CN116310391B - Identification method for tea diseases - Google Patents
Identification method for tea diseases Download PDFInfo
- Publication number
- CN116310391B CN116310391B CN202310558925.7A CN202310558925A CN116310391B CN 116310391 B CN116310391 B CN 116310391B CN 202310558925 A CN202310558925 A CN 202310558925A CN 116310391 B CN116310391 B CN 116310391B
- Authority
- CN
- China
- Prior art keywords
- image
- graph
- node
- nodes
- tea
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
- G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Abstract
The application relates to a method for identifying tea diseases, which comprises the following steps: step one, data preprocessing: collecting tea disease image data, cutting each picture containing a plurality of diseases and background noise, and cutting the shot picture into a single disease blade to form an image; constructing an optimal small sample graph network model and training the model: firstly, embedding a tea disease image into feature vectors, then taking each feature vector as a two-domain node initialization diagram of the tea disease image, and carrying out updating optimization of the diagram according to the built two-domain node initialization diagram, wherein an optimal small sample diagram network model comprises three parts of reasoning from bottom to top, reasoning from top to bottom and jumping connection; and thirdly, carrying out tea disease image identification on the image. The intelligent tea leaf disease identification method reduces manpower and material resources consumed by human disease identification, and carries out tea leaf disease identification by using an intelligent identification technology.
Description
Technical Field
The application relates to the technical field of image processing, in particular to a method for identifying tea diseases.
Background
At present, tea diseases are identified mainly through expert investigation in the related field or artificial judgment of experienced farmers, the mode is relatively subjective, a large amount of manpower and material resources are consumed, in order to respond to the national green control call, and technologies such as big data, artificial intelligence and the like are applied to the green control of intelligent agriculture, how to provide a method for better monitoring and identifying tea diseases in a tea garden, and can monitor healthy tea and give different disease control suggestions to the diseased tea, so that the tea diseases can be found out in time, and the large-area diffusion of the diseases can be avoided as much as possible, and the tea disease identification method for reducing pesticide use to achieve the green control purpose is a technical problem which needs to be solved by the technicians in the field.
Disclosure of Invention
(1) Technical problem to be solved
The embodiment of the application provides a tea disease identification method, which comprises the steps of data preprocessing, constructing an optimal small sample graph network model, training the model, carrying out tea disease image identification on images and the like.
(2) Technical proposal
The embodiment of the application provides a method for identifying tea diseases, which comprises the following steps:
step one, data preprocessing: collecting tea disease image data, cutting each picture containing a plurality of diseases and background noise, and cutting the shot picture into a single disease blade to form an image;
constructing an optimal small sample graph network model and training the model: firstly, embedding a tea disease image into feature vectors, and then taking each feature vector as a two-domain node initialization map of the tea disease image; carrying out updating optimization on the graph according to the constructed two-domain node initialization graph, wherein the optimal small sample graph network model comprises three parts, namely bottom-up reasoning, top-down reasoning and jump connection;
thirdly, carrying out tea disease image recognition on the image, inputting the image to be recognized into a trained network model, calculating the distance between the image and other images by using a distance measurement function of the network model to measure the similarity between the image to be detected and the known image, and using the maximum similarity as a recognition result.
Further, the first step further includes: the method comprises the steps of small sample learning, wherein a data set in the small sample learning comprises a support set S and a query set Q, the support set S comprises N different classes, and each class comprises K marked samples; the query set Q contains Q samples to be classified.
Further, the dual-domain node initialization map is used for converting the RGB image obtained by the data preprocessing into a digital gray image, obtaining a dual-domain image pair composed of the RGB image and the digital gray image to serve as input of an embedded network, then respectively inputting the dual-domain image into two independent embedded modules to extract features, obtaining the RGB image and the digital gray image features after the features are embedded into the modules, and splicing the RGB image features and the digital gray image features to serve as nodes of the map.
Further, the second step specifically includes: after all characteristic representations of a certain task target sample are given, a full-communication diagram is firstly constructedWherein; />；/>Respectively representing a double-node set and an edge set in the full communication graph; representing the number of all samples in the task; then aggregating multi-domain image information through updating and jumping connection of the graph, and finally predicting the recognition result of the query set according to the updated edge feature set; in each top-down reasoning block, a graph update layer is firstly proposed to explore the correlation among all nodes in the same layer; then sampling the support nodes, wherein the reserved nodes of all layers are fully connected; graph feed combining upsampling recovery with graph feed providing the same number of nodes mainly through a jump connection in a bottom-up reasoning processUpdating a line graph; each bottom-up upsampling recovery block consists of a single picture update layer and a picture upsampling layer; the graph update layer aggregates information from adjacent nodes, and the graph up-sampling layer is responsible for recovering the nodes of the pooled clusters of the graph; in addition, jump connection exists between the same-level blocks from bottom to top and from top to bottom, so that low-level node characteristics in the bottom to top blocks and high-level node characteristics in the top to bottom blocks are effectively fused; finally is a softmax function that maps node features to node labels, which is used to optimize the overall network model using cross entropy element loss.
