Real-time rice resistance monitoring system and method

Abstract

The invention relates to a rice real-time resistance monitoring system and a method thereof, and solves the technical problems of poor real-time performance, poor precision and complexity, an acquisition system comprises a data acquisition module and a transmission module, and the data acquisition module comprises a sensor and a video and image transmission inlet; the transmission module can be used for transmitting rice graph or video data of the data acquisition module; the video and image transmission inlet can be used for uploading rice graphs or video data; the data analysis and processing module is used for evaluating and analyzing rice disease and pest resistance data according to rice graph or video data; the display module is used for displaying resistance monitoring result characteristics in a rice virtual space; and the control and decision module is used for formulating prevention and control measures according to the analysis result and providing an executable technical scheme, so that the problems are well solved, and the method can be applied to rice resistance monitoring.

Description

Technical Field

[0001] The invention relates to the field of rice monitoring, and in particular to a real-time rice resistance monitoring system and method. Background Art

[0002] As the world's largest rice producer and consumer, rice cultivation plays an important role in agricultural production in China. Rice is not only one of China's main food crops, but also an important source of food security and farmers' income. However, there are many challenges in the rice cultivation process, including monitoring growth conditions and preventing and controlling pests and diseases. Traditional rice monitoring and pest identification methods rely on manual operations, which consumes manpower and material resources, and lacks timeliness and accuracy.

[0003] The present invention provides a new rice real-time resistance monitoring system and method, which can solve the above technical problems. Summary of the invention

[0004] The technical problem to be solved by the present invention is the poor real-time performance and poor precision in the prior art. A new rice real-time resistance monitoring system and method are provided, which has the characteristics of good real-time performance, high precision and convenient use.

[0005] In order to solve the above technical problems, the technical solutions adopted are as follows:

[0006] A rice real-time resistance monitoring system, comprising:

[0007] A data acquisition module and a transmission module, wherein the data acquisition module includes a sensor, a video and image transmission port; the transmission module can be used to transmit the rice graphics or video data of the data acquisition module; the video and image transmission port can be used to upload the rice graphics or video data;

[0008] Data analysis and processing module, used to evaluate and analyze rice pest resistance data based on rice graphics or video data;

[0009] A display module, used for displaying the characteristics of resistance monitoring results in a rice virtual space;

[0010] The control and decision-making module is used to formulate and implement prevention and control measures based on the analysis results;

[0011] The data analysis and processing module performs the following algorithm steps:

[0012] Step 1: Data preparation, including data collection, graphic preprocessing and graphic standards;

[0013] Step 2: Initialization, including YOLO V8 initialization and re-melting signal honey algorithm initialization;

[0014] The YOLO V8 initialization includes: initializing the weight parameter W of the YOLO V8 model 0 , you can use the weights of a pre-trained model or random initialization;

[0015] Set the hyperparameters of YOLO V8, including learning rate α∈(0.001-0.1), batch size batch_size, and number of training rounds epochs;

[0016] The initialization of the remelting signal honey algorithm includes: determining the number of signal honeys m, defining the position of the signal honey as a vector X = {x 1 , x 2 , ..., x m}, where x i Indicates the position of the i-th pheromone, corresponding to a set of parameters of the YOLO V8 model; sets the number of iterations T, pheromone volatility coefficient ρ, and pheromone intensity Q;

[0017] Step three, complete feature extraction and target detection;

[0018] Feature extraction includes: assuming that the size of the input image is H×W×3, H=height, W=width, 3=number of channels); input image I, and extract features through the C2f structure of YOLO V8: the C2f structure consists of multiple convolutional layer operations, batch normalization layers and activation function layers; after multiple convolution operations, the image is gradually extracted to obtain a feature map F with a size of h×w×d, where h represents the feature map height, w represents the feature map width, and d represents the feature dimension of the feature map;

[0019] The convolutional layer operation is F out =W*F in +b, where W is the convolution kernel weight, F in is the input feature map, b is the bias term;

[0020] The batch normalization layer is Where μ and σ are the mean and standard deviation respectively, and ∈ is a small predefined constant;

[0021] The feature extraction function uses the formula F = f C2f (I, W); where f C2f The function representing the C2f structure, W is the parameter of YOLO V8; the activation function layer is completed using the formula f(x)=max(0,x);

[0022] Object detection includes: on the feature map F, using the detection head of YOLO V8 to detect objects, the detection head includes a classification branch and a regression branch; for each grid cell, predicting multiple bounding boxes and corresponding category probabilities and confidences;

[0023] Assume that the predicted bounding box set is B = {b 1 , b 2 , ..., b k}, the category probability set is C = {c 1 , c 2 , ..., c k}, the confidence set is P = {p 1 , p 2 , ..., p k}, where k is the number of predicted bounding boxes;

[0024] The bounding box prediction can be expressed as b = (t x , t y , t w , t h ), where t x , t y , t w , t h is the offset relative to the grid unit, and the actual bounding box coordinates are obtained through the conversion formula: y =(σ(t y )+c y )×sb x =(σ(t x )+c x )×s, where c x , c y is the coordinate of the grid cell, s is the grid size, p w , p h is the width and height of the prior box;

[0025] The category probability is calculated by the softmax function: where z i is the output of the classification branch;

[0026] The confidence represents the probability that the predicted bounding box contains the target, which is calculated by the sigmoid function: p = σ(z), where z is the output of the regression branch;

[0027] Step 4: Calculate the fitness function value to complete the fitness evaluation:

[0028] For each bee's position x i Corresponding to a set of YOLO V8 parameters, the YOLO V8 model using this parameter is tested on the validation set; define the fitness function value f(x i ) = AP, which means the average accuracy of the model on the validation set. P(R) is the correlation curve between precision and recall;

[0029] For each image in the validation set, the bounding box predicted by the model is compared with the annotated true bounding box to calculate the precision and recall.

