is combined to generate three-dimensional operation instructions to drive the weeding actuator to perform targeted removal.

[0087] Furthermore, the quantitative discrimination logic for the monocotyledonous texture features and dicotyledonous texture features is as follows: Step S1: Calculate the texture gradient field; Obtain the under-membrane perspectral texture feature map, and calculate the gradient information of each pixel using the Sobel operator. Convolve the texture map with horizontal and vertical convolution kernels respectively to obtain the horizontal and vertical gradient components. Then, calculate the gradient magnitude and gradient direction angle of each pixel using the following formula: ; ; in, For horizontal gradient components; For vertical gradient components; The gradient magnitude of the pixel; The gradient direction angle; It is the arctangent function.

[0088] Step S2: Construct the directional gradient histogram; Considering the symmetry of leaf vein texture, the gradient direction angle is... The value range is mapped to the [0, 180) degree interval; this interval is evenly divided into N directions to build a histogram (for example, N=18, that is, each interval is 10 degrees). Traverse all pixels in the texture map whose gradient magnitude is greater than the preset noise threshold, and accumulate their gradient magnitudes into the corresponding direction interval to obtain a directional gradient histogram vector containing 18 values.

[0089] Step S3: Monocot texture feature discrimination (parallelism detection); Calculate the peak ratio in the histogram. Find the maximum value in the histogram vector and its corresponding directional interval, and find the second largest value; calculate the main peak significance ratio based on the maximum value divided by the second largest value, where the second largest value is the second largest value. Judgment criterion: If the ratio is greater than a preset parallelism threshold (for example, set to 2.5 in this embodiment), then the region is determined to show significant parallel texture directionality, consistent with the parallel vein characteristics of monocotyledonous plants (such as corn).

[0090] Step S4: Dicotyledonous texture feature discrimination (isotropic detection); Calculate the directional distribution entropy of the histogram. First, perform probability normalization on the histogram vector to obtain the probability distribution; then, use the Shannon entropy formula to calculate the information entropy of the texture based on the probability distribution. Criterion: Theoretically, the entropy value is maximum when the texture direction is completely random (for 18 intervals, the maximum entropy is approximately 4.17). If the calculated entropy value is greater than a preset isotropic threshold, it indicates that the texture direction is uniformly distributed with no obvious dominant direction, consistent with the reticulate leaf vein characteristics of dicotyledonous plants (such as soybeans).

[0091] Furthermore, the quantitative definition and calculation process of the two-point dispersion tension characteristics are as follows: Principal stress field decomposition: Eigenvalue decomposition is performed on the stress distribution tensor; for each grid node, the eigenvalues ​​and eigenvectors of its stress tensor matrix are solved to obtain the magnitude of the first principal stress (maximum tensile stress) and its corresponding direction vector.

[0092] Bimodal extremum detection: A two-dimensional nonmaximum suppression algorithm is executed on the principal stress amplitude layer; a neighborhood window (e.g., 3*3 or 5*5 grid) is set to search for local maxima points and filter out the set of all candidate extrema points whose stress amplitude is greater than the background noise threshold.

[0093] Two-point spatial distance discrimination: Calculate the Euclidean distance between any two points in the candidate set; search for whether there exists at least one pair of extreme points whose physical distance satisfies the biological statistical range of cotyledon unfolding in dicotyledonous plants. In this embodiment, the distance threshold is set to [15mm, 40mm]; if there is no pair of points that satisfy this distance constraint, it is determined that the two-point feature is not present.

[0094] Verification of arched stress distribution: For point pairs that meet the distance condition, extract the stress distribution curve along the path connecting the two points; calculate the ratio of the stress value at the midpoint of the path to the peak stress value, and detect the curve shape. If there is a continuous non-zero stress distribution between the two points, and the stress value at the midpoint of the line is lower than the peak values ​​at both ends, forming an M-shaped or arch-shaped stress profile, i.e., a stress saddle point exists, then the existence of an arched stress distribution is confirmed.

[0095] Horizontal tension component identification: For extreme point regions, decompose the principal stress direction vector; calculate its component in the horizontal plane and its component perpendicular to the mulch film. Calculate the anisotropy ratio by dividing the component in the horizontal plane by the component perpendicular to the mulch film; when the anisotropy ratio is greater than a preset tension threshold, it indicates that the dominant force in this region includes significant lateral tensile force, which is significantly different from the force characteristics of monocotyledonous plants, which are mainly vertically piercing, thus confirming it as the mechanical behavior of dicotyledonous plants emerging from the soil.

[0096] This invention explicitly hard-codes disclosed physical / biological prior knowledge, such as parallel texture and single-point force, into the network structure, greatly improving the network's interpretability. Through a mask-gated routing mechanism, a filter is applied to the network; when identifying corn, the network automatically disables its sensitivity to mesh textures, solving the problem of broadleaf weeds in soybean strips misleading corn identification. During inference, for any pixel in the image, only one branch is actually in a highly activated state, resulting in more rational allocation of computational resources and effectively improving the accuracy of weed identification.