Further, the optimal small sample graph network model comprises a graph reasoning updating layer, a node pooling clustering layer, an up-sampling restoring layer and a jump connecting layer.
Further, the graph inference update layer includes adjacency matrix learning, graph rolling, and nonlinear multi-layer perceptron mapping, the adjacency matrix learning edges in the form of adjacency matrices from the known node representations through information transfer; and then carrying out graph rolling operation to update and infer a graph by combining the characteristics of the two-domain nodes, aggregating the learned low-level and high-level relations between the nodes by utilizing the low-level and high-level adjacent matrixes, aggregating the characteristics of each node from the adjacent nodes, connecting the two characteristic matrixes generated by the operations, and changing the dimension to obtain a new characteristic matrix.
Further, the node pooling clustering comprises intra-class node pooling and inter-class node pooling, firstly, an intra-class node pooling clustering layer is carried out, K nodes in each mode are reduced to single nodes, and then, an inter-class node pooling clustering layer is carried out.
Further, in the graph up-sampling recovery layer, each node is marked to reserve an index of a selected node in the original feature matrix, and the pooled nodes are put back to the original positions in the graph according to the reserved node indexes to generate the output new two-domain node features.
Further, in the optimal small sample graph network model, jump connection is realized by using simple splicing between a bottom-up reasoning block and a top-down reasoning block with the same node number, and node characteristics in the top-down reasoning block are directly added into the top-down reasoning block.
(3) Advantageous effects
The embodiment of the application is based on a dual-domain graph neural network, combines the tea disease image information of different domains mainly through the graph neural network, extracts the characteristics of each channel in the tea disease images of different domains, designs intra-class node pooling and inter-class node pooling clustering layers, and performs layered learning on intra-class and inter-class nodes.
For top-down reasoning, the downsampled graph is restored to the original size by utilizing the graph upsampling restoring layer, jump connection is also provided, multi-level characteristics are fused, the accuracy of similarity calculation of edges on node similarity measurement is improved, and finally node classification is carried out, so that the accuracy of tea disease identification is improved. According to the embodiment of the application, manpower and material resources consumed by manually identifying the diseases are reduced, and the intelligent identification technology is used for identifying the tea diseases.
Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the application.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings that are needed in the embodiments of the present application will be briefly described below, and it is obvious that the drawings described below are only some embodiments of the present application, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a flowchart of a method for identifying tea diseases according to the present application.
FIG. 2 is an embedded block diagram of a method for identifying tea leaf diseases according to the present application.
Fig. 3 is an algorithm network structure diagram of a method for identifying tea diseases according to the present application.
Fig. 4 is a node class inner pooling cluster proposed by the identification method for tea diseases.
Fig. 5 is a pooling cluster among node classes proposed by the identification method for tea diseases.
Fig. 6 is a flowchart of another identification method for tea diseases according to the present application.
Detailed Description
Embodiments of the present application are described in further detail below with reference to the accompanying drawings and examples. The following detailed description of the embodiments and the accompanying drawings are provided to illustrate the principles of the application and are not intended to limit the scope of the application, i.e., the application is not limited to the embodiments described, but covers any modifications, substitutions and improvements in parts, components and connections without departing from the spirit of the application.
It should be noted that, without conflict, the embodiments of the present application and features of the embodiments may be combined with each other.
The present application will be described in detail with reference to fig. 1 to 6 in conjunction with the embodiments.
The embodiment of the application provides a method for identifying tea diseases.
At present, the image neural network mainly takes RGB images as nodes, and omits information possibly contained in tea disease images in other image areas. The embodiment of the application is based on a dual-domain graph neural network, and the characteristics of each channel in the tea disease images in different domains are extracted mainly by combining the tea disease image information in different domains through the graph neural network. Meanwhile, in order to better combine the characteristics of the tea leaf disease images in different domains, a double-domain node is constructed by utilizing a mode of combining nodes in the graph neural network, and each node respectively represents the characteristics of different domains. The optimal small sample graph network model comprises three parts, namely bottom-up reasoning, top-down reasoning and jump connection, so that the optimal small sample graph network model is used for realizing efficient learning of multi-level relations. For top-down reasoning, the embodiment of the application designs an intra-class node pooling clustering layer and an inter-class node pooling clustering layer to perform hierarchical learning on intra-class and inter-class nodes. For top-down reasoning, the downsampled graph is restored to the original size using a graph upsampling restoration layer. The embodiment of the application also comprises jump connection, merges multi-level characteristics, improves the accuracy of node similarity measurement by edge similarity calculation, and finally classifies the nodes, thereby improving the accuracy of tea disease identification.