[0030] Accuracy calculation formula: Among them, TP is a true positive example and FP is a false positive example;

[0031] Recall calculation formula: Where FN is a false counterexample;

[0032] Step 5: Complete the movement of the pheromone and the update of pheromone. The movement of the pheromone is that the pheromone selects the moving direction according to the pheromone concentration and heuristic information, including:

[0033] Assume that the message i is at the current position x i Select to move to new position x′ i The probability is P i，j , the calculation formula is: Where τ(x j ) is the position x j The pheromone concentration, η(x j ) is the heuristic information, α and β are weight parameters, and N is the set of optional positions of the message;

[0034] Xinmi chooses to move to a new position based on probability and updates its own position: x′ i =x i +step×dir, where step is the moving step length, which can be adjusted according to the problem, and dir is the moving direction, which can be randomly generated or determined according to certain rules;

[0035] The pheromone update includes pheromone volatilization and pheromone enhancement;

[0036] The pheromone volatilization includes: the pheromone will gradually volatilize over time, and the updated pheromone concentration is τ(x j )=(1-ρ)τ(x j ), where ρ is the pheromone volatility coefficient.

[0037] Pheromone enhancement includes: increasing pheromone according to the fitness value of the honey; for honey i, if it finds a new position x′ i If the corresponding fitness value is better, add pheromone at this position. Where Q is the pheromone intensity, f(x′ i ) is the fitness value of the signal honey in the new position.

[0038] Step 6: Model optimization and iteration: Apply the parameters corresponding to the best position found by Xinmi to the YOLO V8 model: Compare the fitness values ​​of all Xinmi to find the best position x best, which is the position with the largest fitness value; the parameters corresponding to the best position are used as the new parameters of the YOLO V8 model;

[0039] Repeat steps 3 to 5 for multiple iterations until the stopping condition is met and the model optimization is completed;

[0040] Step 7: Model evaluation: Use the test data set to calculate the accuracy, recall, and average precision of the optimized model to complete the evaluation;

[0041] Accuracy calculation formula: Its TP is the true positive sample (the number of correctly predicted positive samples), and FP is the false positive sample (the number of incorrectly predicted positive samples);

[0042] Recall calculation formula: Where FN is the false negative examples (the number of incorrectly predicted negative samples);

[0043] Calculate the average precision based on the precision and recall Where P(R) is the precision-recall curve;

[0044] Step 8: Calculate the rice resistance expression Let rice resistance be R, where N diseased is the number of damaged rice detected, N total is the total number of rice, which is obtained by detecting the affected rice through the YOLO V8 model and then counting the number.

[0045] Working principle of the present invention: The present invention can use remote sensing data or visual data of sensors, or data uploaded by users and searched on the Internet as data graphics sources. On this basis, the present invention creatively sets a rice resistance detection and analysis algorithm, which can realize rice resistance monitoring and analysis efficiently and accurately.

[0046] The data collection in the above scheme includes collecting a large number of image datasets D containing different rice disaster situations:

[0047] D={I 1 , I 2 , ..., I n}, where I i represents the i-th image;

[0048] The graphic preprocessing includes image resizing and normalization. The image resizing is H×W×3. The normalization formula is: Where I is the original image pixel value, I min and I max are the minimum and maximum values ​​of the image pixel values ​​respectively;

[0049] The image annotation includes using an annotation tool to annotate the disaster area in each image with a bounding box and category information to obtain an annotation data set L = {(B 1 , C 1 ), (B 2 , C 2 ),...,(B n , C n )}; Among them, B i represents the set of annotated bounding boxes in the i-th image, C i Represents the corresponding category set;

[0050] The bounding box annotation can be in the form of upper left corner and lower right corner coordinates, that is, b = (x 1 ,y 1 , x 2 ,y 2 ), where (x 1 ,y 1 ) is the coordinate of the upper left corner, (x 2 ,y 2 ) is the lower right corner coordinate.

[0051] The stopping condition in step six is ​​that the maximum number of iterations T is reached or the fitness value no longer increases.

[0052] The display module also displays rice pest data information, including pest type, pest location, and placement measures.

[0053] The present invention also provides a method for real-time resistance monitoring of rice, the method comprising:

[0054] Step 1: The data acquisition module and the transmission module upload the rice resistance monitoring pattern through user upload or sensor upload;

[0055] Step 2: Complete data preparation through data collection and data preprocessing, build the algorithm model and initialize the algorithm model, including YOLO V8 initialization and re-melting signal honey algorithm initialization;

[0056] Step three, complete feature extraction and target detection;

[0057] Feature extraction includes: assuming that the size of the input image is H×W×3, H=height, W=width, 3=number of channels); input image I, and extract features through the C2f structure of YOLO V8: the C2f structure consists of multiple convolutional layer operations, batch normalization layers and activation function layers; after multiple convolution operations, the image is gradually extracted to obtain a feature map F with a size of h×w×d, where h represents the feature map height, w represents the feature map width, and d represents the feature dimension of the feature map;

[0058] The convolutional layer operation is F out=W*F in +b, where W is the convolution kernel weight, F in is the input feature map, b is the bias term;

[0059] The batch normalization layer is Where μ and σ are the mean and standard deviation respectively, and ∈ is a small predefined constant;

[0060] The feature extraction function uses the formula F = f C2f (I, W); where f C2f The function representing the C2f structure, W is the parameter of YOLO V8; the activation function layer is completed using the formula f(x)=max(0,x);

[0061] Object detection includes: on the feature map F, using the detection head of YOLO V8 to detect objects, the detection head includes a classification branch and a regression branch; for each grid cell, predicting multiple bounding boxes and corresponding category probabilities and confidences;

[0062] Assume that the predicted bounding box set is B = {b 1 , b 2 , ..., b k}, the category probability set is C = {c 1 , c 2 , ..., c k}, the confidence set is P = {p 1 , p 2 , ..., p k}, where k is the number of predicted bounding boxes;

[0063] The bounding box prediction can be expressed as b = (t x , t y , t w , t h ), where t x , t y , t w , t h is the offset relative to the grid unit, and the actual bounding box coordinates are obtained through the conversion formula: y =(σ(t y )+c y )×sb x =(σ(t x )+c x )×s, where c x , c y is the coordinate of the grid cell, s is the grid size, p w , p h is the width and height of the prior box;

[0064] The category probability is calculated by the softmax function: where z i is the output of the classification branch;

[0065] The confidence represents the probability that the predicted bounding box contains the target, which is calculated by the sigmoid function: p = σ(z), where z is the output of the regression branch;

[0066] Step 4: Calculate the fitness function value to complete the fitness evaluation:

[0067] For each bee's position x i Corresponding to a set of YOLO V8 parameters, the YOLO V8 model using this parameter is tested on the validation set; define the fitness function value f(x i ) = AP, which means the average accuracy of the model on the validation set. P(R) is the correlation curve between precision and recall;

[0068] For each image in the validation set, the bounding box predicted by the model is compared with the annotated true bounding box to calculate the precision and recall.

