Method for ai image recognition and grading of leaf diseases and pests of field crops

Abstract

The present application relates to the technical field of pest image analysis, and particularly relates to an AI image recognition and grading method for field crop leaf diseases and pests, comprising the following steps: under the irradiation of a fixed light source in a field, multiple polarized reflection images are synchronously collected at a predetermined angle interval around the crop leaves; a pixel polarization degree matrix of a leaf area in each polarized reflection image is extracted; the pixel polarization degree matrix is input into a polarization transmission model, and an abnormal cuticle coefficient map is output, and a region exceeding a preset abnormal threshold in the abnormal cuticle coefficient map is marked as a highlighted display region; an infection type template library is matched according to a distribution mode of the highlighted display region in the abnormal coefficient map; an infection intensity value is calculated in combination with a diffusion gradient of the highlighted region, and a disease and pest grade is output. The present application characterizes the optical mutation region boundary of a lesion area, not only improves the physical interpretation of disease detection, but also provides a distinguishable feature space for different infection mechanisms (such as a fungal growth layer and a pest piercing point).

Description

Technical Field

[0001] This invention relates to the field of pest and disease image analysis technology, and in particular to an AI image recognition and classification method for pests and diseases on the leaves of field crops. Background Technology

[0002] With the development of precision agriculture and green prevention and control concepts, early identification and grading of pests and diseases have become key links in ensuring stable and high yields of field crops. At present, field pest and disease identification mainly relies on manual inspection or visual recognition methods based on RGB images. However, in practical applications, there are common problems such as low identification accuracy, poor ability to distinguish types, and poor environmental adaptability.

[0003] Traditional image recognition methods mainly rely on texture, color, or shape features, which are difficult to deal with interference from lesions and leaf veins, reflection interference under different lighting conditions, and morphological similarities between diseases. At the same time, these methods generally ignore the physical structural changes of the leaf surface and cannot establish an interpretable model from pathological behavior to image response, resulting in type recognition relying on model experience and insufficient generalization ability.

[0004] At the level of pathological detection, some studies have attempted to introduce multispectral imaging, thermal infrared imaging and other means to enhance the perception ability, but they are costly, the equipment is complex and they are not sensitive to changes in the microstructure of the cuticle, making it difficult to promote them widely in the field. In addition, existing pest and disease classification methods are mostly based on area or color intensity, lacking dynamic diffusion modeling related to pathogen transmission mechanisms, and cannot effectively distinguish different transmission behaviors such as fungal diffusion, insect piercing and sucking or bacterial infiltration. Summary of the Invention

[0005] This invention provides an AI image recognition and classification method for leaf diseases and pests of field crops. The image recognition method, which combines physical mechanism modeling and feature discrimination, can accurately identify, determine the type and grade of leaf diseases and pests of field crops in the field environment, forming an intelligent recognition closed loop from image acquisition to prevention and control decision-making.

[0006] An AI-based image recognition and classification method for leaf diseases and pests in field crops includes the following steps:

[0007] S1. Under the illumination of a fixed light source in the field, multiple polarization reflection images are simultaneously acquired around the crop leaves at predetermined angular intervals; the pixel polarization degree matrix of the leaf region in each polarization reflection image is extracted.

[0008] S2, stratum corneum refraction field reconstruction: Input the pixel polarization degree matrix into the polarization transmission model, output the stratum corneum anomaly coefficient map, and mark the areas in the stratum corneum anomaly coefficient map that exceed the preset anomaly threshold as highlighted display areas; the polarization transmission model inverts the stratum corneum microstructure through a ray tracing algorithm;

[0009] S3 matches the infection type template library based on the distribution pattern of the highlighted areas in the anomaly coefficient graph; calculates the infection intensity value by combining the diffusion gradient of the highlighted areas, and outputs the pest and disease level.

[0010] Optionally, S1 includes the deployment of a polarization image acquisition device, specifically including setting a fixed laser light source in the normal direction of the leaf to be tested, with the incident direction of the light source forming a 45° angle with the normal of the leaf surface;

[0011] A polarization camera is deployed along a circular track with a fixed radius centered on the blade, the plane of which is parallel to the blade surface.

