CN114266965A - Bitter orange pest and disease identification method based on multitask learning - Google Patents
Bitter orange pest and disease identification method based on multitask learning Download PDFInfo
- Publication number
- CN114266965A CN114266965A CN202111514958.9A CN202111514958A CN114266965A CN 114266965 A CN114266965 A CN 114266965A CN 202111514958 A CN202111514958 A CN 202111514958A CN 114266965 A CN114266965 A CN 114266965A
- Authority
- CN
- China
- Prior art keywords
- network
- bitter orange
- pest
- training
- model
- 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.)
- Pending
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a method for identifying diseases and insect pests of bitter orange based on multi-task learning, which trains an identification network by adopting two tasks of identification and reconstruction. The method comprises the steps of carrying out image classification training on an bitter orange pest and disease data set by adopting a plurality of pre-training models, selecting a pre-training model with better generalization capability and carrying out fine tuning on the pre-training model to serve as a main shared network of a multi-task learning model, then designing a branch task network for bitter orange pest and disease classification and feature reconstruction, and finally training an image recognition model and carrying out classification recognition on bitter orange pests and diseases by utilizing the bitter orange pest and disease data set through weight adjustment on different tasks. The invention performs multi-task coordination optimization on the recognition model based on the idea of multi-task learning to obtain a better pest and disease recognition network model, thereby improving the recognition accuracy.
Description
Technical Field
The invention relates to the technical field of bitter orange pest and disease identification, in particular to a method for identifying bitter orange pest and disease based on multitask learning.
Background
The fructus aurantii is a dried immature fruit of Citrus aurantium L (Citrus aurantium L.) of Citrus of Rutaceae and its cultivar, and is used as an important Chinese medicinal material for treating food stagnation, inhibiting gastrointestinal motility, boosting pressure, strengthening heart, resisting ulcer, shock, relieving pain and other pharmacological effects, and is also an important raw material for medicinal material market, clinical application and the like. Therefore, the healthy and stable development of the bitter orange planting industry has important practical significance in the aspects of medicine, agriculture and the like. The key to the stable development of the bitter orange planting industry lies in the quality guarantee of the bitter orange, in each growth period of the bitter orange, due to factors such as the biological attribute of the bitter orange and seasonal changes of the external environment, the bitter orange is easy to cause fruit, branch and leaf diseases and become a delicious food for various pests, if a large amount of infection and transmission of the diseases and insect pests of the bitter orange occur, the quality of the bitter orange can be seriously threatened, and further the development of the bitter orange planting industry is greatly restricted, so that the timely discovery and treatment of the diseases and insect pests of the bitter orange are very important. However, the basic knowledge of pest control is popularized to rural farmers since the time when the professor uses fish is not as good as the time when the professor uses fish and the time when sheep death is difficult to be mended is late, so that the professor enjoys the era of red profit brought by the high-speed development of the internet and teaches that the problem is solved by the information acquisition skill of the high-speed developed internet.
The traditional identification and management of the diseases and insect pests of the bitter orange completely depends on manual observation experience, diagnosticians need to have higher professional level and rich experience, and planting and maintenance personnel of the bitter orange have limited professional level relative to professionals and cannot make timely and effective judgment in case of diseases and insect pests, so that the disease and insect pest propagation causes the expansion of the infection area of a planting place of the bitter orange to influence the yield and quality of the bitter orange. Image recognition technology has been developed for many years, but the traditional part of planting industry still does not enjoy the great benefits brought by the image recognition technology, which shows that the information technology developed at a high speed and the lagging planting production mode in partial remote areas of China are important contradictions restricting the economic development of China.
The image recognition technology is widely applied to various aspects of people's life at present, the image recognition technology is mature at present, the recognition rate and the accuracy rate can be very high, and theoretically, the image recognition technology can also be applied to recognition of relevant products in the agricultural planting industry. In the field of pest and disease identification, experts and scholars at home and abroad carry out a great deal of research, which mainly focuses on two aspects: the first is the traditional machine learning method; the second is a deep learning approach. The traditional machine learning method mainly focuses on feature extraction and selection, such as extracting edge features of disease and insect scars. The method usually needs complex characteristic engineering, has low identification precision and is difficult to identify the plant diseases and insect pests with complex backgrounds. The Deep learning method combines the advantages of a Deep Network, the Deep characteristics of a training model for learning the pest and disease picture are better than those of a traditional method, a Deep Convolutional Neural Network (DCNN) is the most fierce Deep learning research field at present, the Deep Convolutional Neural Network has obvious performance advantages in the field of recognition and classification, but the training model is high in cost and cost, and a huge data set is required to be used as a support to obtain a good recognition effect.
Compared with other crops, the yield of the fructus aurantii is low, and the cognition degree of people on the fructus aurantii is not high, so that the pest and disease identification research on the fructus aurantii at home and abroad is few at present, and more attention is paid to the pest and disease identification of citrus in the existing work. Although bitter orange has many similarities with citrus, there are some differences in the main categories and the pathogenesis of diseases and insect pests. Therefore, the research on the method for identifying the diseases and insect pests of the bitter orange has important significance and application prospect.
Most of the previous crop disease and pest identification methods are carried out based on a single-task learning mode, but the yield of the bitter orange is low compared with other crops, and the cognition degree of people to the bitter orange is not very high, so that the collected picture data set of the bitter orange disease and pest is low, and due to the limitation of the number of images in the data set, under the single-task learning mode, the over-fitting phenomenon often occurs to a trained network model, and then the higher classification performance cannot be achieved on a test set.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide the bitter orange pest and disease identification method based on the multitask learning, and the recognition model is subjected to multitask coordination optimization based on the multitask learning idea to obtain a better pest and disease identification network model, so that the recognition accuracy is improved.
In order to achieve the purpose, the invention adopts the following specific technical scheme:
a method for identifying diseases and insect pests of fructus aurantii based on multitask learning specifically comprises the following steps:
step 1, taking an bitter orange pest and disease damage data set as training data, performing feature extraction on the training data through a plurality of pre-training models, and adding a full-connection layer to perform classification training on extracted feature vectors;
step 2, comparing and analyzing classification performance and generalization capability of all pre-training models on the bitter orange pest and disease damage data set, and selecting the pre-training model with the optimal generalization capability to construct a trunk sharing network;
step 3, a hard parameter sharing mechanism in multi-task learning is adopted to construct a first branch task network for bitter orange pest classification and a second branch task network for feature reconstruction, the input of a backbone sharing network is training data, and feature vectors of the training data are output to the first branch task network and the second branch task network; setting the structures and loss functions of a first branch task network and a second branch task network;
step 4, training and optimizing a first branch task network and a second branch task network, and storing an bitter orange pest and disease identification model formed by a trunk sharing network, the first branch task network and the second branch task network;
and 5, inputting the pest and disease damage image of the fructus aurantii into the pest and disease damage identification model of the fructus aurantii to obtain an identification result.
Preferably, the pre-trained model in step 2 comprises 4 dense connection blocks, and each dense connection block is connected with each other through a transition layer.
Preferably, in step 2, the backbone sharing network is connected with a global average pooling layer and a full connection layer on the feature output layer of the pre-training model.
Preferably, the step 3 of setting the structure and the loss function of the first branch task network specifically includes:
and setting the network connection of the first branch task to be connected with a 20-dimensional full connection layer, and calculating classification loss through the output of a SoRmax layer, wherein cross entropy is used as a loss function in the network training process.
Preferably, the step 3 of setting the structure and the loss function of the second branch task network specifically includes:
setting a second branch task network to comprise four full-connection layers, and then connecting a full-inverse local average pooling layer to complete feature reconstruction, wherein the mean square error is used as a loss function in the network training process.
Preferably, in step 4, the first branch task network and the second branch task network are jointly optimized, the weight is updated in the gradient back propagation process, and the final loss of the bitter orange pest identification model is represented as follows:
Ltotal＝αLcls+βLfeaturewherein L isclsAnd LfeatureAnd the loss functions of the bitter orange pest classification task and the characteristic reconstruction task are respectively expressed, and alpha and beta are corresponding two task weights.
Preferably, α has a preferred value of 1.8 and β has a preferred value of 0.2.
Preferably, the models in step 1 and step 4 are trained using the same hyper-parameters.
The invention has the beneficial effects that: based on the idea of multi-task learning, multi-task coordination optimization is carried out on the recognition model, a better pest and disease recognition network model is obtained, and the recognition accuracy is further improved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a flow chart of a method for identifying diseases and insect pests of fructus Aurantii based on multitask learning according to the present invention;
fig. 2 is a schematic diagram of a bitter orange pest identification model.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Other embodiments, which can be derived by one of ordinary skill in the art from the embodiments given herein without any creative effort, shall fall within the protection scope of the present invention.
In deep learning, multi-task learning is divided into a hard parameter sharing mechanism and a soft parameter sharing mechanism according to different parameter sharing modes of a hidden layer of a deep network. The hard parameter sharing mechanism is realized by sharing a hidden layer among all tasks and simultaneously reserving output layers of a plurality of specific tasks, the risk of overfitting of the model can be reduced to a certain extent, and the generalization capability of the model is stronger when more tasks are learned at the same time. Therefore, the hard parameter sharing mechanism in the multi-task learning is common in practical application. Each task in the soft parameter sharing mechanism has its own parameters and model, which encourages model parameter similarity by regularizing the distance between model parameters, for example, regularizing using the L2 distance. The effects of multi-task learning are caused by a variety of working mechanisms, including implicit data augmentation, attention mechanism, eavesdropping, signature biasing, regularization, and the like.
As shown in fig. 1, the invention provides a method for identifying diseases and insect pests of fructus aurantii based on multitask learning, which adopts two task training identification networks of image identification and image reconstruction, and specifically comprises the following steps:
step 1, taking an bitter orange pest and disease damage data set as training data, performing feature extraction on the training data through a plurality of pre-training models, and adding a full-connection layer to perform classification training on extracted feature vectors;
the input bitter orange pest data set can be subjected to feature extraction by removing a bottleneck layer by selecting a commonly used VGG16, Inception V3 and DenseNet121 pre-training model.
Step 2, comparing and analyzing classification performance and generalization capability of all pre-training models on the bitter orange pest and disease damage data set, and selecting the pre-training model with the optimal generalization capability to construct a trunk sharing network;
the feature vector is connected with the global average pooling layer as input again, then a 20-dimensional full-connection layer is connected, a Softmax activation function is used for classification, the model is trained, the classification accuracy of three pre-training network structures on a data set is contrastively analyzed, finally, a DenseNet121 pre-training model with good generalization capability is selected as a feature extraction part of a trunk sharing network in the multi-task bitter orange pest identification model for bitter orange pests, then the global average pooling layer is connected with the feature output layer, space parameters are reduced, the model is enabled to be more robust, and finally, a 1024-dimensional full-connection layer is connected.
The DenseNet network structure model comprises 4 dense connection blocks (dense blocks), and all the dense connection blocks are connected together through a Transition Layer (Transition Layer). The dense connection block is a module comprising a plurality of layers, each layer is connected with all the layers in the previous layer in one dense connection block and serves as the input of the next layer, the feature diagram of each layer is the same in size, and a dense connection (dense connection) mode is adopted between the layers. For a network of L layers, DenseNet contains L (L +1)/2 connections in total.
The input of the model is x0Layer 1 transportOut is x1，Hl(. a) is a non-linear transformation function (non-linear transformation), which is a combinatorial operation that may include a series of Batch Normalization (Batch Normalization), activation function (ReLU), Pooling (Pooling), and convolution (Conv) operations. The relation of the l-1 layer and the l layer is shown in formula (1). Wherein, [ x ]0，x1，…，xl-1]Representing a cascade of 0 to l-1 level outputs in a dense block.
xl＝Hl([x0，x1，…，xl-1]) (1)
In addition, the dense connection between the inner layers and the layers of the dense connection blocks in the DenseNet structure is also beneficial to the propagation of information and gradient in the network, so that a deeper convolutional neural network is trained, and the problem of gradient disappearance of a deep network is solved.
The transitional connection layer is mainly used for connecting two adjacent dense connection blocks and reducing the size of a feature map, and the Transition layer comprises a convolution of 1 x 1 and an average pooling of 2 x 2. In addition, the Transition layer can play a role of a compression model, so that the network statistical efficiency is improved.
Step 3, a hard parameter sharing mechanism in multi-task learning is adopted to construct a first branch task network for bitter orange pest classification and a second branch task network for feature reconstruction, the input of a backbone sharing network is training data, and feature vectors of the training data are output to the first branch task network and the second branch task network; setting the structures and loss functions of a first branch task network and a second branch task network;
the invention designs a multitask deep network model for identifying the diseases and insect pests of fructus aurantii as shown in figure 2 by adopting a common hard parameter sharing mechanism in multitask learning. The model constructed by the method comprises three parts, namely a trunk sharing network, a first branch task network for bitter orange pest classification and a second branch task network for feature reconstruction, and the accuracy rate of bitter orange pest identification is improved by utilizing the complementary characteristics between the two tasks.
The main shared network is shared by two single task networks of bitter orange pest classification and feature reconstruction, the partial network is input as a constructed bitter orange pest image data set, and then the output result of the data set is used as the input of two subsequent tasks through the main shared network.
The bitter orange pest classification task network is connected with a 20-dimensional full connection layer, and then classification loss is calculated through output of a Softmax layer. The cross entropy is used as a loss function in the network training process, and is specifically defined as shown in formula (2).
Where N is the number of categories of the dataset, yicIs a flag variable whose value may be 0 or 1, 1 if the identified pest class is the same as the class of data set sample i, otherwise 0. p is a radical oficIs the prediction probability for a prediction sample i belonging to the true class c.
The feature reconstruction task comprises full connection layers with 256 neurons, 218 neurons, 256 neurons and 1024 neurons, and then a full inverse local average pooling layer is connected to complete feature reconstruction. The mean square error is used as a loss function in the network training process, and is defined as formula (3).
Wherein n is the number of pest data set samples.
And yiRespectively, a predicted value and an actual value of the feature.
Step 4, training and optimizing a first branch task network and a second branch task network, and storing an bitter orange pest and disease identification model formed by a trunk sharing network, the first branch task network and the second branch task network;
jointly optimizing the first branch task network and the second branch task network, updating the weight in the gradient back propagation process, and expressing the final loss of the bitter orange pest and disease identification model as follows:
Ltotal＝αLcls+βLfeaturewherein L isclsAnd LfeatureAnd the loss functions of the bitter orange pest classification task and the characteristic reconstruction task are respectively expressed, and alpha and beta are corresponding two task weights. Through multiple times of training, the classification accuracy can reach the best when alpha is 1.8 and beta is 0.2.
And 5, inputting the pest and disease damage image of the fructus aurantii into the pest and disease damage identification model of the fructus aurantii to obtain an identification result.
Preferably, the models in step 1 and step 4 are trained using the same hyper-parameters. Some of the main parameter settings are as follows: all model networks in the experiment use the Adam algorithm as an optimization algorithm, the training Batch Size (Batch Size) is set to be 32, and the initial learning rate is 0.0001. Aiming at a multitask model, because the problem of inconsistent learning among tasks is easy to exist in multitask learning, the weight of multitask loss is manually adjusted for multiple times, so that the optimal task learning effect is obtained, the initial weight values of the two tasks are set to be 1, the weight sum is 2, the damage weight of the bitter orange pest classification task is finally set to be 1.8 through comparison of multiple experimental results, the optimal result is obtained when the characteristic reconstruction task weight is 0.2, and the comparison of the models is based on the performances of the tasks on a test set.
Verification result
In order to verify the effectiveness of the method provided by the invention on identifying the diseases and insect pests of the fructus aurantii, firstly, the disease and insect pest image is preprocessed. Because the sources of the data sets of the bitter orange plant diseases and insect pests are different, image data can be influenced by different factors such as image size, illumination difference and the like, and in order to reduce the influence of the factors on work such as image feature extraction, image identification and the like in a later stage, the original images in the data sets of the bitter orange need to be preprocessed. Firstly, reading an image from a data set, decoding the image to obtain a three-dimensional matrix corresponding to the image, then adjusting the size of the image by using a bilinear interpolation method, setting the size of the image to be 224x224, and finally performing normalization processing on the image. Limited by the number of data sets, according to the training set: test set 8: the ratio of 2 randomly partitions the data set of fructus aurantii. The training set data is 1627 pictures, and the test set data is 406 pictures.
The method comprises the steps of collecting an bitter orange pest and disease damage image on the internet through a crawler script written by Python, obtaining the image from a book related to bitter orange and collecting the image in a bitter orange planting field through photographing equipment, then carrying out manual examination and screening on the collected data, cleaning the image of non-bitter orange pest and disease damage and the image with poor quality, finally classifying the screened bitter orange pest and disease damage image, and finally constructing a bitter orange pest and disease damage data set (Citrus 20 for short) containing 2033 pictures of 2 major and 20 minor classes. Detailed information of the collected bitter orange pest data set is shown in table 1.
Table 1 data set detail information table
| Name of disease and pest of fructus Aurantii | Categories | Number/sheets of pictures |
| Scale insect | 0 | 54 |
| Ulcer disease | 1 | 394 |
| Leaf miner | 2 | 182 |
| Anthracnose | 3 | 244 |
| Sooty mould | 4 | 118 |
| Scab of sore | 5 | 187 |
| Sand skin disease | 6 | 184 |
| Red spider | 7 | 37 |
| Ericerus pela | 8 | 28 |
| Lack of potassium | 9 | 21 |
| Iron deficiency | 10 | 24 |
| Lack of zinc | 11 | 29 |
| Lack of magnesium | 12 | 19 |
| Foot rot | 13 | 52 |
| Aphids | 14 | 54 |
| Brown longhorn beetle | 15 | 33 |
| Yellow spiny moth | 16 | 28 |
| Yellow spot disease | 17 | 89 |
| Yellow dragon disease | 18 | 238 |
| Bemisia spinosa (L.) Kuntze | 19 | 18 |
(1) Comparison experiment of generalization ability of pre-training model
In order to compare the generalization ability of VGG16, Inception V3 and DenseNet121 pre-trained models on the constructed Citrus20 dataset, a basic model generalization ability comparison experiment was performed on the Citrus20 dataset, and the experimental results are shown in Table 2. As can be seen from table 2, on the Citrus20 dataset, the DenseNet121 pre-training model has better generalization capability than the VGG16 and inclusion v3 pre-training models for test set classification accuracy. Therefore, the DenseNet121 pre-training model is used as a backbone shared reference network of the bitter orange pest identification model.
TABLE 2 evaluation results of Citrus20 basic pre-training model experiment
| Model (model) | Test set classification accuracy | Run time (s/Batch Size) |
| VGG16 | 0.7414 | 624 |
| InceptionV3 | 0.7488 | 193 |
| DenseNet121 | 0.8251 | 383 |
(2) Comparison experiment of single task model and multi-task model
In order to verify whether the performance of a multitask learning model is improved compared with a single task or not, on a Citrus20 data set, a DenseNet121 pre-training model is used as a reference to construct a fructus aurantii pest classification single-task network model ST-DenseNet121, the network structure and the super-parameter setting are the same as the fructus aurantii pest classification tasks in the multitask recognition model provided by the invention, then pre-trained VGG16 and Incepiton V3-based fructus aurantii pest image fine-tuning classification models ST-VGG16 and ST-Incepiton V3 and other paper models are constructed and compared with the multitask recognition model MT-CNN provided by the invention, and the experimental result is shown in Table 3.
TABLE 3 evaluation results of the single-task training model and the multi-task training model
| Model (model) | Test set classification accuracy | Run time (s/Batch Size) |
| ST-VGG16 | 0.7562 | 428 |
| ST-InceptionV3 | 0.7783 | 190 |
| Simplified DenseNet | 0.8414 | 165 |
| BridgeNet-19 | 0.8462 | 310 |
| ST-DenseNet121 | 0.8399 | 380 |
| MT-CNN | 0.8522 | 383 |
As can be seen from the experimental results in table 3, on the Citrus20 dataset, compared with the training result of a single task, the multi-task recognition model MT-CNN provided by the present invention can effectively improve the classification accuracy, which further illustrates that multiple tasks can increase the constraint on the bottom layer feature extraction by sharing one feature extraction network, and the generalization performance of the model can be better improved by balancing and restricting different tasks, and the multi-task network model is more time-consuming than the single task, but shows obvious performance advantages.
In conclusion, the method for identifying the diseases and insect pests of the bitter orange based on the multitask learning combines the collaborative optimization mode of the multitask learning, combines the image reconstruction task and the tasks for identifying the diseases and insect pests of the bitter orange, and performs model training together, so that the constraint on the extraction of the bottom layer features can be increased, and the generalization performance of the model can be better improved through mutual balance and restriction among different tasks.
In light of the foregoing description of the preferred embodiments of the present invention, those skilled in the art can now make various alterations and modifications without departing from the scope of the invention. The technical scope of the present invention is not limited to the contents of the specification, and must be determined according to the scope of the claims.
Claims (8)
1. A method for identifying diseases and insect pests of fructus aurantii based on multitask learning is characterized by comprising the following steps:
step 1, taking an bitter orange pest and disease damage data set as training data, performing feature extraction on the training data through a plurality of pre-training models, and adding a full-connection layer to perform classification training on extracted feature vectors;
step 2, comparing and analyzing classification performance and generalization capability of all pre-training models on the bitter orange pest and disease damage data set, and selecting the pre-training model with the optimal generalization capability to construct a trunk sharing network;
step 3, a hard parameter sharing mechanism in multi-task learning is adopted to construct a first branch task network for bitter orange pest classification and a second branch task network for feature reconstruction, the input of a backbone sharing network is training data, and feature vectors of the training data are output to the first branch task network and the second branch task network; setting the structures and loss functions of a first branch task network and a second branch task network;
step 4, training and optimizing a first branch task network and a second branch task network, and storing an bitter orange pest and disease identification model formed by a trunk sharing network, the first branch task network and the second branch task network;
and 5, inputting the pest and disease damage image of the fructus aurantii into the pest and disease damage identification model of the fructus aurantii to obtain an identification result.
2. The method for identifying fructus aurantii pests and diseases based on multitask learning according to claim 1, wherein the pre-trained model in the step 2 comprises 4 dense connecting blocks, and each dense connecting block is connected with each other through a transition layer.
3. The method for identifying bitter orange pests and diseases based on multitask learning according to claim 2, wherein in the step 2, a global average pooling layer and a full connection layer are connected to a trunk sharing network at a feature output layer of a pre-training model.
4. The method for identifying bitter orange pests and diseases based on multitask learning according to claim 1, wherein the step 3 of setting the structure and the loss function of the first branch task network specifically refers to the following steps:
and setting the network connection of the first branch task to be a 20-dimensional full connection layer, and calculating classification loss through the output of the Softmax layer, wherein cross entropy is used as a loss function in the network training process.
5. The method for identifying bitter orange pests and diseases based on multitask learning according to claim 4, wherein the step 3 of setting the structure and the loss function of the second branch task network specifically refers to the following steps:
setting a second branch task network to comprise four full-connection layers, and then connecting a full-inverse local average pooling layer to complete feature reconstruction, wherein the mean square error is used as a loss function in the network training process.
6. The method for identifying fructus aurantii pests and diseases based on multitask learning according to claim 5, wherein the first branch task network and the second branch task network are jointly optimized in the step 4, the weight is updated in the gradient back propagation process, and the final loss of the fructus aurantii pest and disease identification model is expressed as follows:
Ltotal＝αLcls+βLfeaturewherein L isclsAnd LfeatureAnd the loss functions of the bitter orange pest classification task and the characteristic reconstruction task are respectively expressed, and alpha and beta are corresponding two task weights.
7. The method for identifying bitter orange pest and disease damage based on multitask learning as claimed in claim 6, wherein the preferred value of alpha is 1.8, and the preferred value of beta is 0.2.
8. The method for identifying bitter orange pests and diseases based on multitask learning as claimed in claim 1, wherein the models in the step 1 and the step 4 are trained by using the same hyper-parameters.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111514958.9A CN114266965A (en) | 2021-12-10 | 2021-12-10 | Bitter orange pest and disease identification method based on multitask learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111514958.9A CN114266965A (en) | 2021-12-10 | 2021-12-10 | Bitter orange pest and disease identification method based on multitask learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114266965A true CN114266965A (en) | 2022-04-01 |
Family
ID=80827137
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111514958.9A Pending CN114266965A (en) | 2021-12-10 | 2021-12-10 | Bitter orange pest and disease identification method based on multitask learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114266965A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114819091A (en) * | 2022-05-07 | 2022-07-29 | 杭州电子科技大学 | Multi-task network model training method and system based on self-adaptive task weight |
| CN117058536A (en) * | 2023-07-19 | 2023-11-14 | 中公高科养护科技股份有限公司 | Pavement disease identification method, device and medium based on double-branch network |
-
2021
- 2021-12-10 CN CN202111514958.9A patent/CN114266965A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114819091A (en) * | 2022-05-07 | 2022-07-29 | 杭州电子科技大学 | Multi-task network model training method and system based on self-adaptive task weight |
| CN114819091B (en) * | 2022-05-07 | 2024-04-16 | 杭州电子科技大学 | Multi-task network model training method and system based on self-adaptive task weight |
| CN117058536A (en) * | 2023-07-19 | 2023-11-14 | 中公高科养护科技股份有限公司 | Pavement disease identification method, device and medium based on double-branch network |
| CN117058536B (en) * | 2023-07-19 | 2024-04-30 | 中公高科养护科技股份有限公司 | Pavement disease identification method, device and medium based on double-branch network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2022160771A1 (en) | Method for classifying hyperspectral images on basis of adaptive multi-scale feature extraction model | |
| CN110309841A (en) | A kind of hickory nut common insect pests recognition methods based on deep learning | |
| CN111696101A (en) | Light-weight solanaceae disease identification method based on SE-Inception | |
| CN114266965A (en) | Bitter orange pest and disease identification method based on multitask learning | |
| CN107122809A (en) | Neural network characteristics learning method based on image own coding | |
| CN106919951A (en) | A kind of Weakly supervised bilinearity deep learning method merged with vision based on click | |
| Mijwil et al. | Utilizing the genetic algorithm to pruning the C4. 5 decision tree algorithm | |
| CN111127423B (en) | Rice pest and disease identification method based on CNN-BP neural network algorithm | |
| Baser et al. | Tomconv: An improved cnn model for diagnosis of diseases in tomato plant leaves | |
| CN111524140B (en) | Medical image semantic segmentation method based on CNN and random forest method | |
| Maria et al. | Cauliflower disease recognition using machine learning and transfer learning | |
| Yu et al. | A recognition method of soybean leaf diseases based on an improved deep learning model | |
| CN113344077A (en) | Anti-noise solanaceae disease identification method based on convolution capsule network structure | |
| CN113221913A (en) | Agriculture and forestry disease and pest fine-grained identification method and device based on Gaussian probability decision-level fusion | |
| Xue et al. | Grafting transformer on automatically designed convolutional neural network for hyperspectral image classification | |
| CN115474939A (en) | Autism spectrum disorder recognition model based on deep expansion neural network | |
| Tanwar et al. | Red rot disease prediction in sugarcane using the deep learning approach | |
| Prasher et al. | Potato leaf disease prediction using RMSProp, Adam and SGD optimizers | |
| Zhang et al. | Appearance quality classification method of Huangguan pear under complex background based on instance segmentation and semantic segmentation | |
| Borman et al. | Classification of Medicinal Wild Plants Using Radial Basis Function Neural Network with Least Mean Square | |
| Wang et al. | Research on identification of corn disease occurrence degree based on improved resnext network | |
| Pareek et al. | Clustering based segmentation with 1D-CNN model for grape fruit disease detection | |
| Li et al. | Fine-grained classification of grape leaves via a pyramid residual convolution neural network | |
| CN110210562A (en) | Image classification method based on depth network and sparse Fisher vector | |
| Shakil et al. | Addressing agricultural challenges: An identification of best feature selection technique for dragon fruit disease recognition |
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 |