[0069] Accuracy calculation formula: Among them, TP is a true positive example and FP is a false positive example;

[0070] Recall calculation formula: Where FN is a false counterexample;

[0071] Step 5: Complete the movement of the pheromone and the update of pheromone. The movement of the pheromone is that the pheromone selects the moving direction according to the pheromone concentration and heuristic information, including:

[0072] Assume that the message i is at the current position x i Select to move to new position x′ i The probability is P i，j , the calculation formula is: Where τ(x j ) is the position x j The pheromone concentration, η(x j ) is the heuristic information, α and β are weight parameters, and N is the set of optional positions of the message;

[0073] Xinmi chooses to move to a new position based on probability and updates its own position: x′ i =x i +step×dir, where step is the moving step length, which can be adjusted according to the problem, and dir is the moving direction, which can be randomly generated or determined according to certain rules;

[0074] The pheromone update includes pheromone volatilization and pheromone enhancement;

[0075] The pheromone volatilization includes: the pheromone will gradually volatilize over time, and the updated pheromone concentration is τ(x j )=(1-ρ)τ(x j ), where ρ is the pheromone volatility coefficient.

[0076] Pheromone enhancement includes: increasing pheromone according to the fitness value of the honey; for honey i, if it finds a new position x′ i If the corresponding fitness value is better, add pheromone at this position. Where Q is the pheromone intensity, f(x′ i ) is the fitness value of the signal honey in the new position.

[0077] Step 6: Model optimization and iteration: Apply the parameters corresponding to the best position found by Xinmi to the YOLO V8 model: Compare the fitness values ​​of all Xinmi to find the best position x best , which is the position with the largest fitness value; the parameters corresponding to the best position are used as the new parameters of the YOLO V8 model;

[0078] Repeat steps 3 to 5 for multiple iterations until the stopping condition is met and the model optimization is completed;

[0079] Step 7: Model evaluation: Use the test data set to calculate the accuracy, recall, and average precision of the optimized model to complete the evaluation;

[0080] Accuracy calculation formula: Its TP is the true positive sample (the number of correctly predicted positive samples), and FP is the false positive sample (the number of incorrectly predicted positive samples);

[0081] Recall calculation formula: Where FN is the false negative examples (the number of incorrectly predicted negative samples);

[0082] Calculate the average precision based on the precision and recall Where P(R) is the precision-recall curve;

[0083] Step 8: Calculate the rice resistance expression Let rice resistance be R, where N diseased is the number of damaged rice detected, N total is the total number of rice, which is obtained by detecting the affected rice through the YOLO V8 model and then counting the number.

[0084] Similar to the principle of the system, this invention can realize rice resistance monitoring efficiently, accurately and conveniently.

[0085] Furthermore, the YOLO V8 initialization includes initializing the weight parameter W of the YOLO V8 model. 0 ;

[0086] Set the hyperparameters of YOLO V8, including learning rate α∈(0.001-0.1), batch size batch_size, and number of training rounds epochs;

[0087] The remelting signal honey algorithm initialization includes: determining the number of signal honeys m, defining the position of the signal honey as a vector X={x 1 , x 2 , ..., x m}, where x i Indicates the position of the i-th pheromone, corresponding to a set of parameters of the YOLO V8 model; sets the number of iterations T, pheromone volatility coefficient ρ, and pheromone intensity Q.

[0088] In step 2, data collection includes collecting a large number of image datasets D containing different rice disaster situations:

[0089] D={I 1 , I 2 , ..., I n}, where I i represents the i-th image;

[0090] The graphic preprocessing includes image resizing and normalization. The image resizing is H×W×3. The normalization formula is: Where I is the original image pixel value, I min and I max are the minimum and maximum values ​​of the image pixel values ​​respectively;

[0091] The image annotation includes using an annotation tool to annotate the disaster area in each image with a bounding box and category information to obtain an annotation data set L = {(B 1 , C 1 ), (B 2 , C 2 ),...,(B n , C n )}; Among them, B i represents the set of annotated bounding boxes in the i-th image, C i Represents the corresponding category set;

[0092] The bounding box annotation can be in the form of upper left corner and lower right corner coordinates, that is, b = (x 1 ,y 1 , x 2 ,y 2 ), where (x 1 ,y 1) is the coordinate of the upper left corner, (x 2 ,y 2 ) is the lower right corner coordinate. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] The present invention is further described below with reference to the accompanying drawings and examples.

[0094] Figure 1 , in Example 1 DETAILED DESCRIPTION

[0095] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples. It should be understood that the specific examples described herein are only used to explain the present invention and are not used to limit the present invention.

[0096] Example 1

[0097] This example provides a rice real-time resistance monitoring system, comprising:

[0098] A data acquisition module and a transmission module, wherein the data acquisition module includes a sensor, a video and image transmission port; the transmission module can be used to transmit the rice graphics or video data of the data acquisition module; the video and image transmission port can be used to upload the rice graphics or video data;

[0099] Data analysis and processing module, used to evaluate and analyze rice pest and disease resistance data based on rice graphics or video data;

[0100] A display module, used for displaying the characteristics of resistance monitoring results in a rice virtual space;

[0101] The control and decision-making module is used to formulate and implement prevention and control measures based on the analysis results;

[0102] The data analysis and processing module performs the following algorithm steps:

[0103] Step 1: Data preparation, including data collection, graphic preprocessing and graphic standards.

[0104] Step 2: Initialization, including YOLO V8 initialization and re-melting signal honey algorithm initialization;

[0105] The YOLO V8 initialization includes: initializing the weight parameter W of the YOLO V8 model 0 , you can use the weights of a pre-trained model or random initialization;

[0106] Set the hyperparameters of YOLO V8, including learning rate α∈(0.001-0.1), batch size batch_size, and number of training rounds epochs;

[0107] The initialization of the remelting signal honey algorithm includes: determining the number of signal honeys m, defining the position of the signal honey as a vector X = {x 1 , x 2 , ..., x m}, where x i Indicates the position of the i-th pheromone, corresponding to a set of parameters of the YOLO V8 model; sets the number of iterations T, pheromone volatility coefficient ρ, and pheromone intensity Q;

[0108] Step three, complete feature extraction and target detection;

[0109] Feature extraction includes: assuming that the size of the input image is H×W×3, H=height, W=width, 3=number of channels); input image I, and extract features through the C2f structure of YOLO V8: the C2f structure consists of multiple convolutional layer operations, batch normalization layers and activation function layers; after multiple convolution operations, the image is gradually extracted to obtain a feature map F with a size of h×w×d, where h represents the feature map height, w represents the feature map width, and d represents the feature dimension of the feature map;

[0110] The convolutional layer operation is F out =W*F in +b, where W is the convolution kernel weight, F in is the input feature map, b is the bias term;

[0111] The batch normalization layer is Where μ and σ are the mean and standard deviation respectively, and ∈ is a small predefined constant;

[0112] The feature extraction function uses the formula F = f C2f (I,W); where f C2f The function representing the C2f structure, W is the parameter of YOLO V8; the activation function layer is completed using the formula f(x)=max(0,x);

[0113] Object detection includes: on the feature map F, using the detection head of YOLO V8 to detect objects, the detection head includes a classification branch and a regression branch; for each grid cell, predicting multiple bounding boxes and corresponding category probabilities and confidences;

[0114] Assume that the predicted bounding box set is B = {b 1 , b 2 , ..., b k}, the category probability set is C = {c 1 , c 2 , ..., c k}, the confidence set is P = {p 1 , p 2 , ..., p k}, where k is the number of predicted bounding boxes;

[0115] The bounding box prediction can be expressed as b = (t x , t y , t w , t h ), where t x , t y , t w , t h is the offset relative to the grid unit, and the actual bounding box coordinates are obtained through the conversion formula: y =(σ(t y )+c y )×sb x =(σ(t x )+c x )×s, where c x , c y is the coordinate of the grid cell, s is the grid size, p w , p h is the width and height of the prior box;

[0116] The category probability is calculated by the softmax function: where z i is the output of the classification branch;

[0117] The confidence represents the probability that the predicted bounding box contains the target, which is calculated by the sigmoid function: p = σ(z), where z is the output of the regression branch;

[0118] Step 4: Calculate the fitness function value to complete the fitness evaluation:

[0119] For each bee's position x i Corresponding to a set of YOLO V8 parameters, the YOLO V8 model using this parameter is tested on the validation set; define the fitness function value f(x i ) = AP, which means the average accuracy of the model on the validation set. P(R) is the correlation curve between precision and recall;

[0120] For each image in the validation set, the bounding box predicted by the model is compared with the annotated true bounding box to calculate the precision and recall.

[0121] Accuracy calculation formula: Among them, TP is a true positive example and FP is a false positive example;

[0122] Recall calculation formula: Where FN is a false counterexample;

[0123] Step 5: Complete the movement of the pheromone and the update of pheromone. The movement of the pheromone is that the pheromone selects the moving direction according to the pheromone concentration and heuristic information, including:

[0124] Assume that the message i is at the current position x i Select to move to new position x′ i The probability is P i，j , the calculation formula is: Where τ(x j ) is the position x j The pheromone concentration, η(x j ) is the heuristic information, α and β are weight parameters, and N is the set of optional positions of the message;

[0125] Xinmi chooses to move to a new position based on probability and updates its own position: x′ i =x i +step×dir, where step is the moving step length, which can be adjusted according to the problem, and dir is the moving direction, which can be randomly generated or determined according to certain rules;

[0126] The pheromone update includes pheromone volatilization and pheromone enhancement;

[0127] The pheromone volatilization includes: the pheromone will gradually volatilize over time, and the updated pheromone concentration is τ(x j )=(1-ρ)τ(x j ), where ρ is the pheromone volatility coefficient.

[0128] Pheromone enhancement includes: increasing pheromone according to the fitness value of the honey; for honey i, if it finds a new position x′ i If the corresponding fitness value is better, add pheromone at this position. Where Q is the pheromone intensity, f(x′ i ) is the fitness value of the signal honey in the new position.

[0129] Step 6: Model optimization and iteration: Apply the parameters corresponding to the best position found by Xinmi to the YOLO V8 model: Compare the fitness values ​​of all Xinmi to find the best position x best , which is the position with the largest fitness value; the parameters corresponding to the best position are used as the new parameters of the YOLO V8 model;

[0130] Repeat steps 3 to 5 for multiple iterations until the stopping condition is met and the model optimization is completed;

[0131] Step 7: Model evaluation: Use the test data set to calculate the accuracy, recall, and average precision of the optimized model to complete the evaluation;

[0132] Accuracy calculation formula: Its TP is the true positive sample (the number of correctly predicted positive samples), and FP is the false positive sample (the number of incorrectly predicted positive samples);

[0133] Recall calculation formula: Where FN is the false negative examples (the number of incorrectly predicted negative samples);

[0134] Calculate the average precision based on the precision and recall Where P(R) is the precision-recall curve;

[0135] Step 8: Calculate the rice resistance expression Let rice resistance be R, where N diseased is the number of damaged rice detected, N total is the total number of rice, which is obtained by detecting the affected rice through the YOLO V8 model and then counting the number.

[0136] Working principle of the present invention: The present invention can use remote sensing data or visual data of sensors, or data uploaded by users and searched on the Internet as data graphics sources. On this basis, the present invention creatively sets a rice resistance detection and analysis algorithm, which can realize rice resistance monitoring and analysis efficiently and accurately.

[0137] The data collection in the above scheme includes collecting a large number of image datasets D containing different rice disaster situations:

[0138] D={I 1 , I 2 , ..., I n}, where I i represents the i-th image;

[0139] The graphic preprocessing includes image resizing and normalization. The image resizing is H×W×3. The normalization formula is: Where I is the original image pixel value, I min and I max are the minimum and maximum values ​​of the image pixel values ​​respectively;

[0140] The image annotation includes using an annotation tool to annotate the disaster area in each image with a bounding box and category information to obtain an annotation data set L = {(B 1 , C 1 ), (B 2 , C 2 ),...,(B n , C n )}; Among them, B i represents the set of annotated bounding boxes in the i-th image, C i Represents the corresponding category set;

[0141] The bounding box annotation can be in the form of upper left corner and lower right corner coordinates, that is, b = (x 1 ,y 1 , x 2 ,y 2 ), where (x 1 ,y 1 ) is the coordinate of the upper left corner, (x 2 ,y 2 ) is the lower right corner coordinate.

[0142] The stopping condition in step six is ​​that the maximum number of iterations T is reached or the fitness value no longer increases.

[0143] The display module also displays rice pest data information, including pest type, pest location, and placement measures.

[0144] This example provides a comprehensive solution for image, video and online real-time recognition. The rice pest and disease system provides three recognition methods. Unlike other visual recognition systems that only recognize images, this system can not only efficiently recognize images, but also perform real-time recognition of rice pests and diseases on videos and mobile app online screens provided by users. Users only need to provide materials, and the system can identify the corresponding pests and diseases, lowering the threshold for use, and provide a pest and disease comparison table to help farmers take corresponding prevention and control measures for the corresponding pests and diseases.

[0145] The algorithm of this system is faster and more accurate, making the recognition work more efficient and greatly reducing the time cost of model training and the error cost of recognition.

[0146] Based on the recognition accuracy of the YOLO V8 model, the material can be shot by a mobile phone, camera, or drone. The acquired material can be directly imported into the system for recognition in batches, eliminating the step of secondary processing of the material and lowering the threshold for using the rice pest and disease identification system.

[0147] The system recognition is highly scalable: the system is not only used to identify rice pests and diseases, but can also be used for other crops or is not limited to agronomy. In the future, other items can be identified by simply replacing the training set configuration file, without having to change the interface and other systems. This reduces the user's learning cost for different systems and greatly facilitates the user's operating experience.

[0148] This example also provides a method for real-time resistance monitoring of rice, the method comprising:

[0149] Step 1: The data acquisition module and the transmission module upload the rice resistance monitoring pattern through user upload or sensor upload;

[0150] Step 2: Complete data preparation through data collection and data preprocessing, build the algorithm model and initialize the algorithm model, including YOLO V8 initialization and re-melting signal honey algorithm initialization;

[0151] Step three, complete feature extraction and target detection;

[0152] Feature extraction includes: assuming that the size of the input image is H×W×3, H=height, W=width, 3=number of channels); input image I, and extract features through the C2f structure of YOLO V8: the C2f structure consists of multiple convolutional layer operations, batch normalization layers and activation function layers; after multiple convolution operations, the image is gradually extracted to obtain a feature map F with a size of h×w×d, where h represents the feature map height, w represents the feature map width, and d represents the feature dimension of the feature map;

[0153] The convolutional layer operation is F out =W*F in +b, where W is the convolution kernel weight, F in is the input feature map, b is the bias term;

[0154] The batch normalization layer is Where μ and σ are the mean and standard deviation respectively, and ∈ is a small predefined constant;

[0155] The feature extraction function uses the formula F = f C2f (I, W); where f C2f The function representing the C2f structure, W is the parameter of YOLO V8; the activation function layer is completed using the formula f(x)=max(0,x);

[0156] Object detection includes: on the feature map F, using the detection head of YOLO V8 to detect objects, the detection head includes a classification branch and a regression branch; for each grid cell, predicting multiple bounding boxes and corresponding category probabilities and confidences;

[0157] Assume that the predicted bounding box set is B = {b 1 , b 2 , ..., b k}, the category probability set is C = {c 1 , c 2 , ..., c k}, the confidence set is P = {p 1 , p 2 , ..., p k}, where k is the number of predicted bounding boxes;

[0158] The bounding box prediction can be expressed as b = (t x , t y , t w, t h ), where t x , t y , t w , t h is the offset relative to the grid unit, and the actual bounding box coordinates are obtained through the conversion formula: y =(σ(t y )+c y )×sb x =(σ(t x )+c x )×s, where c x , c y is the coordinate of the grid cell, s is the grid size, p w , p h is the width and height of the prior box;

[0159] The category probability is calculated by the softmax function: where z i is the output of the classification branch;

[0160] The confidence represents the probability that the predicted bounding box contains the target, which is calculated by the sigmoid function: p = σ(z), where z is the output of the regression branch;

[0161] Step 4: Calculate the fitness function value to complete the fitness evaluation:

[0162] For each bee's position x i Corresponding to a set of YOLO V8 parameters, the YOLO V8 model using this parameter is tested on the validation set; define the fitness function value f(x i ) = AP, which means the average accuracy of the model on the validation set. P(R) is the correlation curve between precision and recall;

[0163] For each image in the validation set, the bounding box predicted by the model is compared with the annotated true bounding box to calculate the precision and recall.

[0164] Accuracy calculation formula: Among them, TP is a true positive example and FP is a false positive example;

[0165] Recall calculation formula: Where FN is a false counterexample;

[0166] Step 5: Complete the movement of the pheromone and the update of pheromone. The movement of the pheromone is that the pheromone selects the moving direction according to the pheromone concentration and heuristic information, including:

[0167] Assume that the message i is at the current position xi Select to move to new position x′ i The probability is P i，j , the calculation formula is: Where τ(x j ) is the position x j The pheromone concentration, η(x j ) is the heuristic information, α and β are weight parameters, and N is the set of optional positions of the message;

[0168] Xinmi chooses to move to a new position based on probability and updates its own position: x′ i =x i +step×dir, where step is the moving step length, which can be adjusted according to the problem, and dir is the moving direction, which can be randomly generated or determined according to certain rules;

[0169] The pheromone update includes pheromone volatilization and pheromone enhancement;

[0170] The pheromone volatilization includes: the pheromone will gradually volatilize over time, and the updated pheromone concentration is τ(x j )=(1-ρ)τ(x j ), where ρ is the pheromone volatility coefficient.

[0171] Pheromone enhancement includes: increasing pheromone according to the fitness value of the honey; for honey i, if it finds a new position x′ i If the corresponding fitness value is better, add pheromone at this position. Where Q is the pheromone intensity, f(x′ i ) is the fitness value of the signal honey in the new position.

