An intelligent wolfberry pest image recognition method

By using an intelligent image recognition method for wolfberry diseases and pests, and employing ensemble learning and transfer learning to construct an anomaly recognition model, the problem of large fluctuations in the accuracy of manual visual identification was solved. This enabled efficient identification and monitoring of wolfberry diseases and pests, thereby improving crop yield and quality.

CN118485915BActive Publication Date: 2024-11-22INST OF AGRI ECONOMICS & INFORMATION TECH NINGXIA ACAD OF AGRI & FORESTRY SCI (NINGXIA AGRI SCI & TECH LIBRARY)
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202410494876.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-23
Publication Date
2024-11-22
Estimated Expiration
2044-04-23

AI Technical Summary

Technical Problem

In the current technology, the identification of diseases and pests of wolfberry mainly relies on manual visual inspection, which leads to large fluctuations in the accuracy of identification and poor stability.

Method used

An intelligent image recognition method for wolfberry diseases and pests is adopted. Through data collection, preprocessing, model building and verification, and by using ensemble learning and transfer learning methods, an anomaly recognition model is constructed to improve the recognition accuracy and efficiency.

Benefits of technology

It enables rapid diagnosis and monitoring of diseases and pests in wolfberry, improving crop yield and quality, and enhancing the accuracy and efficiency of identification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118485915B_ABST
    Figure CN118485915B_ABST
Patent Text Reader

Abstract

The present invention provides an intelligent wolfberry pest image recognition method, belonging to the field of image recognition technology; determining a target acquisition object, constructing an environmental information data acquisition system according to the characteristics of the target acquisition object, and obtaining an original image group of the target acquisition object; preprocessing the original image group to obtain a standard image group, the standard image group comprising wolfberry images without pests and diseases and wolfberry images with pests and diseases; filtering the standard image group for impurities based on a filtering model to obtain a standard data set of the target acquisition object; constructing an abnormality recognition model of the target acquisition object according to a machine learning algorithm and in combination with the standard data set; obtaining a wolfberry pest verification image and inputting it into the abnormality recognition model to obtain an abnormality recognition result, and using an evaluation index to evaluate the performance of the abnormality recognition model to determine the effectiveness of the model. Accurate recognition and evaluation of wolfberry pests and diseases are achieved, and the efficiency of pest disease evaluation is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to an intelligent wolfberry pest and disease image recognition method. Background Art

[0002] At present, the main method for diagnosing wolfberry diseases and pests is visual inspection, which relies on visual field inspections by growers and agricultural technicians. This method is completely dependent on their professional level, judgment experience and subjective differences, resulting in large fluctuations in the accuracy of identifying wolfberry diseases and pests and poor stability.

[0003] Therefore, the present invention proposes an intelligent wolfberry pest and disease image recognition method. Summary of the invention

[0004] The present invention provides an intelligent wolfberry pest and disease image recognition method, which is used to improve the recognition accuracy and efficiency of wolfberry pest and disease images by using ensemble learning and transfer learning methods through steps such as data collection, preprocessing, model construction and verification.

[0005] In one aspect, the present invention provides an intelligent wolfberry pest image recognition method, comprising:

[0006] Step 1: Determine the target acquisition object, build an environmental information data acquisition system according to the characteristics of the target acquisition object, and obtain the original image group of the target acquisition object;

[0007] Step 2: preprocessing the original image group to obtain a standard image group, wherein the standard image group includes wolfberry images without pests and diseases and wolfberry images with pests and diseases;

[0008] Step 3: Perform impurity filtering on the standard image group based on the filtering model to obtain a standard data set of the target acquisition object;

[0009] Step 4: Construct an abnormality recognition model for the target acquisition object based on a machine learning algorithm and in combination with a standard data set;

[0010] Step 5: Obtain the wolfberry verification image and input it into the abnormality recognition model to obtain the abnormality recognition result, and use the evaluation index to evaluate the performance of the abnormality recognition model to determine the effectiveness of the abnormality recognition model.

[0011] On the other hand, determining the target acquisition object, building an environmental information data acquisition system according to the characteristics of the target acquisition object, and obtaining the original image group of the target acquisition object, including:

[0012] Determine the target collection objects according to the types of pests and diseases suffered by wolfberry plants and plant types;

[0013] According to the research needs of pests and diseases and the characteristics of the target collection objects, the collection equipment is matched, wherein all the collection equipment constitutes the environmental information data collection system;

[0014] An original image group of the target acquisition object is acquired based on the acquisition device.

