AI image recognition and grading method for field crop leaf diseases and insect pests

Abstract

The invention relates to the technical field of disease and insect pest image analysis, in particular to an AI image recognition and grading method for field crop leaf disease and insect pests, which comprises the following steps: under the irradiation of a field fixed light source, synchronously acquiring a plurality of polarized reflection images around crop leaves at preset angle intervals; extracting a pixel polarization degree matrix of a leaf area in each polarization reflection image; inputting the pixel polarization degree matrix into a polarization transmission model, outputting a cuticle anomaly coefficient graph, and marking an area exceeding a preset anomaly threshold in the cuticle anomaly coefficient graph as a highlight display area; matching an infection type template library according to a highlight display area distribution mode in the abnormal coefficient graph; and calculating an infection intensity value by combining the diffusion gradient of the highlight area, and outputting a pest grade. According to the method, the boundary of the optical mutation region of the focus region is depicted, so that the physical interpretation of disease detection is improved, and distinguishable feature spaces are provided for different infection mechanisms (such as fungal growth layers and insect pest piercing and sucking points).

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of pest image analysis, in particular to an AI image recognition and grading method for field crop leaf diseases and pests. BACKGROUND

[0002] With the development of fine agriculture and green prevention and control concept, early identification and grading of diseases and pests have become a key link to ensure stable and high yield of field crops. At present, field disease and pest identification mainly relies on manual patrol or visual recognition method based on RGB images, but in actual application, there are problems such as low recognition accuracy, poor type distinguishing ability and poor environmental adaptability.

[0003] Traditional image recognition methods mainly rely on texture, color or shape features, which are difficult to deal with the interference of disease spots and leaf veins, the interference of reflection under different light conditions and the similar morphology between diseases. At the same time, such methods generally ignore the physical structure changes of the leaf surface layer, and cannot establish an interpretable model from pathological behavior to image response, resulting in that the type recognition depends on model experience and the generalization ability is insufficient.

[0004] At the pathological detection level, some studies try to introduce multispectral imaging, thermal infrared imaging and other means to enhance the perception ability, but the cost is high, the equipment is complex and it is not sensitive to the microstructure changes of the cuticle layer, which is difficult to be widely popularized in the field. In addition, the existing disease and pest grading methods are mainly based on area or color intensity, lack of dynamic diffusion modeling related to pathogen transmission mechanism, and cannot effectively distinguish different transmission behaviors such as fungal diffusion, pest piercing or bacterial infiltration. SUMMARY

[0005] The application provides an AI image recognition and grading method for field crop leaf diseases and pests, which combines physical mechanism modeling and feature discrimination image recognition method, can accurately identify, type determine and grade output the field crop leaf diseases and pests in the field environment, and forms an intelligent recognition closed loop from image acquisition to prevention and control decision.

[0006] The AI image recognition and grading method for field crop leaf diseases and pests comprises the following steps: S1. Under the irradiation of a fixed light source in the field, a plurality of polarized reflection images are synchronously collected at a predetermined angle interval around the crop leaf; a pixel polarization degree matrix of the leaf area in each polarized reflection image is extracted; S2. Cuticle refraction field reconstruction: input the pixel polarization degree matrix into a polarization transmission model to output a cuticle abnormality coefficient map, and mark the area in the cuticle abnormality coefficient map exceeding a preset abnormality threshold as a highlighted area; the polarization transmission model inverses the microstructure of the cuticle layer through a ray tracing algorithm; S3, matches the infection type template library according to the distribution pattern of the highlighted area in the anomaly coefficient map; calculates the infection intensity value based on the diffusion gradient of the highlighted area, and outputs the pest and disease grade.

[0007] 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 measured, with the incident direction of the light source forming an angle of 45° with the leaf surface normal; The polarization camera is deployed along a circular track with a fixed radius centered on the leaf, and the plane of the circular track is parallel to the leaf surface.

[0008] Optionally, the synchronous acquisition of the multiple polarized reflection images includes driving the polarization camera to move stepwise along a circular track at 15° angular intervals, briefly pausing at each angular position and triggering the following operations: a) Start the laser light source to emit continuous pulse laser; b) Synchronously capture reflection images in multiple polarization directions using the four focal plane polarization sensors of the polarization camera; c) Generate an original polarization image group at the current angular position.