Specifically, in order to solve the technical problems, the application is realized by the following procedures: the method mainly comprises three flows of data preprocessing, optimal small sample graph network model training and image recognition.
1. Data preprocessing
Because the collected tea disease data is different in shooting equipment and shooting environment, the problems that one image of the collected image contains multiple diseases, the background noise is large and the like are caused, one image containing multiple diseases and the background noise is needed to be cut, the shot image is cut into a single disease blade to be one image, and the background noise is smaller. In order to avoid the problem that the number of the acquired original RGB images is small, the data of the original RGB images is expanded, the data set is expanded in a rotation, symmetry, retinex image enhancement mode and the like, and the expanded images can enable the disease recognition algorithm not to cause poor results due to insufficient data sets. According to the embodiment of the application, the characteristics of the disease image are fully acquired by combining the image gray domain information, and gray conversion is carried out after the original RGB image is acquired to obtain the digital gray domain picture, so that the RGB and gray double domain picture is acquired as a data set of a disease recognition algorithm.
After the data set is built, the data set is required to be divided, and in order to solve the problem of poor recognition effect caused by insufficient samples, the embodiment of the application solves the problem of insufficient samples by utilizing a unique training strategy of N-way K-shot in a small sample learning algorithm. The small sample classification task aims to train a classifier with good classification performance under the condition that only a small quantity of training samples are needed to participate in training. The data set in the small sample learning algorithm is mainly divided into two parts: (1) support set S: n different classes, each class containing K label samples, (2) query set Q: comprising q waitsClassifying the sample. Specifically, in each training round of the element learning of the small sample learning algorithm, the N-way K-shot classification task is subdivided into a training stage and a testing stage, and the whole data set sample contains C different categories in total, includingAnd is also provided with. In the training phase, a training set sample is given, which contains the support set +.>And a query setThe network is trained, and the network parameters are continuously updated through back propagation, so that the classification capability of the classifier in the training set is improved. The test phase, then the test set sample is given, which also contains the support set +.>And query set->The purpose is to use the classification model trained in the training phase in the support set +.>At this time, a small number of support set samples are used to sample the query set +.>Accurately maps to the correct label to realize accurate and efficient identification task.
After the data set is divided according to the small sample learning training algorithm strategy, the recognition algorithm can be input for network training to obtain an optimal algorithm model.
2. Network model training for optimal small sample graph
When the embodiment of the application is used for training the optimal small sample graph network model, a two-domain node initialization graph is constructed. Specifically, the step of constructing the two-domain node initialization map includes: firstly, embedding a tea disease image into feature vectors, and then taking each feature vector as a two-domain node initialization map of the tea disease image. In order to increase structural information contained in the nodes, multi-dimensional information in the samples is fully utilized, and the embodiment of the application passes the original RGB image through a preprocessing module. The module converts the RGB image into the digital gray image to obtain a double-domain image pair consisting of the RGB image and the digital gray image as the input of an embedded network, then the double-domain image is respectively input into two independent embedded modules to extract the characteristics, the RGB image and the digital gray image characteristics are obtained after the embedded modules are used for being embedded, the RGB image characteristics and the digital gray image characteristics are spliced to be used as the nodes of the image, and therefore the built image can better utilize the double-domain information of the image to update the characteristics. As shown in fig. 2. This figure illustrates the initialization of the graph in the 2-way-2-shot small sample classification task. Firstly, feature extraction is carried out on the tea disease RGB and gray scale double-domain image through a CNN module. And then, splicing the extracted RGB features and gray features to construct nodes, wherein the similarity between the nodes is used as an edge construction initialization diagram.
And performing optimal small sample graph network model training, wherein the optimal small sample graph network model training mainly performs continuous training on an algorithm to obtain an optimal network model, and then performs tea disease identification. The method mainly carries out updating optimization of the graph according to the built two-domain node initialization graph, and an optimal small sample graph network model mainly comprises three parts of bottom-up reasoning, top-down reasoning and jump connection. The embodiment of the application is mainly inspired by a U-Net network, as shown in fig. 3, and the problem of classification task of 2-way-2-shot small samples is illustrated by taking two bottom-up and two top-down graph update blocks as examples in the figure. Circles and squares represent two different support set class nodes, and five-pointed star represents a query set sample node. Each node represents the extracted characteristics of the RGB image and the digital gray image, integrates the information of the multi-domain image through updating and jumping connection of the image, and finally classifies the query set image. Specifically, given all characteristic representations of a task target sample, full communication is constructed firstDrawing of the figureWherein->；Representing a set of dual nodes and a set of edges in the full communication graph, respectively. />Indicating the number of all samples in the task. And then aggregating multi-domain image information through updating and jumping connection of the graph, and finally predicting the recognition result of the query set according to the updated edge feature set.
