Pest and disease identification system and method

Abstract

The invention relates to a pest and disease damage identification system and method, and solves the technical problems of large identification difference and non-uniform treatment schemes. A plurality of image acquisition modules, an edge image processing module and a thermal inductance image acquisition module are adopted; the plurality of edge image processing modules are connected to the pest and disease damage identification module and the growth state identification module through the heterogeneous execution mimicry judgment unit; the pest and disease identification unit is also connected with a user interaction module; the disease and insect pest identification module comprises a plurality of heterogeneous redundant sub-identification modules, the heterogeneous redundant sub-identification modules are mutually backed up and debugged and work in parallel, a watchdog unit is arranged in each heterogeneous redundant sub-identification module, the watchdog unit periodically updates a count value, otherwise, the heterogeneous redundant sub-identification modules are automatically reset, and the heterogeneous redundant sub-identification modules are automatically reset. Errors are reported to other heterogeneous redundancy sub-identification modules; according to the technical scheme, the heterogeneous execution mimicry judgment unit is internally provided with an online mimicry judgment device and a synchronizer, the problem is well solved, and the method can be used for pest and disease damage identification.

Description

Technical Field

[0001] The present invention relates to the field of pest and disease identification, and in particular to a pest and disease identification system and method. Background Art

[0002] With the continuous advancement of agricultural modernization, efficient and accurate identification of plant diseases and pests is crucial to ensuring crop yield and quality. Traditional plant disease and pest identification mainly relies on manual observation, but this method has many limitations.

[0003] On the one hand, manual identification is subjective and unstable. Different people may have different criteria for identifying pests and diseases, and manual observation is easily affected by factors such as personal experience and vision, resulting in inaccurate identification results. Furthermore, manual identification is inefficient and requires a significant amount of time and labor for large-scale farmland or gardening.

[0004] On the other hand, traditional pest and disease identification methods can only be used when symptoms are more obvious, which can lead to a delay in detecting early-stage pests and diseases. However, early intervention and prevention can significantly reduce the spread and severity of pests and diseases.

[0005] Furthermore, with climate change and changes in the agricultural ecological environment, new pests and diseases continue to emerge, and traditional identification methods are unable to adapt to these changes. Furthermore, the symptoms of some pests and diseases can be similar to those caused by other non-pest and disease factors, which can easily lead to misdiagnosis.

[0006] While some pest and disease identification methods based on image recognition have emerged in recent years, these methods also present challenges. For example, image acquisition quality and environmental factors can significantly impact recognition results. Problems such as insufficient lighting and blurred images can lead to recognition errors. Furthermore, the accuracy and stability of existing recognition algorithms need to be improved, especially for pest and disease identification in complex backgrounds, which are susceptible to interference.

[0007] Furthermore, current pest and disease identification systems often lack a comprehensive consideration of plant growth status. Plant growth is closely related to the occurrence and development of pests and diseases. However, existing systems typically focus only on pest and disease characteristics while ignoring the overall plant growth status, which can lead to misidentification or omission of pests and diseases.

[0008] In summary, the present invention provides a pest and disease identification system and method, which can solve the above technical problems. Summary of the Invention

[0009] The technical problem to be solved by the present invention is the poor confidentiality and complex solutions in the prior art. A new pest and disease identification system is provided, which has the characteristics of high confidentiality, high identification accuracy and simple operation.

[0010] In order to solve the above technical problems, the technical solutions adopted are as follows:

[0011] A pest identification system, comprising:

[0012] Multiple image acquisition modules for acquiring image information of crops or plants;

[0013] An edge image processing module and a thermal image acquisition module connected to the image acquisition module. The edge image processing module is used to pre-process and analyze the collected visual images and thermal images. The thermal image acquisition module is connected through a light discrimination unit and is used to collect thermal images for supplementation when the light intensity is lower than a predefined threshold.

[0014] The plurality of edge image processing modules are connected to the pest and disease identification module and the growth status identification module through the heterogeneous execution mimicry judgment unit;

[0015] The pest identification unit is also connected to a user interaction module, which is responsible for providing an interactive interface between the user and the system, including a display module, an operation module, and an alarm module;

[0016] The pest and disease identification module includes multiple heterogeneous redundant sub-identification modules. The heterogeneous redundant sub-identification modules back up each other, check each other's errors, and work in parallel. Each heterogeneous redundant sub-identification module is equipped with a watchdog unit. The watchdog unit regularly updates the count value. Otherwise, it will automatically reset the heterogeneous redundant sub-identification module and report an error to other heterogeneous redundant sub-identification modules.

[0017] The heterogeneous execution mimetic decision unit has an online mimetic decider and a synchronizer built in;

[0018] Step 1: edge image processing module connection validity judgment, including:

[0019] When the edge image processing module initiates a task to the sub-identification module, the heterogeneous execution mimic judgment unit removes the authentication information from the connection instruction of the edge image processing module based on the connection information between the edge image processing module and the sub-identification module, obtains the judgment instruction information, and then adds it to the instruction similarity judgment queue;

[0020] If the judgment instruction information from different sub-recognition modules is consistent, the edge image processing module is trusted in real time, and the data of the edge processing module is added to the valid data set;

[0021] If the judgment instruction information from different sub-recognition modules is not completely consistent, the validity of the edge processing module is determined under the principle of minority obeys majority. If it is valid, the data of the edge processing module is added to the valid data set;

[0022] If the judgment instruction information from different sub-recognition modules is completely inconsistent, the data of the edge image processing module is discarded and marked as a fault;

[0023] Step 2: edge image processing module image data validity judgment, including:

[0024] When multiple edge image processing modules send data to the sub-identification module, the heterogeneous execution mimic judgment unit removes the verification information from the edge image processing module's data based on the connection information between the edge image processing module and the sub-identification module, obtains the data judgment instruction information, and adds it to the instruction similarity judgment queue;

[0025] If it is determined that the data judgment instruction information from different edge image processing modules is consistent, the current data of the edge image processing module is trusted in real time and the current data of the edge image processing module is added to the valid data set;

[0026] If the data judgment instruction information from different edge image processing modules is not completely consistent, the validity of the edge image processing module is judged under the principle of minority obeys majority. If it is valid, the current data of the edge image processing module is added to the valid data set;

[0027] If it is determined that the data judgment instruction information from different sub-recognition modules are completely inconsistent, the current data will be discarded.

[0028] The present invention operates by employing multiple image acquisition modules to collect image information of crops or plants. The edge image processing module preprocesses and analyzes visual and thermal images to extract useful feature information. The thermal image acquisition module provides a supplementary function in low-light conditions, ensuring the system can acquire comprehensive image data. The heterogeneous execution mimicry judgment unit, based on the connection information between the edge image processing module and the sub-identification module, removes verification information from the connection instructions and image data to obtain judgment instruction information, which is then added to the instruction similarity judgment queue. Using different judgment principles, such as unanimous trust and majority rule, the validity of the edge image processing module is determined, and valid data is added to the valid dataset to improve data reliability. Multiple heterogeneous redundant sub-identification modules operate in parallel, utilizing their own algorithms and models to identify pests and diseases. Simultaneously, the growth status recognition module combines image information and other data to determine the growth status of the plant. These two modules work together to comprehensively consider the plant's pest and disease status and growth status, improving recognition accuracy. The user interaction module provides an interface for users to interact with the system, allowing them to easily understand system operation and take appropriate measures. The watchdog unit monitors the heterogeneous redundant sub-identification modules and automatically resets and reports errors when a fault occurs, ensuring the stability and reliability of the system.