[0009] Optionally, the generation of the pixel polarization degree matrix includes performing the following steps on the original polarization image group at each angular position: a) Calculate the Stokes parameter vector for each pixel ; b) Calculate the pixel polarization degree based on the Stokes parameter vector: ; c) All pixels in the leaf area The values ​​are constructed into an m×n matrix and the pixel polarization degree matrix is ​​output.

[0010] Optionally, the S2 specifically includes: S21, Polarization Transmission Modeling: The polarization degree maps collected at each angle 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 is constructed, including the interaction process of light attenuation and scattering. S22, microstructure parameter inversion: Initialize the refractive index distribution in the leaf cuticle as the simulation basis, perform simulation iterations angle by angle, and continuously adjust the refractive index distribution according to the image fitting error until the error meets the convergence condition or the maximum number of iterations is reached; S23, anomaly coefficient generation: When the inversion error reaches the set convergence condition or iteration upper limit, the optimization is terminated. According to the final refractive index distribution, the anomaly coefficient of each spatial position is calculated, and the area where the anomaly coefficient exceeds the preset anomaly threshold is marked as a highlighted area, and the output is a stratum corneum anomaly coefficient map.

[0011] Optionally, the simulation iteration specifically includes: extracting a measured polarimetric map of the current angle; emitting light rays of the corresponding direction in the refractive index simulation field; tracking the propagation path of the light rays in the medium and recording the polarization change thereof; comparing the simulation result with the measured map, and minimizing the deviation between them according to a preset target.

[0012] Optionally, the polarization transmission model is represented as: ; wherein, represents the radiation brightness along the direction at the position , is a total attenuation coefficient, is a scattering coefficient, is a polarization phase function, describing the scattering probability from the incident direction to the exit direction , represents the arc length along the propagation path of the light rays, represents the full space integral symbol.

[0013] Optionally, the highlight display area distribution mode in S3 is represented as a spatial distribution descriptor, a binary mask image of the highlight area is generated based on the highlight display area extracted from the cuticle anomaly coefficient map; the binary mask image is input into a pre-trained convolutional neural network, the center of gravity coordinates, the radial distribution histogram and the number of connected domains are extracted to constitute the spatial distribution descriptor; the spatial distribution descriptor is matched with each type of template in the preset template library of the infection type in terms of similarity, and the similarity score is evaluated according to the preset weight; when the similarity score exceeds a similarity threshold, the type corresponding to the template with the highest score is taken as the infection type recognition result of the current image.

[0014] Optionally, the infection intensity value calculation includes: constructing a spatial gradient field of the anomaly coefficient image of the highlight display area; extracting a direction gradient histogram information based on the spatial gradient field to form a diffusion mode vector for representing the lesion diffusion morphology; according to the determined infection type, selecting a corresponding intensity calculation method, and combining the area, gradient distribution and connected structure information to calculate the infection intensity value.

[0015] 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; matching the corresponding pest grade according to the intensity interval of different types.

[0016] The beneficial effects of the present application are: ​The present application, by means of multi-angle polarization image acquisition and Stokes parameter derivation, constructs a pixel-level polarization degree matrix, and combines a pulse laser synchronization mechanism and a four-way focal plane sensor, realizes high-fidelity acquisition of the optical microstructure of the surface layer of the leaf, targets the 532nm wavelength laser to enhance the reflection difference of the lesion area, combines dust compensation and dynamic focusing algorithm, and can stably extract the abnormal area under the condition of strong noise and complex background in the field, avoiding the problem that the lesion boundary is blurred and the structure is disturbed in the traditional RGB and multispectral images.

[0017] The present application introduces the light ray tracing algorithm and the polarization transmission model into the agricultural pathological image analysis, realizes the inversion of the physical parameters from the image polarization data to the microstructure by reconstructing the refractive index distribution field of the cuticle and constructing the abnormal coefficient graph, and introduces the gradient field, the direction entropy and the curvature compensation mechanism on this basis, which can accurately depict the boundary of the optical mutation area of the lesion area, not only improves the physical interpretation of the disease detection, but also provides a distinguishable feature space for different infection mechanisms (such as fungal growth layer and insect stinging point).