In each top-down reasoning block, the embodiment of the application first adopts the graph updating layer proposed by us to explore the correlation between all nodes (including the support node and the query node) in the same layer. The support nodes are then sampled using a layer of node pooling clusters (intra-class node pooling and inter-class node pooling), with the reserved nodes of all layers being fully connected. Several bottom-up reasoning blocks are stacked to learn relationships hierarchically. In the bottom-up reasoning process, the graph is updated mainly by providing the graph of the same number of nodes through jump connection and combining the graph recovered by up-sampling. Each bottom-up upsampling recovery block consists of a single picture update layer and a picture upsampling layer. The graph update layer aggregates information from neighboring nodes, and the graph up-sampling layer is responsible for recovering the nodes of the graph clustered by pooling. In addition, jump connection exists between the same-level blocks from bottom to top and from top to bottom, and the low-level node characteristics in the bottom to top blocks and the high-level node characteristics in the top to bottom blocks are effectively fused. Finally is the softmax function in the algorithm, which maps node features to node labels. The use of cross entropy element loss is used to optimize the overall network model.
The proposed intra-class node pooling clustering attempts to extract intra-class commonalities for each class and causes the underlying graph update inference layer to learn the relationship between the inferred class average features and the query. The purpose of the proposed inter-class node pooling clustering is to preserve the indistinguishable nodes so that the underlying graph update inference layer can learn the relationships between the preserved nodes and the query nodes. In the graph update inference layer, nodes can obtain information from other nodes, and therefore, even if some nodes are reduced, the characteristics of the reduced nodes can be preserved in the next graph update inference layer. The idea of the intra-class node pooling and inter-class node pooling clustering layers is to preserve nodes that need further utilization. The following is a detailed description of the various blocks in the algorithm block diagram.
(1) Graph inference update layer
The purpose of the graph inference update layer is to explore the relationships between nodes in the same layer. Typically consists of three parts, adjacency matrix learning, graph rolling, and nonlinear multi-layer perceptron (MLP) mapping. Wherein the adjacency matrix learning learns edges in the form of adjacency matrices from the representation of the known nodes by information transfer. The learning of the adjacency matrix is particularly important when the input data is a graph structure but the metrics are not known. The adjacency matrix is calculated in such a way that the similarity between the two-domain nodes is mainly used as the value at each position of the matrix. After the adjacency matrix is adopted, graph convolution operation is carried out by combining the characteristics of the two-domain nodes to update and infer the graph, the low-level and high-level relations learned between the nodes are aggregated by utilizing the low-level and high-level adjacency matrices, the characteristics of each node are aggregated from the adjacent nodes, then two characteristic matrices generated by the GCN operation are connected, the dimension is changed by using a nonlinear MLP, a new characteristic matrix is obtained, each graph inference update layer is provided with an MLP, and the different graph inference update layers do not share the parameters of the MLP.
(2) Node pooling clustering
Node pooling clusters are mainly divided into intra-class node pooling and inter-class node pooling. Pooling operations are typically used to reduce the size of feature maps. The embodiment of the application provides a learnable two-domain node pooling clustering layer for downsampling a graph and learning a hierarchical representation of the graph. The two-domain node pooling clustering layer only performs downsampling on the support nodes, and reserves all query nodes, so as to learn the relation between the multi-stage support nodes and all query nodes, and further learn to perform meaningful feature embedding on all query nodes. In addition, node pooling clustering, and in particular, intra-class node pooling clustering layers, aim to reduce the number of nodes in a particular manner, while inter-class node pooling clustering layers aim to reduce the number of nodes in all ways. For an N-way K-shot small sample learning task algorithm, we firstly perform intra-class node pooling clustering layers, reduce K nodes in each mode to a single node, and then perform inter-class node pooling clustering layers.
(3) On-map sampling recovery layer
In order to restore the pooled feature representation to the original resolution, an upsampling operation is performed, the graph upsampling restoration layer being the inverse of the node pooled clustering layer. In the present algorithm, embodiments of the present application design a graph up-sampling recovery layer in the bottom-up reasoning block to perform the up-sampling operation, which is the inverse of the down-sampling operation. In the up-sampling recovery layer, each node is labeled to preserve the index of the selected node in the original feature matrix. From the saved node index, we put the pooled nodes back to the original position in the graph to generate the output new two-domain node feature.