[0015] On the other hand, the process of preprocessing the original image group to obtain the standard image group includes:

[0016] Performing data enhancement on the original image group to generate an enhanced image group;

[0017] A first dimension processing library is established according to a preset image standard size, and the enhanced image group is introduced into the library for processing. If the aspect ratio of any enhanced image does not conform to the preset image standard size, padding and cropping operations are performed on the corresponding enhanced image to obtain a first image group.

[0018] On the other hand, after obtaining the first image group, the method further includes:

[0019] Performing pixel grayscale processing on any first image of the first image group to obtain a grayscale image, determining a grayscale value of each pixel in the grayscale image, and determining a grayscale level of a corresponding pixel based on a value-level mapping table;

[0020] Counting the number of pixels at the same gray level in the gray image respectively, and constructing a gray distribution map according to the gray level order;

[0021] Determine the grayscale mean of the grayscale distribution diagram ;

[0022] ;

[0023] ;

[0024] in, represents the probability density distribution value of the pixel at the jth gray level in the corresponding gray image, n represents the number of gray levels, Represents the level value of the j-th gray level; Represents the proportion of the jth gray level determined based on the gray distribution diagram; represents the weight of the position where the j-th gray level is determined in the gray distribution diagram; Indicates all The maximum value in ; Indicates all The maximum value in ; Indicates all The maximum value among them; e represents a constant, and its value is 2.7; represents the average value of all grayscale values ​​in the jth grayscale level determined based on the grayscale distribution diagram; max represents the maximum value symbol; represents the variance of all grayscale values ​​in the jth grayscale level determined based on the grayscale distribution diagram;

[0025] All pixels of the grayscale image are de-averaged according to the grayscale mean to obtain standard pixels, and a standard image corresponding to the first image is formed, thereby forming a standard image group.

[0026] On the other hand, the standard image group is impurity filtered based on the filtering model to obtain a standard data set of the target acquisition object, including:

[0027] Extracting a number of wolfberry-related images from the historical data set as positive impurity filtering samples, and extracting a number of non-wolfberry-related images from the historical data set as negative impurity filtering samples;

[0028] Building a filtering recognition network, and performing filtering training on the positive magazine filtering samples and the negative impurity filtering samples to obtain a filtering model;

[0029] The standard image group is imported into the filtering model to screen out all standard images that meet the requirements of wolfberry pests and diseases as the standard data set.

[0030] On the other hand, according to the machine learning algorithm and in combination with the standard data set, an abnormality recognition model of the target acquisition object is constructed, including:

[0031] We extracted wolfberry pest and disease images of different preset analysis categories from the wolfberry image library in equal proportions, performed data preprocessing on the images of each category, and obtained a public data set;

[0032] Develop a deep learning model using ensemble learning and transfer learning methods, and train and identify the public data set;

[0033] According to the comparison between the training results and the recognition results, the fully connected layer of the deep learning model is adjusted, and the feature extraction part of the deep learning model is used as the feature extractor;

[0034] Propose a first weight to the performance parameter of the feature extractor, and at the same time, set a second weight to different types of recognition processes by using a voting method based on the training results of the adjusted deep learning model and the feature comparison results of the standard data set;

[0035] The first weight, the second weight, and the adjusted deep learning model are integrated to obtain an abnormality recognition model for the target acquisition object.

[0036] On the other hand, before using the evaluation index to evaluate the performance of the anomaly recognition model, it includes:

[0037] Extracting the abnormal feature y identified by the abnormal network recognition layer of each abnormal type in the abnormal recognition model for the wolfberry verification image and the preset recognition feature x of the corresponding abnormal network recognition layer, and calculating the similarity value between the abnormal feature y and the preset recognition feature x;

[0038]

[0039] in, Represents the similarity value between the preset recognition feature x of the corresponding abnormal network recognition layer and the abnormal feature y of the corresponding abnormal type existing in the wolfberry verification image, , They represent the feature brightness means of the preset identification feature x and the abnormal feature y respectively, They represent the feature brightness variance of the preset identification feature x and the abnormal feature y respectively, represents the feature brightness covariance of the preset recognition feature x and the abnormal feature y, and represents the stability constant.

[0040] On the other hand, the performance of the anomaly recognition model is evaluated using evaluation indicators to determine the effectiveness of the anomaly recognition model, including:

[0041] A first feature quantity N1 whose statistical similarity value is greater than or equal to a preset value of the corresponding abnormal network identification layer, and a second feature quantity N2 whose statistical similarity value is less than a preset value of the corresponding abnormal network identification layer, wherein the evaluation index is the first feature quantity and the second feature quantity;

[0042] Calculate the total evaluation score S based on the first feature number N1 and the second feature number N2;

[0043] ;

[0044] in, 1 indicates the preset evaluation weight, represents the sum of all similarity values ​​under the first feature quantity N1; represents the sum of all preset values ​​under the first feature quantity N1; represents the sum of all preset values ​​under the second feature quantity N2;

[0045] If the total evaluation score is greater than the preset score, the anomaly recognition model is judged to be highly effective; otherwise, the anomaly recognition model is judged to be insufficiently effective and a warning is issued.