[0018] The present application proposes a type discrimination mechanism driven by spatial distribution descriptor, realizes AI feature coding of the spatial form of diseases and pests by modeling the gravity center of the highlighted area, the direction histogram and the connectivity, and simultaneously fuses the HOG feature, the gradient statistics and the infection intensity calculation model, adapts to the diffusion behavior characteristics of different disease types, and outputs real-time grade value and fine prevention suggestion. The closed-loop structure from image to microstructure to behavior mode to grade decision runs through perception, modeling and application, and forms the trinity of "physical quantization + identification + agricultural intervention". BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only a part of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0020] Fig. 1 It is a recognition grading method flow chart of the embodiment of the present application. Fig. 2 It is a cuticle refractive field reconstruction method schematic diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0021] The present application will be described in detail below in combination with the drawings and specific embodiments. For some known technologies, other alternative ways can also be adopted by those skilled in the art to implement them; and the drawings are only used to more specifically describe the embodiments, and are not intended to specifically limit the present application.

[0022] As Figs. 1-2 shown in the figure, the AI image recognition and grading method for field crop leaf diseases and pests includes the following steps: S1, under the irradiation of a fixed light source in the field, multiple polarized reflection images are synchronously collected at a predetermined angle interval around the crop leaf; the pixel polarization degree matrix of the leaf area in each polarized reflection image is extracted; S2, cuticle refraction field reconstruction: input the pixel polarization degree matrix into the polarization transmission model, output the cuticle anomaly coefficient map, mark the area in the cuticle anomaly coefficient map that exceeds the preset anomaly threshold as the highlighted area; the polarization transmission model inverses the cuticle microstructure through the ray tracing algorithm; S3, according to the distribution mode of the highlighted area in the anomaly coefficient map, match the infection type template library; combine the highlighted area diffusion gradient to calculate the infection intensity value, and output the disease and pest grade.

[0023] S1 specifically includes: S11, polarization image acquisition device deployment: set a fixed laser light source with a wavelength of 532 nm in the normal direction of the leaf to be measured, the incident direction of the light source forms a 45° angle with the leaf normal; deploy a polarization camera along a circular track with the leaf as the center and a radius of 30 cm, the plane of the track is parallel to the leaf surface.

[0024] S12: multi-angle image synchronous acquisition: drive the polarization camera to step along the track at an angle interval of 15°, and stop at each angle position for 200 ms. At each angle position, the following operations are performed synchronously: a) start the laser light source and emit pulsed laser for 10 ms; b) use the four focal plane polarization sensors built-in the polarization camera to synchronously capture the reflection images of four polarization directions of 0°, 45°, 90° and 135° respectively; c) merge the four images collected at this angle position into an original polarization image group.

[0025] S13, pixel polarization degree matrix generation: for each angle position of the original polarization image group, the following processing is performed: a) calculate the Stokes parameter vector of each pixel point: ; ; ; Wherein: represents the pixel gray value of the corresponding polarization direction; b) based on the Stokes parameter, calculate the polarization degree of each pixel point : ; c) The DoP values of all pixels in the leaf region are grouped to form a matrix, and the output is the pixel DoP matrix at this angle position; repeat the above acquisition and calculation process until the data acquisition and processing of 24 angle positions are completed. Further, the construction of the pixel DoP matrix is as follows: c) The DoP values of all pixels in the leaf region are grouped to form a matrix, and the output is the pixel DoP matrix at this angle position; repeat the above acquisition and calculation process until the data acquisition and processing of 24 angle positions are completed. Further, the construction of the pixel DoP matrix is as follows: I. Input data structure: At each angle position, four polarization direction images are obtained: , which are all gray-scale images with a size of . For example: , (the rest are the same) ; II. DoP calculation logic: for each pixel , the DoP value is calculated according to the Stokes parameter: ; ; ; and further obtain: ; III. Construction of the pixel DoP matrix: the of all pixels in the leaf region is arranged according to its image space to form a matrix: ; where is the pixel range of the extracted leaf region (ROI) on the image, which is not equal to the entire image size , but the actual effective region after mask cropping.

[0026] IV. Finally, the matrix is the pixel DoP matrix at this angle position, which is used as the input of the subsequent refraction field reconstruction model. The polarization camera generates a polarization matrix at angle position, and 24 angles are collected, each of which outputs a matrix .

[0027] The acquisition of four polarization direction reflection images is based on the minimum sampling requirement of the Stokes parameter model. The Stokes parameter is a set of standard parameters that describe the state of partially polarized light, defined as: : total light intensity; : ° and 90° polarization direction intensity difference; : 45° and 135° polarization direction intensity difference; : Describes circular polarization (often used in more complex systems, this scheme is not used).