[0012] Optionally, the synchronous acquisition of multiple polarization reflection images includes driving the polarization camera to move along a circular track at 15° angular intervals, briefly pausing at each angular position and triggering the following operation:

[0013] a) Start the laser source to emit continuous pulsed laser light;

[0014] b) Simultaneously capture reflection images in multiple polarization directions using four focal plane polarization sensors of a polarization camera;

[0015] c) Generate the original polarization image set for the current angular position.

[0016] Optionally, the generation of the pixel polarization matrix includes performing the following on the original polarization image group for each angular position:

[0017] a) Calculate the Stokes parameter vector for each pixel. ;

[0018] b) Calculate pixel polarization degree based on Stokes parameter vector: ;

[0019] c) All pixels within the leaf area The values ​​are constructed into an m×n matrix, and the output is the pixel polarization degree matrix.

[0020] Optionally, S2 specifically includes:

[0021] S21, Polarization Transmission Modeling: The polarization degree maps collected from various angles are sequentially combined into a three-dimensional polarization data volume; based on the propagation law of scattered light in the stratum corneum, a polarization transmission model including light attenuation and scattering interaction processes is constructed.

[0022] S22, Microstructure parameter inversion: Initialize the refractive index distribution in the blade cuticle as the simulation basis, perform simulation iterations from angle to angle, and continuously adjust the refractive index distribution according to the image fitting error until the error meets the convergence condition or reaches the maximum number of iterations.

[0023] S23, Anomaly Coefficient Generation: When the inversion error reaches the set convergence condition or iteration limit, the optimization is terminated. Based on the final refractive index distribution, the anomaly coefficient of each spatial location is calculated. Areas with anomaly coefficients exceeding the preset anomaly threshold are marked as highlighted areas, and the output is a stratum corneum anomaly coefficient map.

[0024] Optionally, the simulation iteration specifically includes:

[0025] Extract the measured degree of polarization map for the current angle;

[0026] Emit light rays in the corresponding direction in a refractive index simulation field;

[0027] Track the propagation path of light in a medium and record its polarization changes;

[0028] The simulation results are compared with the measured figures, and the deviation between the two is minimized according to the preset target.

[0029] Optionally, the polarization transmission model is expressed as:

[0030] ;

[0031] in, Indicates position Along the direction Radiance The total attenuation coefficient is... The scattering coefficient is... The polarization phase function describes the incident direction. To the direction of launch The scattering probability, This represents the arc length along the path of light propagation. This represents the integral symbol for the entire space.

[0032] Optionally, the distribution pattern of the highlighted display area in S3 is represented as a spatial distribution descriptor. Based on the highlighted display area extracted from the stratum corneum abnormality coefficient map, a binary mask image of the highlighted area is generated. The binary mask image is input into a pre-trained convolutional neural network to extract the centroid coordinates, radial distribution histogram, and number of connected components to form a spatial distribution descriptor. The spatial distribution descriptor is matched with each type of template in a preset infection type template library for similarity, and the similarity score is evaluated according to a preset weight. When the similarity score exceeds the similarity threshold, the type corresponding to the template with the highest score is taken as the infection type recognition result of the current image.

[0033] Optionally, the calculation of the infection intensity value includes: constructing a spatial gradient field of the abnormal coefficient image for the highlighted display area; extracting directional gradient histogram information based on the spatial gradient field to form a diffusion pattern vector, which is used to characterize the lesion diffusion morphology; selecting the corresponding intensity calculation method according to the determined infection type, and calculating the infection intensity value by combining the area, gradient distribution, and connectivity information.

[0034] Optionally, S3 further includes a grade mapping output, specifically including inputting the identified infection type and the corresponding infection intensity value into a grade mapping table; and matching the corresponding pest and disease grade according to the intensity range of different types.