[0046] The present invention provides an intelligent wolfberry pest and disease image recognition method, which is used to improve the recognition accuracy and efficiency of wolfberry pest and disease images by using ensemble learning and transfer learning methods through steps such as data collection, preprocessing, model construction and verification. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0048] Figure 1 It is a flow chart of an intelligent wolfberry pest and disease image recognition method provided by an embodiment of the present invention;

[0049] Figure 2 It is a flowchart of building an anomaly recognition model for a target acquisition object according to a machine learning algorithm and a standard data set provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] Embodiment 1:

[0052] like Figure 1 As shown, an intelligent wolfberry pest image recognition method provided by an embodiment of the present invention includes:

[0053] Step 1: Determine the target acquisition object, build an environmental information data acquisition system according to the characteristics of the target acquisition object, and obtain the original image group of the target acquisition object;

[0054] Step 2: preprocessing the original image group to obtain a standard image group, wherein the standard image group includes wolfberry images without pests and diseases and wolfberry images with pests and diseases;

[0055] Step 3: Perform impurity filtering on the standard image group based on the filtering model to obtain a standard data set of the target acquisition object;

[0056] Step 4: Construct an abnormality recognition model for the target acquisition object based on a machine learning algorithm and in combination with a standard data set;

[0057] Step 5: Obtain the wolfberry verification image and input it into the abnormality recognition model to obtain the abnormality recognition result, and use the evaluation index to evaluate the performance of the abnormality recognition model to determine the effectiveness of the abnormality recognition model.

[0058] In this embodiment, the target collection object refers to the wolfberry plant or wolfberry fruit that needs to be identified for pests and diseases.

[0059] In this embodiment, the characteristics include the growth status, morphological characteristics, possible manifestations of diseases and insect pests, etc. of the wolfberry plant or fruit.

[0060] In this embodiment, the environmental information data collection system includes a monocular camera, a light source, a computer, a power supply, a bracket, a connecting line and other equipment, and takes photos of the bird's-eye view of wolfberry pests and diseases in an indoor environment.

[0061] In this embodiment, the original image group refers to a set of unprocessed images obtained from the environmental information data acquisition system, and these images capture the real scene of the target acquisition object.

[0062] In this embodiment, preprocessing is a series of processing steps to reduce data noise, improve image quality, highlight features, etc. The main process includes: denoising, smoothing, contrast enhancement, size adjustment, standardization adjustment and other operations.

[0063] In this embodiment, the standard image group is a set of standardized image data obtained after preprocessing and annotation, and includes wolfberry images without pests and diseases and wolfberry images with pests and diseases.

[0064] In this embodiment, pests and diseases refer to diseases or damages caused by the invasion of bacteria, viruses, fungi or insects and other pests on plants, including diseases caused by viruses, bacteria and fungi, and invasion of insects, borers and other pests.

[0065] In this embodiment, the wolfberry image is an image taken of a wolfberry plant, wolfberry fruit, or a wolfberry planting environment.

[0066] In this embodiment, the filtering model extracts some samples from the constructed wolfberry pest and disease dataset, and combines other non-wolfberry pictures (including trees, portraits, and other pictures) searched on the Internet to construct a dataset for training the filtering model.

[0067] In this embodiment, impurity filtering refers to the operation of filtering out irrelevant images (such as trees, flowers, etc.) using a filtering model.

[0068] In this embodiment, the standard data set refers to a data set for machine learning training that is constructed after preprocessing and impurity filtering.

[0069] In this embodiment, the machine learning algorithm includes an algorithm of integrated learning and transfer learning, such as Figure 2 As shown. Ensemble learning is a machine learning method whose basic idea is to improve the overall prediction performance by combining the prediction results of multiple models. The basic idea of ​​transfer learning is to use a well-trained model in one field to help solve problems in another related field.

[0070] In this embodiment, the abnormality recognition model is a deep learning algorithm model constructed by integrating learning and transfer learning methods, which is used to identify the types of pests and diseases in wolfberry images.

[0071] In this embodiment, the wolfberry verification image refers to the wolfberry image data used to verify the performance of the abnormal recognition model.

[0072] In this embodiment, the abnormal recognition result is the recognition result of each verification image. The result can indicate what kind of pests and diseases the verification image is judged to have.