[0028] In most agricultural or industrial vision systems, circular polarization (often used in more complex systems, this scheme is not used). ) is not easily accessible and unnecessary, so three parameters are sufficient to build a complete linear polarization information model, which requires the acquisition of four polarization direction images: 0°, 45°, 90°, 135°.

[0029] In order to accurately reconstruct the pixel-level degree of polarization DoP, according to the Stokes parameters, the degree of polarization (DoP) can be calculated: ; The infection of pests and diseases in leaves will change the microstructure of the cuticle, and then affect the polarization maintaining ability of light. The polarization degree of the reflected light is usually higher or disordered in the diseased area, while the polarization of the healthy tissue is more uniform and stable. The calculation of DoP can reveal the structural optical differences and is the key basis for the reconstruction of the refractive field. Compared with the direct use of grayscale images, the polarization degree image is more sensitive to the changes in the cuticle, which helps to amplify the differences between the diseased area and the background, form a high-light heterogeneous area, and facilitate template matching and grading. The polarization information is more stable relative to the incident angle, avoiding uneven illumination caused by leaf morphology, thereby improving the spatial robustness of the model, especially suitable for complex field environments.

[0030] S2 specifically comprises the following steps: S21, polarization transmission modeling: stack the pixel polarization degree matrix corresponding to the 24 angle positions in the order of the acquisition angle, and construct a three-dimensional polarization data body: ; Based on Mie scattering theory, a polarization transmission model (polarization control equation) of photon transmission in the cuticle is established: ; describes the attenuation and scattering behavior of light in the cuticle during propagation, and is the physical modeling basis for the entire refractive field reconstruction; wherein, represents the radiation intensity along the direction at position , is the total attenuation coefficient, is the scattering coefficient, is the polarization phase function, which describes the scattering probability from the incident direction to the exit direction , represents the arc length along the light propagation path, represents the full space integral symbol, which is integrated over all possible incident directions on the unit sphere, covering the entire spherical solid angle 4𝜋 steradians. The polarization phase function combines the angle factor and the wavelength response: ; wherein, is the scattering angle (i.e. , It represents the wavelength-dependent gain factor. The polarization response gain of the lesion tissue at 532 nm is about 2.3 times.

[0031] Basic optical parameters of healthy leaves: By consulting existing plant spectrum databases and agricultural optical data, the optical properties of the cuticle of healthy field crop leaves at a wavelength of 532nm were selected as a basic reference: The total attenuation coefficient of the stratum corneum is: ; The scattering coefficient is dominant and is approximately: ; The absorption coefficient is: This indicates that the main energy loss caused by the wax layer on the leaf epidermis during short-path propagation of laser light is scattering. Considering that fungal infection (such as powdery mildew and rust) can lead to structural disorder among stratum corneum cells and increased water content, the light scattering effect is significantly enhanced, and the measured polarization degree is increased. Therefore, enhanced optical parameter settings are used in the diseased area: the empirical enhancement ratio is set to 1.6-1.8 times that of healthy tissue: ; .

[0032] The optical properties of the stratum corneum at a wavelength of 532 nm are shown in the following table based on measured data.

[0033] Table 1 Measured optical properties of stratum corneum

[0034] The above data are the results of laboratory integrating sphere measurement and image inversion fitting. The field test sites cover three field crops: rice, rapeseed and corn.

[0035] S22, microstructure parameter inversion: Initialize the stratum corneum refractive index distribution field as: ; Use ray tracing algorithm to perform inversion iteration: S221, extracting the current angle from the three-dimensional polarization data volume The measured polarization degree diagram of: The three-dimensional polarization data volume is a data set formed by stacking the polarization matrix of pixels collected at different angles in the order of angles. , where each frame Corresponding to Acquisition angle ; The extraction method is as follows: According to the set acquisition angle sequence, directly press the from Index the corresponding frame; the frame is regarded as the current angle The measured polarization degree diagram under ; S222, in the current refractive index distribution field , a virtual light beam is emitted along the direction , simulating the propagation of light in a specific direction; S223, simulate the propagation and scattering of each light in the cuticle, calculate the corresponding theoretical polarization degree: , specifically, along the current angle direction A set of incident light rays are generated on the surface of the leaf, covering the entire pixel area; when each light 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 of each interaction are calculated; record all the direction and energy information of the light from the incident to the exit, and integrate the exit polarization state of multiple light rays at the same point to deduce the simulated polarization degree value of the corresponding pixel , the method is consistent with the measured polarization degree (using Stokes vector to deduce); after all the light ray simulation is completed, arrange the of each pixel point into a two-dimensional matrix with the same size as the measured image, which is the theoretical polarization degree map; S224, construct and minimize the cost function: ; minimize the error by iterative optimization , update the refractive index distribution field.