[0029] In the above scheme, for optimization, the pest and disease identification module further performs the following algorithm steps to complete the pest and disease damage identification:

[0030] Step 1: Initialize multiple heterogeneous redundant sub-identification modules, determine the number of sub-identification modules to be N, and assign unique identifiers to each sub-identification module, denoted as ID_1, ID_2, ..., ID_N; set the bee algorithm parameters, the number of bees to be M, the pheromone volatility coefficient to be rho, ranging from 0 to 1, and the pheromone intensity to be Q; set the de-ignition algorithm parameters, and the denoising intensity parameter to be alpha; start the watchdog unit and set the count value update period to T;

[0031] Step 2: Collect rice pest and disease image samples and divide them into training set, validation set and test set; digitally process the images, including floating point algorithm, integer method, shift method and average method;

[0032] Then, filtering is performed to remove noise in the image, and filtering algorithms including mean filtering and median filtering can be used;

[0033] Then enhancement processing is performed to highlight the features of the image, including contrast enhancement and brightness adjustment;

[0034] Finally, segmentation processing is performed to separate the rice in the image from the background, and pattern recognition methods and book recognition are used to extract and analyze the internal correlation and abstract model;

[0035] Step 3: The convolutional neural network system is trained on rice diseases and insect pests respectively;

[0036] Select the convolution kernel and perform convolution operation on the image. After the convolution step, a feature map is formed. Configure the parameters of the full convolution network, including the parameters of the convolution layer, pooling layer, and upsampling layer. The convolution formula is:

[0037] Output feature map = ∑(input image * convolution kernel);

[0038] Calculate the feature map size to extract rice pest and disease characteristics. The feature map size calculation formula is:

[0039] Output feature map size = (input image size - convolution kernel size + 2 * padding) / step size + 1;

[0040] The cross-stage partial network of the C2F structure is used to segment and fuse the feature maps, accurately capturing the characteristics of different growth stages and complexity of rice through multi-level feature fusion;

[0041] Precision, recall, F1 score, and accuracy are used as performance metrics to determine the accuracy of the classification model:

[0042] The precision calculation formula is: Precision = TP / (TP + FP), where TP is the true positive, which is the number of samples that are actually infected with pests and diseases and are correctly identified as such; FP is the false positive, which is the number of samples that are actually free of pests and diseases but are incorrectly identified as such.

[0043] The recall calculation formula is: Recall = TP / (TP+FN); FN is a false negative example;

[0044] The F1 score calculation formula is: 2*(precision*recall) / (precision+recall);

[0045] The accuracy calculation formula is: Accuracy = (TP + TN) / (TP + TN + FP + FN); TN is the true negative example, which refers to the number of samples that are correctly judged as negative examples; FN is the false negative example, which refers to the number of samples that actually have pests and diseases but are mistakenly identified as not having pests and diseases;

[0046] The cross-stage partial network of the C2F structure can be determined as follows:

[0047] (1) Multi-branch convolution

[0048] Let the input feature tensor be X.

[0049] Create multiple branches for convolution operations. For example, there can be three branches, each performing a different convolution operation.

[0050] Branch 1: Y1 = W1 * ReLU (BN (W 11 *X)), where W 11 , W1 is the convolution kernel weight, * represents the convolution operation, BN represents the batch normalization operation, and ReLU is the activation function.

[0051] Branch 2: Y2=W2*ReLU(BN(W 21 *X)).

[0052] Branch 3: Y3=W3*ReLU(BN(W 31 *X)).

[0053] (2) Feature Fusion

[0054] The outputs of the three branches are fused. Different fusion methods can be used, such as concatenation or addition.

[0055] Splicing and fusion: F concat =concat(Y1,Y2,Y3)

[0056] Additive Fusion: F add =Y1+Y2+Y3.

[0057] (3) Residual connection:

[0058] If residual connections are used, the input feature tensor can be added to the fused features.

[0059] Let the output after residual connection be Z. If additive fusion is used and there is residual connection, then Z = X + F add If splicing fusion is used and there is a residual connection, the convolutional features can be properly adjusted before adding them to the input, i.e. Z = X + W4*F concat , where W4 is the convolution kernel weight used to adjust the concatenated features.

[0060] (4) Output processing

[0061] The output Z after the residual connection is further processed, such as through activation function and batch normalization.

[0062] O = BN(ReLU(W5*Z)), where W5 is the convolution kernel weight used for output processing. The final output is the feature tensor O after processing by the C3f structure.

[0063] Step 4: Select the optimal sub-recognition module combination based on the comprehensive evaluation value, assign the pest and disease data to be identified to the optimal sub-recognition module combination, and each sub-recognition module independently performs pest and disease identification, including:

[0064] The bee algorithm is used for feature extraction. The formula for calculating pheromone concentration in the bee algorithm is:

[0065] Tau_ij(t)=(1-rho)Tau_ij(t-1)+DeltaTau_ij, where Tau_ij(t) represents the pheromone concentration from feature i to feature j at time t, and rho is the pheromone volatility coefficient;

[0066] DeltaTau_ij is the increment of pheromone left by the bee on the path in this cycle, and the calculation formula is DeltaTau_ij = Q / L_k, where Q is the pheromone intensity and L_k is the length of the path walked by the kth bee;

[0067] The de-noising algorithm is used to denoise the data. The calculation formula of the new eigenvalue after denoising in the de-noising algorithm is:

[0068] NewFeatureValue=FeatureValue(1-alpha), where FeatureValue is the original feature value and alpha is the denoising strength parameter;

[0069] The sub-recognition module summarizes the recognition results and checks for errors. If there are any discrepancies in the results, the error checking mechanism is activated. The new pheromone concentration calculation formula in the error checking mechanism is:

[0070] Tau_ij(t)=(1-rho)Tau_ij(t-1)+DeltaTau_ij', where DeltaTau_ij' is the pheromone increment left by the bees on the path during the error checking process;

[0071] At the same time, the adjusted denoising intensity parameter alpha' is used for denoising, and the new feature value calculation formula is: NewFeatureValue'=FeatureValue(1-alpha');

[0072] The watchdog unit monitors the operation of the sub-identification modules. If a fault is found, it will automatically reset and report the error to other sub-identification modules, re-evaluate the sub-identification modules and generate a new sub-module combination, redistribute the faulty module data, and finally output the pest and disease nuisance identification results.

[0073] The preferred solution utilizes multiple heterogeneous redundant sub-identification modules, image acquisition modules, edge image processing modules, thermal image acquisition modules, and a heterogeneous execution mimicry decision unit. This combines a defuzzified trapezoidal function, a deep belief network load prediction algorithm for classification recognition, and a gray potential recognition algorithm to design a plant growth status recognition algorithm, achieving efficient and accurate pest and disease infestation identification. Each sub-identification module operates in parallel, providing mutual backup and error checking. A watchdog unit ensures proper module operation, promptly reporting and resetting any issues, redistributing data from the faulty module, and reassessing and generating a new sub-module combination.

[0074] Furthermore, the floating-point algorithm processes the pixel values ​​of the image according to specific floating-point operation rules, and the formula is:

[0075] (new pixel value) = f(original pixel value), where f is a specific floating-point operation function;

[0076] The integer method adjusts the pixel value through integer operations. The formula is: (new pixel value) = g (original pixel value), where g is an integer operation function;

[0077] The shift method performs a shift operation on the pixel value according to the shift rule. The formula is (new pixel value) = h (original pixel value), where h is the shift function;

[0078] The average method calculates the average value of adjacent pixels as the new pixel value. The formula is (new pixel value) = (pixel value_1 + pixel value_2 + ... + pixel value_n) / n, where n is the number of pixels participating in the average.

[0079] In the preferred solution, diversified processing: a variety of different image pixel value processing methods are provided, including floating-point arithmetic, integer method, shift method and average method. The appropriate processing method can be selected according to different image characteristics and requirements, which improves the flexibility and adaptability of image processing. Precise adjustment: The floating-point algorithm can make fine adjustments to the pixel value through specific floating-point operation functions to adapt to scenarios with high image accuracy requirements. The integer method is suitable for situations that require fast processing and relatively low accuracy requirements. Data optimization: The shift method can quickly change the bit-level representation of the pixel value to achieve specific data optimization effects, which may play an important role in certain specific image processing tasks. Noise reduction and smoothing: The average method can effectively reduce the noise in the image, make the image smoother, and improve the image quality by calculating the average value of adjacent pixels as the new pixel value. Combinable use: These different processing methods can be used in combination according to actual conditions, further enhancing the effect and flexibility of image processing.

[0080] Furthermore, for each pixel point, the mean filter takes the average value of the pixel values ​​in its surrounding neighborhood as the new pixel value of the point. The formula is:

[0081] (new pixel value) = (pixel_value_1 + pixel_value_2 + ... + pixel_value_k) / k, where k is the number of pixels in the neighborhood;

[0082] Median filtering takes the median of the pixel values ​​in the neighborhood as the new pixel value, that is, sorts the pixel values ​​in the neighborhood and takes the middle value as the new pixel value;

[0083] Contrast enhancement enhances the contrast by adjusting the grayscale range of the image. The formula is:

[0084] (new pixel value) = a*(original pixel value) + b, where a and b are adjustment parameters;

[0085] Brightness adjustment emphasizes features by increasing or decreasing the overall brightness of the image. The formula is:

[0086] (new pixel value) = (original pixel value) + c, where c is the brightness adjustment parameter;

[0087] Threshold segmentation selects a suitable threshold value, classifies pixels with values ​​greater than the threshold value into one category, and pixels with values ​​less than the threshold value into another category, thereby achieving image segmentation. The formula is:

[0088] (New pixel value) = d*(original pixel value), where d is the value determined based on the threshold.