[0073] In this embodiment, the performance evaluation is to verify and evaluate the constructed anomaly recognition model to determine its effectiveness and reliability in practical applications.

[0074] In this embodiment, effectiveness refers to the ability of the anomaly recognition model in detecting wolfberry pests and diseases.

[0075] The working principle and beneficial effects of the above technical solution are: by building an environmental information data collection system, preprocessing images, building an abnormality recognition model and other steps, rapid diagnosis and monitoring of wolfberry pests and diseases can be achieved, thereby improving crop yield and quality.

[0076] Embodiment 2:

[0077] On the basis of the above-mentioned embodiment 1, a target acquisition object is determined, an environmental information data acquisition system is constructed according to the characteristics of the target acquisition object, and an original image group of the target acquisition object is obtained, including:

[0078] Determine the target collection objects according to the types of pests and diseases suffered by wolfberry plants and plant types;

[0079] According to the research needs of pests and diseases and the characteristics of the target collection objects, the collection equipment is matched, wherein all the collection equipment constitutes the environmental information data collection system;

[0080] An original image group of the target acquisition object is acquired based on the acquisition device.

[0081] In this embodiment, the acquisition equipment includes specific equipment such as a monocular camera, a light source, a computer, a power supply, and a sensor.

[0082] The working principle and beneficial effects of the above technical solution are: using the environmental information data collection system and machine learning technology to conduct real-time monitoring and identification of diseases and pests of wolfberry plants to improve yield and quality.

[0083] Embodiment 3:

[0084] On the basis of the above-mentioned embodiment 1, the process of preprocessing the original image group to obtain the standard image group includes:

[0085] Performing data enhancement on the original image group to generate an enhanced image group;

[0086] A first dimension processing library is established according to a preset image standard size, and the enhanced image group is introduced into the library for processing. If the aspect ratio of any enhanced image does not conform to the preset image standard size, padding and cropping operations are performed on the corresponding enhanced image to obtain a first image group.

[0087] In this embodiment, data enhancement is performed through operations such as rotation, flipping, scaling, cropping, and adding noise to expand the training data set and improve the versatility of the training data.

[0088] In this embodiment, the enhanced image group is a set of new images related to the original images generated by data enhancement technology.

[0089] In this embodiment, the preset image standard size is an ideal image size preset during image processing or model training.

[0090] In this embodiment, the first dimension processing library is a library for storing and processing image data, and is used to perform operations such as preprocessing, enhancing, and resizing the image to meet a preset standard image size.

[0091] In this embodiment, the aspect ratio refers to the proportional relationship between the width and height of an image or object.

[0092] In this embodiment, padding is to expand the size of the image by adding extra pixels (blank pixels or pixels of a specific color) around the image when the size of the image is smaller than a preset standard size, wherein the filling operation utilizes nearest neighbor interpolation to identify all pixels around the target pixel as neighboring pixel points, and fills the edge of the target enhanced image according to the pixel arrangement rule and the original logic of the image.

[0093] In this embodiment, cropping is to adjust the size and shape of the image through a cropping operation when the aspect ratio of the image does not match the standard size.

[0094] In this embodiment, the first image group is a set of images that have been data enhanced and padded and cropped according to a preset standard image size.

[0095] The working principle and beneficial effects of the above technical solution are: through data enhancement and size adjustment, image consistency and compliance with standard sizes are ensured, model stability and generalization ability are improved, resources are saved and processing efficiency is improved.

[0096] Embodiment 4:

[0097] On the basis of the above embodiment 3, after obtaining the first image group, the method further includes:

[0098] Performing pixel grayscale processing on any first image of the first image group to obtain a grayscale image, determining a grayscale value of each pixel in the grayscale image, and determining a grayscale level of a corresponding pixel based on a value-level mapping table;

[0099] Counting the number of pixels at the same gray level in the gray image respectively, and constructing a gray distribution map according to the gray level order;

[0100] Determine the grayscale mean of the grayscale distribution diagram ;

[0101] ;

[0102] ;

[0103] in, represents the probability density distribution value of the pixel at the jth gray level in the corresponding gray image, n represents the number of gray levels, Represents the level value of the j-th gray level; Represents the proportion of the jth gray level determined based on the gray distribution diagram; represents the weight of the position where the j-th gray level is determined in the gray distribution diagram; Indicates all The maximum value in ; Indicates all The maximum value in ; Indicates all The maximum value among them; e represents a constant, and its value is 2.7; represents the average value of all grayscale values ​​in the jth grayscale level determined based on the grayscale distribution diagram; max represents the maximum value symbol; represents the variance of all grayscale values ​​in the jth grayscale level determined based on the grayscale distribution diagram;

[0104] All pixels of the grayscale image are de-averaged according to the grayscale mean to obtain standard pixels, and a standard image corresponding to the first image is formed, thereby forming a standard image group.