[0036] S23, abnormal coefficient generation: Stop iteration when the error meets any of the following conditions: condition one: ; condition two: reach the maximum number of iterations; Calculate the abnormal coefficient at each spatial position : ; Output the cuticle abnormal coefficient map, where: marked as a high-light abnormal area.

[0037] The basic refractive index of the cuticle of a healthy plant leaf is about According to existing optical research and measured data, the refractive index fluctuation caused by slight water evaporation and environmental fluctuations is usually ±0.05, and the tissue destruction, water accumulation or mycelium infection caused by diseases and pests often makes the refractive index rise to above 1.6, set , which means: , which is beyond the range of natural physiological fluctuations, ensuring that the detected area belongs to abnormal optical behavior, not physiological disturbance.

[0038] And in field tests, it is found that: If the threshold is lower than 0.20, it is easy to produce high-light artifacts in the vein and margin areas (due to natural curvature and thickness difference); If the threshold is higher than 0.30, part of the early lesion area cannot be identified, reducing the sensitivity.

[0039] The experimental results show that 0.25 is the optimal compromise threshold considering the lesion recognition rate and misidentification rate, especially suitable for diseases such as rust, aphids, and mold spots that cause local microstructure disturbance.

[0040] In the inversion process, after the error function E tends to be stable with iteration, the descending amplitude becomes smaller, but the calculation resource consumption increases significantly. Through experimental evaluation: When E<0.1, most abnormal features can be roughly located; Further optimization to E<0.05 can improve about 15-20% of the lesion boundary resolution; If forced to optimize to E<0.01, although the numerical error is lower, the number of iteration rounds is significantly increased, which is low in cost performance.

[0041] Therefore, 0.05 is taken as the error convergence threshold, achieving the engineering optimization between accuracy and efficiency.

[0042] S3 specifically includes the following steps: S31, infection mode matching: S311, based on the highlighted area satisfying the extraction from the cuticle anomaly coefficient map to form a binary mask: highlighted mask ; specifically including traversing all pixel positions in the entire image, marking any pixel point whose anomaly coefficient is greater than the threshold as "1", indicating an abnormal area; the rest is marked as "0", indicating the background; finally, a binary mask image with the same size as the original image is generated, which only contains "0" and "1" two pixel values, used to highlight the lesion area; the binary mask will be input into the CNN model as input for further spatial structure analysis and infection type recognition.

[0043] S312, input the binary mask into a pre-trained convolutional neural network (CNN) to extract spatial distribution descriptors , including: : center of gravity coordinates of the highlighted area (reflecting spatial offset); : radial distribution histogram (reflecting texture structure); : number of local connected domains (reflecting discrete degree); S313, for each preset infection type in the infection type template library, calculate the weighted cosine similarity score between the descriptor and the template vector : ; wherein, represents the weight vector of each component, Represents the cosine similarity function.

[0044] S314, will The type of template with the highest score As the result of infection type determination.

[0045] S32, infection intensity calculation: S321, mask the highlighted area Calculate the anomaly coefficient gradient field in the region: ; S322, extracting a histogram of oriented gradients (HOG) from the gradient field image and outputting a diffusion pattern vector; S323, according to the infection type , select the corresponding intensity calculation method for real-time intensity value Calculation: Fungal models (such as powdery mildew and rust): ; Pest models (such as aphids and plant lice): ; Bacterial models (such as soft rot and angular leaf spot): ; in, Indicates the area of ​​the highlighted region, Represents the total detection area of ​​the leaf, mean, std, and entropy are the gradient mean, standard deviation, and HOG image entropy respectively. Indicates the number of connected sub-regions in the highlighted area, It is the directional gradient histogram extracted above, which is the specific expression of the aforementioned “diffusion pattern vector”.