[0172] Step 6: Model optimization and iteration: Apply the parameters corresponding to the best position found by Xinmi to the YOLO V8 model: Compare the fitness values ​​of all Xinmi to find the best position x best , which is the position with the largest fitness value; the parameters corresponding to the best position are used as the new parameters of the YOLO V8 model;

[0173] Repeat steps 3 to 5 for multiple iterations until the stopping condition is met and the model optimization is completed;

[0174] Step 7: Model evaluation: Use the test data set to calculate the accuracy, recall, and average precision of the optimized model to complete the evaluation;

[0175] Accuracy calculation formula: Its FP is the true positive (the number of positive samples predicted correctly), and FP is the false positive (the number of positive samples predicted incorrectly);

[0176] Recall calculation formula: Where FN is the false negative examples (the number of incorrectly predicted negative samples);

[0177] Calculate the average precision based on the precision and recall Where P(R) is the precision-recall curve;

[0178] Step 8: Calculate the rice resistance expression Let rice resistance be R, where N diseased is the number of damaged rice detected, N total is the total number of rice, which is obtained by detecting the affected rice through the YOLO V8 model and then counting the number.

[0179] Similar to the principle of the system, this invention can realize rice resistance monitoring efficiently, accurately and conveniently.

[0180] Furthermore, the YOLO V8 initialization includes initializing the weight parameter W of the YOLO V8 model. 0 ;

[0181] Set the hyperparameters of YOLO V8, including learning rate α∈(0.001-0.1), batch size batch_size, and number of training rounds epochs;

[0182] The remelting signal honey algorithm initialization includes: determining the number of signal honeys m, defining the position of the signal honey as a vector X={x 1 , x 2 , ..., x m}, where x i Indicates the position of the i-th pheromone, corresponding to a set of parameters of the YOLO V8 model; sets the number of iterations T, pheromone volatility coefficient ρ, and pheromone intensity Q.

[0183] In step 2, data collection includes collecting a large number of image datasets D containing different rice disaster situations:

[0184] D={I 1 , I 2 , ..., I n}, where I i represents the i-th image;

[0185] The graphic preprocessing includes image resizing and normalization. The image resizing is H×W×3. The normalization formula is: Where I is the original image pixel value, I min and I max are the minimum and maximum values ​​of the image pixel values ​​respectively;

[0186] The image annotation includes using an annotation tool to annotate the disaster area in each image with a bounding box and category information to obtain an annotation data set L = {(B 1 , C 1 ), (B 2 , C 2 ),...,(B n , C n )}; Among them, B i represents the set of annotated bounding boxes in the i-th image, C i Represents the corresponding category set;

[0187] The bounding box annotation can be in the form of upper left corner and lower right corner coordinates, that is, b = (x 1 ,y 1 , x 2 ,y 2 ), where (x 1 ,y 1 ) is the coordinate of the upper left corner, (x 2 ,y 2 ) is the lower right corner coordinate.

[0188] As a preferred solution for this example, the C2f structure of YOLO V8 can be further improved to form C3f. Input: image feature tensor. Output: feature tensor processed by the C3f structure. Specifically including:

[0189] (1) Staged feature processing

[0190] The input feature tensor is processed in multiple stages. Each stage contains a series of convolution, batch normalization and activation function operations.

[0191] Suppose the input feature is x. After one stage of transformation, it can be expressed as F(x)=W*ReLU(BN(W*x)), where W is the convolution kernel weight, BN represents batch normalization operation, and ReLU is the activation function.

[0192] For example, convolution kernels of different sizes can be used to extract features at different levels, batch normalization is used to stabilize the training process, and activation functions such as ReLU introduce nonlinearity, allowing the model to learn more complex feature representations.

[0193] (2) Residual Connection

[0194] Drawing on the residual structure of the C3 module, residual connections are introduced in each stage.

[0195] Assuming the input feature is x, and the feature after a series of transformations is F(x), the output is y=x+F(x). This makes it easier for the network to learn the identity mapping, helps alleviate the gradient vanishing problem, and improves the training effect and performance of the model.

[0196] (3) Feature Fusion

[0197] Similar to the ELAN idea of ​​YOLO V8v7, features at different stages are integrated.

[0198] Assume that the characteristics of multiple stages are f 1 , f 2 , ..., f n , the fusion can be expressed as F fusion =concat(f 1 , f 2 , ..., f n ) or F fusion =f 1 +f 2 +...+f n ,The specific fusion method can be selected according to actual conditions.

[0199] The features at different stages are fused by concatenation or addition to enrich the expression of features.

[0200] (4) Output processing

[0201] The fused feature tensor is further processed, such as using additional convolutional layers for feature adjustment and refinement.

[0202] Finally, the feature tensor processed by the C2f structure is output for subsequent tasks such as target detection.

[0203] In summary, the C2f structure effectively improves the feature extraction capability of the YOLOv8v8 model through operations such as staged processing, residual connection, and feature fusion, providing strong support for tasks such as accurate target detection and rice disaster resistance detection.

[0204] Although the above describes an illustrative specific embodiment 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 embodiment. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, all inventions and creations using the concept of the present invention are protected.

