Artificial intelligence image recognition method and system for detecting plant diseases and insect pests of medicinal plants

Abstract

The invention discloses an artificial intelligence image recognition method and system for detecting plant diseases and insect pests of medicinal plants, and particularly relates to the technical field of plant disease and insect pest recognition. The method comprises the following steps: acquiring a leaf image, a stem image and a medicinal part image of a medicinal plant to obtain multi-part pre-processed image data; respectively extracting texture change features, color gradient features and scab edge morphological features of the leaf area and tissue texture detail features, epidermis damage mode features and hidden scab dark fringe features of the medicinal part area based on the multi-part pre-processed image data; integrating feature data of the leaf area and the medicinal part area by utilizing spatial structure mapping, and constructing cross-part associated feature data; analyzing the internal mapping relation between the visual symptoms of the leaves and the actual damage risk of the medicinal parts according to the cross-part associated feature data, generating the damage risk index of the medicinal parts, and judging the risk level of plant diseases and insect pests, so as to accurately detect the plant diseases and insect pests of the medicinal plants.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pest and disease identification, and more particularly to an artificial intelligence image recognition method and system for detecting diseases and pests of medicinal plants. BACKGROUND

[0002] Medicinal plants are prone to be attacked by diseases and pests during planting. In order to quickly and accurately find and evaluate the occurrence degree of diseases and pests of medicinal plants, it is usually only to identify and diagnose the visual symptoms of diseases and pests on the easily observed parts such as leaves of medicinal plants.

[0003] However, the prior art usually only relies on the visual symptoms of diseases and pests on the leaves or above-ground parts of medicinal plants for disease and pest identification and risk assessment, without considering the spatial non-co-location characteristics between the visual symptoms of diseases and pests on the leaves and other parts of medicinal plants and the actual damage degree of the medicinal parts of medicinal plants, resulting in deviation between the identification result and the real damage risk of the medicinal parts, and it is difficult to accurately reflect the actual quality risk of medicinal materials. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide an artificial intelligence image recognition method and system for detecting diseases and pests of medicinal plants to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0006] An artificial intelligence image recognition method for detecting diseases and pests of medicinal plants, comprising the following steps:

[0007] S1: collecting leaf images, stem images and medicinal part images of medicinal plant plants, and performing illumination correction processing and texture enhancement processing to generate multi-part preprocessed image data;

[0008] S2: based on the multi-part preprocessed image data, analyzing the texture changes, color gradient and lesion edge morphology of the leaf area of the medicinal plant, and constructing leaf phenotype feature data;

[0009] S3: based on the multi-part preprocessed image data, analyzing the tissue texture details, epidermis damage mode and latent lesion dark lines of the medicinal part area, and constructing medicinal part phenotype feature data;

[0010] S4: spatial structure mapping and integration of the leaf phenotype feature data and the medicinal part phenotype feature data to generate cross-part associated feature data;

[0011] S5: inputting the cross-part associated feature data into a pre-trained medicinal part damage risk prediction model, analyzing the mapping relationship between the leaf visual symptoms and the damage risk of the medicinal part, and generating a medicinal part damage risk index;

[0012] S6: According to the medicinal plant part damage risk index, the disease and pest risk level is determined by using an artificial intelligence classification model, and a disease and pest detection result is output.

[0013] In a preferred embodiment, S1, in particular:

[0014] Under standard light conditions, leaf images, stem images and medicinal plant part images of medicinal plants are collected;

[0015] The leaf images, stem images and medicinal plant part images of the medicinal plants are respectively subjected to light correction and texture enhancement processing;

[0016] Based on the leaf images, stem images and medicinal plant part images after light correction and texture enhancement processing, multi-part preprocessed image data is obtained.

[0017] In a preferred embodiment, S2, in particular:

[0018] Based on the multi-part preprocessed image data, the gray value change statistical characteristics of the pixel points in the leaf image are calculated, and texture change parameters are obtained;

[0019] Based on the multi-part preprocessed image data, the color value gradient distribution of the leaf image in the color space is calculated, and color gradient features are obtained;

[0020] Based on the multi-part preprocessed image data, the edge contour of the lesion area in the leaf image is detected, and the shape feature parameters of the edge contour are calculated, and lesion edge morphology features are obtained;

[0021] The texture change parameters, color gradient features and lesion edge morphology features are integrated to construct leaf phenotype feature data.

[0022] In a preferred embodiment, S3, in particular:

[0023] Based on the multi-part preprocessed image data, the tissue texture feature parameters of the pixel points in the medicinal plant part image are calculated, and tissue texture detail features are obtained;

[0024] Based on the multi-part preprocessed image data, the epidermis damage area in the medicinal plant part image is detected, and the morphology distribution features of the epidermis damage area are extracted, and epidermis damage mode features are obtained;

[0025] Based on the multi-part preprocessed image data, the contrast of the dark line area in the medicinal plant part image is enhanced, and the texture direction features of the dark line area are extracted, and hidden lesion dark line features are obtained;

[0026] The tissue texture detail features, epidermis damage mode features and hidden lesion dark line features are integrated to construct medicinal plant part phenotype feature data.

[0027] In a preferred embodiment, S4 specifically refers to:

[0028] Based on the physiological structural characteristics of medicinal plants, a spatial structural mapping relationship is established between leaf phenotypic characteristic data and medicinal part phenotypic characteristic data;

[0029] The texture variation parameters, color gradient features, and lesion edge morphology features in the leaf phenotypic feature data are integrated with the tissue texture detail features, epidermal damage pattern features, and latent lesion dark pattern features in the phenotypic feature data of medicinal parts.

[0030] By mapping spatial structure relationships, the integrated feature data is correlated to generate cross-part correlated feature data.