[0046] S33, level mapping output: according to the infection type and intensity value , find the corresponding level and prevention and control recommendations in the table below, and output the pest and disease level and prevention and control instructions: Table 2 Infection type and intensity value query classification mapping table

[0047] The preset infection type template library summarizes and matches different types of pests and diseases at the spatial feature level. By comparing the similarity between the descriptors extracted from the real-time image and the descriptor vectors of each type in the template library, the potential cause of the diseased area (fungal, insect, or bacterial infection) is determined. The template is constructed as follows: Step 1: Standard sample collection: Collect pest and disease image samples of various field crop leaves, covering the following infection types: Fungi (e.g. powdery mildew, rust), insects (e.g. aphids, planthoppers), bacteria (e.g. angular leaf spot, soft rot); at least 100 images with typical lesion characteristics are selected as the basic dataset for each type of infection Step 2: Abnormal area extraction and descriptor generation: Perform the preceding steps (abnormal coefficient map generation → highlight mask extraction) on each standard image, and extract the following three spatial distribution descriptors through CNN:

[0048] Take the average or statistical median of multiple samples of the same type to form a template descriptor group for that type.

[0049] Step 3: Each template type is represented by a triple, with the following structure: The storage format can be JSON or a vector database, facilitating fast retrieval and similarity calculation.

[0050] In the recognition process, the real-time extracted descriptor is calculated with the weighted cosine similarity of each template: Set the similarity threshold to 0.85, and the template with the highest similarity is the determination result of the current detection of the infection type.

[0051] The convolutional neural network (CNN) used in the present application is a lightweight feature extraction structure, which encodes the spatial distribution features in binary images. Its network structure includes the following: 1. Input layer: receives binary image masks with a standardized size of 64x64 pixels; 2. Multi-layer convolution module: uses multiple convolution kernels with different receptive fields (e.g. 3x3, 5x5) to extract local structure, edge direction, and concentration features; each layer is followed by a ReLU activation function and batch normalization; 3. Spatial attention module: emphasizes the target area deviating from the image center through a weighting mechanism, improving the sensitivity to lesion displacement; 4. Global pooling layer: performs maximum pooling and average pooling on the spatial dimension respectively, compressing the feature vector to a fixed length; 5. Output embedding layer: through three parallel output channels, the final network features are mapped and encoded as: Target barycenter coordinate vector (describing the spatial distribution of the highlight area center); Radial direction histogram vector (describing the distribution tendency of the bright area relative to the center); Connected domain structure statistics (describing the fragmentation or aggregation of the region).

[0052] Finally, the CNN outputs a combination vector composed of the above three types of features in a nested structure as a spatial distribution descriptor.

[0053] The mask image input and feature extraction process is as follows: Image preprocessing: The extracted highlight area binary mask is standardized to a uniform size. If the original image area is insufficient, edge padding or equal scaling processing can be performed. The input image only contains 0 and 1 values, representing the abnormal area and the background.

[0054] Feature automatic learning and coding: The network first extracts the basic texture features of the shape, size, direction, etc. of the highlight area in the local image; the middle layer strengthens the learning of the center of gravity shift or the concentrated shape of the lesion through the spatial focusing mechanism; the network at the end converts the feature map into three interpretable numerical structures, which are used to measure the center of gravity position, directional distribution, and connected domain structure.

[0055] Output spatial distribution descriptor: The three channels output by the network correspond to: The relative position coordinates of the current highlight area in the image; The direction mode statistical vector of the highlight area radiating outward; Statistical results of whether the abnormal area is discrete, strip or spot type distribution.

[0056] These outputs serve as inputs for the feature template matching module and participate in subsequent infestation type recognition and matching determination.

[0057] The network uses a large number of typical pest mask images labeled in the design stage for pre-training, and each type of training sample is associated with a real descriptor value labeled by artificial labeling as a training label; the loss function considers the center deviation, direction distribution difference, and connected domain number error, etc. dimensions to ensure that the model has stable extraction capability for the three types of features.

[0058] The present application covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present application. In order to make the public have a thorough understanding of the present application, specific details are described in the following preferred embodiments of the present application, and the present application can also be completely understood without the description of these details. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits, etc. are not described in detail.

[0059] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principle of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be regarded as the protection scope of the present application.

Claims

1. AI image recognition and classification method for field crop leaf pests and diseases, characterized by: The following steps are involved: S1, under the illumination of a fixed light source in the field, synchronously collect multiple polarized reflection images at predetermined angle intervals around the crop leaves; extract the pixel polarization degree matrix of the leaf area in each polarized reflection image; S2, stratum corneum refraction field reconstruction: inputting the pixel polarization degree matrix into the polarization transmission model, outputting a stratum corneum anomaly coefficient map, marking the areas in the stratum corneum anomaly coefficient map that exceed a preset anomaly threshold as highlighted areas; the polarization transmission model inverts the stratum corneum microstructure using a ray tracing algorithm; S3, matches the infection type template library according to the distribution pattern of the highlighted area in the anomaly coefficient map; calculates the infection intensity value based on the diffusion gradient of the highlighted area, and outputs the pest and disease grade.

2. The AI ​​image recognition and classification method for field crop leaf pests and diseases according to claim 1, characterized in that: Said 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 measured, with the incident direction of the light source forming a 45° angle with the leaf surface normal; The polarization camera is deployed along a circular track with a fixed radius centered on the leaf, and the plane of the circular track is parallel to the leaf surface.

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

4. The AI ​​image recognition and classification method for field crop leaf pests and diseases according to claim 3, characterized in that: The generation of the pixel polarization degree matrix includes performing the following steps on the original polarization image group at each angular position: a) Calculate the Stokes parameter vector for each pixel ; b) Calculate the pixel polarization degree based on the Stokes parameter vector: ; c) All pixels in the leaf area The values ​​are constructed into an m×n matrix and the pixel polarization degree matrix is ​​output.

5. The AI ​​image recognition and classification method for field crop leaf pests and diseases according to claim 1, characterized in that , the S2 specifically includes: S21, Polarization Transmission Modeling: The polarization degree maps collected at each angle 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 is constructed, including the interaction process of light attenuation and scattering. S22, microstructure parameter inversion: Initialize the refractive index distribution in the leaf cuticle as the simulation basis, perform simulation iterations angle by angle, and continuously adjust the refractive index distribution according to the image fitting error until the error meets the convergence condition or the maximum number of iterations is reached; S23, anomaly coefficient generation: When the inversion error reaches the set convergence condition or iteration upper limit, the optimization is terminated. According to the final refractive index distribution, the anomaly coefficient of each spatial position is calculated, and the area where the anomaly coefficient exceeds the preset anomaly threshold is marked as a highlighted area, and the output is a stratum corneum anomaly coefficient map.

6. The AI ​​image recognition and classification method for field crop leaf pests and diseases according to claim 5, characterized in that: The simulation iteration specifically includes: Extract the measured polarization degree map at the current angle; Launch light in the corresponding direction in the refractive index simulation field; Track the path of light in a medium and record its polarization changes; Compare the simulation results with the measured images and minimize the deviation between the two according to the preset goals.

7. The AI ​​image recognition and classification method for field crop leaf pests and diseases according to claim 5, characterized in that: The polarization transmission model is expressed as: ； in, Indicates location Along the direction The radiance, is the total attenuation coefficient, is the scattering coefficient, is the polarization phase function, describing the incident direction To the outgoing direction The scattering probability, represents the arc length along the light propagation path, Represents the full-space integral symbol.

8. The AI ​​image recognition and classification method for field crop leaf pests and diseases according to claim 7, characterized in that: The distribution pattern of the highlighted area in S3 is represented as a spatial distribution descriptor. Based on the highlighted 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, and the center of gravity coordinates, radial distribution histogram, and number of connected domains are extracted 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 a 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 used as the infection type recognition result of the current image.

9. The AI ​​image recognition and classification method for field crop leaf pests and diseases according to claim 8, characterized in that: The infection intensity value calculation includes: constructing a spatial gradient field of the abnormal coefficient image of the highlighted area; extracting directional gradient histogram information based on the spatial gradient field to form a diffusion pattern vector for characterizing the diffusion morphology of the lesion; selecting the corresponding intensity calculation method according to the determined infection type, and calculating the infection intensity value in combination with the regional area, gradient distribution, and connectivity structure information.

10. The AI ​​image recognition and classification method for field crop leaf pests and diseases according to claim 9, characterized in that: The S3 also includes a level mapping output, specifically including inputting the identified infection type and the corresponding infection intensity value into a grading mapping table; and matching the corresponding pest and disease level according to different types of intensity intervals.

Images