(4) Jump connection
In the present algorithm, embodiments of the present application implement a jump connection with a simple splice between a bottom-up reasoning block and a top-down reasoning block with the same number of nodes. Node features in the top-down reasoning block are added directly to the top-down reasoning block. It should be noted here that some two-domain intra-class nodes and inter-class nodes are reduced by intra-class node pooling clusters and inter-class node pooling clusters. The hopping connections in the present algorithm may also pass the characteristics of the reduced nodes to the lower layers. Further, in the graph update inference layer, a node may receive messages from other nodes. Even if some nodes are reduced, the features of the reduced nodes can be preserved in the graph update inference layer below.
3. Image recognition function
After an optimal model is obtained by training an algorithm, a tea disease image recognition function can be performed, wherein the recognition function is a core function of the recognition system, and when an unknown tea disease image is input into an optimal small sample graph network model, the type of the disease can be rapidly and accurately recognized by the recognition method provided by the embodiment of the application. At this time, the image to be identified is input into the trained network model, the distance between the image and other images is finally calculated by using the distance measurement function of the network model to measure the similarity between the image to be identified and the known image, and the similarity is used as the identification result with the maximum similarity
The application is illustrated in the following by way of further examples: the whole identification process in the embodiment of the application is a set of systematic and standard operation flow. Referring to fig. 1, the method mainly comprises three stages: the first stage is data preprocessing; the second stage is optimal network training; the third stage is an image recognition function.
Data preprocessing (one)
The data preprocessing comprises cutting of tea disease pictures, data enhancement and gray scale transformation. Because of different shooting equipment and shooting environments, the resolution of the acquired pictures is different, and some acquired images contain some tea disease images and some images contain complicated noise backgrounds, so that the images need to be cut. Since the same-size image is required to be sent in when the optimal network is trained, the cut-out images with different sizes can be adjusted to 84×84 image sizes through bilinear interpolation. In order to prevent the poor result of the model during training caused by the too small amount of the acquired image data, the data expansion can be carried out on the cut data, the expansion method mainly comprises rotation, symmetry, retinex enhancement and the like, and the network can not cause the poor result due to the insufficient amount of the data during training after the data expansion. In order to fully utilize multi-domain information of the images, not only the information of RGB images but also the information of digital gray images are utilized when the network is trained, the cut and expanded images need to be subjected to gray level transformation, and then the two-domain images are sent to the network to extract features.
(II) optimal network training
To better verify the validity of the algorithm herein, embodiments of the present application are configured as a CPU: i9-12900K, GPU: experiments were performed on a GeForce RTX3090-OX24G apparatus, with the following specific experimental settings:
1. embedded module structure
The embedded network of the experiment selects two network structures Convent and ResNet12 which are mainstream at present. The invent mainly comprises 4 convolution blocks with different sizes, and each convolution block comprises a Conv-BN-ReLU structure; resNet12 is mainly formed by combining 4 ResNet12 residual blocks through jump connection. In the algorithm model, two independent embedded network frames are needed to extract sample characteristics due to the dual-node characteristic of the network, and the final node characteristic is a 2×128-dimensional characteristic vector.
2. Training process
Before training starts, the data samples are preprocessed to obtain digital gray images corresponding to RGB sample images. According to the training method provided in the small sample learning based on the graph neural network, all experiments of the algorithm model adopt an Adam optimizer, the initial learning rate is 5 multiplied by 10 < -4 >, the weight attenuation is 10 < -6 >, and the final result adopts the average accuracy rate of every 1000 rounds.
3. Evaluation
The experimental evaluation of the algorithm network model is carried out on a standard small sample learning data set miniImageNet, tieredImageNet and a self-collected tea disease data set, and the adopted evaluation method follows the evaluation method proposed in the small sample learning based on the graph neural network to give the average accuracy (%) of the random extraction task and a 95% confidence interval.
4. Results of the main experiments
In order to verify the effectiveness of the network model, several different small sample learning methods are compared on the miniImageNet dataset, the Tieredimagenet dataset and the tea disease dataset acquired by the user, including small sample learning methods with graph structures and non-graph structures. According to experimental results, the classification accuracy of 5way-1shot and 5way-5shot in the miniImageNet data set and the TieredImageNet data set is improved compared with that of the current advanced graph network model. In addition, in the collected tea disease data set, the experimental results of 5way-1shot and 5way-5shot show that the use of the double-domain image for identifying tea diseases is obviously better than the use of the single-domain RGB image. In summary, the proposed tea disease classification algorithm of the two-domain graph neural network has better classification performance compared with the task with the graph structure in the current advanced small sample learning task.