[0089] Furthermore, the calculation formula of the comprehensive evaluation value EvaluationValue_i is:

[0090] =omega_1Accuracy_i-omega_2(1 / MTBF_i)+omega_3*ResponseTime_i;

[0091] Where EvaluationValue_i is the comprehensive evaluation value of sub-identification module i, omega_1, omega_2, and omega_3 are weight coefficients, Accuracy_i is the accuracy of sub-identification module i, MTBF_i is the mean time between failures of sub-identification module i, and ResponseTime_i is the response time of sub-identification module i.

[0092] Furthermore, the watchdog unit monitors the operation of the sub-identification module, and if the count value is not updated on time, the sub-identification module is automatically reset and an error is reported to other sub-identification modules;

[0093] After receiving the error message, other sub-identification modules restart the evaluation of the sub-identification unit, collect the operating data of the remaining normal sub-identification modules in the current stage, recalculate the comprehensive evaluation value according to the full life cycle valuation and identification algorithm, regenerate a new sub-module unit combination, and reallocate the data originally allocated to the faulty sub-identification module to the sub-identification module in the new combination for processing.

[0094] The present invention also provides a method for identifying pests and diseases, the method comprising:

[0095] Step A: Determine the validity of the edge image processing module connection. When the edge image processing module initiates a task to the sub-identification module, the heterogeneous execution mimicry determination unit removes the authentication information from the edge image processing module's connection instruction based on the connection information between the edge image processing module and the sub-identification module to obtain the determination instruction information, and then adds the information to the instruction similarity determination queue.

[0096] If the judgment instruction information from different sub-recognition modules is consistent, the edge image processing module is trusted in real time, and the data of the edge processing module is added to the valid data set;

[0097] If the judgment instruction information from different sub-recognition modules is not completely consistent, the validity of the edge processing module is determined under the principle of minority obeys majority. If it is valid, the data of the edge image processing module is added to the valid data set;

[0098] If the judgment instruction information from different sub-recognition modules is completely inconsistent, the data of the edge image processing module is discarded and marked as a fault;

[0099] Step B, edge image processing module image data validity judgment: When multiple edge image processing modules initiate data to the sub-identification module, the heterogeneous execution mimicry judgment unit removes the verification information from the edge image processing module data based on the connection information between the edge image processing module and the sub-identification module, obtains data judgment instruction information, and adds it to the instruction similarity judgment queue.

[0100] If it is determined that the data judgment instruction information from different edge image processing modules is consistent, the current data of the edge image processing module is trusted in real time and the current data of the edge image processing module is added to the valid data set;

[0101] If the data judgment instruction information from different edge image processing modules is not completely consistent, the validity of the edge image processing module is judged under the principle of minority obeys majority. If it is valid, the current data of the edge image processing module is added to the valid data set;

[0102] If the data judgment instruction information from different sub-recognition modules is completely inconsistent, the current data is discarded;

[0103] Step C: Initialize multiple heterogeneous redundant sub-identification modules, determine the number of sub-identification modules to be N, assign unique identifiers to each sub-identification module, and record them as ID_1, ID_2, ..., ID_N; set the parameters of the bee algorithm, the number of bees to be M, the pheromone volatility coefficient to be rho, ranging from 0 to 1, and the pheromone intensity to be Q; set the de-ignition algorithm parameters, and the denoising intensity parameter to be alpha; start the watchdog unit and set the count value update period to T;

[0104] Step D, collecting rice pest and disease image samples and dividing them into training set, validation set and test set; performing digital processing on the images, including floating point algorithm, integer method, shift method and average method;

[0105] Then, filtering is performed to remove noise in the image, and filtering algorithms including mean filtering and median filtering can be used;

[0106] Then enhancement processing is performed to highlight the features of the image, including contrast enhancement and brightness adjustment;

[0107] Finally, segmentation processing is performed to separate the rice in the image from the background, and pattern recognition methods and book recognition are used to extract and analyze the internal correlation and abstract model;

[0108] Step E: training the neural network system based on convolution operation on rice diseases and insect pests respectively;

[0109] Select a convolution kernel and perform a convolution operation on the image. The convolution formula is: (output feature map) = ∑ (input image * convolution kernel). After the convolution step, a feature map is formed. Configure the parameters of the full convolution network, including the parameters of the convolution layer, pooling layer, and upsampling layer;

[0110] Calculate the feature map size to extract rice pest and disease characteristics. The feature map size calculation formula is:

[0111] (Output feature map size) = (input image size - convolution kernel size + 2 * padding) / stride + 1;

[0112] The cross-stage partial network of the C3F structure is used to segment and fuse feature maps to enhance feature expression capabilities. The multi-level feature fusion accurately captures the characteristics of rice at different growth stages and complexity.

[0113] Precision, recall, F1 score, and accuracy are used as performance metrics to determine the accuracy of the classification model:

[0114] The accuracy calculation formula is: Precision = TP / (TP + FP);

[0115] The recall calculation formula is: Recall = TP / (TP + FN);

[0116] The F1 score calculation formula is: 2*(precision*recall) / (precision+recall);

[0117] The accuracy calculation formula is: Accuracy = (TP + TN) / (TP + TN + FP + FN).

[0118] Step F, selecting the optimal sub-recognition module combination based on the comprehensive evaluation value, assigning the pest and disease data to be identified to the optimal sub-recognition module combination, and each sub-recognition module independently performs pest and disease identification, including:

[0119] The bee algorithm is used for feature extraction. The formula for calculating pheromone concentration in the bee algorithm is:

[0120] Tau_ij(t)=(1-rho)Tau_ij(t-1)+DeltaTau_ij, where Tau_ij(t) represents the pheromone concentration from feature i to feature j at time t, and rho is the pheromone volatility coefficient;

[0121] DeltaTau_ij is the increment of pheromone left by the bee on the path in this cycle, and the calculation formula is DeltaTau_ij = Q / L_k, where Q is the pheromone intensity and L_k is the length of the path walked by the kth bee;

[0122] The de-noising algorithm is used to denoise the data. The calculation formula of the new eigenvalue after denoising in the de-noising algorithm is:

[0123] NewFeatureValue=FeatureValue(1-alpha), where FeatureValue is the original feature value and alpha is the denoising strength parameter;

[0124] The sub-recognition module summarizes the recognition results and checks for errors. If there are any discrepancies in the results, the error checking mechanism is activated. The new pheromone concentration calculation formula in the error checking mechanism is:

[0125] Tau_ij(t)=(1-rho)Tau_ij(t-1)+DeltaTau_ij', where DeltaTau_ij' is the pheromone increment left by the bees on the path during the error checking process;

[0126] At the same time, the adjusted denoising intensity parameter alpha' is used for denoising, and the new feature value calculation formula is: NewFeatureValue'=FeatureValue(1-alpha');

[0127] The watchdog unit monitors the operation of the sub-identification modules. If a fault is found, it automatically resets and reports the error to other sub-identification modules, re-evaluates the sub-identification modules and generates a new sub-module combination, redistributing the faulty module data.

[0128] Step F, result optimization, combines the plant growth status recognition results with the pest and disease recognition module to optimize the pest and disease recognition results, and finally outputs the pest and disease recognition results, including:

[0129] The collected image data is fuzzy processed using a defuzzified trapezoidal function to reduce noise and uncertainty. The defuzzified trapezoidal function can be expressed as:

[0130] Where a, b, c, and d are the parameters of the trapezoidal function;

[0131] A deep belief network load forecasting algorithm based on classification and recognition is used to make a preliminary prediction of the plant growth status;

[0132] A deep belief network composed of multiple stacked restricted Boltzmann machines (RBMs) is called, and unsupervised learning methods and supervised learning methods are used for fusion training to obtain growth status prediction results;

[0133] The features in the growth status prediction results are integrated with the features in the pest and disease identification results. Based on experience and data analysis, different weights are assigned to the growth status features and pest and disease characteristics. According to the comprehensive evaluation health index, the pest and disease identification results are optimized to finally obtain the pest and disease damage identification results.

[0134] This algorithm leverages plant growth status predictions to further optimize pest and disease identification, improving accuracy and reliability. By integrating growth status information with pest and disease characteristics, it comprehensively assesses plant health and provides stronger support for precise pest and disease control.

[0135] Furthermore, calling the unsupervised learning method and the supervised learning method for fusion training includes:

[0136] First, each RBM is pre-trained using unsupervised learning to extract features from the image;

[0137] Then, supervised learning is used to fine-tune the entire deep belief network to improve the accuracy of prediction. During the prediction process, the extracted features are input into the deep belief network to obtain the prediction results of the plant growth status.