[0031] In a preferred embodiment, S5 specifically refers to:

[0032] The labeled leaf visual symptom level data and the corresponding actual damage level data of the medicinal parts were used as training data.

[0033] Establish a risk prediction model for damage to medicinal parts based on training data;

[0034] Cross-part correlation feature data is input into the medicinal part damage risk prediction model, regression calculation is performed from feature data to the actual degree of damage to the medicinal part, and the medicinal part damage risk index representing the degree of damage to the medicinal part is output.

[0035] In a preferred embodiment, S6 specifically refers to:

[0036] Establish a correspondence between the risk indicators of damage to medicinal parts and predefined pest and disease risk level labels;

[0037] Based on the labeled risk indicators of damage to medicinal parts and the risk levels of pests and diseases, train an artificial intelligence classification model;

[0038] The risk indicators of damage to medicinal parts are input into an artificial intelligence classification model. The model then performs risk level classification on the risk indicators of damage to medicinal parts and outputs the detection results of diseases and pests of medicinal plants.

[0039] On the other hand, the present invention provides an artificial intelligence image recognition system for detecting diseases and pests of medicinal plants, comprising:

[0040] Imaging processing module: Acquires images of leaves, stems, and medicinal parts of medicinal plants, and performs illumination correction and texture enhancement processing to generate pre-processed image data of multiple parts;

[0041] Leaf feature module: Based on preprocessed image data from multiple parts, the texture changes, color gradients and lesion edge morphology of the leaf area of ​​medicinal plants are analyzed to construct leaf phenotypic feature data;

[0042] Medicinal Part Feature Module: Based on preprocessed image data from multiple parts, the module analyzes the tissue texture details, epidermal damage patterns, and latent lesions and dark lines in the medicinal part area to construct phenotypic feature data of the medicinal part.

[0043] Spatial mapping module: Performs spatial structure mapping and integration of leaf phenotypic feature data and medicinal part phenotypic feature data to generate cross-part association feature data;

[0044] Risk assessment module: Input cross-part association feature data into a pre-trained medicinal part damage risk prediction model, analyze the mapping relationship between leaf visual symptoms and medicinal part damage risk, and generate medicinal part damage risk indicators.

[0045] Level determination module: Based on the risk indicators of damage to medicinal parts, the module uses an artificial intelligence classification model to determine the risk level of pests and diseases and outputs the detection results.

[0046] The technical effects and advantages of the artificial intelligence image recognition method and system for detecting diseases and pests of medicinal plants according to the present invention are as follows:

[0047] By applying unified illumination correction and texture enhancement to images of medicinal plant leaves, stems, and medicinal parts, the interference of ambient light fluctuations and differences in imaging conditions on the recognition process is reduced. By extracting texture variation parameters, color gradient features, and lesion edge morphology features, leaf phenotypic feature data is formed, enabling a refined representation of the visual symptoms of diseases and pests on medicinal plant leaves. By extracting tissue texture details, epidermal damage patterns, and latent lesion dark patterns, phenotypic feature data of medicinal parts is formed, accurately reflecting the potential damage status of medicinal parts to diseases and pests. Finally, leaf phenotypic features are established using the physiological structural characteristics of medicinal plants. The spatial structure mapping relationship between phenotypic feature data and phenotypic feature data of medicinal parts is used to generate cross-part association feature data, which reflects the structural connection between leaf symptoms and the state of medicinal parts. Based on the cross-part association feature data, a mapping relationship between leaf visual symptoms and the risk of damage to medicinal parts is constructed, and the risk index of damage to medicinal parts is output, transforming complex multi-part features into quantitative risk indicators that can be used for judgment. The risk level of pests and diseases is determined according to the risk index of damage to medicinal parts and the results of pest and disease detection are output, so that the results of pest and disease identification directly correspond to the risk of damage to medicinal parts, which is closer to the needs of medicinal material quality assessment and prevention and control decisions. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of an artificial intelligence image recognition method for detecting diseases and pests of medicinal plants according to the present invention;

[0049] Figure 2 This is a schematic diagram of the structure of an artificial intelligence image recognition system for detecting diseases and pests of medicinal plants according to the present invention. Detailed Implementation

[0050] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0051] Example 1

[0052] Figure 1 This invention provides an artificial intelligence image recognition method for detecting diseases and pests in medicinal plants, comprising the following steps:

[0053] S1: Collect images of leaves, stems, and medicinal parts of medicinal plants, and perform illumination correction and texture enhancement processing to generate multi-part preprocessed image data;

[0054] S2: Based on multi-site preprocessed image data, analyze the texture changes, color gradients and lesion edge morphology of the leaf area of ​​medicinal plants to construct leaf phenotypic feature data;

[0055] S3: Based on preprocessed image data from multiple sites, the tissue texture details, epidermal damage patterns, and latent lesion dark lines in the medicinal part area are analyzed to construct phenotypic feature data of the medicinal part.

[0056] S4: Spatial structure mapping and integration of leaf phenotypic feature data and medicinal part phenotypic feature data are performed to generate cross-part association feature data;

[0057] S5: Input cross-part association feature data into a pre-trained medicinal part damage risk prediction model, analyze the mapping relationship between leaf visual symptoms and medicinal part damage risk, and generate medicinal part damage risk index.

[0058] S6: Based on the risk indicators of damage to medicinal parts, use an artificial intelligence classification model to determine the risk level of pests and diseases and output the detection results of pests and diseases.

[0059] S1: Acquire images of leaves, stems, and medicinal parts of the medicinal plant, and perform illumination correction and texture enhancement processing to generate multi-part preprocessed image data, including:

[0060] Images of leaves, stems, and medicinal parts of medicinal plants were collected under standard lighting conditions.