In the embodiment of the application, a double-domain map neural network tea disease recognition algorithm based on small sample learning is provided, the network adopts images of different domains of samples to construct sub-nodes, and remarkable high-frequency information in a digital gray level image is used as supplement. The multi-dimensional information existing in images of different domains of the sample is fully utilized, and the expression capability of the node characteristics to the sample is improved. In addition, in order to fully utilize multidimensional information among different domains of a sample in the updating process of the graph structure, a double-domain hierarchical graph neural network is designed, node features and edge features are aggregated into richer neighborhood information through alternate updating of double nodes and edges, and the expression capacity of the updated node features and the distance measurement capacity of the edge features to the nodes are enhanced. Meanwhile, the network consists of three parts of reasoning from bottom to top, reasoning from top to bottom and jump connection, so that the efficient learning of the multi-level relation is realized. For top-down reasoning, we design intra-class node pooling clustering and inter-class node pooling clustering layers to learn intra-class and inter-class nodes hierarchically. For top-down reasoning, the downsampled graph is restored to the original size using a graph upsampling restoration layer. And the jump connection is also provided, the multi-level characteristics are fused, and finally the node classification is carried out. A large number of experiments show that in the field of small sample learning, the classification performance of the algorithm provided by the application is superior to that of other advanced Graph Neural Network (GNN) methods.
(III) image recognition function
The identification function is a core function of the identification system, and when an unknown tea disease image is input into the system, the type of the disease can be quickly and accurately identified through the identification method provided by the embodiment of the application. At this time, the image to be identified is required to be input into the trained network model, and finally the distance between the image and other images is calculated by using the distance measurement function of the network model to measure the similarity between the image to be identified and the known image, and the similarity is used as an identification result.
Specifically, the identification method for tea diseases provided by the embodiment of the application comprises the following specific operation steps:
s1: the mobile image acquisition equipment comprises a camera and an unmanned aerial vehicle for acquiring tea disease data;
s2: preprocessing the acquired picture to obtain an RGB and gray double-domain image and a data set with enhanced data;
s3: the code of the method is continuously trained to obtain an optimal network model and a processed tea disease picture;
s4: selecting disease pictures to be identified by using the whole identification system for classification;
s5: obtaining a final classification result by utilizing an identification function in an identification system;
s6: and printing the final identification result in a display window.
In summary, the embodiment of the application mainly combines the tea disease image information of different domains through the graph neural network, extracts the characteristics of each channel in the tea disease images of different domains, designs intra-class node pooling clustering and inter-class node pooling clustering layers, and performs hierarchical learning on intra-class and inter-class nodes. For top-down reasoning, the downsampled graph is restored to the original size by utilizing the graph upsampling restoring layer, jump connection is also provided, multi-level characteristics are fused, the accuracy of similarity calculation of edges on node similarity measurement is improved, and finally node classification is carried out, so that the accuracy of tea disease identification is improved. According to the embodiment of the application, manpower and material resources consumed by manually identifying the diseases are reduced, and the intelligent identification technology is used for identifying the tea diseases.
For embodiments of the method, reference may be made to the description of parts of embodiments of the apparatus. The application is not limited to the specific steps and structures described above and shown in the drawings. Also, a detailed description of known method techniques is omitted here for the sake of brevity.
The above description is only an example of the present application and is not limited to the present application. Various modifications and alterations of this application will become apparent to those skilled in the art without departing from the scope of this application. Any modification, equivalent replacement, improvement, etc. which come within the spirit and principles of the application are to be included in the scope of the claims of the present application.
Claims (7)
1. A method for identifying tea diseases, comprising the steps of:
step one, data preprocessing: collecting tea disease image data, cutting each picture containing a plurality of diseases and background noise, and cutting the shot picture into a single disease blade to form an image;
constructing an optimal small sample graph network model and training the model: firstly, embedding a tea disease image into feature vectors, then taking each feature vector as a two-domain node initialization diagram of the tea disease image, and carrying out updating optimization of the diagram according to the built two-domain node initialization diagram, wherein an optimal small sample diagram network model comprises three parts of reasoning from bottom to top, reasoning from top to bottom and jumping connection;
thirdly, carrying out tea disease image recognition on the image, inputting the image to be recognized into a trained network model, calculating the distance between the image and other images by using a distance measurement function of the network model to measure the similarity between the image to be detected and the known image, and using the maximum similarity as a recognition result;
the double-domain node initializing graph is used for converting an RGB image obtained by preprocessing the data into a digital gray image, obtaining a double-domain image pair consisting of the RGB image and the digital gray image to serve as input of an embedded network, then respectively inputting the double-domain image into two independent embedded modules to extract characteristics, obtaining the RGB image and the digital gray image characteristics after passing through the embedded modules, and splicing the RGB image characteristics and the digital gray image characteristics to serve as nodes of the graph;
the second step specifically comprises the following steps: given a certain task target sampleAfter all the characteristics are expressed, a full-communication diagram is firstly constructedWherein; />；The method comprises the steps of carrying out a first treatment on the surface of the Respectively representing a double-node set and an edge set in the full communication graph; t represents the number of all samples in the task; then aggregating multi-domain image information through updating and jumping connection of the graph, and finally predicting the recognition result of the query set according to the updated edge feature set; in each top-down reasoning block, a graph update layer is firstly proposed to explore the correlation among all nodes in the same layer; then sampling the support nodes, wherein the reserved nodes of all layers are fully connected; in the bottom-up reasoning process, the graphs with the same number of nodes are provided through jump connection, and then the graphs are updated by combining the graphs recovered by up-sampling; each bottom-up upsampling recovery block consists of a single picture update layer and a picture upsampling layer; the graph update layer aggregates information from adjacent nodes, and the graph up-sampling layer is responsible for recovering the nodes of the pooled clusters of the graph; in addition, there is a jump connection between the same level blocks from bottom up and top down; finally is a softmax function that maps node features to node labels, which is used to optimize the overall network model using cross entropy element loss.