[0138] Combined with the grey potential recognition algorithm, the growth status of plants is further analyzed to discover the potential laws and changing trends in the growth status of plants;

[0139] The calculation formula of the grey potential recognition algorithm is:

[0140]

[0141] where x (0) (i) is the original data sequence, X (1) (k) is the cumulative generated sequence;

[0142] By analyzing the cumulatively generated sequence, the gray potential value of the plant growth status can be obtained, thereby judging the growth trend and health status of the plant. BRIEF DESCRIPTION OF THE DRAWINGS

[0143] The present invention will be further described below with reference to the accompanying drawings and examples.

[0144] Figure 1 , schematic diagram of the pest and disease identification system in the embodiment. DETAILED DESCRIPTION

[0145] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0146] Example 1

[0147] This embodiment provides a pest and disease identification system, which includes:

[0148] Multiple image acquisition modules for acquiring image information of crops or plants;

[0149] An edge image processing module and a thermal image acquisition module connected to the image acquisition module. The edge image processing module is used to pre-process and analyze the collected visual images and thermal images. The thermal image acquisition module is connected through a light discrimination unit and is used to collect thermal images for supplementation when the light intensity is lower than a predefined threshold.

[0150] The plurality of edge image processing modules are connected to the pest and disease identification module and the growth status identification module through the heterogeneous execution mimicry judgment unit;

[0151] The pest identification unit is also connected to a user interaction module, which is responsible for providing an interactive interface between the user and the system, including a display module, an operation module, and an alarm module;

[0152] The pest and disease identification module includes multiple heterogeneous redundant sub-identification modules. The heterogeneous redundant sub-identification modules back up each other, check each other's errors, and work in parallel. Each heterogeneous redundant sub-identification module is equipped with a watchdog unit. The watchdog unit regularly updates the count value. Otherwise, it will automatically reset the heterogeneous redundant sub-identification module and report an error to other heterogeneous redundant sub-identification modules.

[0153] The heterogeneous execution mimetic decision unit has an online mimetic decider and a synchronizer built in;

[0154] Step 1: edge image processing module connection validity judgment, including:

[0155] When the edge image processing module initiates a task to the sub-identification module, the heterogeneous execution mimic judgment unit removes the authentication information from the connection instruction of the edge image processing module based on the connection information between the edge image processing module and the sub-identification module, obtains the judgment instruction information, and then adds it to the instruction similarity judgment queue;

[0156] If the judgment instruction information from different sub-recognition modules is consistent, the edge image processing module is trusted in real time, and the data of the edge processing module is added to the valid data set;

[0157] If the judgment instruction information from different sub-recognition modules is not completely consistent, the validity of the edge processing module is determined under the principle of minority obeys majority. If it is valid, the data of the edge processing module is added to the valid data set;

[0158] If the judgment instruction information from different sub-recognition modules is completely inconsistent, the data of the edge image processing module is discarded and marked as a fault;

[0159] Step 2: edge image processing module image data validity judgment, including:

[0160] When multiple edge image processing modules send data to the sub-identification module, the heterogeneous execution mimic judgment unit removes the verification information from the edge image processing module's data based on the connection information between the edge image processing module and the sub-identification module, obtains the data judgment instruction information, and adds it to the instruction similarity judgment queue;

[0161] If it is determined that the data judgment instruction information from different edge image processing modules is consistent, the current data of the edge image processing module is trusted in real time and the current data of the edge image processing module is added to the valid data set;

[0162] If the data judgment instruction information from different edge image processing modules is not completely consistent, the validity of the edge image processing module is judged under the principle of minority obeys majority. If it is valid, the current data of the edge image processing module is added to the valid data set;

[0163] If it is determined that the data judgment instruction information from different sub-recognition modules are completely inconsistent, the current data will be discarded.

[0164] This embodiment uses multiple image acquisition modules to collect image information of crops or plants. The edge image processing module preprocesses and analyzes visual and thermal images to extract useful feature information. The thermal image acquisition module provides a supplementary function when lighting is insufficient, ensuring the system can acquire comprehensive image data. The heterogeneous execution mimicry judgment unit, based on the connection information between the edge image processing module and the sub-identification module, removes verification information from the connection instructions and image data, obtains judgment instruction information, and adds it to the instruction similarity judgment queue. Using different judgment principles, such as unanimous trust and majority rule, the validity of the edge image processing module is determined, and valid data is added to the valid dataset to improve data reliability. Multiple heterogeneous redundant sub-identification modules operate in parallel, utilizing their own algorithms and models to identify pests and diseases. Simultaneously, the growth status recognition module combines image information and other data to determine the growth status of the plant. The two modules work together to comprehensively consider the plant's pest and disease status and growth status, improving recognition accuracy. The user interaction module provides an interface for users to interact with the system, making it easy for them to understand system operation and take appropriate measures. The watchdog unit monitors the heterogeneous redundant sub-identification modules and automatically resets and reports errors when a fault occurs, ensuring the stability and reliability of the system.

[0165] Preferably, the pest and disease identification module performs the following algorithm steps to complete pest and disease damage identification:

[0166] Step 1: Algorithm initialization

[0167] 1.1. Determine the number of submodules and identifier allocation

[0168] 1.1.1. Define the number of heterogeneous redundant sub-identification modules as N.

[0169] 1.1.2. Assign a unique identifier to each sub-identification module, denoted as ID_1, ID_2, ..., ID_N.

[0170] 1.2. Bee Algorithm Parameter Setting

[0171] 1.2.1. Determine the number of bees as M, which is set according to actual conditions, for example, it can be set to a certain positive integer.

[0172] 1.2.2. Set the pheromone volatility coefficient to rho, which usually ranges from 0 to 1. The specific value is determined according to the actual application scenario.

[0173] 1.2.3. Set the pheromone intensity to Q. You can set the specific value according to your needs.

[0174] 1.3. De-ignition algorithm parameter setting

[0175] 1.3.1. Determine the denoising strength parameter as alpha, and its value range is determined according to the specific situation.

[0176] Step 2: Data collection and preprocessing

[0177] 2.1. Pest and disease data collection

[0178] 2.1.1. Based on deep learning principles, collect a large number of image samples of rice diseases and pests, ensuring that the samples cover the various manifestations of different diseases and pests to ensure diversity and representativeness.

[0179] 2.1.2. Divide the collected image samples into training set, validation set and test set.

[0180] 2.2. Image digital processing

[0181] 2.2.1. When the image is input by the scanner, it is digitized using floating point arithmetic, integer method, shift method and average method.

[0182] 2.2.1.1. Floating-point algorithm: The pixel values ​​of the image are processed according to specific floating-point operation rules. The formula is (new pixel value) = f(original pixel value), where f is a specific floating-point operation function.

[0183] 2.2.1.2. Integer method: The pixel value is adjusted through integer operations. The formula is (new pixel value) = g (original pixel value), where g is an integer operation function.

[0184] 2.2.1.3. Shift method: Shift the pixel value according to the shift rule. The formula is (new pixel value) = h (original pixel value), where h is the shift function.

[0185] 2.2.1.4. Average value method: Calculate the average value of adjacent pixels as the new pixel value. The formula is (new pixel value) = (pixel value_1 + pixel value_2 + ... + pixel value_n) / n, where n is the number of pixels participating in the average.

[0186] 2.3. Image filtering

[0187] 2.3.1. Perform filtering to remove noise from the image. Common filtering algorithms such as mean filtering or median filtering can be used.

[0188] 2.3.1.1. Mean filtering: For each pixel point, take the average value of the pixel values ​​in a certain neighborhood around it as the new pixel value of the point. The formula is (new pixel value) = (pixel value_1 + pixel value_2 + ... + pixel value_k) / k, where k is the number of pixels in the neighborhood.

[0189] 2.3.1.2. Median filtering: Take the median of the pixel values ​​in the neighborhood as the new pixel value, that is, sort the pixel values ​​in the neighborhood and take the middle value as the new pixel value.

[0190] 2.4. Image Enhancement Processing

[0191] 2.4.1. Perform enhancement processing to highlight the features of the image. Methods such as contrast enhancement and brightness adjustment can be used.

[0192] 2.4.1.1. Contrast enhancement: The contrast is enhanced by adjusting the grayscale range of the image. The formula is (new pixel value) = a*(original pixel value) + b, where a and b are adjustment parameters.

[0193] 2.4.1.2. Brightness adjustment: Highlight features by increasing or decreasing the overall brightness of the image. The formula is (new pixel value) = (original pixel value) + c, where c is the brightness adjustment parameter.