[0035] The beneficial effects of this invention are:

[0036] This invention constructs a pixel-level polarization matrix through multi-angle polarization image acquisition and Stokes parameter derivation. Combined with a pulsed laser synchronization mechanism and a four-directional focal plane sensor, it achieves high-fidelity acquisition of the optical microstructure of the leaf surface. A 532nm fixed-wavelength laser is used to target and enhance the reflection difference in the lesion area. Combined with dust compensation and dynamic focusing algorithms, it can stably extract abnormal areas under strong noise and complex background conditions in the field, avoiding the problems of blurred lesion boundaries and difficulty in distinguishing structural disturbances in traditional RGB and multispectral images.

[0037] This invention introduces ray tracing algorithms and polarization transmission models into agricultural pathological image analysis. By reconstructing the cuticle refractive index distribution field and constructing anomaly coefficient maps, it achieves the inversion of physical parameters from image polarization data to microstructures. Based on this, gradient field, directional entropy, and curvature compensation mechanisms are introduced to accurately characterize the optical abrupt change region boundary of the lesion area. This not only improves the physical interpretability of disease detection but also provides a distinguishable feature space for different infection mechanisms (such as fungal growth layers and insect piercing and sucking points).

[0038] This invention proposes a spatial distribution descriptor-driven type discrimination mechanism. Through highlight region centroid, orientation histogram, and connectivity modeling, it achieves AI feature encoding of the spatial morphology of pests and diseases. Simultaneously, it integrates HOG features, gradient statistics, and an infection intensity calculation model to adapt to the spread behavior characteristics of different disease types, outputting real-time grade values ​​and refined control suggestions. This closed-loop structure, from image to microstructure to behavioral patterns to grade decision-making, permeates perception, modeling, and application, forming an innovative three-in-one field intelligent identification system integrating "physical quantification + recognition + agricultural intervention." Attached Figure Description

[0039] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of the identification and classification method according to an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of the stratum corneum refractive field reconstruction method according to an embodiment of the present invention. Detailed Implementation

[0042] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0043] like Figures 1-2 As shown, the AI ​​image recognition and classification method for leaf diseases and pests in field crops includes the following steps:

[0044] S1. Under the illumination of a fixed light source in the field, multiple polarization reflection images are simultaneously acquired around the crop leaves at predetermined angular intervals; the pixel polarization degree matrix of the leaf region in each polarization reflection image is extracted.

[0045] S2, stratum corneum refraction field reconstruction: Input the pixel polarization degree matrix into the polarization transmission model and output the stratum corneum anomaly coefficient map. Mark the areas in the stratum corneum anomaly coefficient map that exceed the preset anomaly threshold as highlighted areas; The polarization transmission model inverts the microstructure of the stratum corneum through the ray tracing algorithm;

[0046] S3 matches the infection type template library based on the distribution pattern of the highlighted areas in the anomaly coefficient graph; calculates the infection intensity value by combining the diffusion gradient of the highlighted areas, and outputs the pest and disease level.

[0047] S1 specifically includes:

[0048] S11, Deployment of polarization image acquisition device: A fixed laser source with a wavelength of 532nm is set in the normal direction of the leaf to be tested, and the incident direction of the light source is at a 45° angle with the normal of the leaf surface; a polarization camera is deployed along a circular track with a radius of 30cm centered on the leaf, and the plane of the track is parallel to the leaf surface.

[0049] S12: Multi-angle image synchronous acquisition: Drive the polarization camera to move along the track in 15° angular intervals, pausing for 200ms at each angular position. Perform the following operations synchronously at each angular position:

[0050] a) Start the laser source and emit a pulsed laser beam lasting 10ms;

[0051] b) Using the four focal plane polarization sensors built into the polarization camera, reflective images in four polarization directions of 0°, 45°, 90°, and 135° are captured simultaneously.

[0052] c) Combine the four images acquired at this angle position into a raw polarization image group.