2. A method for identifying tea diseases according to claim 1, wherein step one further comprises: the method comprises the steps of small sample learning, wherein a data set in the small sample learning comprises a support set S and a query set Q, the support set S comprises N different classes, and each class comprises K marked samples; the query set Q contains Q samples to be classified.
3. The method for identifying tea leaf diseases according to claim 1, wherein the optimal small sample graph network model comprises a graph inference update layer, a node pooling cluster, an up-sampling restoration layer and a jump connection layer.
4. A method of identifying tea leaf diseases as claimed in claim 3 wherein the graph inference update layer comprises adjacency matrix learning, graph convolution and nonlinear multi-layer perceptron mapping, the adjacency matrix learning edges in the form of adjacency matrices from a representation of known nodes by information transfer; and then carrying out graph rolling operation to update and infer a graph by combining the characteristics of the two-domain nodes, aggregating the learned low-level and high-level relations between the nodes by utilizing the low-level and high-level adjacent matrixes, aggregating the characteristics of each node from the adjacent nodes, connecting the two characteristic matrixes generated by the operations, and changing the dimension to obtain a new characteristic matrix.
5. A method for identifying tea leaf diseases according to claim 3, wherein the node pooling clustering comprises intra-class node pooling and inter-class node pooling, wherein the intra-class node pooling clustering layer is performed first, K nodes of each mode are reduced to a single node, and then the inter-class node pooling clustering layer is performed.
6. A method of identifying tea leaf diseases as claimed in claim 3 wherein in the on-map sample recovery layer, each node is marked to retain an index of a selected node in the original feature matrix, and the pooled nodes are returned to their original positions in the map in accordance with the stored node indices to generate the output new double-domain node features.
7. A method for identifying tea leaf diseases according to claim 3, wherein in the optimal small sample graph network model, jump connection is implemented by using simple concatenation between a bottom-up reasoning block and a top-down reasoning block with the same number of nodes, and node features in the top-down reasoning block are directly added to the bottom-up reasoning block.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310558925.7A CN116310391B (en) | 2023-05-18 | 2023-05-18 | Identification method for tea diseases |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310558925.7A CN116310391B (en) | 2023-05-18 | 2023-05-18 | Identification method for tea diseases |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116310391A CN116310391A (en) | 2023-06-23 |
| CN116310391B true CN116310391B (en) | 2023-08-15 |
Family
ID=86798139
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310558925.7A Active CN116310391B (en) | 2023-05-18 | 2023-05-18 | Identification method for tea diseases |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116310391B (en) |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020047739A1 (en) * | 2018-09-04 | 2020-03-12 | 安徽中科智能感知大数据产业技术研究院有限责任公司 | Method for predicting severe wheat disease on the basis of multiple time-series attribute element depth features |
| CN111047606A (en) * | 2019-12-05 | 2020-04-21 | 北京航空航天大学 | Pathological full-section image segmentation algorithm based on cascade thought |
| EP3739504A1 (en) * | 2019-05-16 | 2020-11-18 | Basf Se | System and method for plant disease detection support |
| CN112434662A (en) * | 2020-12-09 | 2021-03-02 | 安徽大学 | Tea leaf scab automatic identification algorithm based on multi-scale convolutional neural network |
| CN113221942A (en) * | 2021-03-24 | 2021-08-06 | 安徽大学 | Tea disease identification algorithm under small sample based on deep migration and Cayley-Klein measurement |
| CN113297427A (en) * | 2021-05-10 | 2021-08-24 | 中南大学 | Graph classification method based on U-shaped nested network |
| CN113989639A (en) * | 2021-10-20 | 2022-01-28 | 华南农业大学 | Hyperspectral image analysis processing method-based automatic litchi disease identification method and device |
| CN114022872A (en) * | 2021-09-24 | 2022-02-08 | 中国海洋大学 | Multi-crop leaf disease identification method based on dynamic neural network |
| CN115761502A (en) * | 2022-11-17 | 2023-03-07 | 西安电子科技大学 | SAR image change detection method based on hybrid convolution |
| CN116129196A (en) * | 2023-03-03 | 2023-05-16 | 西安电子科技大学 | Tea plant diseases and insect pests classification method, system, equipment and medium based on small sample learning |
