CN115953631B - Long-tail small sample sonar image classification method and system based on deep migration learning - Google Patents
Long-tail small sample sonar image classification method and system based on deep migration learning Download PDFInfo
- Publication number
- CN115953631B CN115953631B CN202310045636.7A CN202310045636A CN115953631B CN 115953631 B CN115953631 B CN 115953631B CN 202310045636 A CN202310045636 A CN 202310045636A CN 115953631 B CN115953631 B CN 115953631B
- Authority
- CN
- China
- Prior art keywords
- training
- classifier
- neural network
- convolutional neural
- network 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 74
- 238000013508 migration Methods 0.000 title description 9
- 230000005012 migration Effects 0.000 title description 9
- 238000012549 training Methods 0.000 claims abstract description 160
- 238000005070 sampling Methods 0.000 claims abstract description 35
- 238000013526 transfer learning Methods 0.000 claims abstract description 24
- 238000012216 screening Methods 0.000 claims abstract description 16
- 238000013138 pruning Methods 0.000 claims abstract description 8
- 238000013527 convolutional neural network Methods 0.000 claims description 66
- 230000004927 fusion Effects 0.000 claims description 14
- 239000013598 vector Substances 0.000 claims description 14
- 238000004590 computer program Methods 0.000 claims description 7
- 230000006870 function Effects 0.000 claims description 7
- 238000012935 Averaging Methods 0.000 claims description 6
- 238000012795 verification Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 3
- 230000014759 maintenance of location Effects 0.000 claims description 2
- 230000000717 retained effect Effects 0.000 claims 2
- 238000004422 calculation algorithm Methods 0.000 description 6
- 238000012952 Resampling Methods 0.000 description 4
- 238000013459 approach Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 238000012360 testing method Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 238000003062 neural network model Methods 0.000 description 3
- 238000001514 detection method Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000007635 classification algorithm Methods 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Classifications
-
- 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
- Y02A90/30—Assessment of water resources
Landscapes
- Measurement Of Velocity Or Position Using Acoustic Or Ultrasonic Waves (AREA)
Abstract
The invention discloses a long-tail small sample sonar image classification method and a long-tail small sample sonar image classification system based on deep transfer learning. The method is based on the prior deep transfer learning method, utilizes a two-stage decoupling training scheme to perform fine adjustment in transfer learning (firstly, training of a feature indicator and then training of a classifier based on a balanced sub-training set), and uses different balance settings during fine adjustment of the feature indicator and the classifier to simultaneously improve the representation capacity and the classification capacity of a model and relieve the problem of long tail distribution of data; constructing a plurality of balanced sub-training sets by adopting a multi-balanced sampling strategy, then training a plurality of balanced classifiers on the sub-training sets, and integrating the classifiers through integrated learning to reduce the loss of data information and improve the model performance under the problem of small samples; and finally, screening classifiers with better classification performance by an integrated pruning method so as to further improve the performance and efficiency of the model.
Description
Technical Field
The invention relates to the technical field of sonar image recognition, in particular to a long-tail small-sample sonar image classification method and system based on deep migration learning.
Background
Sonar imaging is an important underwater detection means, and can be used for detection in environments with weak light and turbid water. The automatic target classification of the underwater sonar images increases challenges due to the fact that the sonar images are large in noise, single in color, sparse in characteristics and the like, and a few good solutions to the problems are provided. In earlier studies, some conventional classification algorithms were used for classification of sonar images. For example using methods based on decision trees, markov random fields, sparse representation. In recent years, more and more researchers have proposed methods for solving the problem of classification of sonar image objects using deep convolutional neural networks, and have achieved interesting results. Among them, the classification effect of the deep transfer learning-based method proves to be more excellent than that of the conventional classification method. However, long tail distribution of sonar image datasets presents a significant challenge to deep transfer learning algorithms. The existing methods do not pay special attention to the long tail distribution problem of the data set, so that when the long tail distribution data set in the real world is faced, the methods tend to train the deep convolutional neural network towards the direction of distinguishing sonar image samples as head classes as much as possible, and therefore the classification accuracy of the head classes is high, the classification accuracy of the tail classes is low, and the classification effect is not ideal.
An important and effective approach to solving the long tail distribution problem in the past has been the class rebalancing approach, which can generally be divided into two families, including resampling and re-weighting. Resampling is an intuitive method for balancing data distribution that is widely used in the field, including undersampling of most categories in a dataset, oversampling of few categories, and class-balanced sampling combining both. But the undersampling discards precious data, so that the generalization capability of the network is weakened to a certain extent, and particularly for sonar images with very little data; while oversampling re-uses the same data, resulting in overfitting of the tail class. Class-balanced sampling samples based on the number of samples per class in the dataset, but it is still a combination of undersampling and oversampling in nature, which cannot avoid the above drawbacks. Another important approach is re-weighting, which performs class balancing by increasing the contribution of the misclassification of a minority class to the penalty function, or decreasing the penalty of misclassification of a majority class. This approach reduces to some extent the excessive attention of the network to the majority class during training, but also causes some loss of representation capacity to the model.
Disclosure of Invention
The invention aims to provide a long-tail small sample sonar image classification method and system based on deep transfer learning, which are used for solving the problem of poor classification performance when a sonar image of a long-tail small sample is faced in the prior art.
In order to achieve the above object, the present invention provides the following solutions:
the invention provides a long-tail small sample sonar image classification method based on deep transfer learning, which comprises the following steps:
training a feature indicator in the pre-trained convolutional neural network model by adopting a sonar image training set to obtain a convolutional neural network model after primary training;
performing balanced sampling on the sonar image training set by adopting a multi-balanced sampling strategy to obtain a plurality of balanced sub-training sets;
training the classifier in the convolutional neural network model after primary training by using a plurality of sub-training sets respectively to obtain a plurality of classifier after secondary training;
screening and fusing the multiple classifiers after secondary training to obtain a fused classifier;
replacing the classifier in the convolutional neural network model after primary training with the fusion classifier to obtain a convolutional neural network model after fusion;
and inputting the sonar image to be detected into the fused convolutional neural network model to obtain a classification result.
Optionally, the performing balance sampling on the sonar image training set by adopting a multi-balance sampling strategy to obtain a plurality of balanced sub-training sets specifically includes:
determining sample types contained in a sonar image training set; the sample categories include a head category and a tail category;
respectively carrying out replaced random sampling on samples of each category in the sonar image training set, and respectively forming sub-training sets by the samples of each category obtained by random sampling; each of the sub-training sets includes N samples.
Optionally, the determining the sample category contained in the sonar image training set specifically includes:
calculating the average value of the number of samples contained in each category in the sonar image training set;
determining a class containing a number of samples greater than the average value as a head class;
a class containing a number of samples no greater than the average value is determined as a tail class.
Optionally, the value of N is an average value of the number of samples contained in each tail class in the sonar image training set.
Optionally, the training the classifier in the convolutional neural network model after the primary training by using the plurality of sub-training sets respectively to obtain a plurality of classifiers after the secondary training specifically includes:
respectively inputting a plurality of sub-training sets into a feature indicator in a convolutional neural network model after one training to obtain a plurality of groups of feature vectors;
and respectively inputting a plurality of groups of feature vectors into the classifiers in the convolutional neural network model after primary training, and respectively carrying out secondary training on the classifiers in the convolutional neural network model after primary training according to the classification result to obtain a plurality of classifiers after secondary training.
Optionally, the screening and fusing are performed on the multiple classifiers after the secondary training to obtain a fused classifier, which specifically includes:
inputting the verification data set into a feature indicator in the convolutional neural network model after one training to obtain a second feature vector; the verification data set is acquired in the sonar image training set in a replaced random sampling mode;
respectively inputting the second feature vectors into a plurality of the secondarily trained classifiers, and obtaining the prediction probability of each secondarily trained classifier through a softMax function;
selecting a preset number of secondary trained classifiers with larger accuracy of the prediction probability and geometric average of the prediction probability;
and fusing the trained classifiers with the preset number by adopting a parameter averaging method to obtain a fused classifier.
Optionally, training the feature indicator in the pre-trained convolutional neural network model by using a sonar image training set to obtain a trained convolutional neural network model, and further including:
and (3) pre-training the convolutional neural network model by adopting an ImageNet image data set to obtain a pre-trained convolutional neural network model.
A long-tail small sample sonar image classification system based on deep transfer learning, the system being applied to the method described above, the system comprising:
the primary training module is used for training the feature indicators in the pre-trained convolutional neural network model by adopting a sonar image training set to obtain a convolutional neural network model after primary training;
the balanced sampling module is used for carrying out balanced sampling on the sonar image training set by adopting a multi-balanced sampling strategy to obtain a plurality of balanced sub-training sets;
the training module is used for training the classifier in the convolutional neural network model after primary training by utilizing the plurality of sub-training sets respectively to obtain a plurality of classifier after secondary training;
the screening and fusing module is used for screening and fusing the plurality of classifier after the secondary training to obtain a fused classifier;
the classifier replacement module is used for replacing the classifier in the convolutional neural network model after primary training with the fusion classifier to obtain a convolutional neural network model after fusion;
and the classification module is used for inputting the sonar images to be detected into the fused convolutional neural network model to obtain classification results.
An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method described above when executing the computer program.
A computer readable storage medium having stored thereon a computer program which when executed by a processor implements the method described above.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
the invention discloses a long-tail small sample sonar image classification method and a system based on deep transfer learning, wherein the method comprises the following steps: performing balanced sampling on the sonar image training set by adopting a multi-balanced sampling strategy to obtain a plurality of balanced sub-training sets; training the classifier in the convolutional neural network model after primary training by using a plurality of sub-training sets respectively to obtain a plurality of classifier after secondary training; screening and fusing the multiple classifiers after secondary training to obtain a fused classifier; replacing the classifier in the convolutional neural network model after primary training with the fusion classifier to obtain a convolutional neural network model after fusion; and inputting the sonar image to be detected into the fused convolutional neural network model to obtain a classification result. The method is based on the prior deep transfer learning method, utilizes a two-stage decoupling training scheme to perform fine adjustment in transfer learning (firstly, training of a feature indicator and then training of a classifier based on a balanced sub-training set), and uses different balance settings during fine adjustment of the feature indicator and the classifier to simultaneously improve the representation capacity and the classification capacity of a model and relieve the problem of long tail distribution of data; constructing a plurality of balanced sub-training sets by adopting a multi-balanced sampling strategy, then training a plurality of balanced classifiers on the sub-training sets, and integrating the classifiers through integrated learning to reduce the loss of data information and improve the model performance under the problem of small samples; and finally, screening classifiers with better classification performance by an integrated pruning method so as to further improve the performance and efficiency of the model.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings that are needed in the embodiments will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present invention, and that 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 long-tail small sample sonar image classification method based on deep transfer learning provided by an embodiment of the invention;
fig. 2 is a schematic diagram of a long-tail small sample sonar image classification method based on a balance integrated transfer learning algorithm provided by an embodiment of the present invention;
FIG. 3 is a schematic diagram of a comparison of a conventional resampling strategy for long tail distribution data provided by an embodiment of the present invention with a multi-balanced sampling strategy of the present invention;
fig. 4 is a schematic diagram of comparison between a method for directly fusing all classifiers and a method for fusing part of classifiers after integrated screening according to an embodiment of the present invention, where (a) in fig. 4 is the performance of the method for directly fusing all classifiers, and (b) in fig. 4 is the performance of the method for fusing part of classifiers after integrated screening according to the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The invention aims to provide a long-tail small sample sonar image classification method and system based on deep transfer learning, which are used for solving the problem of poor classification performance when a sonar image of a long-tail small sample is faced in the prior art.
In order that the above-recited objects, features and advantages of the present invention will become more readily apparent, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description.
The embodiment of the invention provides a long-tail small sample sonar image classification method based on deep transfer learning, which comprises the following steps:
and carrying out balanced sampling on the sonar image training set by adopting a multi-balanced sampling strategy to obtain a plurality of balanced sub-training sets.
And training the classifier in the convolutional neural network model after primary training by using a plurality of sub-training sets respectively to obtain a plurality of classifier after secondary training.
And screening and fusing the multiple classifiers after the secondary training to obtain a fused classifier.
And replacing the classifier in the convolutional neural network model after primary training with the fusion classifier to obtain the convolutional neural network model after fusion.
And inputting the sonar image to be detected into the fused convolutional neural network model to obtain a classification result.
The training process of the feature indicator in the convolutional neural network model is actually one time in the embodiment of the invention, and is specifically as follows:
and pre-training the convolutional neural network model to obtain a pre-trained convolutional neural network model.
And training the feature indicators in the pre-trained convolutional neural network model by adopting a sonar image training set to obtain the convolutional neural network model after one training.
In one embodiment, as shown in fig. 1 and 2, the method specifically includes:
step 1: pre-training a deep convolutional neural network model on a large image dataset;
training a deep convolutional neural network model (such as ResNet) on a large image dataset (such as ImageNet), calculating classification loss by using a cross entropy loss function, calculating prediction probability by using a softMax function, evaluating classification performance by using accuracy, and optimizing model parameters by using a random gradient descent algorithm.
Step 2: and adopting transfer learning, and performing fine tuning on the pre-trained deep convolutional neural network model by a two-stage decoupling training method.
The characteristic indicator of the deep convolutional neural network model (namely, parameters of the convolutional neural network part) is finely tuned by adopting an original long-tail distributed data set, and a small learning rate (for example, 0.01) is set during fine tuning; the parameters of the feature presenter are fixed, and a balanced sub-training set is adopted to perform fine tuning on the classifier (namely, the fully connected neural network part) of the deep convolutional neural network model (specific steps refer to step 3-step 6, which are not described herein in detail), and a smaller learning rate (for example, 0.1) is set during fine tuning.
Step 3: performing balanced sampling on the sonar image training set distributed on the original long tail by adopting a multi-balanced sampling strategy to obtain a plurality of balanced sub-training sets, as shown in fig. 3;
the head and tail categories of the dataset are determined by the following method: calculating an average of the number of samples contained in each category in the datasetWill contain a number of samples greater than + ->Is divided into head classes, less than +.>Is divided into tail classes, < >>The calculation method of (1) is as follows:
wherein C is the total category number in the dataset, n i The number of training samples contained in the ith class.
Randomly sampling the samples of each category in the sonar image data set with a put back, and setting the sampling number N of each category as the average value of the sample numbers contained in each tail category in the data set The calculation method of (1) is as follows:
wherein ,Ctail For the number of tail categories in the dataset,the number of training samples contained for the ith tail category; the sampling process is repeated T times to obtain T balance sub-training sets with the number of samples of N multiplied by C, wherein C is the total category number in the data set. In the present invention, the value of T is set according to the unbalance factor IF of the data set: T≡IF; the unbalance factor is calculated as follows:
wherein ,nmax and nmin The maximum and minimum number of category samples in the dataset, respectively.
Step 4: training a plurality of classifiers respectively using a plurality of balanced sub-training sets;
and respectively inputting the T balanced sub-training sets into the feature extractor of the deep convolutional neural network model which is pre-trained and finely tuned by adopting the two-stage decoupling training method in the previous step to obtain T groups of feature vectors, and then respectively inputting the T groups of feature vectors into the T classifiers to finely tune.
Step 5: screening the multiple classifiers obtained through training by adopting an integrated pruning method, and reserving the classifier with better performance;
as shown in fig. 4, first, a set of verification data sets is obtained by randomly sampling the samples of each class in the sonar image data set with a set of sampling numbers that are the number of training samples contained in each class; then, the verification data set is input to the frontThe feature vector is obtained from a feature indicator of the depth convolution neural network model which is subjected to pre-training and fine adjustment by adopting a two-stage decoupling training method; further, the feature vectors are respectively input into T classifiers, and T prediction probability outputs are obtained through a softMax function; finally, respectively calculating the accuracy and geometric mean (biometricmean) of the T prediction probability outputs, sequencing the T classifiers according to the results of the two evaluation indexes, and reserving the classifiers with average ranks of top rho, wherein the reserved number of classifiers is T save The method comprises the following steps:
where ρ ε (0, 1) is the retention ratio of the classifier ρ is set to 0.6 in the present invention.
Step 6: integrating and learning information of a plurality of classifiers;
adopting a parameter averaging method to ensure that T remained after the integrated pruning operation save Averaging parameters of the classifiers to obtain a new classifier; the parameter averaging method comprises the following steps: the parameters of a new classifier are obtained by the following formula
wherein ,Tsave For the number of remaining training sets of balancing weights,parameters of the classifier reserved for the i-th.
Step 7: in the test stage, inputting the images into the fused deep convolutional neural network model to obtain a classification result;
inputting the test image into a feature indicator of the depth convolution neural network model which is pre-trained and finely tuned by adopting a two-stage training method in the previous step to obtain a feature vector of the test image;
inputting the feature vectors of the test images into the fused classifier to obtain a predictive value;
inputting the predictive value into a softMax function to obtain predictive probability;
and outputting the category corresponding to the maximum prediction probability as a prediction category.
The present embodiment has performed experiments on three sonar image datasets (KLSG, LTSID, and FLSMDD) of long-tail small samples, the total number of samples, the number of samples of each class, and the unbalance factor (IF) of the three datasets are shown in table 1; because the number of samples in the data set is small, in order to reduce experimental errors and improve the effectiveness of experimental results, we perform fifteen-fold cross validation.
The running environment of the embodiment is a small server, a Graphic Processor (GPU) is Indeltoid RTX2080Ti, a Central Processing Unit (CPU) is Intel to strong E3-1275v6, an operating system is Ubuntu20.04, codes run in a Python3.6 environment, and a neural network model is built based on PyTorr 1.9;
table 1 data set information table
The long-tail small sample sonar image classification method (BETL) based on the balance integrated transfer learning algorithm and various existing methods provided by the invention are subjected to a comparison experiment, and the obtained comparison experiment results are shown in a table 2.
Table 2 comparative experimental results
Various existing methods include: common deep migration learning method (DTL), deep migration learning method+resampling (DTL-RS), deep migration learning method+reweighting (DTL-RW), deep migration learning method+focus loss (DTL-Focal), common convolutional neural network+support vector machine (CNN-SVM), two-stage decoupling training (CE-DRS), deep migration learning method+two-stage decoupling training (DTL-CE-DRS), weight balancing (LTR-WB), deep migration learning method+weight balancing (DTL-LTR-WB); as evaluation indexes, a geometric mean (Gmeasan) and a global average F1 score (Macro-F1) were used. As can be seen from the results in table 2, compared with the existing method, the classification effect of the BETL method provided by the invention in three data sets is improved, and the effectiveness of the proposed method is verified.
In addition, the embodiment of the invention also provides electronic equipment, which comprises: the long-tail small sample sonar image classification method based on the balance integrated transfer learning algorithm is realized when the processor executes the program.
The embodiment of the invention also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the long-tail small sample sonar image classification method based on the balance integrated transfer learning algorithm.
Compared with the prior art, the invention has the remarkable advantages that: (1) The invention utilizes a two-stage decoupling training scheme to carry out fine adjustment in transfer learning, uses different balance settings during fine adjustment of the feature indicator and the classifier, and improves the representation capacity and the classification capacity of the model; (2) Constructing a plurality of balanced sub-training sets through a multi-balanced sampling strategy, then training a plurality of balanced classifiers on the sub-training sets, and fusing the classifiers through ensemble learning to reduce the loss of data information; (3) The classifier with better classification performance is screened by the integrated pruning method, so that the performance and efficiency of the model are further improved.
In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant points refer to the description of the method section.
The principles and embodiments of the present invention have been described herein with reference to specific examples, the description of which is intended only to assist in understanding the methods of the present invention and the core ideas thereof; also, it is within the scope of the present invention to be modified by those of ordinary skill in the art in light of the present teachings. In view of the foregoing, this description should not be construed as limiting the invention.
Claims (6)
1. The long-tail small sample sonar image classification method based on deep transfer learning is characterized by comprising the following steps of:
training a feature indicator in the pre-trained convolutional neural network model by adopting a sonar image training set to obtain a convolutional neural network model after primary training;
performing balanced sampling on the sonar image training set by adopting a multi-balanced sampling strategy to obtain a plurality of balanced sub-training sets, wherein the method specifically comprises the following steps:
calculating the average value of the number of samples contained in each category in the sonar image training set; determining a class containing a number of samples greater than the average value as a head class; determining a class containing a number of samples not greater than the average value as a tail class; the sample categories include a head category and a tail category;
respectively carrying out replaced random sampling on samples of each category in the sonar image training set, and respectively forming sub-training sets by the samples of each category obtained by random sampling; each sub-training set comprises N samples, and the value of N is the average value of the number of samples contained in each tail category in the sonar image training set The calculation method of (1) is as follows:
wherein ,for the average value of the number of samples contained in each tail class in the sonar image training set, C tail For the number of tail categories->The number of training samples contained for the ith tail category;
training the classifier in the convolutional neural network model after primary training by using a plurality of sub-training sets respectively to obtain a plurality of classifier after secondary training;
screening and fusing a plurality of classifier after secondary training to obtain a fused classifier, which specifically comprises the following steps:
screening the multiple classifiers obtained through training by adopting an integrated pruning method, and reserving the classifier with better performance and integrating the information of the multiple classifiers by adopting integrated learning;
the method for screening the plurality of classifiers obtained by training by adopting the integrated pruning method comprises the following steps of: inputting the verification data set into a feature indicator in the convolutional neural network model after one training to obtain a second feature vector; the verification data set is acquired in the sonar image training set in a replaced random sampling mode; respectively inputting the second feature vectors into a plurality of the secondarily trained classifiers, and obtaining the prediction probability of each secondarily trained classifier through a softMax function; selecting a preset number of secondary trained classifiers with larger accuracy of the prediction probability and geometric average of the prediction probability;
the number of retained post-secondary trained classifiers is:
wherein ,Tsave For the number of retained post-training classifiers, where ρ ε (0, 1]The retention ratio for the classifier;
the information integrating a plurality of classifiers by adopting the integrated learning specifically comprises:
adopting a parameter averaging method to ensure that T remained after the integrated pruning operation save Averaging parameters of the classifier after the secondary training to obtain a fusion classifier, wherein the parameters of the fusion classifier are as follows:
wherein ,for fusing the parameters of the classifier +.>Parameters of the classifier reserved for the i-th;
replacing the classifier in the convolutional neural network model after primary training with the fusion classifier to obtain a convolutional neural network model after fusion;
and inputting the sonar image to be detected into the fused convolutional neural network model to obtain a classification result.
2. The long-tail small sample sonar image classification method based on deep transfer learning of claim 1, wherein the training of the classifier in the convolutional neural network model after the primary training by using the plurality of sub-training sets respectively, to obtain a plurality of classifier after the secondary training, specifically comprises:
respectively inputting a plurality of sub-training sets into a feature indicator in a convolutional neural network model after one training to obtain a plurality of groups of feature vectors;
and respectively inputting a plurality of groups of feature vectors into the classifiers in the convolutional neural network model after primary training, and respectively carrying out secondary training on the classifiers in the convolutional neural network model after primary training according to the classification result to obtain a plurality of classifiers after secondary training.
3. The long-tail small sample sonar image classification method based on deep transfer learning of claim 1, wherein the training of the feature indicator in the pre-trained convolutional neural network model by using the sonar image training set to obtain a trained convolutional neural network model further comprises:
and (3) pre-training the convolutional neural network model by adopting an ImageNet image data set to obtain a pre-trained convolutional neural network model.
4. A long-tail small sample sonar image classification system based on deep transfer learning, characterized in that the system is applied to the method of any one of claims 1-3, the system comprising:
the primary training module is used for training the feature indicators in the pre-trained convolutional neural network model by adopting a sonar image training set to obtain a convolutional neural network model after primary training;
the balanced sampling module is used for carrying out balanced sampling on the sonar image training set by adopting a multi-balanced sampling strategy to obtain a plurality of balanced sub-training sets;
the training module is used for training the classifier in the convolutional neural network model after primary training by utilizing the plurality of sub-training sets respectively to obtain a plurality of classifier after secondary training;
the screening and fusing module is used for screening and fusing the plurality of classifier after the secondary training to obtain a fused classifier;
the classifier replacement module is used for replacing the classifier in the convolutional neural network model after primary training with the fusion classifier to obtain a convolutional neural network model after fusion;
and the classification module is used for inputting the sonar images to be detected into the fused convolutional neural network model to obtain classification results.
5. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method of any one of claims 1 to 3 when executing the computer program.