[0061] Based on the cultivation characteristics of medicinal plants, standard sampling positions and angles for leaves, stems, and medicinal parts are determined to ensure that the acquired images of leaves, stems, and medicinal parts reflect the plant's growth status and pest and disease symptoms. The sampling positions are selected based on the vertical height stratification principle of the medicinal plant. This principle divides the plant's vertical height from the ground into several height intervals. For example, the plant can be divided into three standard height intervals: upper, middle, and lower. Each height interval corresponds to a fixed position for sampling leaf, stem, and medicinal part images. The standard sampling angle is the angle between the lens of the sampling device and the target area of ​​the medicinal plant, set perpendicular to the surface of the target area. This ensures that the acquired leaf, stem, and medicinal part images are free from significant distortion or deformation.

[0062] Standard lighting conditions refer to setting uniform and stable illumination conditions on the surface of the target area of ​​the medicinal plant to be photographed. Multiple light sources are arranged around the target area, and the power, emission angle, color temperature, and illuminance of each light source are uniformly set. The power parameter is determined by ensuring that the light intensity when the light emitted by the light source reaches the target area is within a reasonable range, such as 500 to 800 lux. The emission angle is determined by ensuring that the emission coverage angle of the light source is uniformly distributed around the target area, such as using an emission angle of 120 to 150 degrees. The color temperature is determined by selecting a color temperature close to natural sunlight, such as 5500K to 6500K. The illuminance is determined by measuring the uniformity of illumination within the target area using an illuminance sensor, ensuring that the illuminance difference at any measurement point within the target area does not exceed 10%. Under standard lighting conditions, high-resolution digital image acquisition equipment was used to acquire images of leaves, stems, and medicinal parts of the medicinal plant. Each target area was captured at least five times to ensure the stability and consistency of the acquired leaf, stem, and medicinal part image data. Leaf image acquisition focused on healthy leaves and suspected lesion areas in the middle of the plant; stem image acquisition covered the main stem and branch nodes; and medicinal part image acquisition targeted the roots, fruits, or specific medicinal organs.

[0063] Illumination correction and texture enhancement were performed on images of leaves, stems, and medicinal parts of medicinal plants, respectively.

[0064] The illumination correction process combines a white balance algorithm based on the gray-world assumption with histogram matching. The original RAW image is converted to a linear RGB color space, and the average RGB value of the entire image is calculated. Based on the gray-world assumption, the gain coefficient of each color channel is adjusted to make the overall image tone more neutral. The gain coefficient is calculated as follows: using the green channel as a reference, the gain coefficient of the red channel equals the average value of the green channel divided by the average value of the red channel, and the gain coefficient of the blue channel equals the average value of the green channel divided by the average value of the blue channel. Histogram matching is applied to align the corrected image with a pre-stored standard illumination reference image. The standard illumination reference image is a grayscale card image taken under the same standard illumination conditions, with an RGB value of 128. Histogram matching includes calculating the cumulative distribution function of the corrected image and mapping it to the cumulative distribution function of the standard illumination reference image. An interpolation algorithm is used to adjust the intensity value of each pixel to ensure a consistent brightness distribution in the image.

[0065] Texture enhancement processing employs a multi-scale Gaussian filtering and Laplacian pyramid fusion. The input image is converted to grayscale, and a weighted average method is used to calculate the grayscale values, with weight coefficients of 0.299 for the red channel, 0.587 for the green channel, and 0.114 for the blue channel. A Gaussian pyramid is constructed with four layers. Each layer is downsampled using a 5x5 Gaussian convolution kernel with a standard deviation of 1.0. For each layer of the Gaussian pyramid, a Laplacian pyramid is calculated by subtracting the upsampled result of the previous layer's Gaussian image from the current layer's Gaussian image, using bilinear interpolation. Finally, the images from each layer of the Laplacian pyramid are weighted and fused, with weight coefficients allocated based on the layer number: 0.1 for the bottom layer, 0.2 for the second layer, 0.3 for the third layer, and 0.4 for the top layer. The fused image is then reconstructed into an enhanced texture image through inverse transformation. Enhance leaf vein texture in leaf images, strengthen epidermal cracks in stem images, and improve the visibility of latent lesions in medicinal part images.

[0066] Based on the leaf images, stem images, and medicinal part images after illumination correction and texture enhancement, multi-part preprocessed image data is obtained;

[0067] All image data of medicinal plant leaves, stems, and medicinal parts, after undergoing illumination correction and texture enhancement, are compiled and sequentially numbered to obtain multi-part preprocessed image data. For example, each processed image is sequentially numbered according to its planting batch number, collection time number, and image data type to distinguish and uniquely identify it.

[0068] S2, based on preprocessed image data from multiple locations, analyzes the texture changes, color gradients, and lesion edge morphology of the leaf region of medicinal plants to construct leaf phenotypic feature data, including:

[0069] Based on preprocessed image data from multiple parts, statistical features of grayscale value changes of pixels in leaf images are calculated to obtain texture change parameters.

[0070] The image data of medicinal plant leaves was converted from RGB color space to grayscale space to obtain grayscale image data of the medicinal plant leaf region. Statistical calculations were performed on the grayscale value changes of each pixel in the grayscale image data, including calculating the mean, standard deviation, variability, and local root mean square error of the grayscale values ​​of the pixels in the leaf region within a 3×3 pixel neighborhood window. The mean, standard deviation, variability, and local root mean square error of the statistically calculated grayscale values ​​were normalized. Finally, the mean, standard deviation, variability, and local root mean square error of the normalized grayscale values ​​were defined as texture variation parameters to represent the texture detail features of the surface of the medicinal plant leaf region.