[0105] In this embodiment, pixel grayscale processing refers to the process of converting a color image into a grayscale image.

[0106] In this embodiment, the grayscale image is an image that contains only black and white colors but no color, and the color value of each pixel represents the brightness of the pixel.

[0107] In this embodiment, the grayscale value refers to the brightness and grayscale level of each pixel in the image.

[0108] In this embodiment, the value-level mapping table is a mapping table that defines the relationship between pixel grayscale values ​​and grayscale levels.

[0109] In this embodiment, the grayscale is used to describe the quantization level of the grayscale value of a pixel in an image. The grayscale is usually used to indicate the degree of grayness of a pixel.

[0110] In this embodiment, the grayscale distribution diagram is a statistical diagram used to represent the number of pixels at different grayscale levels in a grayscale image.

[0111] In this embodiment, the grayscale mean refers to the average value of each grayscale level in the grayscale distribution diagram.

[0112] In this embodiment, the probability density distribution value is the proportion of the number of pixels at each gray level in the entire image, which is also the probability density of each gray level.

[0113] In this embodiment, the level value is a constant value of the gray level in the gray image.

[0114] In this embodiment, the appearance position weight is a preset value indicating the degree of influence of the position where the gray level appears in the gray level distribution diagram on the entire distribution.

[0115] In this embodiment, the de-averaging process is an operation of subtracting the grayscale mean from the G, R, and B values ​​corresponding to each pixel in the grayscale image.

[0116] In this embodiment, the standard pixel point refers to the pixel point after the de-averaging process.

[0117] In this embodiment, the standard image refers to an image constructed by standard pixels.

[0118] The working principle and beneficial effects of the above technical solution are: grayscale processing is performed on the first image, a grayscale distribution map is constructed, the grayscale mean is calculated, and de-meaning processing is performed to improve the image processing effect and quality.

[0119] Embodiment 5:

[0120] On the basis of the above-mentioned embodiment 1, the standard image group is impurity filtered based on the filtering model to obtain the standard data set of the target acquisition object, including:

[0121] Extracting a number of wolfberry-related images from the historical data set as positive impurity filtering samples, and extracting a number of non-wolfberry-related images from the historical data set as negative impurity filtering samples;

[0122] Building a filtering recognition network, and performing filtering training on the positive magazine filtering samples and the negative impurity filtering samples to obtain a filtering model;

[0123] The standard image group is imported into the filtering model to screen out all standard images that meet the requirements of wolfberry pests and diseases as the standard data set.

[0124] In this embodiment, the positive impurity filtering samples refer to sample images related to healthy wolfberry plants or wolfberry plants damaged by diseases and insect pests.

[0125] In this embodiment, the negative impurity filtering samples are sample images that do not contain wolfberry or are irrelevant to wolfberry.

[0126] In this embodiment, the filtering and recognition network refers to a neural network model for filtering and recognizing images.

[0127] In this embodiment, filtering training refers to the process of training the filtering recognition network using positive impurity filtering samples and negative impurity filtering samples.

[0128] The working principle and beneficial effects of the above technical solution are: using the filtering recognition network to train historical image data, to achieve automatic screening of images that meet the requirements of wolfberry pests and diseases, to improve work efficiency, and to enhance data quality and mining capabilities.

[0129] Embodiment 6:

[0130] On the basis of the above-mentioned embodiment 1, according to the machine learning algorithm and in combination with the standard data set, an abnormality recognition model of the target acquisition object is constructed, such as Figure 2 As shown, including:

[0131] We extracted wolfberry pest and disease images of different preset analysis categories from the wolfberry image library in equal proportions, performed data preprocessing on the images of each category, and obtained a public data set;

[0132] Develop a deep learning model using ensemble learning and transfer learning methods, and train and identify the public data set;

[0133] According to the comparison between the training results and the recognition results, the fully connected layer of the deep learning model is adjusted, and the feature extraction part of the deep learning model is used as the feature extractor;

[0134] Propose a first weight to the performance parameter of the feature extractor, and at the same time, set a second weight to different types of recognition processes by using a voting method based on the training results of the adjusted deep learning model and the feature comparison results of the standard data set;

[0135] The first weight, the second weight, and the adjusted deep learning model are integrated to obtain an abnormality recognition model for the target acquisition object.

[0136] In this embodiment, the wolfberry image library refers to an image database containing various wolfberry pests and diseases images, which is used to study and analyze various pests and diseases that wolfberry plants suffer from.