[0053] S13, Pixel polarization matrix generation: For each group of original polarized images at each angular position, perform the following processing:

[0054] a) Calculate the Stokes parameter vector for each pixel:

[0055] ;

[0056] ;

[0057] ;

[0058] in: These represent the pixel grayscale values ​​corresponding to the polarization direction;

[0059] b) Calculate the polarization degree of each pixel based on the Stokes parameters. : ;

[0060] c) Combining the DoP values ​​of all pixels within the leaf area The matrix output is the pixel polarization degree matrix at that angular position; repeat the above acquisition and calculation process until data acquisition and processing at 24 angular positions are completed. Furthermore, the pixel polarization degree matrix is ​​constructed as follows:

[0061] I. Input Data Structure: For each angular position, four polarization direction images will be acquired: They are all of size A grayscale image. For example:

[0062] (The same applies to the rest);

[0063] II. DoP Calculation Logic: For each pixel... Calculate the DoP value based on the Stokes parameters:

[0064] ;

[0065] ;

[0066] ;

[0067] Therefore, we get: ;

[0068] III. Construction of the pixel polarization matrix: This involves all pixels within the blade region. of Arranged according to their image space, they form a The matrix:

[0069] ;

[0070] in The extracted region of interest (ROI) is the pixel range in the image, and is not equal to the entire image size. It is not the actual effective area after mask clipping.

[0071] Fourth, this matrix is ​​the pixel polarization degree matrix at that angular position, which is used as the input for the subsequent refraction field reconstruction model. Polarizing camera at angle The polarization matrix generated at the given location was acquired from 24 angles, with one matrix output for each angle. .

[0072] Obtaining reflection images in four polarization directions stems from the minimum sampling requirement of the Stokes parameter model. The Stokes parameters are a set of standard parameters describing the state of partially polarized light, defined as:

[0073] Total light intensity;

[0074] : The intensity difference between the 90° and 90° polarization directions;

[0075] The intensity difference between the 45° and 135° polarization directions;

[0076] : Describes circular polarization (often used in more complex systems, not used in this scheme).

[0077] In most agricultural or industrial vision systems, circular polarization ( These three parameters are neither easy to obtain nor necessary. This is sufficient to construct a complete linear polarization information model, which requires acquiring images in four polarization directions: 0°, 45°, 90°, and 135°.

[0078] To accurately reconstruct the pixel-level polarization degree (DoP), the polarization degree (DoP) can be calculated based on the Stokes parameters: ;

[0079] Pests and diseases infecting leaves alter the microstructure of the cuticle, thus affecting the ability to maintain light polarization. Diseased areas typically reflect light with higher or disordered polarization, while healthy tissue exhibits more uniform and stable polarization. Calculating the degree of polarization (DoP) reveals structural optical differences, providing crucial information for subsequent refractive field reconstruction. Compared to directly using grayscale images, the degree of polarization map is more sensitive to cuticle changes, helping to amplify the differences between lesion areas and the background, creating bright, heterogeneous regions that facilitate template matching and grading. The polarization information is more stable relative to the angle of incidence, avoiding uneven illumination caused by leaf morphology, thereby improving the model's spatial robustness, especially suitable for complex field environments.

[0080] S2 specifically includes the following steps:

[0081] S21, Polarization Transmission Modeling: Stack the pixel polarization degree matrices corresponding to the 24 angular positions according to the acquisition angle order to construct a three-dimensional polarization data volume.

[0082] ;

[0083] Based on Mie scattering theory, a polarization transport model (polarization control equation) for photon transmission within the stratum corneum is established:

[0084] The description of light attenuation and scattering behavior during propagation in the stratum corneum forms the physical modeling basis for the entire refraction field reconstruction; among which, Indicates position Along the direction Radiance The total attenuation coefficient is... The scattering coefficient is... The polarization phase function describes the incident direction. To the direction of launch The scattering probability, This represents the arc length along the path of light propagation. Represents the integral symbol for the entire space, over all possible incident directions on the unit sphere. Integrating to cover the entire spherical solid angle of 4°steradians. The polarization phase function combines the angle factor and wavelength response: ;in, The scattering angle (i.e.) , This represents the wavelength-dependent gain factor; the polarization response gain of the lesion tissue at 532 nm is approximately 2.3 times.

[0085] Basic optical parameters of healthy leaves: By consulting existing plant spectral databases and agricultural optical data, the optical characteristics of the cuticle of healthy field crop leaves at a wavelength of 532 nm were selected as the basic reference.

[0086] The total attenuation coefficient of the stratum corneum is: The scattering coefficient is dominant and is approximately: The absorption coefficient is: This reflects that the main energy loss caused by the waxy layer of the leaf epidermis during laser propagation in short paths is scattering.