-
2023
- 2023-05-18 CN CN202310558925.7A patent/CN116310391B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020047739A1 (en) * | 2018-09-04 | 2020-03-12 | 安徽中科智能感知大数据产业技术研究院有限责任公司 | Method for predicting severe wheat disease on the basis of multiple time-series attribute element depth features |
| EP3739504A1 (en) * | 2019-05-16 | 2020-11-18 | Basf Se | System and method for plant disease detection support |
| CN111047606A (en) * | 2019-12-05 | 2020-04-21 | 北京航空航天大学 | Pathological full-section image segmentation algorithm based on cascade thought |
| CN112434662A (en) * | 2020-12-09 | 2021-03-02 | 安徽大学 | Tea leaf scab automatic identification algorithm based on multi-scale convolutional neural network |
| CN113221942A (en) * | 2021-03-24 | 2021-08-06 | 安徽大学 | Tea disease identification algorithm under small sample based on deep migration and Cayley-Klein measurement |
| CN113297427A (en) * | 2021-05-10 | 2021-08-24 | 中南大学 | Graph classification method based on U-shaped nested network |
| CN114022872A (en) * | 2021-09-24 | 2022-02-08 | 中国海洋大学 | Multi-crop leaf disease identification method based on dynamic neural network |
| CN113989639A (en) * | 2021-10-20 | 2022-01-28 | 华南农业大学 | Hyperspectral image analysis processing method-based automatic litchi disease identification method and device |
| CN115761502A (en) * | 2022-11-17 | 2023-03-07 | 西安电子科技大学 | SAR image change detection method based on hybrid convolution |
| CN116129196A (en) * | 2023-03-03 | 2023-05-16 | 西安电子科技大学 | Tea plant diseases and insect pests classification method, system, equipment and medium based on small sample learning |
Non-Patent Citations (1)
| Title |
|---|
| 基于多尺度特征拼接的小样本茶叶病害分类;张艳 等;安徽大学学报(自然科学版);第46卷(第05期);58-63 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116310391A (en) | 2023-06-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112150493B (en) | Semantic guidance-based screen area detection method in natural scene | |
| CN108052966A (en) | Remote sensing images scene based on convolutional neural networks automatically extracts and sorting technique | |
| CN113609889B (en) | High-resolution remote sensing image vegetation extraction method based on sensitive characteristic focusing perception | |
| CN108596240B (en) | Image semantic segmentation method based on discriminant feature network | |
| CN108596243A (en) | The eye movement for watching figure and condition random field attentively based on classification watches figure prediction technique attentively | |
| CN114419413A (en) | Method for constructing sensing field self-adaptive transformer substation insulator defect detection neural network | |
| CN114494195A (en) | Small sample attention mechanism parallel twinning method for fundus image classification | |
| CN110110724A (en) | The text authentication code recognition methods of function drive capsule neural network is squeezed based on exponential type | |
| CN112950780A (en) | Intelligent network map generation method and system based on remote sensing image | |
| CN115797781A (en) | Crop identification method and device, computer equipment and storage medium | |
| CN116452810A (en) | Multi-level semantic segmentation method and device, electronic equipment and storage medium | |
| CN114510594A (en) | Traditional pattern subgraph retrieval method based on self-attention mechanism | |
| CN105809200A (en) | Biologically-inspired image meaning information autonomous extraction method and device | |
| CN116310391B (en) | Identification method for tea diseases | |
| CN115330759B (en) | Method and device for calculating distance loss based on Hausdorff distance | |
| CN116523877A (en) | Brain MRI image tumor block segmentation method based on convolutional neural network | |
| CN112818982B (en) | Agricultural pest image detection method based on depth feature autocorrelation activation | |
| CN114998702A (en) | Entity recognition and knowledge graph generation method and system based on BlendMask | |
| CN114511732A (en) | Citrus spotted disease and insect pest fine-grained image identification method | |
| CN114049500A (en) | Image evaluation method and system based on meta-learning reweighting network pseudo label training | |
| Awad et al. | Multicomponent image segmentation: a comparative analysis between a hybrid genetic algorithm and self‐organizing maps | |
| Zhang et al. | Bayesian Layer Graph Convolutioanl Network for Hyperspetral Image Classification | |
| Mayuri et al. | Artificial Neural Network (ANN) With Chan-Vese (CV) Algorithm-Based Plant Disease Detection And Classification | |
| CN117649631B (en) | Client image processing method and system based on improved convolutional neural network | |
| Li et al. | Spatial-spectral hyperspectral classification based on learnable 3D group convolution |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |