CN115019175B - Pest identification method based on migration element learning - Google Patents
Pest identification method based on migration element learning Download PDFInfo
- Publication number
- CN115019175B CN115019175B CN202210738451.XA CN202210738451A CN115019175B CN 115019175 B CN115019175 B CN 115019175B CN 202210738451 A CN202210738451 A CN 202210738451A CN 115019175 B CN115019175 B CN 115019175B
- Authority
- CN
- China
- Prior art keywords
- learning
- pest
- sample
- feature extraction
- category
- 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
- 241000607479 Yersinia pestis Species 0.000 title claims abstract description 104
- 238000000034 method Methods 0.000 title claims abstract description 56
- 238000013508 migration Methods 0.000 title claims abstract description 24
- 230000005012 migration Effects 0.000 title claims abstract description 24
- 238000000605 extraction Methods 0.000 claims abstract description 64
- 239000013598 vector Substances 0.000 claims abstract description 30
- 238000005070 sampling Methods 0.000 claims abstract description 5
- 230000006870 function Effects 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 10
- 238000003780 insertion Methods 0.000 claims description 3
- 230000037431 insertion Effects 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 238000012935 Averaging Methods 0.000 claims description 2
- 230000007704 transition Effects 0.000 claims 2
- 102100023593 Fibroblast growth factor receptor 1 Human genes 0.000 description 7
- 101000827746 Homo sapiens Fibroblast growth factor receptor 1 Proteins 0.000 description 7
- 238000012549 training Methods 0.000 description 7
- 238000012546 transfer Methods 0.000 description 6
- 238000002474 experimental method Methods 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 238000001514 detection method Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 241000238631 Hexapoda Species 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000013526 transfer learning Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/188—Vegetation
-
- 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
-
- 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
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/761—Proximity, similarity or dissimilarity measures
-
- 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
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Medical Informatics (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a pest identification method based on migration element learning, which comprises the following steps: s1: constructing a pest data set, and randomly sampling to respectively form a support set and a query set; an improved feature extraction network is obtained; s2: updating parameters of the improved feature extraction network; s3: obtaining feature vectors of samples in a support set and a query set; s4: calculating to obtain a prototype of each category in the support set; s5: and calculating the class probability of each sample in the query set to be used as a pest identification result. The invention provides a pest identification method based on migration element learning, which solves the problem that the existing pest identification method has lower accuracy in distinguishing new type pests from old type pests.
Description
Technical Field
The invention relates to the technical field of image recognition, in particular to a pest recognition method based on migration element learning.
Background
With the development of deep learning technology, in intelligent control of agricultural diseases and insect pests, it is becoming increasingly popular to adopt a deep learning model to identify and classify the insect pests. Most pest identification methods based on small sample learning at present need to train a deep model on a common class (basic class) with enough samples, and then transfer the model to a new class with only a few examples for small sample learning. The model network has a simple structure, and is often insufficient for effectively expressing information contained in pest images, so that the accuracy is low when new and old types of pests are distinguished.
In the prior art, as disclosed in 2021-03-09, the publication number is CN112464971A, and the convolutional attention model is tested by using a test set to construct a pest detection model, so that the pest detection model has high detection precision and strong robustness, but when a new type of pests and an old type of pests are mixed, the type of the pests cannot be accurately identified as known or unknown.
Disclosure of Invention
The invention provides a pest identification method based on migration element learning, which aims to overcome the technical defect that the existing pest identification method has lower accuracy in distinguishing new type pests from old type pests.
In order to solve the technical problems, the technical scheme of the invention is as follows:
a pest identification method based on migration element learning comprises the following steps:
S1: acquiring pest image samples to construct a pest data set, and randomly sampling the pest data set to respectively form a support set and a query set;
The method comprises the steps of combining an existing feature extraction network with a feature extraction network with an improved spatial attention module;
s2: updating parameters of the improved feature extraction network by a meta learning method based on migration;
S3: respectively inputting the support set and the query set into the updated feature extraction network, and correspondingly obtaining feature vectors of each sample in the support set and the query set;
s4: calculating according to the feature vectors of the samples in the support set to obtain a prototype of each category in the support set;
S5: and calculating the class probability of each sample in the query set according to the prototype of each class in the support set and the feature vector of each sample in the query set to obtain the pest identification result.
In the scheme, the characteristic extraction capability of the existing characteristic extraction network is improved by introducing the spatial attention module, and the pest image recognition performance is improved; meanwhile, the parameter of the improved feature extraction network is updated based on the meta-learning method of migration, so that the feature extraction capability of the feature extraction network can be maintained, the feature extraction capability of the feature extraction network also has the capability of distinguishing new category features after migration, and the accuracy of distinguishing new category pests from old category pests is improved.
Preferably, in step S1, the pest data set includes M types of pest images, N types of pest images are randomly extracted from the pest data set, N is smaller than M, K pest images are randomly extracted from each extracted type to form a supporting set, and q non-repeated pest images are randomly extracted from the extracted types to form a query set.
Preferably, the feature extraction network is improved by combining ResNeSt a 101 with a spatial attention module, specifically: the insertion of spatial attention modules at the SAM locations before 3 x 3 global maximization of ResNeSt a 101 and/or at the last SAM location of each neck of each layer of ResNeSt a results in an improved feature extraction network.
Preferably, the spatial attention module is a NAM spatial attention module or CBAM spatial attention module.
Preferably, the method comprises the steps of,
The spatial attention module inserted at the SAM location prior to 3 x 3 global maximum pooling of ResNeSt101 is of the same or different type as the spatial attention module inserted at the last SAM location of each neck of each layer of ResNeSt 101;
the spatial attention module inserted at the last SAM location of each neck of each layer ResNeSt is of the same type.
In the above scheme, the improved feature extraction network SA-ResNeSt101 reserves the network depth consistent with other residual networks, can fully express the information contained in the pest images, and uses a plurality of groups of modules such as 3×3 convolution, channel attention, space attention and the like in the network, so that not only the information of each channel in the images, but also the information of each pixel in each channel can be focused, the focus area of the pest images can be focused, and the expression capability of the network can be improved on the premise of not greatly increasing the calculated amount, thereby improving the expression capability of the feature extraction network.
Preferably, step S2 specifically includes:
A1: setting a learning batch and a learning rate;
A2: randomly extracting N 1 categories from the pest data set, randomly extracting K 1 pest images from each extracted category to form a learning support set, and randomly extracting q 1 non-repeated pest images from the extracted categories to form a learning query set;
a3: inputting the learning support set and the learning query set into an improved feature extraction network, and correspondingly obtaining feature vectors of each sample in the learning support set and the learning query set;
A4: calculating according to the feature vectors of the samples in the learning support set to obtain a prototype of each category in the learning support set;
a5: calculating according to the prototype of each category in the learning support set and the feature vector of each sample in the learning query set to obtain the category probability of each sample in the learning query set;
a6: calculating a loss value between the class probability of each sample and the corresponding class label;
A7: it is determined whether the loss value is converged or not,
If yes, completing the small sample learning;
if not, updating the parameters of the improved feature extraction network, and returning to the step A2.
In the scheme, unlike the main stream meta-learning and transfer learning method, the transfer-based meta-learning (small sample learning) method not only can maintain the feature extraction capability of SA-ResNeSt101, but also has the distinguishing capability of new category features after transfer after the meta-learning process of a small number of samples, so that the method is more effective than the main stream small sample learning method in pest identification.
Preferably, a cross entropy loss function is used to calculate the loss value between the class probability of each sample and its corresponding class label.
Preferably, before inputting the pest image sample into the feature extraction network, further comprising performing data enhancement processing on the pest image sample; the data enhancement process includes: random aspect ratio clipping, random scaling, random horizontal flipping, random vertical flipping, random rotation, color dithering, gaussian blurring.
Preferably, the prototype of class c is calculated by the following formula:
Wherein, An i-th sample representing a support set category c, c=1, 2,..Representing the eigenvector of sample x i.
Preferably, the probability of category c is calculated by the following formula:
Where x j represents the j-th sample in the query set, y j represents the class label for sample x j, j=1, 2,..q, The eigenvectors representing samples x j, τ representing the hyper-parameters, sim representing the function that calculates the similarity between the two vectors, E n representing the prototype of support-set class N, n=1, 2.
Compared with the prior art, the technical scheme of the invention has the beneficial effects that:
The invention provides a pest identification method based on migration element learning, which improves the feature extraction capacity of the existing feature extraction network by introducing a spatial attention module, and improves the pest image identification performance; meanwhile, the parameter of the improved feature extraction network is updated based on the meta-learning method of migration, so that the feature extraction capability of the feature extraction network can be maintained, the feature extraction capability of the feature extraction network also has the capability of distinguishing new category features after migration, and the accuracy of distinguishing new category pests from old category pests is improved.
Drawings
FIG. 1 is a flow chart of the steps performed in the technical scheme of the invention;
FIG. 2 is a pre-training flow chart of the improved feature extraction network of the present invention;
FIG. 3 is a schematic diagram of an improved feature extraction network in accordance with the present invention;
FIG. 4 is a schematic diagram of the NAM space attention module according to the present invention;
FIG. 5 is a schematic diagram of a CBAM spatial attention module according to the present invention;
FIG. 6 is a flow chart of meta learning based on migration in the present invention.
Detailed Description
The drawings are for illustrative purposes only and are not to be construed as limiting the present patent;
For the purpose of better illustrating the embodiments, certain elements of the drawings may be omitted, enlarged or reduced and do not represent the actual product dimensions;
It will be appreciated by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted.
The technical scheme of the invention is further described below with reference to the accompanying drawings and examples.
Example 1
As shown in fig. 1, a pest identification method based on transfer element learning includes the following steps:
S1: acquiring pest image samples to construct a pest data set, and randomly sampling the pest data set to respectively form a support set and a query set;
The method comprises the steps of combining an existing feature extraction network with a feature extraction network with an improved spatial attention module;
s2: updating parameters of the improved feature extraction network by a meta learning method based on migration;
S3: respectively inputting the support set and the query set into the updated feature extraction network, and correspondingly obtaining feature vectors of each sample in the support set and the query set;
s4: calculating according to the feature vectors of the samples in the support set to obtain a prototype of each category in the support set;
S5: and calculating the class probability of each sample in the query set according to the prototype of each class in the support set and the feature vector of each sample in the query set to obtain the pest identification result.
In the specific implementation process, the feature extraction capability of the existing feature extraction network is improved by introducing a spatial attention module, and the pest image recognition performance is improved; meanwhile, the parameter of the improved feature extraction network is updated based on the meta-learning method of migration, so that the feature extraction capability of the feature extraction network can be maintained, the feature extraction capability of the feature extraction network also has the capability of distinguishing new category features after migration, and the accuracy of distinguishing new category pests from old category pests is improved.
Example 2
A pest identification method based on migration element learning comprises the following steps:
S1: acquiring pest image samples to construct a pest data set, and randomly sampling the pest data set to respectively form a support set and a query set;
More specifically, the pest data set includes M classes of pest images, N classes are randomly extracted from the pest data set, N is less than M, K pest images are randomly extracted from each extracted class to form a supporting set, and q non-repeated pest images are randomly extracted from the extracted classes to form a query set.
Feature extraction network with improved combination of existing feature extraction network with spatial attention module
As shown in fig. 2, the improved feature extraction network is pre-trained: and adding a full connection layer at the output end of the SA-ResNeSt to serve as a classifier, adopting cross entropy as a loss function, and setting super parameters such as batch, learning rate and the like. Firstly taking a pest image sample for training as input to obtain corresponding prediction as output, then calculating a loss value between the prediction and a label of the sample, repeating the process until the loss value is not reduced, and removing a classifier after training is finished to obtain SA-ResNeSt with initialization parameters.
More specifically, as shown in fig. 3, the feature extraction network improved by incorporating ResNeSt a 101 with a spatial attention module is specifically: the insertion of spatial attention modules at the SAM locations before 3 x 3 global maximization of ResNeSt a 101 and/or at the last SAM location of each neck of each layer of ResNeSt a results in an improved feature extraction network.
More specifically, as shown in fig. 4-5, the spatial attention module is a NAM spatial attention module or CBAM spatial attention module.
More specifically, the method comprises the steps of,
The spatial attention module inserted at the SAM location prior to 3 x 3 global maximum pooling of ResNeSt101 is of the same or different type as the spatial attention module inserted at the last SAM location of each neck of each layer of ResNeSt 101;
the spatial attention module inserted at the last SAM location of each neck of each layer ResNeSt is of the same type.
In the specific implementation process, the improved feature extraction network SA-ResNeSt101 keeps the network depth consistent with other residual networks, can fully express information contained in pest images, and uses a plurality of groups of modules such as 3X 3 convolution, channel attention, space attention and the like in the network, so that not only is the information of each channel in the image focused, but also the information of each pixel in each channel can be focused, the focus area of the pest images can be focused, and the expression capability of the network can be improved on the premise of not greatly increasing the calculated amount, thereby improving the expression capability of the feature extraction network.
S2: updating parameters of the improved feature extraction network by a meta learning method based on migration;
More specifically, as shown in fig. 6, step S2 specifically includes:
A1: setting a learning batch and a learning rate;
A2: randomly extracting N 1 categories from the pest data set, randomly extracting K 1 pest images from each extracted category to form a learning support set, and randomly extracting q 1 non-repeated pest images from the extracted categories to form a learning query set;
a3: inputting the learning support set and the learning query set into an improved feature extraction network, and correspondingly obtaining feature vectors of each sample in the learning support set and the learning query set;
A4: calculating according to the feature vectors of the samples in the learning support set to obtain a prototype of each category in the learning support set;
more specifically, a prototype of category c is calculated by the following formula:
Wherein, An i-th sample representing a learning support set category c, c=1, 2, N, y i represents the class label of sample x i, avg ()'s function represents averaging multiple vectors,/>Representing the eigenvector of sample x i.
A5: calculating according to the prototype of each category in the learning support set and the feature vector of each sample in the learning query set to obtain the category probability of each sample in the learning query set;
More specifically, the probability of category c is calculated by the following formula:
Where x j represents the j-th sample in the learning query set, y j represents the class label for sample x j, j=1, 2,..q, Representative of the feature vector of sample x j, τ representative of the hyper-parameters, sim representative of the function that calculates the similarity between the two vectors, E n representative of the prototype of learning support set class N, n=1, 2.
A6: calculating a loss value between the class probability of each sample and the corresponding class label;
More specifically, a cross entropy loss function is used to calculate the loss value between the class probability of each sample and its corresponding class label.
In a specific implementation, the cross entropy loss function is:
Where p (y i＝c|xi) represents the class probability for sample x i, y i' represents the class label for sample x i, n represents the number of samples of the learning query set, and log (-) is a logarithmic function.
A7: it is determined whether the loss value is converged or not,
If yes, completing the small sample learning;
if not, updating the parameters of the improved feature extraction network, and returning to the step A2.
In the specific implementation process, unlike the main stream meta-learning and transfer learning method, the transfer-based meta-learning (small sample learning) method not only can maintain the feature extraction capability of SA-ResNeSt101, but also has the distinguishing capability of new category features after transfer after the meta-learning process of a small number of samples, so that the method is more effective than the main stream small sample learning method in pest identification.
S3: respectively inputting the support set and the query set into the updated feature extraction network, and correspondingly obtaining feature vectors of each sample in the support set and the query set;
s4: calculating according to the feature vectors of the samples in the support set to obtain a prototype of each category in the support set;
more specifically, a prototype of category c is calculated by the following formula:
Wherein, An i-th sample representing a support set category c, c=1, 2,..Representing the eigenvector of sample x i.
S5: and calculating the class probability of each sample in the query set according to the prototype of each class in the support set and the feature vector of each sample in the query set to obtain the pest identification result.
More specifically, the probability of category c is calculated by the following formula:
Where x j represents the j-th sample in the query set, y j represents the class label for sample x j, j=1, 2,..q, The eigenvectors representing samples x j, τ representing the hyper-parameters, sim representing the function that calculates the similarity between the two vectors, E n representing the prototype of support-set class N, n=1, 2.
More specifically, before inputting the pest image sample into the feature extraction network, further comprising performing data enhancement processing on the pest image sample; the data enhancement process includes: random aspect ratio clipping, random scaling, random horizontal flipping, random vertical flipping, random rotation, color dithering, gaussian blurring.
Example 3
In this example, two relatively complete pest data sets, aug-D0 and semi-IP102, were constructed, and the two pest data sets were divided into a training data set, a test data set, and a verification data set, with specific statistical data shown in table 1.
TABLE 1 statistics of aug-D0 and semi-IP102 pest data sets
The aug-D0 and semi-IP102 are adopted to carry out experiments on the pest identification method (the method) based on the migration element learning, and the experimental results are as follows:
The comparison of the results of the small sample learning experiments on aug-IP102 with the conventional model is shown in Table 2;
TABLE 2
During training, the number of types extracted in each round is set to be 5, the number of samples of the support set extracted in each round is set to be 5, and the number of samples of the query set is set to be 15; in the test, the evaluation index is that the accuracy (accuracy) (%) of the samples of each type are set to be 5, the number of samples of each type of the support set is set to be 1, 5 and 10, the number of samples of the query set is set to be 15, and the confidence interval is 95%;
the calculation formula of the average recognition accuracy is as follows:
Where Num correct represents the predicted correct number and Num represents the total number of samples in the query set.
The comparison of the small sample learning experiment results with the traditional model on the semi-IP102 data set is shown in Table 3;
TABLE 3 Table 3
| Model name | 1-shot | 5-shot | 10-shot |
| Matching network | 42.66％±0.42％ | 55.06％±0.39％ | 60.61％±0.38％ |
| Prototype network | 38.23％±0.36％ | 54.96％±0.36％ | 59.96％±0.34％ |
| Relationship network | 37.74％±0.40％ | 51.35％±0.41％ | 55.61％±0.39％ |
| Primitive base line network | 34.07％±0.35％ | 46.90％±0.55％ | 51.83％±0.38％ |
| SA-ResNeSt101 | 57.62％±0.53％ | 72.90％±0.30％ | 76.80％±0.30％ |
During training, the number of types extracted in each round is set to be 5, the number of samples of the support set extracted in each round is set to be 5, and the number of samples of the query set is set to be 15; during testing, the evaluation index is that the accuracy (accuracy) (%) of each extracted category is set to be 5, the number of samples of each extracted support set is set to be 1, 5 and 10, the number of images of the query set is set to be 15, and the confidence interval is 95%.
As can be seen from tables 2 and 3, the effect of SA-ResNeSt101 in the present embodiment achieves higher recognition accuracy than the prior art on the dataset composed of the semi-IP102 and aug-D0 new categories.
A spatial attention module is added to the SAM location of ResNeSt a 101, wherein ResNeSt a in the attention type represents a network of ResNeSt a without any spatial attention module added, C-SAM represents a spatial attention module of CBAM and N-SAM represents a spatial attention module of NAM. Of the spatial attention module positions, B represents the addition of Spatial Attention Module (SAM) at the 3 convolutions before ResNeSt main way, E represents the addition of Spatial Attention Module (SAM) at the last position of each neck of ResNeSt, and the training and testing patterns are consistent with tables 2 and 3.