[0071] Based on preprocessed image data from multiple parts, the color value gradient distribution of the leaf image in the color space is calculated to obtain color gradient features;

[0072] The image data of medicinal plant leaves is converted from RGB color space to HSV color space to obtain the color component image data of the medicinal plant leaf area in HSV color space. The gradient of each pixel in the H hue channel and S saturation channel in HSV color space is calculated in both horizontal and vertical directions. The calculation method includes, but is not limited to, using the Sobel gradient operator to perform convolution operations on the H hue channel and S saturation channel respectively, with the convolution kernel size selected as 3×3 or 5×5. Based on the calculated gradient of each pixel, the magnitude distribution and direction distribution of the gradient are statistically analyzed and statistically analyzed using histograms. The number of distribution intervals of the histogram is set to, but is not limited to, 16 to 32 intervals. Finally, the statistical histogram of gradient magnitude and direction distribution is defined as the color gradient feature, which is used to represent the subtlety and directionality of color changes in the medicinal plant leaf area.

[0073] Based on preprocessed image data from multiple parts, the edge contour of the lesion area in the leaf image is detected, and the shape feature parameters of the edge contour are calculated to obtain the morphological features of the lesion edge.

[0074] Image binarization segmentation methods are used to extract lesion regions from medicinal plant leaf image data. Segmentation methods include, but are not limited to, Otsu thresholding or adaptive thresholding. After obtaining the binarized image data of the lesion regions, the Canny edge detection method is used to determine the edge contours of the lesion regions. The high and low threshold parameters of the Canny edge detection method are set as follows: First, a histogram of gray values ​​of all edge pixels in the binary image data of the lesion regions is plotted. Based on the distribution pattern of gray values ​​in the gray value histogram, high and low thresholds are selected. For example, the high threshold is set to the gray value with a cumulative frequency of 95% in the gray value histogram. The low threshold is set to 50% of the high threshold. After determining the edge contour of the lesion region, the shape feature parameters of each lesion region edge contour are calculated, including but not limited to the area, perimeter, compactness (compactness is defined as 4π * area divided by the square of perimeter) and concavity / convexity of the edge contour (concavity / convexity is defined as the ratio of the actual contour area of ​​the lesion region to the area of ​​the circumscribed convex hull of the lesion region). The shape feature parameters of each lesion region are standardized. Finally, the shape feature parameters of the standardized lesion edge contour are defined as the lesion edge morphology features, which are used to represent the appearance geometric characteristics of the lesion region.

[0075] By integrating texture variation parameters, color gradient features, and lesion edge morphology features, leaf phenotypic feature data are constructed.

[0076] Texture variation parameters, color gradient features, and lesion edge morphology features are converted into feature vectors of a unified dimension, forming leaf region texture variation feature vector, leaf region color gradient feature vector, and leaf region lesion edge morphology feature vector. These feature vectors are then concatenated and combined to form a high-dimensional leaf phenotypic feature vector. To eliminate differences in numerical ranges between different feature vectors, the leaf phenotypic feature vector is normalized. Finally, the normalized leaf phenotypic feature vector is defined as leaf phenotypic feature data, used to characterize the visual symptoms of leaf diseases and pests.

[0077] S3, based on preprocessed image data from multiple sites, analyzes the tissue texture details, epidermal damage patterns, and latent lesion dark lines in the medicinal part areas to construct phenotypic feature data of the medicinal parts, including:

[0078] Based on preprocessed image data from multiple parts, the tissue texture feature parameters of pixels in the medicinal part image are calculated to obtain the tissue texture detail features.

[0079] Grayscale conversion processing is performed on the image data of medicinal parts of medicinal plants to convert color image data into single-channel grayscale image data. Taking each pixel in the grayscale image data as the center, the contrast, homogeneity, correlation, and energy parameters of the texture in the neighborhood of each pixel are calculated using the gray-level co-occurrence matrix method. The neighborhood size of the gray-level co-occurrence matrix is ​​set to a range including but not limited to 5×5 pixels to 7×7 pixels, and the grayscale level is set to a uniform 64 grayscale levels during the calculation. The obtained contrast, homogeneity, correlation, and energy parameters are linearly normalized. Finally, the normalized texture parameters are combined to form tissue texture detail features to represent the tissue texture detail information of the medicinal part region of the medicinal plant.

[0080] Based on preprocessed image data from multiple sites, the epidermal damage area in the image of medicinal part is detected, and the morphological distribution features of the epidermal damage area are extracted to obtain epidermal damage pattern features.

[0081] Based on image data of medicinal parts of medicinal plants, the epidermal damage area is extracted using image binarization segmentation methods. These methods include, but are not limited to, the maximum inter-class variance thresholding method or the adaptive local thresholding method. The threshold is set by calculating the grayscale threshold that maximizes the inter-class variance in the overall grayscale histogram of the image data. Each independent connected region is marked in the binarized image data, and area, aspect ratio, and roundness parameters are calculated for each epidermal damage area. Finally, the location coordinates of each damage area are recorded to identify its distribution within the medicinal part region of the medicinal plant. The area, aspect ratio, roundness, and spatial location parameters of all damage areas are combined to form an epidermal damage pattern feature, representing the epidermal damage status of the medicinal part region of the medicinal plant.

[0082] Based on multi-site preprocessed image data, the contrast of dark texture areas in the images of medicinal parts is enhanced, and the texture direction features of the dark texture areas are extracted to obtain the features of latent lesions and dark textures.

[0083] Based on image data of medicinal parts of medicinal plants, a high-frequency enhancement filter is used to improve the local contrast of dark texture areas. The kernel size of the filter is set according to factors including, but not limited to, image resolution and texture scale. For example, setting the kernel size to a 3×3 to 5×5 matrix makes the grayscale difference between the dark texture area and the background area more obvious in the enhanced image data. The texture direction of the enhanced dark texture area is extracted using a directional filter bank method. The directional filter bank method includes, but is not limited to, using Gabor filter banks for convolution processing. The number of directions of the Gabor filter bank includes, but is not limited to, 4 or 8 directions, with the directions evenly spaced between 0 degrees and 180 degrees. Finally, by statistically analyzing the frequency distribution of the direction with the maximum energy in each directional filter response value, the dominant texture direction feature of the dark texture area is formed to represent the latent lesion texture distribution feature of the medicinal parts of medicinal plants.

[0084] By integrating the detailed features of tissue texture, the features of epidermal damage patterns, and the features of latent lesions and dark lines, phenotypic feature data of medicinal parts are constructed.

[0085] Tissue texture detail features, epidermal damage pattern features, and latent lesion dark pattern features are represented as numerical feature vectors, forming tissue texture detail feature vectors, epidermal damage pattern feature vectors, and latent lesion dark pattern feature vectors of medicinal parts. These vectors are then combined in a predetermined order to form a unified high-dimensional feature vector of medicinal part phenotype. The combined high-dimensional feature vector of medicinal part phenotype is then standardized. Finally, the standardized high-dimensional feature vector of medicinal part phenotype is defined as the phenotypic feature data of medicinal part to represent the potential damage status of medicinal parts of medicinal plants.

[0086] S4, spatial structure mapping and integration of leaf phenotypic feature data and medicinal part phenotypic feature data to generate cross-part association feature data, including:

[0087] Based on the physiological structural characteristics of medicinal plants, a spatial structural mapping relationship is established between leaf phenotypic characteristic data and medicinal part phenotypic characteristic data;

[0088] Using the main stem of the medicinal plant as a spatial reference axis, the three-dimensional position parameters of the leaf region and the medicinal part region in the plant's spatial coordinate system are recorded. The methods for determining these three-dimensional position parameters include, but are not limited to, using a three-dimensional spatial coordinate measurement method. Specifically, the center point of the main stem at the bottom of the medicinal plant is taken as the origin of the spatial coordinates, the vertically upward direction of the main stem is taken as the longitudinal axis, and the plane perpendicular to the longitudinal axis corresponding to the main stem cross-section is taken as the horizontal reference plane. The height coordinates of the leaf region and the medicinal part region along the longitudinal axis, as well as their radial distance coordinates and circumferential angle coordinates on the horizontal reference plane, are determined. The three-dimensional position parameters are standardized and recorded in numerical vector form. The phenotypic feature data of the medicinal plant leaves and the phenotypic feature data of the medicinal parts are spatially matched according to the three-dimensional position parameters to determine the mapping relationship between the phenotypic feature data of the medicinal parts corresponding to each leaf phenotypic feature data, ultimately forming a spatial structure mapping relationship.

[0089] The texture variation parameters, color gradient features, and lesion edge morphology features in the leaf phenotypic feature data are integrated with the tissue texture detail features, epidermal damage pattern features, and latent lesion dark pattern features in the phenotypic feature data of medicinal parts.

[0090] Based on the spatial structure mapping data, texture variation parameters, color gradient features, and lesion edge morphology features corresponding to the same spatial location parameter in the leaf phenotypic feature data are combined to form a local leaf phenotypic feature subset; tissue texture detail features, epidermal damage pattern features, and latent lesion dark pattern features corresponding to the same spatial location parameter in the medicinal part phenotypic feature data are combined to form a local medicinal part phenotypic feature subset; the local leaf phenotypic feature subsets are matched one by one with the local medicinal part phenotypic feature subsets at the corresponding spatial locations to ensure that each set of leaf phenotypic feature data has a one-to-one correspondence with the medicinal part phenotypic feature data at a specific spatial location; each successfully matched set of local leaf phenotypic feature subsets and local medicinal part phenotypic feature subsets are merged to form a cross-part joint feature subset; the cross-part joint feature subset is standardized using a feature normalization method.

[0091] The integrated feature data is correlated through spatial structure mapping to generate cross-part correlated feature data.

[0092] The integrated and standardized cross-part joint feature subset is uniformly represented as a cross-part joint feature vector. Based on the cross-part joint feature vector, the correlation coefficient between each feature component in the feature vector is calculated. The correlation coefficient is calculated using the Pearson correlation coefficient: first, the covariance value between each feature component is calculated, and then the covariance value is divided by the product of the standard deviations of each feature component to obtain the Pearson correlation coefficient between the feature components. According to the correlation coefficient between the feature components, a cross-part correlation coefficient matrix is ​​constructed. Each row of the correlation coefficient matrix represents a feature component of a leaf region, and each column represents the feature component of the corresponding medicinal part region. The matrix elements range from -1 to 1. Feature combinations with strong feature correlation are screened through the correlation coefficient matrix. The screening criterion is that feature combinations with an absolute value of correlation coefficient exceeding 0.75 are defined as effective cross-part feature combinations. Finally, the screened effective cross-part feature combinations are defined as cross-part correlation feature data in the form of numerical feature combinations to represent the structural correlation between the phenotypic feature data of the leaf region of medicinal plants and the phenotypic feature data of the medicinal part region of medicinal plants.

[0093] S5, input cross-part association feature data into a pre-trained medicinal part damage risk prediction model, analyze the mapping relationship between leaf visual symptoms and medicinal part damage risk, and generate medicinal part damage risk indicators, including:

[0094] The labeled leaf visual symptom level data and the corresponding actual damage level data of the medicinal parts were used as training data.

[0095] Leaf visual symptom grading data is the result of technical personnel with pest and disease diagnosis qualifications classifying and labeling leaf lesions based on actual observations of their symptoms using visual diagnostic methods. The method for determining visual symptom grading is the lesion area percentage assessment method, which calculates the percentage of the total area of ​​a single leaf of the medicinal plant occupied by the lesion and grading it according to predetermined rules. For example, lesion area percentage less than 5% is grade 1, 5% to 15% is grade 2, 15% to 30% is grade 3, and greater than 30% is grade 4. The labeling method for the actual damage level data of medicinal parts involves collecting samples from medicinal parts at the actual harvest of the medicinal plant and performing laboratory physicochemical tests to measure the degree of decline in the content of active ingredients and the degree of tissue damage. Methods for detecting active ingredients include, for example, high-performance liquid chromatography (HPLC) for content determination; methods for detecting the degree of tissue damage include, for example, observing the structural integrity of tissue sections under a microscope and calculating the area percentage of tissue damage. Based on the experimental test results, the actual damage level of the medicinal parts is then classified and labeled. For example, Level 1 is defined as an active ingredient content decline of no more than 10% and a tissue damage area percentage of less than 5%; Level 2 is an active ingredient content decline of 10% to 20% and a tissue damage area percentage of 5% to 15%; Level 3 is an active ingredient content decline of 20% to 30% and a tissue damage area percentage of 15% to 25%; and Level 4 is an active ingredient content decline of more than 30% and a tissue damage area percentage of more than 25%. After labeling, the leaf visual symptom level data and the corresponding actual damage level data of the medicinal parts are integrated to form a training dataset.

[0096] Establish a risk prediction model for damage to medicinal parts based on training data;

[0097] The risk prediction model for damage to medicinal parts employs a convolutional neural network (CNN) regression model. This model consists of an input layer, multiple convolutional layers, multiple pooling layers, a fully connected layer, and a regression output layer. The input layer contains normalized cross-part correlation feature data. The number of convolutional layers is determined based on cross-validation methods; for example, three convolutional layers are used, each with a different number of kernels. The kernel size is set to 3×3 or 5×5, and the number of kernels is set to 32, 64, and 128, respectively. The pooling layers use max pooling to compress the spatial feature dimensions. Each pooling layer is set to half the size of the output feature map of the convolutional layer; for example, if the convolutional layer feature map size is 64×64, the corresponding pooling layer feature map size is 32×32. The number of neurons in the fully connected layer ranges from 128 to 256. A linear activation function is used between the fully connected layer and the regression output layer for regression calculation. The training method for the convolutional neural network regression model employs supervised learning. During training, the mean squared error is selected as the loss function, calculated by summing the squared differences between the actual labeled data and the predicted data. The optimization method uses an adaptive moment estimation algorithm for weight optimization and updates. The learning rate setting method for this algorithm includes, but is not limited to, dynamic learning rate adjustment strategies. For example, the initial learning rate is set to 0.001, and the learning rate is gradually adjusted based on the rate of loss reduction. If the loss value decreases by less than 1% for five consecutive periods during model training, the learning rate is adjusted to 50% of the original learning rate. After training and optimization, when the model loss function converges and stabilizes within a preset threshold (e.g., below 0.01), the model training is complete, forming a risk prediction model for damage to medicinal parts.

[0098] Input cross-part association feature data into the medicinal part damage risk prediction model, perform regression calculation from feature data to the actual damage degree of medicinal parts, and output a medicinal part damage risk index representing the damage degree of medicinal parts.

[0099] Cross-site correlation feature data is input into the medicinal part damage risk prediction model. Before inputting into the model, the cross-site correlation feature data is normalized, for example, using a max-min normalization method to ensure that each feature component maps to between 0 and 1. The normalized cross-site correlation feature data is then input into the model, which performs regression calculations. Specifically, the cross-site correlation feature data undergoes feature extraction via convolutional layers, dimensionality reduction via pooling layers, and feature combination via fully connected layers. The result is then used to predict the actual degree of damage to the medicinal part through a regression output layer. The predicted value representing the actual degree of damage to the medicinal part is used as the medicinal part damage risk index.

[0100] S6, based on the risk indicators of damage to medicinal parts, uses an artificial intelligence classification model to determine the risk level of pests and diseases, and outputs the pest and disease detection results, including:

[0101] Establish a correspondence between the risk indicators of damage to medicinal parts and predefined pest and disease risk level labels;

[0102] The risk index for damage to medicinal parts is a predicted value output by a risk prediction model for damage to medicinal parts. The pest and disease risk level labels are several pre-defined levels based on the practical needs of medicinal plant production. The definition method for the pest and disease risk level labels is based on the degree of impact of the actual damage to the medicinal parts on the economic value and quality of the medicinal material. For example, they are divided into four risk levels: Risk Level 1 indicates that the medicinal plant is basically unaffected by pests and diseases, and the quality of the medicinal material is close to or reaches the normal standard; Risk Level 2 indicates that the medicinal plant is slightly affected by pests and diseases, the content of active ingredients in the medicinal material decreases slightly, and the economic value is slightly affected; Risk Level 3 indicates that the medicinal plant is moderately affected by pests and diseases, the content of active ingredients in the medicinal material decreases significantly, and the economic value of the medicinal material decreases significantly; Risk Level 4 indicates that the medicinal plant is severely affected by pests and diseases, the content of active ingredients in the medicinal material decreases drastically, and the material cannot meet the usage standards. The risk index of damage to medicinal parts is matched one by one with the risk level label of pests and diseases. For example, when the risk index of damage to medicinal parts is between 0 and 0.25, it corresponds to risk level 1; when the risk index of damage to medicinal parts is between 0.25 and 0.5, it corresponds to risk level 2; when the risk index of damage to medicinal parts is between 0.5 and 0.75, it corresponds to risk level 3; and when the risk index of damage to medicinal parts is between 0.75 and 1, it corresponds to risk level 4.

[0103] Based on the labeled risk indicators of damage to medicinal parts and the risk levels of pests and diseases, train an artificial intelligence classification model;

[0104] An artificial intelligence classification model was trained using labeled risk indicators for damage to medicinal parts and risk levels of pests and diseases. The AI ​​classification model employed a gradient boosting decision tree model. The gradient boosting decision tree model is an ensemble learning algorithm based on decision trees, which iteratively trains multiple weak classifiers (decision trees) and accumulates the prediction results to achieve high-precision classification.

[0105] The training method is as follows: Using labeled risk indicators of damage to medicinal parts as input features and pest and disease risk level labels as classification targets, an initial decision tree model is constructed. The structural parameters of the initial decision tree model include the maximum tree depth, the minimum number of samples per tree node, and the maximum number of leaf nodes. The maximum tree depth is set by evaluating the performance of the training data at different depth settings through cross-validation; for example, the maximum depth is set between 3 and 6. The minimum number of samples per tree node is set to 1% to 5% of the total number of training data samples. Based on the prediction error of the initial decision tree model, a new decision tree model is constructed using gradient descent, and the residuals are modeled. The learning rate determination method for gradient descent includes, but is not limited to, setting the initial value to 0.1 to 0.3, and based on the model... Fine-tuning of the training effect is performed; for example, when the classification accuracy no longer improves, the learning rate can be reduced to 50% of the current value. This process is repeated iteratively to obtain multiple weak classifiers for the decision tree. The total number of weak classifiers is determined by continuously increasing the number of decision trees until the classification accuracy on the validation set stabilizes; for example, the number of weak classifiers is determined to be between 50 and 150. Finally, the outputs of all weak classifiers are combined in a weighted manner to obtain the final classification result. The weight of each weak classifier is determined based on its error performance during training; weak classifiers with lower error performance have higher weight coefficients. The weight coefficients are determined by using a negative gradient method for adaptive weight allocation to minimize the overall prediction error, thus completing the training of the gradient boosting decision tree model.

[0106] The risk indicators of damage to medicinal parts are input into an artificial intelligence classification model. The artificial intelligence classification model performs risk level classification on the risk indicators of damage to medicinal parts and outputs the detection results of diseases and pests of medicinal plants.

[0107] The risk index of damage to medicinal parts is input into the trained gradient boosting decision tree model. Each weak classifier of the decision tree predicts the category of the risk index of damage to medicinal parts. The output of each weak classifier is a level in the risk level label of medicinal plant diseases and pests. Finally, the gradient boosting decision tree model sums the prediction results of all weak classifiers of the decision trees with weights and outputs the final risk level classification result according to the category with the highest cumulative weight.

[0108] The pest and disease risk level classification results output by the gradient boosting decision tree model represent the pest and disease detection results for medicinal plants. For example, a pest and disease risk level of 3 indicates the current pest and disease risk status of the medicinal plant. The model also outputs the cumulative predicted probability or weight value corresponding to the risk level; for example, a pest and disease risk level of 3 with a cumulative weight value of 0.68 indicates the reliability of the classification result. These output pest and disease detection results are used to guide the development of actual pest and disease control strategies for medicinal plants.

[0109] Example 2

[0110] The difference between Embodiment 2 and Embodiment 1 is that this embodiment introduces an artificial intelligence image recognition system for detecting diseases and pests of medicinal plants.

[0111] Figure 2 A schematic diagram of the structure of an artificial intelligence image recognition system for detecting diseases and pests of medicinal plants is provided. The artificial intelligence image recognition system for detecting diseases and pests of medicinal plants includes:

[0112] Imaging processing module: Acquires images of leaves, stems, and medicinal parts of medicinal plants, and performs illumination correction and texture enhancement processing to generate pre-processed image data of multiple parts;

[0113] Leaf feature module: Based on preprocessed image data from multiple parts, the texture changes, color gradients and lesion edge morphology of the leaf area of ​​medicinal plants are analyzed to construct leaf phenotypic feature data;

[0114] Medicinal Part Feature Module: Based on preprocessed image data from multiple parts, the module analyzes the tissue texture details, epidermal damage patterns, and latent lesions and dark lines in the medicinal part area to construct phenotypic feature data of the medicinal part.

[0115] Spatial mapping module: Performs spatial structure mapping and integration of leaf phenotypic feature data and medicinal part phenotypic feature data to generate cross-part association feature data;

[0116] Risk assessment module: Input cross-part association feature data into a pre-trained medicinal part damage risk prediction model, analyze the mapping relationship between leaf visual symptoms and medicinal part damage risk, and generate medicinal part damage risk indicators.

[0117] Level determination module: Based on the risk indicators of damage to medicinal parts, the module uses an artificial intelligence classification model to determine the risk level of pests and diseases and outputs the detection results.