[0097] Example 2: This invention also proposes an image recognition-based weed identification system, the structure of which is as follows: Figure 3 As shown, it includes: The raw data acquisition module acquires raw polarization images and depth point cloud data, extracts the feature point set of the mulch film and matches it with a standard agronomic template, and calculates the homography transformation matrix. This module is equipped with a multimodal industrial camera at the physical level. During actual operation, the camera lens is aimed at the field operation area and two sets of data are collected in one exposure: the first set is the original polarization image with a resolution of 2448×2048 pixels. This image is not a single picture, but contains four grayscale intensity maps after grating filtering in four different polarization directions: 0°, 45°, 90° and 135°; the second set is the depth point cloud data corresponding to the pixel coordinates, which records the three-dimensional coordinate value of each sampling point in the field of view relative to the optical center of the camera.

[0098] After the data is loaded into memory, a connected component analysis algorithm is first run to scan the high-brightness areas in the original polarized image, identify the emergence points or pre-drilled holes of dark crops against a white plastic film background, and calculate the geometric centroids of these connected regions to form a set of plastic film feature points. Subsequently, the system retrieves a pre-stored standard agronomic template, which is a virtual two-dimensional numerical matrix storing ideal coordinate points generated according to preset agronomic parameters, such as a regularly arranged matrix with a row spacing of 40 cm and a plant spacing of 12 cm. The module uses a random sampling consensus algorithm to iteratively match the actual extracted feature point set with the ideal template matrix, calculating a homography transformation matrix that maps the horizontal and vertical coordinates of the image pixel coordinate system to the standard agronomic coordinate system. This matrix is ​​a three-row, three-column numerical array.

[0099] The data transformation and recognition module, based on the homography transformation matrix, constructs a spatially aligned multidimensional data field containing optical and depth layers, and generates partitioned mask data that marks the location information of the corn and soybean strips. This module utilizes the homography transformation matrix calculated above to rearrange the original data at the pixel level. For the optical layer, it iterates through each integer coordinate point in the standard coordinate system, finds its floating-point position in the original image through matrix inverse operation, and reads the grayscale values ​​of the four neighboring pixels around that position. A weighted average is then performed according to distance weights to complete bilinear interpolation and generate a distortion-free orthophoto map. For the depth layer, the module multiplies the point cloud coordinates by the transformation matrix, projects them onto a grid in the standard coordinate system, and takes the median of multiple height values ​​falling within the same grid to generate a depth raster map that is strictly aligned with the optical layer, thus constructing a spatially aligned multidimensional data field.

[0100] Simultaneously, the module generates partition mask data based on the predefined geometric boundaries in the standard agronomic template. A two-dimensional array with the same size as the image is allocated in memory. According to the planting plan, the array elements corresponding to the corn planting areas in the first and second rows are assigned the logical number 1, and the array elements corresponding to the soybean planting areas in the third to sixth rows are assigned the logical number 2. This array is the logical gating mask for subsequent processing.

[0101] The multidimensional data processing module performs parallel processing on layers within a spatially aligned multidimensional data field. In the optical layer, it calculates the Stokes vector and uses a polarization difference model to remove the high-polarization reflection component of the mulch film, generating the perspective texture features under the film. In the depth layer, it performs inverse finite element analysis based on the elastic properties of the thin film, transforming the deformation data into a stress distribution tensor feature. This allows for the synthesis of a composite feature matrix that includes internal texture and external stress. This module employs a parallel computing architecture to process optical and mechanical data simultaneously. In the optical processing channel, the module reads the grayscale values ​​of the four polarization directions pixel by pixel. First, it adds the grayscale values ​​of the zero-degree and ninety-degree directions and averages them to obtain the total light intensity value. Then, it calculates the square root of the sum of the squares of the grayscale differences in each direction and divides it by the total light intensity to obtain the linear polarization degree value. Next, it performs a polarization difference operation: subtracting the total light intensity value from the linear polarization degree value and then from the dynamic suppression coefficient. The dynamic suppression coefficient is calculated based on the normal vector provided by the depth layer. When the cosine of the angle between the viewpoint and the normal to the film surface is close to zero (i.e., at the edge), the coefficient automatically increases to strongly remove highlights; otherwise, it decreases. The final output value constitutes the under-film perspective texture feature after removing reflections.