The effectiveness of the attention mechanism on aug-D0 is shown in Table 4;
TABLE 4 Table 4
| Attention type | Spatial attention module location | 1-shot | 5-shot | 10-shot |
| ResNeSt101 | —— | 78.83％±0.51％ | 91.53％±0.26％ | 93.48％±0.21％ |
| N-SAM | B | 80.45％±0.32％ | 92.44％±0.17％ | 94.16％±0.15％ |
| N-SAM | E | 79.18％±0.43％ | 91.89％±0.26％ | 93.80％±0.20％ |
| C-SAM | B | 84.18％±0.43％ | 95.00％±0.14％ | 96.37％±0.10％ |
| C-SAM | E | 77.56％±0.50％ | 90.25％±0.27％ | 92.31％±0.22％ |
| N-SAM,N-SAM | B,E | 78.40％±0.43％ | 91.78％±0.28％ | 93.88％±0.21％ |
| N-SAM,C-SAM | B,E | 78.19％±0.46％ | 92.13％±0.29％ | 94.11％±0.22％ |
| C-SAM,N-SAM | B,E | 79.75％±0.46％ | 92.44％±0.26％ | 94.30％±0.22％ |
| C-SAM,C-SAM | B,E | 78.09％±0.46％ | 91.11％±0.28％ | 93.10％±0.24％ |
The effectiveness of the attention mechanism on semi-IP102 is shown in Table 5;
TABLE 5
As can be seen from tables 4 and 5, the improved feature extraction network SA-ResNeSt101 enhances the expression capability of the pest image features, and effectively improves the recognition accuracy of the pest images.
The results of the open world study experiments on aug-D0 are compared to those shown in Table 6;
TABLE 6
| Model name | 1-shot | 5-shot | 10-shot |
| Matching network | 55.49％±0.18％ | 70.56％±0.09％ | 74％±0.08％ |
| Prototype network | 47.77％±0.18％ | 67.39％±0.10％ | 70.17％±0.08％ |
| Relationship network | 43.13％±0.18％ | 62.16±0.10％ | 64.01％±0.08％ |
| Primitive base line network | 12.33％±0.05％ | 34.18％±0.09％ | 37.39％±0.07％ |
| SA-ResNeSt101 | 63.44％±0.13％ | 79.37％±0.07％ | 81.86％±0.07％ |
The results of the open world study experiments on semi-IP102 are compared to the results shown in Table 7;
TABLE 7
| Model name | 1-shot | 5-shot | 10-shot |
| Matching network | 34.36％±0.13％ | 46.61％±0.09％ | 51.69％±0.08％ |
| Prototype network | 36.82％±0.15％ | 53.19％±0.09％ | 55.6％±0.08％ |
| Relationship network | 36.46％±0.15％ | 51.11％±0.08％ | 52.7％±0.07％ |
| Primitive base line network | 17.92％±0.08％ | 30.23％±0.08％ | 33.74％±0.07％ |
| SA-ResNeSt101 | 44.87％±0.18％ | 60.01％±0.08％ | 63.58％±0.07％ |
As can be seen from tables 6 and 7, the present method can more accurately distinguish new and old categories as compared to the prior art on an open world dataset consisting of aug-D0 and semi-IP102 new and old categories.
The comparison of the migration-based meta-learning method described in this embodiment with the experimental results of the existing meta-learning on the dataset composed of the aug-D0 and semi-IP102 new classes is shown in table 8;
TABLE 8
The experimental results of the migration-based meta learning method described in this embodiment and the existing meta learning on the open world dataset composed of the new class and the old class of aug-D0 and semi-IP102 are shown in table 9;
TABLE 9
As can be seen from tables 8 and 9, the method further improves pest identification accuracy based on meta-learning compared with the existing meta-learning process on the new data set and the open world data set consisting of semi-IP102 and aug-D0.
It is to be understood that the above examples of the present invention are provided by way of illustration only and not by way of limitation of the embodiments of the present invention. Other variations or modifications of the above teachings will be apparent to those of ordinary skill in the art. It is not necessary here nor is it exhaustive of all embodiments. Any modification, equivalent replacement, improvement, etc. which come within the spirit and principles of the invention are desired to be protected by the following claims.
Claims (7)
1. The pest identification method based on migration element learning is characterized by comprising the following steps of:
S1: acquiring pest image samples to construct a pest data set, and randomly sampling the pest data set to respectively form a support set and a query set;
The method comprises the steps of combining an existing feature extraction network with a feature extraction network with an improved spatial attention module;
s2: updating parameters of the improved feature extraction network by a meta learning method based on migration;
S3: respectively inputting the support set and the query set into the updated feature extraction network, and correspondingly obtaining feature vectors of each sample in the support set and the query set;
s4: calculating according to the feature vectors of the samples in the support set to obtain a prototype of each category in the support set;
S5: calculating according to the prototype of each category in the support set and the feature vector of each sample in the query set to obtain the category probability of each sample in the query set as a pest identification result;
The step S2 specifically comprises the following steps:
A1: setting a learning batch and a learning rate;
A2: randomly extracting N 1 categories from the pest data set, randomly extracting K 1 pest images from each extracted category to form a learning support set, and randomly extracting q 1 non-repeated pest images from the extracted categories to form a learning query set;
a3: inputting the learning support set and the learning query set into an improved feature extraction network, and correspondingly obtaining feature vectors of each sample in the learning support set and the learning query set;
A4: calculating according to the feature vectors of the samples in the learning support set to obtain a prototype of each category in the learning support set;
a5: calculating according to the prototype of each category in the learning support set and the feature vector of each sample in the learning query set to obtain the category probability of each sample in the learning query set;
a6: calculating a loss value between the class probability of each sample and the corresponding class label;
A7: it is determined whether the loss value is converged or not,
If yes, completing the small sample learning;
if not, updating parameters of the improved feature extraction network, and returning to the step A2;
The prototype of category c is calculated by the following formula:
Wherein, The ith sample representing class c in the support set, c=1, 2, …, N, y i represents the class label for sample x i, avg (-) function represents averaging multiple vectors,/>A feature vector representing sample x i;
The probability of category c is calculated by the following formula:
Where x j represents the j-th sample in the query set, y j represents the class label for sample x j, j=1, 2, …, q, Representing the eigenvectors of sample x j, τ representing the hyper-parameters, sim representing the function of computing the similarity between the two vectors, E n representing the prototype of class N in the support set, n=1, 2, …, N.
2. The method for identifying pests based on the shift element learning according to claim 1, wherein in the step S1, the pest data set includes M types of pest images, N types of pest data sets are randomly extracted, N < M, K pest image composition support sets are randomly extracted from each extracted type, and q non-repeated pest image composition query sets are randomly extracted from the extracted types.
3. The pest identification method based on the transition element learning according to claim 1, wherein the feature extraction network improved by combining ResNeSt with the spatial attention module is specifically: the insertion of spatial attention modules at the SAM locations before 3 x 3 global maximization of ResNeSt a 101 and/or at the last SAM location of each neck of each layer of ResNeSt a results in an improved feature extraction network.
4. A method of pest identification based on mobile element learning according to claim 3, wherein the spatial attention module is a NAM spatial attention module or CBAM spatial attention module.
5. The pest identification method based on the transition element learning according to claim 4, wherein,
The spatial attention module inserted at the SAM location prior to 3 x 3 global maximum pooling of ResNeSt101 is of the same or different type as the spatial attention module inserted at the last SAM location of each neck of each layer of ResNeSt 101;
the spatial attention module inserted at the last SAM location of each neck of each layer ResNeSt is of the same type.
6. The method of claim 5, wherein the cross entropy loss function is used to calculate a loss value between the class probability of each sample and its corresponding class label.
7. The pest identification method based on the shift element learning according to claim 1, further comprising performing data enhancement processing on the pest image sample before inputting the pest image sample into the feature extraction network; the data enhancement process includes: random aspect ratio clipping, random scaling, random horizontal flipping, random vertical flipping, random rotation, color dithering, gaussian blurring.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210738451.XA CN115019175B (en) | 2022-06-27 | 2022-06-27 | Pest identification method based on migration element learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210738451.XA CN115019175B (en) | 2022-06-27 | 2022-06-27 | Pest identification method based on migration element learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115019175A CN115019175A (en) | 2022-09-06 |
| CN115019175B true CN115019175B (en) | 2024-06-18 |
Family
ID=83076870
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210738451.XA Active CN115019175B (en) | 2022-06-27 | 2022-06-27 | Pest identification method based on migration element learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115019175B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117709394A (en) * | 2024-02-06 | 2024-03-15 | 华侨大学 | Vehicle track prediction model training method, multi-model migration prediction method and device |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113191229A (en) * | 2021-04-20 | 2021-07-30 | 华南农业大学 | Intelligent visual pest detection method |
| CN114511739A (en) * | 2022-01-25 | 2022-05-17 | 哈尔滨工程大学 | Task-adaptive small sample image classification method based on meta-migration learning |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112270681B (en) * | 2020-11-26 | 2022-11-15 | 华南农业大学 | Method and system for detecting and counting yellow plate pests deeply |
| CN114639000A (en) * | 2022-03-30 | 2022-06-17 | 浙江大学 | Small sample learning method and device based on cross-sample attention aggregation |
-
2022
- 2022-06-27 CN CN202210738451.XA patent/CN115019175B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113191229A (en) * | 2021-04-20 | 2021-07-30 | 华南农业大学 | Intelligent visual pest detection method |
| CN114511739A (en) * | 2022-01-25 | 2022-05-17 | 哈尔滨工程大学 | Task-adaptive small sample image classification method based on meta-migration learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115019175A (en) | 2022-09-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113378632B (en) | Pseudo-label optimization-based unsupervised domain adaptive pedestrian re-identification method | |
| CN110443143B (en) | Multi-branch convolutional neural network fused remote sensing image scene classification method | |
| CN107506761B (en) | Brain image segmentation method and system based on significance learning convolutional neural network | |
| CN110334765B (en) | Remote sensing image classification method based on attention mechanism multi-scale deep learning | |
| CN108416370A (en) | Image classification method, device based on semi-supervised deep learning and storage medium | |
| CN102609681B (en) | Face recognition method based on dictionary learning models | |
| CN111079847B (en) | Remote sensing image automatic labeling method based on deep learning | |
| CN112800876B (en) | Super-spherical feature embedding method and system for re-identification | |
| CN111178120B (en) | Pest image detection method based on crop identification cascading technology | |
| CN111860596B (en) | Unsupervised pavement crack classification method and model building method based on deep learning | |
| CN109583379A (en) | A kind of pedestrian's recognition methods again being aligned network based on selective erasing pedestrian | |
| CN110457677B (en) | Entity relationship identification method and device, storage medium and computer equipment | |
| CN112862849B (en) | Image segmentation and full convolution neural network-based field rice ear counting method | |
| CN111461238A (en) | Model training method, character recognition method, device, equipment and storage medium | |
| CN108681735A (en) | Optical character recognition method based on convolutional neural networks deep learning model | |
| CN104268552B (en) | One kind is based on the polygonal fine classification sorting technique of part | |
| CN113887480B (en) | Burma language image text recognition method and device based on multi-decoder joint learning | |
| CN113486886B (en) | License plate recognition method and device in natural scene | |
| CN116258861B (en) | Semi-supervised semantic segmentation method and segmentation device based on multi-label learning | |
| CN111723815A (en) | Model training method, image processing method, device, computer system, and medium | |
| CN115019175B (en) | Pest identification method based on migration element learning | |
| CN116612307A (en) | Solanaceae disease grade identification method based on transfer learning | |
| CN111310820A (en) | Foundation meteorological cloud chart classification method based on cross validation depth CNN feature integration | |
| CN113033345B (en) | V2V video face recognition method based on public feature subspace | |
| CN114037886A (en) | Image recognition method and device, electronic equipment and readable storage medium |
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 |