[0118] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0119] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0120] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0121] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0122] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0123] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0124] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0125] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0126] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An artificial intelligence image recognition method for detecting pests and diseases of medicinal plants, characterized by, The method comprises the following steps: S1: Collecting leaf images, stem images and medicinal part images of medicinal plant plants, and performing illumination correction processing and texture enhancement processing to generate multi-part preprocessed image data; S2: Based on the multi-part preprocessed image data, the texture change, color gradient and lesion edge morphology of the leaf area of the medicinal plant are analyzed, and the leaf phenotype feature data is constructed; S3: Based on the multi-part preprocessed image data, the tissue texture details, epidermis damage mode and latent lesion dark lines of the medicinal part area are analyzed, and the medicinal part phenotype feature data is constructed; S4: The leaf phenotype feature data and the medicinal part phenotype feature data are integrated by spatial structure mapping to generate cross-part correlation feature data; S5: The cross-part correlation feature data is input into a pre-trained medicinal part damage risk prediction model to analyze the mapping relationship between the leaf visual symptoms and the medicinal part damage risk, and a medicinal part damage risk index is generated; S6: According to the medicinal part damage risk index, a disease and pest risk grade determination is performed by using an artificial intelligence classification model, and a disease and pest detection result is output.

2. The artificial intelligence image recognition method for detecting pests and diseases of medicinal plants according to claim 1, characterized in that, S1, specifically: Collecting leaf images, stem images and medicinal part images of medicinal plants under standard illumination conditions; Performing illumination correction and texture enhancement processing on the leaf images, stem images and medicinal part images of the medicinal plants respectively; Based on the leaf images, stem images and medicinal part images after illumination correction and texture enhancement processing, multi-part preprocessed image data is obtained.

3. The artificial intelligence image recognition method for detecting pests and diseases of medicinal plants according to claim 2, characterized in that, S2, specifically: Based on the multi-part preprocessed image data, the gray value change statistical characteristics of the pixel points in the leaf image are calculated to obtain the texture change parameters; Based on the multi-part preprocessed image data, the color value gradient distribution of the leaf image in the color space is calculated to obtain the color gradient characteristics; Based on the multi-part preprocessed image data, the edge contour of the lesion area in the leaf image is detected, and the shape feature parameters of the edge contour are calculated to obtain the lesion edge morphology characteristics; Integrate the texture change parameters, color gradient characteristics and lesion edge morphology characteristics to construct the leaf phenotype feature data.

4. The artificial intelligence image recognition method for detecting pests and diseases of medicinal plants according to claim 3, characterized in that, S3, specifically: Based on the multi-part preprocessed image data, the tissue texture feature parameters of the pixel points in the medicinal part image are calculated to obtain the tissue texture detail characteristics; Based on the multi-part preprocessed image data, the epidermis damage area in the medicinal part image is detected, and the morphological distribution characteristics of the epidermis damage area are extracted to obtain the epidermis damage mode characteristics; Based on the multi-part preprocessed image data, the contrast of the dark line area in the medicinal part image is enhanced, and the texture direction characteristics of the dark line area are extracted to obtain the latent lesion dark line characteristics; Integrate the tissue texture detail characteristics, epidermis damage mode characteristics and latent lesion dark line characteristics to construct the medicinal part phenotype feature data.

5. The artificial intelligence image recognition method for detecting pests and diseases of medicinal plants according to claim 4, characterized in that, S4, specifically: Based on the physiological structure characteristics of the medicinal plant, the spatial structure mapping relationship between the leaf phenotype feature data and the medicinal part phenotype feature data is established; Integrating the texture change parameters, color gradient features and lesion edge morphology features in the leaf phenotype feature data with the tissue texture detail features, epidermis damage pattern features and latent lesion dark line features in the medicinal part phenotype feature data; Correlating the integrated feature data through the spatial structure mapping relationship to generate cross-part associated feature data.

6. The artificial intelligence image recognition method for detecting pests and diseases of medicinal plants according to claim 5, characterized in that, S5, specifically: Using the labeled leaf visual symptom grade data and the corresponding actual damage degree data of the medicinal part as training data; Establishing a medicinal part damage risk prediction model based on the training data; Inputting the cross-part associated feature data into the medicinal part damage risk prediction model to perform regression calculation of the feature data to the actual damage degree of the medicinal part, and outputting a medicinal part damage risk index representing the damage degree of the medicinal part.

7. The artificial intelligence image recognition method for detecting pests and diseases of medicinal plants according to claim 6, characterized in that, S6, specifically: Establishing a corresponding relationship between the medicinal part damage risk index and the pre-defined pest and disease risk level label; Training an artificial intelligence classification model based on the labeled medicinal part damage risk index and pest and disease risk level; Inputting the medicinal part damage risk index into the artificial intelligence classification model, performing risk level classification processing on the medicinal part damage risk index through the artificial intelligence classification model, and outputting the pest and disease detection result of the medicinal plant.

8. An artificial intelligence image recognition system for detecting pests and diseases of medicinal plants, for implementing the artificial intelligence image recognition method for detecting pests and diseases of medicinal plants according to any one of claims 1-7, characterized in that, It includes: An imaging processing module: collects leaf images, stem images and medicinal part images of medicinal plants, and performs illumination correction processing and texture enhancement processing to generate multi-part preprocessed image data; A leaf feature module: based on the multi-part preprocessed image data, the texture change, color gradient and lesion edge morphology of the medicinal plant leaf region are analyzed, and leaf phenotype feature data is constructed; A medicinal part feature module: based on the multi-part preprocessed image data, the tissue texture detail, epidermis damage pattern and latent lesion dark line of the medicinal part region are analyzed, and medicinal part phenotype feature data is constructed; A spatial mapping module: spatial structure mapping and integration of leaf phenotype feature data and medicinal part phenotype feature data to generate cross-part associated feature data; A risk assessment module: inputting the cross-part associated feature data into the pre-trained medicinal part damage risk prediction model to analyze the mapping relationship between leaf visual symptoms and medicinal part damage risk, and generating a medicinal part damage risk index; A grade determination module: according to the medicinal part damage risk index, using an artificial intelligence classification model to perform pest and disease risk level determination, and outputting a pest and disease detection result.

Images