[0102] In the mechanical processing channel, the module inputs the uplift height data from the depth layer into the inverse finite element solver; it discretizes the mulch film into an elastic mesh composed of nodes, setting the longitudinal mesh stiffness value to be greater than the transverse mesh stiffness value to simulate the tensile tension of the mulch film along the direction of travel. The solver calculates the internal force vector that each mesh node must experience to maintain the current uplift height and shape through iterative calculations. The module extracts three components of this force vector: the magnitude of the force, the divergence of the force field gradient, and the angle between the force direction and the mulch film normal, combining them to form the force distribution tensor feature.

[0103] Finally, the module concatenates the normalized perspective texture values ​​and the force tensor values ​​along the channel dimension to generate a composite feature matrix.

[0104] The weed data judgment module uses partitioned mask data as a logical gate to perform differential discrimination on the composite feature matrix: in the corn strip mask area, it filters based on single-point concentrated vertical force features and monocot texture features; in the soybean strip mask area, it filters based on double-point dispersed tension features and dicot texture features; all raised targets that fail the screening are marked as weed data.

[0105] This module reads logical values ​​from the partition mask data and uses them as switches to select the decision branch; When a pixel region with a mask value of logical number one (corn strip) is scanned, the monocot feature criteria are compared: check whether the force distribution tensor shows a unique extreme center and is vertically upward, and check whether the perspective texture presents parallel stripes. If both conditions are met, it is determined to be corn; if there are only ridges but the force is dispersed or the texture is messy, it is determined to be weeds.

[0106] When a pixel region with a mask value of logical number two (soybean band) is scanned, switch to the dicotyledonous feature standard: check whether the stress distribution tensor shows two separate stress points or an arched linear distribution, and whether it contains a horizontal tension component; at the same time, check whether the perspective texture presents a mesh distribution without a dominant direction. If it meets the criteria, it is identified as a soybean; if it does not meet the above features, it is identified as a weed.

[0107] All locations identified as weeds will be marked and output, completing the identification process.

[0108] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A weed identification method based on image recognition, characterized in that, include: The original polarization image and depth point cloud data are acquired, the feature point set of the mulch film is extracted and matched with the standard agronomic template, and the homography transformation matrix is ​​calculated. Based on the homography transformation matrix, a spatially aligned multidimensional data field containing optical and depth layers is constructed, and partitioned mask data with marked positional information of corn and soybean strips is generated. Parallel processing is performed on the layers within the spatially aligned multidimensional data field. In the optical layer, Stokes vectors are calculated, and the high-polarization reflection component of the mulch film is removed using a polarization difference model to generate the perspective texture features under the film. In the depth layer, inverse finite element analysis is performed based on the elastic properties of the thin film to transform the deformation data into stress distribution tensor features. This allows for the synthesis of a composite feature matrix that includes internal texture and external forces; Using partitioned mask data as a logical gate, differential discrimination is performed on the composite feature matrix: in the corn strip mask region, filtering is based on single-point concentrated vertical force features and monocot texture features; in the soybean strip mask region, filtering is based on double-point dispersed tension features and dicot texture features. Mark all raised targets that fail the screening as weed data.

2. The weed identification method based on image recognition according to claim 1, characterized in that, The process of solving the homography transformation matrix includes: The raw data of the field operation scene were collected using a multimodal camera. The raw polarization image included intensity maps of four polarization directions: 0°, 45°, 90° and 135°, and an image pixel coordinate system was established. The depth point cloud data included three-dimensional coordinate values ​​based on the camera coordinate system. In the raw polarization image, the crop emergence position on the surface of the mulch film was identified, and its centroid coordinates were extracted as the mulch film feature point set. The standard agronomic template is a pre-constructed virtual two-dimensional dot matrix, generated based on preset row spacing and plant spacing parameters, representing an ideal planting distribution without distortion, and establishing a standard agronomic coordinate system. The random sampling consensus algorithm is used to establish a correspondence between the feature point set of the mulch film and the standard agronomic template, eliminating mismatched points caused by mulch film damage and irregular holes, and solving the homography transformation matrix that maps the image pixel coordinate system to the standard agronomic coordinate system using the least squares method.

3. The weed identification method based on image recognition according to claim 1, characterized in that, The process of acquiring the spatially aligned multidimensional data field and partitioned mask data includes: The original polarized image is transformed by perspective projection using a homography transformation matrix, and the pixel gray values ​​are reconstructed by bilinear interpolation to generate an optical layer under orthophoto projection. The depth point cloud data is mapped to the standard agronomic coordinate system through the same transformation matrix and rasterized to generate a depth layer, thus constructing a pixel-level aligned spatially aligned multidimensional data field. The partition mask data is generated based on the geometric boundaries of the standard agronomic template. A binary logic diagram is drawn in the standard agronomic coordinate system. The pixels in the region corresponding to the corn planting row are marked as the first logic value, and the pixels in the region corresponding to the soybean planting row are marked as the second logic value, thus obtaining partition mask data that marks the position information of the corn strip and the soybean strip.