[0137] In this embodiment, equal-proportion extraction refers to an operation of extracting the same number of proportions from different types of image libraries according to certain parameters.

[0138] In this embodiment, the preset analysis categories refer to different categories or types of pests and diseases that are predetermined during the wolfberry pest and disease image analysis process.

[0139] In this embodiment, the public dataset refers to a dataset that is publicly released and available for research and use.

[0140] In this embodiment, the fully connected layer is a basic structure in a deep neural network, usually located in the last few layers of the neural network, and its function is to convert the output of the previous hidden layer (or feature extraction part) into the final output category or score.

[0141] In this embodiment, the feature extractor refers to the part in the deep learning model that is responsible for extracting effective feature representations from raw data.

[0142] In this embodiment, the performance parameter is a parameter that describes the feature extractor, including parameters such as feature expression capability, feature repeatability, feature stability, and feature dimension.

[0143] In this embodiment, the first weight refers to the initial weight coefficient of the performance parameter of the feature extractor. There are k performance parameters in total. Each performance parameter has a preset score based on expert analysis, and the first weight of the performance parameter is obtained based on the proportion of the score of each performance parameter in the total score.

[0144] In this embodiment, the voting method is used to integrate the prediction results of multiple classifiers to make a final classification decision, and the class with the highest probability is finally selected as the result by giving each classifier a probability estimate for each class.

[0145] In this embodiment, the second weight refers to the weight distribution of the performance parameter after the feature comparison result of the training result and the standard data set. The probability of the category feature is decided according to the voting method. Then, the second weight of any performance parameter under the category feature is:

[0146] ;

[0147] in, represents a second weight of the performance parameter, It means there are f performance parameters in total. represents a first weight of the performance parameter, represents the parameter probability obtained by the voting method, represents the allocation weight adjustment constant of the performance parameter.

[0148] In this embodiment, linearly inseparable training sample features are transformed from a low-dimensional input space into a high-dimensional feature space to make them linearly separable, thereby classifying the image features of the training samples into corresponding categories.

[0149] In this embodiment, the deep learning model is a model composed of multiple layers, each layer contains multiple neurons, and these neurons transmit and process information through connections. It is used to learn complex feature representation and pattern recognition tasks, such as image recognition, speech recognition, natural language processing, etc.

[0150] In this embodiment, integration refers to combining multiple adjustment results and parameters together to utilize their respective advantages, thereby improving the performance of the overall model.

[0151] The working principle and beneficial effects of the above technical solution are: using ensemble learning and transfer learning to build a deep learning model to identify wolfberry pests and diseases, improve accuracy, adapt to different situations, and effectively identify anomalies.

[0152] Embodiment 7:

[0153] Based on the above embodiment 6, before using the evaluation index to evaluate the performance of the abnormality recognition model, the following steps are included:

[0154] Extracting the abnormal feature y identified by the abnormal network recognition layer of each abnormal type in the abnormal recognition model for the wolfberry verification image and the preset recognition feature x of the corresponding abnormal network recognition layer, and calculating the similarity value between the abnormal feature y and the preset recognition feature x;

[0155]

[0156] in, Represents the similarity value between the preset recognition feature x of the corresponding abnormal network recognition layer and the abnormal feature y of the corresponding abnormal type existing in the wolfberry verification image, , They represent the feature brightness means of the preset identification feature x and the abnormal feature y respectively, They represent the feature brightness variance of the preset identification feature x and the abnormal feature y respectively, represents the feature brightness covariance of the preset recognition feature x and the abnormal feature y, and represents the stability constant.

[0157] In this embodiment, the abnormal features refer to abnormal image features after the image is processed by the abnormal recognition model, including: irregular spots appearing on wolfberry leaves, abnormal leaf color, deformed leaf edges, etc.

[0158] In this embodiment, the abnormal network identification layer refers to a specific layer in the abnormal detection or abnormal identification model for identifying abnormal situations in images or data. It should be noted that the abnormal identification model includes several layers, and each layer corresponds to an abnormal type.

[0159] In this embodiment, the preset identification features refer to image features that are set in advance during the training process of the abnormality recognition model for identifying specific abnormality types.

[0160] In this embodiment, the similarity value refers to the degree of similarity between the abnormal feature and the preset recognition feature, which is mainly used to determine the accuracy of model recognition.

[0161] In this embodiment, the feature brightness refers to the brightness level of a feature in the image.

[0162] In this embodiment, the stability constant is a predetermined constant used to stabilize the model function.

[0163] The working principle and beneficial effects of the above technical solution are: using the anomaly recognition model, extracting the abnormal features of the image and comparing them with the preset features, calculating the similarity value, and guiding the anomaly recognition. Through the feature brightness statistics and similarity measurement, the accuracy and stability of anomaly detection are improved, and the reliability of the system is enhanced.

[0164] Embodiment 8:

[0165] On the basis of the above-mentioned embodiment 7, the performance of the anomaly recognition model is evaluated by using evaluation indicators to judge the effectiveness of the anomaly recognition model, including:

[0166] A first feature quantity N1 whose statistical similarity value is greater than or equal to a preset value of the corresponding abnormal network identification layer, and a second feature quantity N2 whose statistical similarity value is less than a preset value of the corresponding abnormal network identification layer, wherein the evaluation index is the first feature quantity and the second feature quantity;

[0167] Calculate the total evaluation score S based on the first feature number N1 and the second feature number N2;

[0168] ;

[0169] in, 1 indicates the preset evaluation weight, represents the sum of all similarity values ​​under the first feature quantity N1; represents the sum of all preset values ​​under the first feature quantity N1; represents the sum of all preset values ​​under the second feature quantity N2;

[0170] If the total evaluation score is greater than the preset score, the anomaly recognition model is judged to be highly effective; otherwise, the anomaly recognition model is judged to be insufficiently effective and a warning is issued.

[0171] In this embodiment, the first feature quantity refers to the number of features whose similarity values ​​are greater than a preset value of the abnormal network identification layer.

[0172] In this embodiment, the second feature quantity refers to the number of features whose similarity values ​​are less than a preset value of the abnormal network identification layer.

[0173] In this embodiment, the preset evaluation weight is the importance weight coefficient assigned to different evaluation indicators when evaluating the effectiveness of the anomaly recognition model.

[0174] In this embodiment, the preset score is a preset threshold or standard for the total evaluation score.

[0175] In this embodiment, the total evaluation score is a score of a comprehensive evaluation model calculated based on the first feature number (N1) and the second feature number (N2) and their similarity values ​​and preset values.

[0176] In this embodiment, the evaluation index may also be to use the root mean square error (RMSE), determination coefficient (R2) and mean absolute error (MAE) for the similarity values ​​under each abnormal type to achieve performance prediction of the model, which is helpful to objectively evaluate the effectiveness and reliability of the model, as follows:

[0177]

[0178] When the performance value is greater than a preset threshold, the judgment is reliable, and the value of the preset threshold is generally 0.8.

[0179] The working principle and beneficial effect of the above technical solution are: through feature quantity statistics and evaluation total score calculation, the effectiveness of the anomaly recognition model is objectively evaluated to improve the recognition efficiency.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent wolfberry pest image recognition method, characterized in that: include: Step 1: Determine the target acquisition object, build an environmental information data acquisition system according to the characteristics of the target acquisition object, and obtain the original image group of the target acquisition object; Step 2: preprocessing the original image group to obtain a standard image group, wherein the standard image group includes wolfberry images without pests and diseases and wolfberry images with pests and diseases; Step 3: Perform impurity filtering on the standard image group based on the filtering model to obtain a standard data set of the target acquisition object; Step 4: Construct an abnormality recognition model for the target acquisition object based on a machine learning algorithm and in combination with a standard data set; Step 5: Obtain the wolfberry verification image and input it into the abnormality recognition model to obtain the abnormality recognition result, and use the evaluation index to evaluate the performance of the abnormality recognition model to determine the effectiveness of the abnormality recognition model; Wherein, step 4 comprises: We extracted wolfberry pest and disease images of different preset analysis categories from the wolfberry image library in equal proportions, performed data preprocessing on the images of each category, and obtained a public data set; Develop a deep learning model using ensemble learning and transfer learning methods, and train and identify the public data set; According to the comparison between the training results and the recognition results, the fully connected layer of the deep learning model is adjusted, and the feature extraction part of the deep learning model is used as the feature extractor; Propose a first weight to the performance parameter of the feature extractor, and at the same time, set a second weight to different types of recognition processes by using a voting method based on the training results of the adjusted deep learning model and the feature comparison results of the standard data set; Integrate the first weight, the second weight, and the adjusted deep learning model to obtain an abnormality recognition model of the target acquisition object; Before using the evaluation index to evaluate the performance of the anomaly recognition model, the following steps are included: Extracting the abnormal feature y identified by the abnormal network recognition layer of each abnormal type in the abnormal recognition model for the wolfberry verification image and the preset recognition feature x of the corresponding abnormal network recognition layer, and calculating the similarity value between the abnormal feature y and the preset recognition feature x; in, Represents the similarity value between the preset recognition feature x of the corresponding abnormal network recognition layer and the abnormal feature y of the corresponding abnormal type existing in the wolfberry verification image, , They represent the feature brightness means of the preset identification feature x and the abnormal feature y respectively, They represent the feature brightness variance of the preset identification feature x and the abnormal feature y respectively, represents the feature brightness covariance of the preset recognition feature x and the abnormal feature y, and represents the stability constant; The performance of the anomaly recognition model is evaluated using evaluation indicators to determine the effectiveness of the anomaly recognition model, including: A first feature quantity N1 whose statistical similarity value is greater than or equal to a preset value of the corresponding abnormal network identification layer, and a second feature quantity N2 whose statistical similarity value is less than a preset value of the corresponding abnormal network identification layer, wherein the evaluation index is the first feature quantity and the second feature quantity; Calculate the total evaluation score S based on the first feature number N1 and the second feature number N2; ; in, 1 indicates the preset evaluation weight, represents the sum of all similarity values ​​under the first feature quantity N1; represents the sum of all preset values ​​under the first feature quantity N1; represents the sum of all preset values ​​under the second feature quantity N2; If the total evaluation score is greater than the preset score, the anomaly recognition model is judged to be highly effective; otherwise, the anomaly recognition model is judged to be insufficiently effective and a warning is issued.

2. The intelligent wolfberry pest image recognition method according to claim 1, characterized in that: Determine the target acquisition object, build an environmental information data acquisition system according to the characteristics of the target acquisition object, and obtain the original image group of the target acquisition object, including: Determine the target collection objects according to the types of pests and diseases suffered by wolfberry plants and plant types; According to the research needs of pests and diseases and the characteristics of the target collection objects, the collection equipment is matched, wherein all the collection equipment constitutes the environmental information data collection system; An original image group of the target acquisition object is acquired based on the acquisition device.

3. The intelligent wolfberry pest image recognition method according to claim 1, characterized in that: The process of preprocessing the original image group to obtain the standard image group includes: Performing data enhancement on the original image group to generate an enhanced image group; A first dimension processing library is established according to a preset image standard size, and the enhanced image group is introduced into the library for processing. If the aspect ratio of any enhanced image does not conform to the preset image standard size, padding and cropping operations are performed on the corresponding enhanced image to obtain a first image group.

4. The intelligent wolfberry pest image recognition method according to claim 3 is characterized in that: After obtaining the first image group, the method further includes: Performing pixel grayscale processing on any first image of the first image group to obtain a grayscale image, determining a grayscale value of each pixel in the grayscale image, and determining a grayscale level of a corresponding pixel based on a value-level mapping table; Counting the number of pixels at the same gray level in the gray image respectively, and constructing a gray distribution map according to the gray level order; Determine the grayscale mean of the grayscale distribution diagram ; ; ; in, represents the probability density distribution value of the pixel at the jth gray level in the corresponding gray image, n represents the number of gray levels, Represents the level value of the j-th gray level; Represents the proportion of the jth gray level determined based on the gray distribution diagram; represents the weight of the position where the j-th gray level is determined in the gray distribution diagram; Indicates all The maximum value in ; Indicates all The maximum value in ; Indicates all The maximum value among them; e represents a constant, and its value is 2.7; represents the average value of all grayscale values ​​in the jth grayscale level determined based on the grayscale distribution diagram; max represents the maximum value symbol; represents the variance of all grayscale values ​​in the jth grayscale level determined based on the grayscale distribution diagram; All pixels of the grayscale image are de-averaged according to the grayscale mean to obtain standard pixels, and a standard image corresponding to the first image is formed, thereby forming a standard image group.

5. The intelligent wolfberry pest image recognition method according to claim 1, characterized in that: The standard image group is filtered for impurities based on the filtering model to obtain a standard data set of the target acquisition object, including: Extracting a number of wolfberry-related images from the historical data set as positive impurity filtering samples, and extracting a number of non-wolfberry-related images from the historical data set as negative impurity filtering samples; Building a filtering recognition network, and performing filtering training on the positive impurity filtering samples and the negative impurity filtering samples to obtain a filtering model; The standard image group is imported into the filtering model to screen out all standard images that meet the requirements of wolfberry pests and diseases as the standard data set.

Citation Information

Patent Citations

  • CNN and transfer learning based disease intelligent identification method and system

    AU2020103613A4

  • An image recognition method for electric power equipment based on depth learning and a patrol robot

    CN109389180A

  • Image recognition method based on deep learning and application in rice disease recognition

    CN111833311A

  • Crop disease and pest identification method based on deep learning

    CN113609941A

  • Pepper disease identification method, identification system and computer readable storage medium

    CN114170137A