Claims

1. A real-time rice resistance monitoring system, characterized in that: The system comprises: A data acquisition module and a transmission module, wherein the data acquisition module includes a sensor, a video and image transmission port; the transmission module can be used to transmit the rice graphics or video data of the data acquisition module; the video and image transmission port can be used to upload the rice graphics or video data; Data analysis and processing module, used to evaluate and analyze rice pest and disease resistance data based on rice graphics or video data; A display module, used for displaying the characteristics of resistance monitoring results in a rice virtual space; The control and decision-making module is used to formulate and implement prevention and control measures based on the analysis results; The data analysis and processing module performs the following algorithm steps: Step 1: Data preparation, including data collection, graphic preprocessing and graphic standards; Step 2: Initialization, including YOLO V8 initialization and re-melting signal honey algorithm initialization; The YOLO V8 initialization includes: initializing the weight parameter W0 of the YOLO V8 model, which can use the weight of the pre-trained model or random initialization; Set the hyperparameters of YOLO V8, including learning rate α∈(0.001-0.1), batch size batch_size, and number of training rounds epochs; The initialization of the remelting signal honey algorithm includes: determining the number of signal honeys m, defining the position of the signal honey as a vector X = {x1, x2, ..., x m }, where x i Indicates the position of the i-th pheromone, corresponding to a set of parameters of the YOLO V8 model; sets the number of iterations T, pheromone volatility coefficient ρ, and pheromone intensity Q; Step three, complete feature extraction and target detection; Feature extraction includes: assuming that the size of the input image is H×W×3, H=height, W=width, 3=number of channels); input image I, and extract features through the C2f structure of YOLO V8: the C2f structure consists of multiple convolutional layer operations, batch normalization layers and activation function layers; after multiple convolution operations, the image is gradually extracted to obtain a feature map F with a size of h×w×d, where h represents the feature map height, w represents the feature map width, and d represents the feature dimension of the feature map; The convolutional layer operation is F out =W*F in +b, where W is the convolution kernel weight, F in is the input feature map, b is the bias term; The batch normalization layer is Where μ and σ are the mean and standard deviation respectively, and ∈ is a small predefined constant; The feature extraction function uses the formula F = f C2f (I, W); where f C2f The function representing the C2f structure, W is the parameter of YOLO V8; the activation function layer is completed using the formula f(x)=max(0,x); Object detection includes: on the feature map F, using the detection head of YOLO V8 to detect objects, the detection head includes a classification branch and a regression branch; for each grid cell, predicting multiple bounding boxes and corresponding category probabilities and confidences; Assume that the predicted bounding box set is B = {b1, b2, ..., b k }, the category probability set is C = {c1, c2, ..., c k }, the confidence set is P = {p1, p2, ..., p k }, where k is the number of predicted bounding boxes; The bounding box prediction can be expressed as b = (t x , t y , t w , t h ), where t x , t y , t w , t h is the offset relative to the grid unit, and the actual bounding box coordinates are obtained through the conversion formula: y =(σ(t y )+c y )×sb x =(σ(t x )+c x )×s, where c x , c y is the coordinate of the grid cell, s is the grid size, p w , p h is the width and height of the prior box; The category probability is calculated by the softmax function: where z i is the output of the classification branch; The confidence represents the probability that the predicted bounding box contains the target, which is calculated by the sigmoid function: p = σ(z), where z is the output of the regression branch; Step 4: Calculate the fitness function value to complete the fitness evaluation: For each bee's position x i Corresponding to a set of YOLO V8 parameters, the YOLO V8 model using this parameter is tested on the validation set; define the fitness function value f(x i ) = AP, which means the average accuracy of the model on the validation set. P(R) is the correlation curve between precision and recall; For each image in the validation set, the bounding box predicted by the model is compared with the annotated true bounding box to calculate the precision and recall. Accuracy calculation formula: Among them, TP is a true positive example and FP is a false positive example; Recall calculation formula: Where FN is a false counterexample; Step 5: Complete the movement of the pheromone and the update of pheromone. The movement of the pheromone is that the pheromone selects the moving direction according to the pheromone concentration and heuristic information, including: Assume that the message i is at the current position x i Select to move to new position x′ i The probability is P i，j , the calculation formula is: Where τ(x j ) is the position x j The pheromone concentration, η(x j ) is the heuristic information, α and β are weight parameters, and N is the set of optional positions of the message; Xinmi chooses to move to a new position based on probability and updates its own position: x′ i =x i +step×dir, where step is the moving step length, which can be adjusted according to the problem, and dir is the moving direction, which can be randomly generated or determined according to certain rules; The pheromone update includes pheromone volatilization and pheromone enhancement; The pheromone volatilization includes: the pheromone will gradually volatilize over time, and the updated pheromone concentration is τ(x j )=(1-ρ)τ(x j ), where ρ is the pheromone volatility coefficient. Pheromone enhancement includes: increasing pheromone according to the fitness value of the honey; for honey i, if it finds a new position x′ i If the corresponding fitness value is better, add pheromone at this position. Where Q is the pheromone intensity, f(x′ i ) is the fitness value of the signal honey in the new position. Step 6: Model optimization and iteration. Apply the parameters corresponding to the best position found by Xinmi to the YOLO V8 model: Compare the fitness values ​​of all Xinmi to find the best position xbest, that is, the position with the largest fitness value; use the parameters corresponding to the best position as the new parameters of the YOLO V8 model; Repeat steps 3 to 5 for multiple iterations until the stopping condition is met and the model optimization is completed; Step 7: Model evaluation: Use the test data set to calculate the accuracy, recall, and average precision of the optimized model to complete the evaluation; Accuracy calculation formula: Its TP is the true positive sample (the number of correctly predicted positive samples), and FP is the false positive sample (the number of incorrectly predicted positive samples); Recall calculation formula: Where FN is the false negative examples (the number of incorrectly predicted negative samples); Calculate the average precision based on the precision and recall Where P(R) is the precision-recall curve; Step 8: Calculate the rice resistance expression Let rice resistance be R, where N diseased is the number of damaged rice detected, N total is the total number of rice, which is obtained by detecting the affected rice through the YOLO V8 model and then counting the number.

2. The rice real-time resistance monitoring system according to claim 1, characterized in that: The data collection includes collecting a large number of image datasets D containing different rice disaster situations: D = {I1, I2, ..., I n }, where I i represents the i-th image; The graphic preprocessing includes image resizing and normalization. The image resizing is H×W×3. The normalization formula is: Where I is the original image pixel value, I min and I max are the minimum and maximum values ​​of the image pixel values ​​respectively; The graphic annotation includes using an annotation tool to annotate the disaster area in each image with a bounding box and category information, and obtaining an annotation data set L = {(B1, C1), (B2, C2), ..., (B n , C n )}; Among them, B i represents the set of annotated bounding boxes in the i-th image, C i Represents the corresponding category set; The bounding box annotation may be in the form of upper left corner and lower right corner coordinates, i.e., b = (x1, y1, x2, y2), where (x1, y1) is the upper left corner coordinate and (x2, y2) is the lower right corner coordinate.

3. The rice real-time resistance monitoring system according to claim 1, characterized in that: The stopping condition in step six is ​​that the maximum number of iterations T is reached or the fitness value no longer increases.

4. The rice real-time resistance monitoring system according to claim 1, characterized in that: The display module also displays rice pest and disease data information, including pest and disease type, pest and disease location, and placement measures.