[0194] 2.5. Image segmentation processing

[0195] 2.5.1. Perform segmentation processing to separate the rice in the image from the background. Algorithms such as threshold segmentation or region growing can be used.

[0196] 2.5.1.1. Threshold segmentation: Select an appropriate threshold and classify pixels with values ​​greater than the threshold into one category and pixels with values ​​less than the threshold into another category, thereby achieving image segmentation. The formula is (new pixel value) = d * (original pixel value), where d is the value after threshold judgment.

[0197] 2.5.1.2. Region growing: Starting from one or more seed points, the region is gradually grown based on the similarity of pixels until a certain stopping condition is met.

[0198] Pattern recognition analysis

[0199] 2.6.1. Analyze and identify processed images using pattern recognition techniques to identify inherent connections or abstract models.

[0200] Step 3: Evaluation of the full life cycle valuation identification algorithm

[0201] 3.1. Neural Network Training Based on Convolution Operation

[0202] 3.1.1. Use a convolutional neural network system to train rice diseases and insect pests separately.

[0203] 3.1.1.1. Select a convolution kernel and perform a convolution operation on the image. The convolution formula is: (output feature map) = ∑ (input image * convolution kernel), where "" represents the convolution operation.

[0204] 3.1.1.2. After the convolution step, a feature map is formed.

[0205] 3.1.1.3. Configure the parameters of the fully convolutional network, including the parameters of the convolutional layer, pooling layer, and upsampling layer.

[0206] 3.1.1.3.1. Convolutional layer parameters: These include kernel size, number, and stride. For example, kernel size can be 3x3, 5x5, etc. The number can be set as needed, and stride can be 1 or 2, etc.

[0207] 3.1.1.3.2. Pooling layer parameters: The pooling method can be max pooling or average pooling. The pooling window size and stride also need to be set. For example, the max pooling window size can be 2x2 with a stride of 2.

[0208] 3.1.1.3.3. Upsampling layer parameters: upsampling method and multiple, etc. For example, bilinear interpolation upsampling with a multiple of 2 can be used.

[0209] 3.1.1.4. Calculate the size of the generated feature map to extract features of rice pests and diseases. The formula for calculating the feature map size is: (output feature map size) = (input image size - convolution kernel size + 2 * padding) / stride + 1.

[0210] 3.2. Introducing C3F convolutional feature extraction technology

[0211] 3.2.1.YOLO V8 introduces C3F (CSP-2Fuse) convolutional feature extraction technology.

[0212] 3.2.2. The C3F structure uses a cross-stage partial network (CSPNet) to segment and fuse feature maps, reducing redundant features and enhancing feature expression capabilities to achieve efficient feature extraction.

[0213] 3.2.3. Through multi-level feature fusion, C3F technology can accurately capture the characteristics of rice at different growth stages and complexity, improving the robustness of monitoring and identification.

[0214] 3.2.4. In simplifying the recognition process, the C3F structure reduces computational costs, reduces the number of model parameters and storage requirements, improves model running speed and training efficiency, and ensures high-performance deployment in resource-constrained environments.

[0215] 3.3. Performance metric calculation

[0216] 3.3.1. Use precision, recall, F1 score and accuracy as performance metrics to determine the accuracy of the classification model.

[0217] 3.3.1.1. The formula for calculating precision is (Precision) = TP / (TP + FP).

[0218] 3.3.1.2. The recall rate calculation formula is (Recall) = TP / (TP + FN).

[0219] 3.3.1.3. The F1 score is calculated as 2(precision * recall) / (precision + recall).

[0220] 3.3.1.4. The accuracy calculation formula is (Accuracy) = (TP + TN) / (TP + TN + FP + FN).

[0221] 3.4. Comprehensive Evaluation Sub-Identification Module

[0222] 3.4.1. Based on the performance and failure rate evaluation results, calculate a comprehensive evaluation value for each sub-identification module, denoted as EvaluationValue_i. For example, a weighted summation method can be used, and the calculation formula is EvaluationValue_i = omega_1Accuracy_i - omega_2(1 / MTBF_i) + omega_3*ResponseTime_i, where omega_1, omega_2, and omega_3 are weight coefficients that can be set according to actual needs.

[0223] Step 4: Sub-recognition module selection

[0224] 4.1. Select the optimal sub-recognition module combination based on the comprehensive evaluation value. You can set a threshold and select the sub-recognition module with a comprehensive evaluation value higher than the threshold.

[0225] Step 5: Data distribution

[0226] 5.1. Evenly distribute the pest and disease data to be identified to the selected optimal sub-identification module combination. Assume that the amount of data allocated to each sub-identification module is DataVolume_i = |D| / n, where |D| represents the size of the data set D and n is the number of optimal sub-identification modules.

[0227] Step 6: Sub-identification module processing

[0228] 6.1. Each sub-recognition module independently uses its own recognition algorithm to identify pests and diseases on the assigned data.

[0229] 6.2. For example, the sub-identification module ID_i can select the following methods for identification:

[0230] 6.2.1. Feature extraction using the bee algorithm:

[0231] 6.2.1.1. Construct the solution space of the problem, that is, the possible feature combinations.

[0232] 6.2.1.2. Bees move through the solution space, selecting the next feature based on pheromone concentration and heuristic information. The pheromone concentration is calculated as Tau_ij(t) = (1-rho)Tau_ij(t-1) + DeltaTau_ij, where Tau_ij(t) represents the pheromone concentration from feature i to feature j at time t, rho is the pheromone volatility coefficient, and DeltaTau_ij is the pheromone increment left by the bee on that path during this cycle. The calculation formula is DeltaTau_ij = Q / L_k, where Q is the pheromone intensity and L_k is the length of the path traversed by the kth bee.

[0233] 6.2.2. Using the de-ignition algorithm to denoise data:

[0234] 6.2.2.1. Perform denoising on the extracted features. The new feature value calculation formula is NewFeatureValue=FeatureValue(1-alpha), where FeatureValue is the original feature value and alpha is the denoising strength parameter.

[0235] Step 7: Result summary and error checking

[0236] 7.1. Each sub-recognition module sends the recognition results to other sub-recognition modules for result aggregation.

[0237] 7.2. Compare the recognition results between the sub-recognition modules to check whether there are any differences.

[0238] 7.3. If the identification results of multiple sub-identification modules are consistent, the result is considered reliable and can be used as the final pest and disease identification result.

[0239] 7.4. If there are discrepancies in the recognition results of the sub-recognition modules, the error checking mechanism is activated.

[0240] Step 8, error checking mechanism

[0241] 8.1. For sub-recognition modules with different recognition results, use the honey bee algorithm to perform path search to find the possible causes of the differences.

[0242] 8.1.1. Bees move in the space of steps that may lead to discrepancies, re-evaluate the pheromone concentration at each step, and adjust their search direction.

[0243] 8.1.2. According to the new pheromone concentration calculation formula Tau_ij(t) = (1-rho)Tau_ij(t-1) + DeltaTau_ij', DeltaTau_ij' is the pheromone increment left by the bees on the path during the error checking process. The calculation method is adjusted according to the specific error checking goal.

[0244] 8.2. At the same time, use the de-ignition algorithm to denoise the data with differences:

[0245] 8.2.1. Adjust the denoising intensity parameter alpha again to perform a more refined denoising operation. The new feature value calculation formula is NewFeatureValue'=FeatureValue(1-alpha'), where alpha' is the adjusted denoising intensity parameter.

[0246] Step 9, watchdog unit monitoring

[0247] 9.1. The watchdog unit periodically updates the count value to ensure the normal operation of the sub-identification module.

[0248] 9.2. If the watchdog unit finds that the count value of a sub-identification module is not updated on time, it will automatically reset the sub-identification module and report an error to other sub-identification modules.

[0249] 9.3. After receiving the error message, other sub-recognition modules proceed to the following iterative steps:

[0250] 9.3.1. Restart the evaluation of the sub-identification unit:

[0251] 9.3.1.1. Collect the operating data of the remaining normal sub-recognition modules at the current stage, including recognition accuracy, response time, etc.

[0252] 9.3.1.2. Recalculate the comprehensive assessment value according to the full life cycle valuation identification algorithm.

[0253] 9.3.2. Regenerate a new submodule unit combination:

[0254] 9.3.2.1. Based on the recalculated comprehensive evaluation value, select a new optimal sub-recognition module combination.

[0255] 9.3.2.2. If the new combination is different from the original combination, the data originally assigned to the faulty sub-identification module shall be reallocated to the sub-identification modules in the new combination for processing.

[0256] Step 10, result output

[0257] 10.1. Output the final pest and disease identification results to the user interaction module so that users can promptly understand the pest and disease situation and take appropriate preventive measures. The iterative steps are clearly defined as starting from the watchdog unit monitoring to the discovery of the fault sub-identification module, and ending with the regeneration of a new sub-module unit combination and the completion of data redistribution.

[0258] The present invention also provides a method for identifying pests and diseases, the method comprising:

[0259] Step A: Determine the validity of the edge image processing module connection. When the edge image processing module initiates a task to the sub-identification module, the heterogeneous execution mimicry determination unit removes the authentication information from the edge image processing module's connection instruction based on the connection information between the edge image processing module and the sub-identification module to obtain the determination instruction information, and then adds the information to the instruction similarity determination queue.

[0260] If the judgment instruction information from different sub-recognition modules is consistent, the edge image processing module is trusted in real time, and the data of the edge processing module is added to the valid data set;

[0261] If the judgment instruction information from different sub-recognition modules is not completely consistent, the validity of the edge processing module is determined under the principle of minority obeys majority. If it is valid, the data of the edge image processing module is added to the valid data set;

[0262] If the judgment instruction information from different sub-recognition modules is completely inconsistent, the data of the edge image processing module is discarded and marked as a fault;

[0263] Step B, edge image processing module image data validity judgment: When multiple edge image processing modules initiate data to the sub-identification module, the heterogeneous execution mimicry judgment unit removes the verification information from the edge image processing module data based on the connection information between the edge image processing module and the sub-identification module, obtains data judgment instruction information, and adds it to the instruction similarity judgment queue.

[0264] If it is determined that the data judgment instruction information from different edge image processing modules is consistent, the current data of the edge image processing module is trusted in real time and the current data of the edge image processing module is added to the valid data set;

[0265] If the data judgment instruction information from different edge image processing modules is not completely consistent, the validity of the edge image processing module is judged under the principle of minority obeys majority. If it is valid, the current data of the edge image processing module is added to the valid data set;

[0266] If the data judgment instruction information from different sub-recognition modules is completely inconsistent, the current data is discarded;

[0267] Step C: Initialize multiple heterogeneous redundant sub-identification modules, determine the number of sub-identification modules to be N, assign unique identifiers to each sub-identification module, and record them as ID_1, ID_2, ..., ID_N; set the parameters of the bee algorithm, the number of bees to be M, the pheromone volatility coefficient to be rho, ranging from 0 to 1, and the pheromone intensity to be Q; set the de-ignition algorithm parameters, and the denoising intensity parameter to be alpha; start the watchdog unit and set the count value update period to T;

[0268] Step D, collecting rice pest and disease image samples and dividing them into training set, validation set and test set; performing digital processing on the images, including floating point algorithm, integer method, shift method and average method;

[0269] Then, filtering is performed to remove noise in the image, and filtering algorithms including mean filtering and median filtering can be used;

[0270] Then enhancement processing is performed to highlight the features of the image, including contrast enhancement and brightness adjustment;

[0271] Finally, segmentation processing is performed to separate the rice in the image from the background, and pattern recognition methods and book recognition are used to extract and analyze the internal correlation and abstract model;

[0272] Step E: training the neural network system based on convolution operation on rice diseases and insect pests respectively;

[0273] Select a convolution kernel and perform a convolution operation on the image. The convolution formula is: (output feature map) = ∑ (input image * convolution kernel). After the convolution step, a feature map is formed. Configure the parameters of the full convolution network, including the parameters of the convolution layer, pooling layer, and upsampling layer;

[0274] Calculate the feature map size to extract rice pest and disease characteristics. The feature map size calculation formula is:

[0275] (Output feature map size) = (input image size - convolution kernel size + 2 * padding) / stride + 1;

[0276] The cross-stage partial network of the C3F structure is used to segment and fuse feature maps to enhance feature expression capabilities. The multi-level feature fusion accurately captures the characteristics of rice at different growth stages and complexity.

[0277] Precision, recall, F1 score, and accuracy are used as performance metrics to determine the accuracy of the classification model:

[0278] The accuracy calculation formula is: Precision = TP / (TP + FP);

[0279] The recall calculation formula is: Recall = TP / (TP + FN);

[0280] The F1 score calculation formula is: 2*(precision*recall) / (precision+recall);

[0281] The accuracy calculation formula is: Accuracy = (TP + TN) / (TP + TN + FP + FN).

[0282] Step F, selecting the optimal sub-recognition module combination based on the comprehensive evaluation value, assigning the pest and disease data to be identified to the optimal sub-recognition module combination, and each sub-recognition module independently performs pest and disease identification, including:

[0283] The bee algorithm is used for feature extraction. The formula for calculating pheromone concentration in the bee algorithm is:

[0284] Tau_ij(t)=(1-rho)Tau_ij(t-1)+DeltaTau_ij, where Tau_ij(t) represents the pheromone concentration from feature i to feature j at time t, and rho is the pheromone volatility coefficient;

[0285] DeltaTau_ij is the increment of pheromone left by the bee on the path in this cycle, and the calculation formula is DeltaTau_ij = Q / L_k, where Q is the pheromone intensity and L_k is the length of the path walked by the kth bee;

[0286] The de-noising algorithm is used to denoise the data. The calculation formula of the new eigenvalue after denoising in the de-noising algorithm is:

[0287] NewFeatureValue=FeatureValue(1-alpha), where FeatureValue is the original feature value and alpha is the denoising strength parameter;

[0288] The sub-recognition module summarizes the recognition results and checks for errors. If there are any discrepancies in the results, the error checking mechanism is activated. The new pheromone concentration calculation formula in the error checking mechanism is:

[0289] Tau_ij(t)=(1-rho)Tau_ij(t-1)+DeltaTau_ij', where DeltaTau_ij' is the pheromone increment left by the bees on the path during the error checking process;

[0290] At the same time, the adjusted denoising intensity parameter alpha' is used for denoising, and the new feature value calculation formula is: NewFeatureValue'=FeatureValue(1-alpha');

[0291] The watchdog unit monitors the operation of the sub-identification modules. If a fault is found, it automatically resets and reports the error to other sub-identification modules, re-evaluates the sub-identification modules and generates a new sub-module combination, redistributing the faulty module data.

[0292] Step F, result optimization, combines the plant growth status recognition results with the pest and disease recognition module to optimize the pest and disease recognition results, and finally outputs the pest and disease recognition results, including:

[0293] The collected image data is fuzzy processed using a defuzzified trapezoidal function to reduce noise and uncertainty. The defuzzified trapezoidal function can be expressed as:

[0294] Where a, b, c, and d are the parameters of the trapezoidal function;

[0295] A deep belief network load forecasting algorithm based on classification and recognition is used to make a preliminary prediction of the plant growth status;

[0296] A deep belief network composed of multiple stacked restricted Boltzmann machines (RBMs) is called, and unsupervised learning methods and supervised learning methods are used for fusion training to obtain growth status prediction results;

[0297] The features in the growth status prediction results are integrated with the features in the pest and disease identification results. Based on experience and data analysis, different weights are assigned to the growth status features and pest and disease characteristics. According to the comprehensive evaluation health index, the pest and disease identification results are optimized to finally obtain the pest and disease damage identification results.

[0298] This algorithm leverages plant growth status predictions to further optimize pest and disease identification, improving accuracy and reliability. By integrating growth status information with pest and disease characteristics, it comprehensively assesses plant health and provides stronger support for precise pest and disease control.

[0299] Furthermore, calling the unsupervised learning method and the supervised learning method for fusion training includes:

[0300] First, each RBM is pre-trained using unsupervised learning to extract features from the image;

[0301] Then, supervised learning is used to fine-tune the entire deep belief network to improve the accuracy of prediction. During the prediction process, the extracted features are input into the deep belief network to obtain the prediction results of the plant growth status.

[0302] Combined with the grey potential recognition algorithm, the growth status of plants is further analyzed to discover the potential laws and changing trends in the growth status of plants;

[0303] The calculation formula of the grey potential recognition algorithm is:

[0304]

[0305] where x (0) (i) is the original data sequence, X (1) (k) is the cumulative generated sequence;

[0306] By analyzing the cumulatively generated sequence, the gray potential value of the plant growth status can be obtained, thereby judging the growth trend and health status of the plant.

[0307] Although the above describes the illustrative specific embodiments of the present invention so that those skilled in the art can understand the present invention, the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, all inventions and creations based on the concepts of the present invention are protected.