[0087] Considering that fungal infections (such as powdery mildew and rust) can lead to disordered intercellular structure and increased water content in the cuticle, significantly enhancing light scattering and increasing measured polarization, enhanced optical parameters were used for the diseased area: the empirical enhancement ratio was set to 1.6-1.8 times that of healthy tissue.

[0088] ;

[0089] .

[0090] The optical properties of the stratum corneum at a wavelength of 532nm, based on measured data, are shown in the table below.

[0091] Table 1. Measured Optical Properties of the Stratum Corneum

[0092]

[0093] The above data are a combination of laboratory integrating sphere measurements and image inversion fitting results, and the field test sites cover three field crops: rice, rapeseed, and corn.

[0094] S22, Microstructure parameter inversion: Initialize the stratum corneum refractive index distribution field as follows: Inversion iteration is performed using ray tracing algorithms:

[0095] S221, Extract the current angle from the three-dimensional polarization data volume. Measured polarization degree diagram: A three-dimensional polarization data volume is a data set composed of pixel polarization degree matrices collected from different angular positions stacked in angular order. Each frame Corresponding to the Each collection angle The extraction method is as follows: Based on the pre-set acquisition angle sequence, directly press the index... from The corresponding frame is indexed in the middle; this frame is then regarded as the current angle. Measured polarization diagram below ;

[0096] S222, in the current refractive index distribution field In the middle, along the direction Emits virtual light beams to simulate the propagation of light in a specific direction;

[0097] S223 simulates the propagation and scattering of each ray of light in the stratum corneum, and calculates the corresponding theoretical degree of polarization: Specifically, along the current angle and direction A set of incident rays is generated on the leaf surface, covering the entire pixel area. As each ray propagates in the cuticle, it is tracked point by point according to the Mie scattering model, and the scattering angle, energy attenuation, and polarization change are calculated for each interaction. All directions and energy information of the ray from incident to exit are recorded. The exit polarization states of multiple rays at the same point are combined to deduce the simulated polarization value of the corresponding pixel. The method is consistent with the measured polarization degree (using Stokes vector back-calculation); after all ray simulations are completed, each pixel is... Arranged into a two-dimensional matrix of the same size as the measured image, this is the theoretical polarization diagram;

[0098] S224, Construct and minimize the cost function: Through iterative optimization To minimize error Update the refractive index distribution field.

[0099] S23, Anomaly Coefficient Generation:

[0100] Iteration stops when the error meets any of the following conditions: Condition 1: Condition 2: Reaching the maximum number of iterations;

[0101] In each spatial location Calculate the anomaly coefficient: ;

[0102] Output the stratum corneum abnormality coefficient map, where: Mark as an abnormal area.

[0103] The basic refractive index of the cuticle of a healthy plant leaf is approximately According to existing optical research and experimental data, refractive index fluctuations caused by slight moisture evaporation and environmental fluctuations are typically within ±0.05. However, tissue damage, moisture accumulation, or mycelial infection caused by pests and diseases often raise the refractive index to above 1.6. This is equivalent to the following requirement: This exceeds the range of natural physiological fluctuations, ensuring that the detected area represents abnormal optical behavior rather than physiological disturbance.

[0104] Furthermore, field trials revealed that:

[0105] If the threshold is below 0.20, bright artifacts are easily generated in the leaf vein and leaf edge areas (due to the difference in natural curvature and thickness).

[0106] If the threshold is higher than 0.30, some early lesions cannot be identified, reducing sensitivity.

[0107] The experimental results show that 0.25 is the optimal compromise threshold that balances lesion recognition rate and false recognition rate, and is particularly suitable for diseases such as rust, aphids, and mold that cause local microstructural disturbances.

[0108] During the inversion process, after the error function E stabilizes with iteration, its decrease rate slows down, but the computational resource consumption increases significantly. Experimental evaluation shows that:

[0109] When E < 0.1, most abnormal features can be roughly located;

[0110] Further optimization to E<0.05 can improve lesion boundary resolution by approximately 15–20%;

[0111] If we force optimization to E<0.01, although the numerical error will be lower, the number of iterations will be increased significantly, resulting in low cost-effectiveness.

[0112] Therefore, 0.05 is used as the error convergence threshold to achieve the engineering optimal balance between accuracy and efficiency.

[0113] S3 specifically includes the following steps:

[0114] S31, Infection Pattern Matching:

[0115] S311, based on extracting the highlighted regions from the stratum corneum anomaly coefficient map to construct a binary mask: Highlight mask Specifically, this involves traversing all pixel positions in the entire image, marking any pixel with an anomaly coefficient greater than a threshold as "1" to indicate an abnormal region, and marking the rest as "0" to indicate the background. Finally, a binary mask image of the same size as the original image is generated, which contains only pixel values ​​"0" and "1" to highlight the lesion's bright areas. This binary mask is then used as input to a CNN model for further spatial structure analysis and infection type identification.

[0116] S312, the binary mask Input a pre-trained convolutional neural network (CNN) to extract spatial distribution descriptors. ,include:

[0117] : The centroid coordinates of the highlighted area (reflecting spatial offset);

[0118] Radial distribution histogram (reflecting texture structure);

[0119] : Number of locally connected components (reflecting the degree of dispersion);

[0120] S313, for each preset infection type in the infection type template library Calculate the descriptor and its template vector. Weighted cosine similarity score:

[0121] ;in, This represents the weight vector of each component. This represents the cosine similarity function.

[0122] S314, And the type corresponding to the highest-scoring template As a result of determining the type of infection.

[0123] S32, Infection Intensity Calculation:

[0124] S321, masking for highlight areas The gradient field of anomaly coefficients in the region is calculated:

[0125] ;

[0126] S322, extracts the oriented gradient histogram (HOG) from the gradient field image and outputs the diffusion mode vector;

[0127] S323, based on the determined type of infection Select the corresponding strength calculation method to obtain the real-time strength value. Calculation:

[0128] Fungal models (such as powdery mildew, rust): ;

[0129] Pest models (such as aphids and planthoppers): ;

[0130] Bacterial models (such as soft rot and angular leaf spot): ;

[0131] in, Indicates the area of ​​the highlighted region. This represents the total detection area of ​​the leaf, where mean, std, and entropy are the gradient mean, standard deviation, and HOG image entropy, respectively. This indicates the number of connected subregions in the highlighted area. The previously extracted directional gradient histogram is a specific representation of the aforementioned "diffusion mode vector".

[0132] S33, Grade Mapping Output: Based on Infection Type and Intensity Value Find the corresponding level and prevention and control recommendations in the table below, and output the pest / disease level and control instructions:

[0133] Table 2. Infection Type and Intensity Value Query and Classification Mapping Table

[0134]

[0135] The pre-defined infection type template library performs feature summarization and matching at the spatial feature level to determine different types of pests and diseases. By comparing the similarity between descriptors extracted from real-time images and descriptor vectors of various types in the template library, the potential causes of lesions (fungal, insect, or bacterial infection) are determined. The template construction is as follows:

[0136] Step 1: Standard Sample Collection: Collect image samples of diseases and pests on the leaves of various field crops, covering the following infection types:

[0137] Fungal diseases (such as powdery mildew and rust), insect pests (such as aphids and planthoppers), and bacterial diseases (such as angular leaf spot and soft rot); at least 100 images with typical lesion characteristics should be selected as the basic dataset for each type of infection.

[0138] Step 2: Anomaly Region Extraction and Descriptor Generation: Perform the aforementioned steps (anomaly coefficient map generation → highlight mask extraction) on each standard image, and extract the following three spatial distribution descriptors using a CNN:

[0139]

[0140] The average or statistical median of multiple samples of the same type are used to form a template descriptor subgroup for that type.

[0141] Step 3: Each template type is represented by a triple, with the following structure: The data can be stored in JSON or a vector database, facilitating fast retrieval and similarity calculation.

[0142] In the recognition process, the descriptors extracted in real time Calculate the weighted cosine similarity with each template:

[0143] A similarity threshold of 0.85 is set, and the template with the highest similarity is the result of the infection type determination in this round of detection.

[0144] The convolutional neural network (CNN) used in this invention is a lightweight feature extraction structure that encodes spatial distribution features in binary images. Its network structure includes the following:

[0145] 1. Input layer: Receives a binary image mask of normalized size (64×64 pixels);

[0146] 2. Multi-layer convolutional module: Uses multiple convolutional kernels with different receptive fields (such as 3×3, 5×5) to extract local structure, edge orientation and concentration features; each convolutional layer is followed by a ReLU activation function and batch normalization;

[0147] 3. Spatial Attention Module: Emphasizes target areas off-center from the image through a weighted mechanism, improving sensitivity to lesion offset;

[0148] 4. Global pooling layer: Performs max pooling and average pooling on the spatial dimension respectively, compressing it into a feature vector of fixed length;

[0149] 5. Output Embedding Layer: The final feature maps of the network are encoded into three parallel output channels:

[0150] Target centroid coordinate vector (describes the spatial distribution of the center of the highlighted area);

[0151] Radial histogram vector (describes the distribution tendency of bright areas in an image relative to the center);

[0152] Connectivity structure statistics (describe regional fragmentation or clustering).

[0153] Finally, the CNN outputs a combined vector of the three types of features in a nested structure, which serves as a spatial distribution descriptor.

[0154] The mask image input and feature extraction process is as follows:

[0155] Image preprocessing: The extracted highlight areas are normalized to a uniform size using a binary mask. If the original image area is insufficient, edge filling or proportional scaling can be performed. The input image contains only 0 and 1 values, representing abnormal areas and background.

[0156] Automatic feature learning and encoding: The network first extracts basic texture features such as shape, size, and orientation of the highlighted areas in the local image; the intermediate layer enhances the learning of the centroid shift or concentration pattern of lesions through a spatial focusing mechanism; the network end converts the feature map into three interpretable numerical structures, which are used to measure the centroid position, directional distribution, and connected component structure, respectively.

[0157] The output spatial distribution descriptor corresponds to the three channels of the network output:

[0158] The relative position coordinates of the currently highlighted area in the image;

[0159] Statistical vector of the directional pattern radiating outward from the highlighted area;

[0160] Statistical results regarding whether the abnormal areas exhibit a discrete, striped, or speckled distribution.

[0161] These outputs serve as input to the feature template matching module, participating in subsequent infection type identification and matching determination.

[0162] During the design phase, the network used a large number of labeled typical pest and disease mask images for pre-training. Each type of training sample was associated with manually labeled real descriptor values ​​as training labels. The loss function comprehensively considered dimensions such as center bias, directional distribution differences, and connected component quantity errors to ensure that the model has a stable ability to extract three types of features.

[0163] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0164] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An AI image recognition and classification method for leaf diseases and pests in field crops, characterized in that, Includes the following steps: S1. Under the illumination of a fixed light source in the field, multiple polarization reflection images are simultaneously acquired around the crop leaves at predetermined angular intervals; the pixel polarization degree matrix of the leaf region in each polarization reflection image is extracted. S2, stratum corneum refraction field reconstruction: Input the pixel polarization degree matrix into the polarization transmission model, output the stratum corneum anomaly coefficient map, and mark the areas in the stratum corneum anomaly coefficient map that exceed the preset anomaly threshold as highlighted display areas; The polarization transport model inverts the microstructure of the stratum corneum using a ray tracing algorithm; S2 specifically includes: S21, Polarization Transmission Modeling: The polarization degree maps collected from various angles are sequentially combined into a three-dimensional polarization data volume; based on the propagation law of scattered light in the stratum corneum, a polarization transmission model including light attenuation and scattering interaction processes is constructed. S22, Microstructure parameter inversion: Initialize the refractive index distribution in the blade cuticle as the simulation basis, perform simulation iterations from angle to angle, and continuously adjust the refractive index distribution according to the image fitting error until the error meets the convergence condition or reaches the maximum number of iterations. S23, Anomaly Coefficient Generation: When the inversion error reaches the set convergence condition or iteration limit, the optimization is terminated. Based on the final refractive index distribution, the anomaly coefficient of each spatial location is calculated. Areas with anomaly coefficients exceeding the preset anomaly threshold are marked as highlighted areas, and the output is a stratum corneum anomaly coefficient map. The simulation iteration specifically includes: Extract the measured degree of polarization map for the current angle; Emit light rays in the corresponding direction in a refractive index simulation field; Track the propagation path of light in a medium and record its polarization changes; The simulation results are compared with the measured figures, and the deviation between the two is minimized according to the preset target. S3. Match the infection type template library according to the distribution pattern of the highlighted areas in the anomaly coefficient map; calculate the infection intensity value by combining the diffusion gradient of the highlighted areas, and output the pest and disease level. The distribution pattern of the highlighted areas is represented as a spatial distribution descriptor. Based on the highlighted areas extracted from the cuticle anomaly coefficient map, generate a binary mask image of the highlighted areas; input the binary mask image into a pre-trained convolutional neural network to extract the centroid coordinates, radial distribution histogram, and number of connected components to form a spatial distribution descriptor; perform similarity matching between the spatial distribution descriptor and each type template in the preset infection type template library, and evaluate the similarity score according to the preset weight; when the similarity score exceeds the similarity threshold, the type corresponding to the template with the highest score is taken as the infection type identification result of the current image.