5. A method for real-time resistance monitoring of rice, characterized in that: The method is based on the system according to any one of claims 1 to 4, and the method comprises: Step 1: The data acquisition module and the transmission module upload the rice resistance monitoring pattern through user upload or sensor upload; Step 2: Complete data preparation through data collection and data preprocessing, build the algorithm model and initialize the algorithm model, including YOLO V8 initialization and re-melting signal honey algorithm initialization; Step three, complete feature extraction and target detection; Feature extraction includes: assuming that the size of the input image is H×W×3, H=height, W=width, 3=number of channels); input image I, and extract features through the C2f structure of YOLO V8: the C2f structure consists of multiple convolutional layer operations, batch normalization layers and activation function layers; after multiple convolution operations, the image is gradually extracted to obtain a feature map F with a size of h×w×d, where h represents the feature map height, w represents the feature map width, and d represents the feature dimension of the feature map; The convolutional layer operation is F out =W*F in +b, where W is the convolution kernel weight, F in is the input feature map, b is the bias term; The batch normalization layer is Where μ and σ are the mean and standard deviation respectively, and ∈ is a small predefined constant; The feature extraction function uses the formula F = f C2f (I, W); where f C2f The function representing the C2f structure, W is the parameter of YOLO V8; the activation function layer is completed using the formula f(x)=max(0,x); Object detection includes: on the feature map F, using the detection head of YOLO V8 to detect objects, the detection head includes a classification branch and a regression branch; for each grid cell, predicting multiple bounding boxes and corresponding category probabilities and confidences; Assume that the predicted bounding box set is B = {b1, b2, ..., b k }, the category probability set is C = {c1, c2, ..., c k }, the confidence set is P = {p1, p2, ..., p k }, where k is the number of predicted bounding boxes; The bounding box prediction can be expressed as b = (t x , t y , t w , t h ), where t x , t y , t w , t h is the offset relative to the grid unit, and the actual bounding box coordinates are obtained through the conversion formula: y =(σ(t y )+c y )×sb x =(σ(t x )+c x )×s where c x , c y is the coordinate of the grid cell, s is the grid size, p w , p h is the width and height of the prior box; The category probability is calculated by the softmax function: where z i is the output of the classification branch; The confidence represents the probability that the predicted bounding box contains the target, which is calculated by the sigmoid function: p = σ(z), where z is the output of the regression branch; Step 4: Calculate the fitness function value to complete the fitness evaluation: For each bee's position x i Corresponding to a set of YOLO V8 parameters, the YOLO V8 model using this parameter is tested on the validation set; define the fitness function value f(x i ) = AP, which means the average accuracy of the model on the validation set. P(R) is the correlation curve between precision and recall; For each image in the validation set, the bounding box predicted by the model is compared with the annotated true bounding box to calculate the precision and recall. Accuracy calculation formula: Among them, TP is a true positive example and FP is a false positive example; Recall calculation formula: Where FN is a false counterexample; Step 5: Complete the movement of the signal honey and the pheromone change. The signal honey movement is that the signal honey selects the movement direction according to the pheromone concentration and heuristic information, including: Assume that the message i is at the current position x i Select to move to new position x′ i The probability is P i，j , the calculation formula is: Where τ(x j ) is the position x j The pheromone concentration, η(x j ) is the heuristic information, α and β are weight parameters, and N is the set of optional positions of the message; Xinmi chooses to move to a new position based on probability and updates its own position: x′ i =x i +step×dir, where step is the moving step length, which can be adjusted according to the problem, and dir is the moving direction, which can be randomly generated or determined according to certain rules; The pheromone update includes pheromone volatilization and pheromone enhancement; The pheromone volatilization includes: the pheromone will gradually volatilize over time, and the updated pheromone concentration is τ(x j )=(1-ρ)τ(x j ), where ρ is the pheromone volatility coefficient. Pheromone enhancement includes: adding pheromones according to the fitness value of the signal honey; for signal honey i, if the fitness value corresponding to the new position x′i it finds is better, then adding pheromones at this position, Where Q is the pheromone intensity, f(x′ i ) is the fitness value of the signal honey in the new position. Step 6: Model optimization and iteration. Apply the parameters corresponding to the best position found by Xinmi to the YOLO V8 model: Compare the fitness values ​​of all Xinmi to find the best position xbest, that is, the position with the largest fitness value; use the parameters corresponding to the best position as the new parameters of the YOLO V8 model; Repeat steps 3 to 5 for multiple iterations until the stopping condition is met and the model optimization is completed; Step 7: Model evaluation: Use the test data set to calculate the accuracy, recall, and average precision of the optimized model to complete the evaluation; Accuracy calculation formula: Its TP is the true positive sample (the number of correctly predicted positive samples), and FP is the false positive sample (the number of incorrectly predicted positive samples); Recall calculation formula: Where FN is the false negative examples (the number of incorrectly predicted negative samples); Calculate the average precision based on the precision and recall Where P(R) is the precision-recall curve; Step 8: Calculate the rice resistance expression Let rice resistance be R, where N diseased is the number of damaged rice detected, N total is the total number of rice, which is obtained by detecting the affected rice through the YOLO V8 model and then counting the number.

6. The method for real-time resistance monitoring of rice according to claim 5, characterized in that: The YOLO V8 initialization includes initializing the weight parameter W0 of the YOLO V8 model; Set the hyperparameters of YOLO V8, including learning rate α∈(0.001-0.1), batch size batch_size, and number of training rounds epochs; The remelting signal honey algorithm initialization includes: determining the number of signal honeys m, defining the position of the signal honeys as a vector X={x1, x2, ..., x m }, where x i Indicates the position of the i-th pheromone, corresponding to a set of parameters of the YOLO V8 model; sets the number of iterations T, pheromone volatility coefficient ρ, and pheromone intensity Q.

7. The method for real-time resistance monitoring of rice according to claim 5, characterized in that: In step 2, data collection includes collecting a large number of image datasets D containing different rice disaster situations: D = {I1, I2, ..., I n }, where I i represents the i-th image; The graphic preprocessing includes image resizing and normalization. The image resizing is H×W×3. The normalization formula is: Where I is the original image pixel value, I min and I max are the minimum and maximum values ​​of the image pixel values ​​respectively; The graphic annotation includes using an annotation tool to annotate the disaster area in each image with a bounding box and category information, and obtaining an annotation data set L = {(B1, C1), (B2, C2), ..., (B n , C n )}; Among them, B i represents the set of annotated bounding boxes in the i-th image, C i Represents the corresponding category set; The bounding box annotation may be in the form of upper left corner and lower right corner coordinates, i.e., b = (x1, y1, x2, y2), where (x1, y1) is the upper left corner coordinate and (x2, y2) is the lower right corner coordinate.

Images