Claims

1. A pest and disease identification system, characterized by: The pest identification system includes: Multiple image acquisition modules for acquiring image information of crops or plants; An edge image processing module and a thermal image acquisition module connected to the image acquisition module. The edge image processing module is used to pre-process and analyze the collected visual images and thermal images. The thermal image acquisition module is connected through a light discrimination unit and is used to collect thermal images for supplementation when the light intensity is lower than a predefined threshold. The plurality of edge image processing modules are connected to the pest and disease identification module and the growth status identification module through the heterogeneous execution mimicry judgment unit; The pest identification unit is also connected to a user interaction module, which is responsible for providing an interactive interface between the user and the system, including a display module, an operation module, and an alarm module; The pest and disease identification module includes multiple heterogeneous redundant sub-identification modules. The heterogeneous redundant sub-identification modules back up each other, check each other's errors, and work in parallel. Each heterogeneous redundant sub-identification module is equipped with a watchdog unit. The watchdog unit regularly updates the count value. Otherwise, it will automatically reset the heterogeneous redundant sub-identification module and report an error to other heterogeneous redundant sub-identification modules. The heterogeneous execution mimetic decision unit has a built-in online mimetic decider and synchronizer; Step 1: edge image processing module connection validity judgment, including: When the edge image processing module initiates a task to the sub-identification module, the heterogeneous execution mimic judgment unit removes the authentication information from the connection instruction of the edge image processing module based on the connection information between the edge image processing module and the sub-identification module, obtains the judgment instruction information, and then adds it to the instruction similarity judgment queue; If the judgment instruction information from different sub-recognition modules is consistent, the edge image processing module is trusted in real time, and the data of the edge processing module is added to the valid data set; If the judgment instruction information from different sub-recognition modules is not completely consistent, the validity of the edge processing module is determined under the principle of minority obeys majority. If it is valid, the data of the edge processing module is added to the valid data set; If the judgment instruction information from different sub-recognition modules is completely inconsistent, the data of the edge image processing module is discarded and marked as a fault; Step 2: edge image processing module image data validity judgment, including: When multiple edge image processing modules send data to the sub-identification module, the heterogeneous execution mimic judgment unit removes the verification information from the edge image processing module's data based on the connection information between the edge image processing module and the sub-identification module, obtains the data judgment instruction information, and adds it to the instruction similarity judgment queue; If it is determined that the data judgment instruction information from different edge image processing modules is consistent, the current data of the edge image processing module is trusted in real time and the current data of the edge image processing module is added to the valid data set; If the data judgment instruction information from different edge image processing modules is not completely consistent, the validity of the edge image processing module is judged under the principle of minority obeys majority. If it is valid, the current data of the edge image processing module is added to the valid data set; If it is determined that the data judgment instruction information from different sub-recognition modules are completely inconsistent, the current data will be discarded.

2. The pest identification system according to claim 1, characterized in that: The pest and disease identification module performs the following algorithm steps to complete pest and disease identification: Step 1: Initialize multiple heterogeneous redundant sub-identification modules, determine the number of sub-identification modules to be N, and assign unique identifiers to each sub-identification module, denoted as ID_1, ID_2, ..., ID_N; set the bee algorithm parameters, the number of bees to be M, the pheromone volatility coefficient to be rho, ranging from 0 to 1, and the pheromone intensity to be Q; set the de-ignition algorithm parameters, and the denoising intensity parameter to be alpha; start the watchdog unit and set the count value update period to T; Step 2: Collect rice pest and disease image samples and divide them into training set, validation set and test set; digitally process the images, including floating point algorithm, integer method, shift method and average method; Then, filtering is performed to remove noise in the image, and filtering algorithms including mean filtering and median filtering can be used; Then enhancement processing is performed to highlight the features of the image, including contrast enhancement and brightness adjustment; Finally, segmentation processing is performed to separate the rice in the image from the background, and pattern recognition methods and book recognition are used to extract and analyze the internal correlation and abstract model; Step 3: The convolutional neural network system is trained on rice diseases and insect pests respectively; Select the convolution kernel and perform convolution operation on the image. After the convolution step, a feature map is formed. Configure the parameters of the full convolution network, including the parameters of the convolution layer, pooling layer, and upsampling layer. The convolution formula is: Output feature map = ∑(input image * convolution kernel); Calculate the feature map size to extract rice pest and disease characteristics. The feature map size calculation formula is: Output feature map size = (input image size - convolution kernel size + 2 * padding) / step size + 1; The cross-stage partial network of the C2F structure is used to segment and fuse the feature maps, and the characteristics of different growth stages and complexity of rice are accurately captured through multi-level feature fusion; Precision, recall, F1 score, and accuracy are used as performance metrics to determine the accuracy of the classification model: The precision calculation formula is: Precision = TP / (TP + FP), where TP is the true positive, which is the number of samples that are actually infected with pests and diseases and are correctly identified as such; FP is the false positive, which is the number of samples that are actually free of pests and diseases but are incorrectly identified as such. The recall calculation formula is: Recall = TP / (TP+FN); FN is a false negative example; The F1 score calculation formula is: 2*(precision*recall) / (precision+recall); The accuracy calculation formula is: Accuracy = (TP + TN) / (TP + TN + FP + FN); TN is the true negative example, which refers to the number of samples that are correctly judged as negative examples; FN is the false negative example, which refers to the number of samples that actually have pests and diseases but are mistakenly identified as not having pests and diseases; Step 4: Select the optimal sub-recognition module combination based on the comprehensive evaluation value, assign the pest and disease data to be identified to the optimal sub-recognition module combination, and each sub-recognition module independently performs pest and disease identification, including: The bee algorithm is used for feature extraction. The formula for calculating pheromone concentration in the bee algorithm is: Tau_ij(t)=(1-rho)Tau_ij(t-1)+DeltaTau_ij, where Tau_ij(t) represents the pheromone concentration from feature i to feature j at time t, and rho is the pheromone volatility coefficient; DeltaTau_ij is the increment of pheromone left by the bee on the path in this cycle, and the calculation formula is DeltaTau_ij = Q / L_k, where Q is the pheromone intensity and L_k is the length of the path walked by the kth bee; The de-noising algorithm is used to denoise the data. The calculation formula of the new eigenvalue after denoising in the de-noising algorithm is: NewFeatureValue=FeatureValue(1-alpha), where FeatureValue is the original feature value and alpha is the denoising strength parameter; The sub-recognition module summarizes the recognition results and checks for errors. If there is a discrepancy between the results, the error checking mechanism is activated. The new pheromone concentration calculation formula in the error checking mechanism is: Tau_ij(t)=(1-rho)Tau_ij(t-1)+DeltaTau_ij', where DeltaTau_ij' is the pheromone increment left by the bees on the path during the error checking process; At the same time, the adjusted denoising intensity parameter alpha' is used for denoising, and the new feature value calculation formula is: NewFeatureValue'=FeatureValue(1-alpha'); The watchdog unit monitors the operation of the sub-identification modules. If a fault is found, it will automatically reset and report the error to other sub-identification modules, re-evaluate the sub-identification modules and generate a new sub-module combination, redistribute the faulty module data, and finally output the pest and disease nuisance identification results.

3. The pest identification system according to claim 2, characterized in that: The floating-point algorithm processes the pixel values ​​of the image according to specific floating-point operation rules. The formula is: New pixel value = f (original pixel value), where f is a specific floating-point operation function; The integer method adjusts the pixel value through integer operations. The formula is: (new pixel value) = g (original pixel value), where g is an integer operation function; The shift method performs a shift operation on the pixel value according to the shift rule. The formula is (new pixel value) = h (original pixel value), where h is the shift function; The average value method calculates the average value of adjacent pixels as the new pixel value. The formula is (new pixel value) = (pixel value_1 + pixel value_2 + ... + pixel value_n) / n, where n is the number of pixels participating in the average.

4. The pest identification system according to claim 2, characterized in that: For each pixel point, the mean filter takes the average value of the pixel values ​​in its surrounding neighborhood as the new pixel value of the point. The formula is: New pixel value = (pixel_value_1 + pixel_value_2 + ... + pixel_value_k) / k, where k is the number of pixels in the neighborhood; Median filtering takes the median of the pixel values ​​in the neighborhood as the new pixel value, that is, sorts the pixel values ​​in the neighborhood and takes the middle value as the new pixel value; Contrast enhancement enhances the contrast by adjusting the grayscale range of the image. The formula is: New pixel value = a*(original pixel value)+b, where a and b are adjustment parameters; Brightness adjustment emphasizes features by increasing or decreasing the overall brightness of the image. The formula is: New pixel value = (original pixel value) + c, where c is the brightness adjustment parameter; Threshold segmentation selects a suitable threshold value, classifies pixels with values ​​greater than the threshold value into one category, and pixels with values ​​less than the threshold value into another category, thereby achieving image segmentation. The formula is: New pixel value = d*(original pixel value), where d is the value determined based on the threshold.

5. The pest identification system according to claim 2, characterized in that: The calculation formula of the comprehensive evaluation value EvaluationValue_i is: =omega_1Accuracy_i-omega_2(1 / MTBF_i)+omega_3*ResponseTime_i; Where EvaluationValue_i is the comprehensive evaluation value of sub-identification module i, omega_1, omega_2, and omega_3 are weight coefficients, Accuracy_i is the accuracy of sub-identification module i, MTBF_i is the mean time between failures of sub-identification module i, and ResponseTime_i is the response time of sub-identification module i.

6. The pest identification system according to claim 2, characterized in that: The watchdog unit monitors the operation of the sub-identification module. If the count value is not updated on time, the sub-identification module is automatically reset and an error is reported to other sub-identification modules. After receiving the error message, other sub-identification modules restart the evaluation of the sub-identification unit, collect the operating data of the remaining normal sub-identification modules in the current stage, recalculate the comprehensive evaluation value according to the full life cycle valuation and identification algorithm, regenerate a new sub-module unit combination, and reallocate the data originally allocated to the faulty sub-identification module to the sub-identification module in the new combination for processing.

7. A method for identifying pests and diseases, characterized by: The pest identification method is based on the system according to any one of claims 1 to 6, and the method comprises: Step A: Determine the validity of the edge image processing module connection. When the edge image processing module initiates a task to the sub-identification module, the heterogeneous execution mimicry determination unit removes the authentication information from the edge image processing module's connection instruction based on the connection information between the edge image processing module and the sub-identification module to obtain the determination instruction information, and then adds it to the instruction similarity determination queue. If the judgment instruction information from different sub-recognition modules is consistent, the edge image processing module is trusted in real time, and the data of the edge processing module is added to the valid data set; If the judgment instruction information from different sub-recognition modules is not completely consistent, the validity of the edge processing module is determined under the principle of minority obeys majority. If it is valid, the data of the edge image processing module is added to the valid data set; If the judgment instruction information from different sub-recognition modules is completely inconsistent, the data of the edge image processing module is discarded and marked as a fault; Step B, edge image processing module image data validity judgment: When multiple edge image processing modules initiate data to the sub-identification module, the heterogeneous execution mimicry judgment unit removes the verification information from the edge image processing module data based on the connection information between the edge image processing module and the sub-identification module, obtains data judgment instruction information, and adds it to the instruction similarity judgment queue. If it is determined that the data judgment instruction information from different edge image processing modules is consistent, the current data of the edge image processing module is trusted in real time and the current data of the edge image processing module is added to the valid data set; If the data judgment instruction information from different edge image processing modules is not completely consistent, the validity of the edge image processing module is judged under the principle of minority obeys majority. If it is valid, the current data of the edge image processing module is added to the valid data set; If the data judgment instruction information from different sub-recognition modules is completely inconsistent, the current data is discarded; Step C: Initialize multiple heterogeneous redundant sub-identification modules, determine the number of sub-identification modules to be N, assign unique identifiers to each sub-identification module, and record them as ID_1, ID_2, ..., ID_N; set the parameters of the bee algorithm, the number of bees to be M, the pheromone volatility coefficient to be rho, ranging from 0 to 1, and the pheromone intensity to be Q; set the de-ignition algorithm parameters, and the denoising intensity parameter to be alpha; start the watchdog unit and set the count value update period to T; Step D, collecting rice pest and disease image samples and dividing them into training set, validation set and test set; performing digital processing on the images, including floating point algorithm, integer method, shift method and average method; Then, filtering is performed to remove noise in the image, and filtering algorithms including mean filtering and median filtering can be used; Then enhancement processing is performed to highlight the features of the image, including contrast enhancement and brightness adjustment; Finally, segmentation processing is performed to separate the rice in the image from the background, and pattern recognition methods and book recognition are used to extract and analyze the internal correlation and abstract model; Step E: training the neural network system based on convolution operation on rice diseases and insect pests respectively; Select a convolution kernel and perform a convolution operation on the image. The convolution formula is: (output feature map) = ∑ (input image * convolution kernel). After the convolution step, a feature map is formed. Configure the parameters of the full convolution network, including the parameters of the convolution layer, pooling layer, and upsampling layer; Calculate the feature map size to extract rice pest and disease characteristics. The feature map size calculation formula is: (Output feature map size) = (input image size - convolution kernel size + 2 * padding) / stride + 1; The cross-stage partial network of the C2F structure is used to segment and fuse the feature maps to enhance the feature expression ability. The multi-level feature fusion accurately captures the characteristics of rice at different growth stages and complexity. Precision, recall, F1 score, and accuracy are used as performance metrics to determine the accuracy of the classification model: The accuracy calculation formula is: Precision = TP / (TP + FP); The recall calculation formula is: Recall = TP / (TP + FN); The F1 score calculation formula is: 2*(precision*recall) / (precision+recall); The accuracy calculation formula is: Accuracy = (TP + TN) / (TP + TN + FP + FN). Step F, selecting the optimal sub-recognition module combination based on the comprehensive evaluation value, assigning the pest and disease data to be identified to the optimal sub-recognition module combination, and each sub-recognition module independently performs pest and disease identification, including: The bee algorithm is used for feature extraction. The formula for calculating pheromone concentration in the bee algorithm is: Tau_ij(t)=(1-rho)Tau_ij(t-1)+DeltaTau_ij, where Tau_ij(t) represents the pheromone concentration from feature i to feature j at time t, and rho is the pheromone volatility coefficient; DeltaTau_ij is the increment of pheromone left by the bee on the path in this cycle, and the calculation formula is DeltaTau_ij = Q / L_k, where Q is the pheromone intensity and L_k is the length of the path walked by the kth bee; The de-noising algorithm is used to denoise the data. The calculation formula of the new eigenvalue after denoising in the de-noising algorithm is: NewFeatureValue=FeatureValue(1-alpha), where FeatureValue is the original feature value and alpha is the denoising strength parameter; The sub-recognition module summarizes the recognition results and checks for errors. If there is a discrepancy between the results, the error checking mechanism is activated. The new pheromone concentration calculation formula in the error checking mechanism is: Tau_ij(t)=(1-rho)Tau_ij(t-1)+DeltaTau_ij', where DeltaTau_ij' is the pheromone increment left by the bees on the path during the error checking process; At the same time, the adjusted denoising intensity parameter alpha' is used for denoising, and the new feature value calculation formula is: NewFeatureValue'=FeatureValue(1-alpha'); The watchdog unit monitors the operation of the sub-identification modules. If a fault is found, it automatically resets and reports the error to other sub-identification modules, re-evaluates the sub-identification modules and generates a new sub-module combination, redistributing the faulty module data. Step F, result optimization, combines the plant growth status recognition results with the pest and disease recognition module to optimize the pest and disease recognition results, and finally outputs the pest and disease recognition results, including: The collected image data is fuzzy processed using a defuzzified trapezoidal function to reduce noise and uncertainty. The defuzzified trapezoidal function can be expressed as: Where a, b, c, and d are the parameters of the trapezoidal function; A deep belief network load forecasting algorithm based on classification and recognition is used to make a preliminary prediction of the plant growth status; A deep belief network composed of multiple stacked restricted Boltzmann machines (RBMs) is called, and unsupervised learning methods and supervised learning methods are used for fusion training to obtain growth status prediction results; The features in the growth status prediction results are integrated with the features in the pest and disease identification results. Based on experience and data analysis, different weights are assigned to the growth status features and pest and disease characteristics. According to the comprehensive evaluation health index, the pest and disease identification results are optimized to finally obtain the pest and disease damage identification results.

8. The method for identifying pests and diseases according to claim 7, characterized in that: Calling unsupervised learning methods and supervised learning methods for fusion training includes: First, each RBM is pre-trained using unsupervised learning to extract features from the image; Then, supervised learning is used to fine-tune the entire deep belief network to improve the accuracy of prediction. During the prediction process, the extracted features are input into the deep belief network to obtain the prediction results of the plant growth status. Combined with the grey potential recognition algorithm, the growth status of plants is further analyzed to discover the potential laws and changing trends in the growth status of plants; The calculation formula of the grey potential recognition algorithm is: where x (0) (i) is the original data sequence, X (1) (k) is the cumulative generated sequence; By analyzing the cumulatively generated sequence, the gray potential value of the plant growth status can be obtained, thereby judging the growth trend and health status of the plant.

Images