2. The AI ​​image recognition and classification method for leaf diseases and pests of field crops according to claim 1, characterized in that, S1 includes the deployment of a polarization image acquisition device, specifically including setting a fixed laser light source in the normal direction of the blade to be tested, with the incident direction of the light source forming a 45° angle with the normal of the blade surface. A polarization camera is deployed along a circular track with a fixed radius centered on the blade, the plane of which is parallel to the blade surface.

3. The AI ​​image recognition and classification method for leaf diseases and pests of field crops according to claim 2, characterized in that, The synchronous acquisition of multiple polarization reflection images includes driving the polarization camera to move along a circular track at 15° angular intervals, briefly pausing at each angular position and triggering the following operations: a) Start the laser source to emit continuous pulsed laser light; b) Simultaneously capture reflection images in multiple polarization directions using four focal plane polarization sensors of a polarization camera; c) Generate the original polarization image set for the current angular position.

4. The AI ​​image recognition and classification method for leaf diseases and pests of field crops according to claim 3, characterized in that, The generation of the pixel polarization matrix involves performing the following steps on the original polarization image group at each angular position: a) Calculate the Stokes parameter vector S0, S1, S2 for each pixel; Where S0 represents the total light intensity, S1 represents the intensity difference between the polarization directions of 0° and 90°, and S2 represents the intensity difference between the polarization directions of 45° and 135°. b) Calculate pixel polarization degree based on Stokes parameter vector: c) Construct an m×n matrix from the DoP values ​​of all pixels within the blade region and output the pixel polarization degree matrix.

5. The AI ​​image recognition and classification method for leaf diseases and pests of field crops according to claim 1, characterized in that, The polarization transmission model is expressed as follows: in, Indicates the direction at position r. Radiance, μ t μ is the total attenuation coefficient. s The scattering coefficient is... The polarization phase function describes the incident direction. To the direction of launch The scattering probability, s represents the arc length along the path of light propagation, ∫ 4π This represents the integral symbol for the entire space.

6. The AI ​​image recognition and classification method for leaf diseases and pests of field crops according to claim 1, characterized in that, The calculation of the infection intensity value includes: constructing a spatial gradient field of the abnormal coefficient image for the highlighted area; extracting the directional gradient histogram information based on the spatial gradient field to form a diffusion pattern vector, which is used to characterize the lesion diffusion morphology; selecting the corresponding intensity calculation method according to the determined infection type, and calculating the infection intensity value by combining the area, gradient distribution, and connectivity information.

7. The AI ​​image recognition and classification method for leaf diseases and pests of field crops according to claim 6, characterized in that, The S3 also includes a grade mapping output, specifically including inputting the identified infection type and the corresponding infection intensity value into a grade mapping table; and matching the corresponding pest and disease grade according to the intensity range of different types.

Images