4. The weed identification method based on image recognition according to claim 2, characterized in that, The process of obtaining the under-membrane perspective texture features includes: For each pixel in the optical layer, the Stokes vector and linear polarization degree, representing the polarization state of the light wave, are calculated based on the radiation intensity in the four polarization directions. The polarization difference model is constructed based on the physical principle that the reflected light from the surface of the fully biodegradable mulch film has high polarization characteristics, while the diffuse reflected light from the biological tissue under the film has depolarization characteristics. The specific calculation logic is as follows: the total light intensity value in the Stokes vector is multiplied by the linear polarization degree value, and then multiplied by the suppression coefficient dynamically adjusted based on the ambient light intensity to calculate the high polarization component value representing the specular reflection of the mulch film. Then, the high polarization component value is subtracted from the total light intensity value to obtain the perspective texture intensity value containing only the diffuse reflection information under the film, generating the perspective texture feature under the film that can penetrate the semi-transparent mulch film.

5. The weed identification method based on image recognition according to claim 1, characterized in that, The process of obtaining the force distribution tensor features includes: The raised region in the depth layer is modeled as an isotropic elastic film surface, and the Young's modulus and Poisson's ratio parameters of the mulch film are set. The edge contour of the raised region is used as the relative displacement boundary condition, and the height value in the depth layer is used as the known displacement field input. The nodal force vectors that cause the deformation are solved by the inverse operation of the finite element method. For each mesh node, the direction and magnitude of the force are calculated, and a force distribution tensor describing the stress concentration state is constructed. The force distribution tensor features include three-dimensional feature components: principal stress amplitude, stress gradient divergence, and the angle between the principal stress direction and the normal of the mulch film, which are used to characterize the contact characteristics when an object under the film lifts the mulch film at the physical and mechanical level.

6. The weed identification method based on image recognition according to claim 1, characterized in that, The steps for generating the composite feature matrix include: The underlying perspectral texture features and stress distribution tensor features are numerically normalized to map the texture grayscale values ​​and tensor feature values ​​to a unified dimensionless range. While maintaining consistent spatial resolution, the normalized texture feature channels and stress tensor feature channels are concatenated in the depth dimension to construct a multi-channel composite feature matrix. Each feature vector of the composite feature matrix simultaneously encodes the spectral texture attributes and top membrane mechanical properties of the object under the membrane at that location.

7. The weed identification method based on image recognition according to claim 1, characterized in that, include: When the partition mask data indicates that the current region is a corn belt, the monocotyledon discrimination logic is activated. The single-point concentrated vertical force characteristic is defined as follows: the force distribution tensor shows the existence of a unique stress maximum center, and the principal stress direction is perpendicular to the surface of the mulch film, characterizing the mechanical behavior of the sharp top membrane of the coleoptile of monocotyledonous plants. The monocotyledon texture feature is defined as: in the membrane perspective texture feature, it shows significant parallel texture directionality, characterizing parallel leaf vein features.

8. The weed identification method based on image recognition according to claim 1, characterized in that, include: When the partition mask data indicates that the current region is a soybean belt, the dicotyledonous discrimination logic is activated; The dual-point dispersed tension characteristic is defined as follows: the stress distribution tensor shows two separate stress concentration points, exhibiting a linear arched stress distribution, and there is a significant horizontal tension component, characterizing the mechanical behavior of dicotyledonous plants when they emerge from the soil. The dicotyledonous texture feature is defined as follows: in the under-membrane perspective texture feature, the texture gradient histogram exhibits an isotropic distribution feature without a clear dominant direction, characterizing the reticulate leaf vein feature.

9. A weed identification system based on image recognition, characterized in that, include: The raw data acquisition module acquires raw polarization images and depth point cloud data, extracts the feature point set of the mulch film and matches it with a standard agronomic template, and calculates the homography transformation matrix. The data transformation and recognition module, based on the homography transformation matrix, constructs a spatially aligned multidimensional data field containing optical and depth layers, and generates partitioned mask data that marks the location information of the corn and soybean strips. The multidimensional data processing module performs parallel processing on layers within the spatially aligned multidimensional data field. In the optical layer, it calculates the Stokes vector, uses the polarization difference model to remove the high polarization reflection component of the mulch film, and generates the perspective texture features under the film. In the depth layer, it performs inverse finite element analysis based on the elastic properties of the thin film, transforming the deformation data into the stress distribution tensor features. This allows for the synthesis of a composite feature matrix that includes internal texture and external forces; The weed data judgment module uses partitioned mask data as a logical gate to perform differential discrimination on the composite feature matrix: in the corn strip mask area, it filters based on single-point concentrated vertical force features and monocot texture features; in the soybean strip mask area, it filters based on double-point dispersed tension features and dicot texture features. Mark all raised targets that fail the screening as weed data.