6. A computer readable storage medium having stored thereon a computer program which when executed by a processor implements the method of any of claims 1-3.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310045636.7A CN115953631B (en) | 2023-01-30 | 2023-01-30 | Long-tail small sample sonar image classification method and system based on deep migration learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310045636.7A CN115953631B (en) | 2023-01-30 | 2023-01-30 | Long-tail small sample sonar image classification method and system based on deep migration learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115953631A CN115953631A (en) | 2023-04-11 |
| CN115953631B true CN115953631B (en) | 2023-09-15 |
Family
ID=87287793
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310045636.7A Active CN115953631B (en) | 2023-01-30 | 2023-01-30 | Long-tail small sample sonar image classification method and system based on deep migration learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115953631B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111191735A (en) * | 2020-01-04 | 2020-05-22 | 西安电子科技大学 | Convolutional neural network image classification method based on data difference and multi-scale features |
| CN111738301A (en) * | 2020-05-28 | 2020-10-02 | 华南理工大学 | Long-tail distribution image data identification method based on two-channel learning |
| CN113076873A (en) * | 2021-04-01 | 2021-07-06 | 重庆邮电大学 | Crop disease long-tail image identification method based on multi-stage training |
| CN114283307A (en) * | 2021-12-24 | 2022-04-05 | 中国科学技术大学 | Network training method based on resampling strategy |
| CN114596590A (en) * | 2022-03-15 | 2022-06-07 | 北京信智文科技有限公司 | Single-monkey video motion classification method with long-tail distribution characteristic |
| CN114998651A (en) * | 2022-05-18 | 2022-09-02 | 吾征智能技术(北京)有限公司 | Skin lesion image classification and identification method, system and medium based on transfer learning |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20220156530A1 (en) * | 2020-11-13 | 2022-05-19 | Salesforce.Com, Inc. | Systems and methods for interpolative centroid contrastive learning |
-
2023
- 2023-01-30 CN CN202310045636.7A patent/CN115953631B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111191735A (en) * | 2020-01-04 | 2020-05-22 | 西安电子科技大学 | Convolutional neural network image classification method based on data difference and multi-scale features |
| CN111738301A (en) * | 2020-05-28 | 2020-10-02 | 华南理工大学 | Long-tail distribution image data identification method based on two-channel learning |
| CN113076873A (en) * | 2021-04-01 | 2021-07-06 | 重庆邮电大学 | Crop disease long-tail image identification method based on multi-stage training |
| CN114283307A (en) * | 2021-12-24 | 2022-04-05 | 中国科学技术大学 | Network training method based on resampling strategy |
| CN114596590A (en) * | 2022-03-15 | 2022-06-07 | 北京信智文科技有限公司 | Single-monkey video motion classification method with long-tail distribution characteristic |
| CN114998651A (en) * | 2022-05-18 | 2022-09-02 | 吾征智能技术(北京)有限公司 | Skin lesion image classification and identification method, system and medium based on transfer learning |
Non-Patent Citations (2)
| Title |
|---|
| 张超 ; 陈莹 ; .残差网络下基于困难样本挖掘的目标检测.激光与光电子学进展.2018,(10),全文. * |
| 残差网络下基于困难样本挖掘的目标检测;张超;陈莹;;激光与光电子学进展(10);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115953631A (en) | 2023-04-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113705526B (en) | Hyperspectral remote sensing image classification method | |
| US20160224903A1 (en) | Hyper-parameter selection for deep convolutional networks | |
| CN112613552B (en) | Convolutional neural network emotion image classification method combined with emotion type attention loss | |
| CN110363230B (en) | Stacking integrated sewage treatment fault diagnosis method based on weighted base classifier | |
| CN103366189B (en) | A kind of intelligent classification method of high-spectrum remote sensing | |
| Wu et al. | Centroid transformers: Learning to abstract with attention | |
| CN109993229A (en) | A kind of serious unbalanced data classification method | |
| CN114694028B (en) | Ship weld defect detection method based on convolution confidence generation countermeasure network model | |
| CN111522953B (en) | Marginal attack method and device for naive Bayes classifier and storage medium | |
| CN116192500A (en) | Malicious flow detection device and method for resisting tag noise | |
| CN117315223A (en) | Target detection method based on transducer architecture | |
| CN115952456A (en) | Method, system, program product and storage medium for determining fault diagnosis model | |
| CN112215268A (en) | Method and device for classifying disaster weather satellite cloud pictures | |
| JP2012099070A (en) | Subject detection method and subject detecting device | |
| CN115953631B (en) | Long-tail small sample sonar image classification method and system based on deep migration learning | |
| CN117312920A (en) | Weighting integration unbalance classification method, system, storage medium, equipment and terminal | |
| Peng | Adaptive sampling with optimal cost for class-imbalance learning | |
| JP3222699B2 (en) | Pattern recognition device | |
| CN116363138B (en) | Lightweight integrated identification method for garbage sorting images | |
| Xu et al. | An Efficient Selective Ensemble Learning with Rejection Approach for Classification | |
| Yu et al. | ANEDL: Adaptive Negative Evidential Deep Learning for Open-Set Semi-supervised Learning | |
| CN112733963B (en) | General image target classification method and system | |
| Bao | Image classification under class-imbalanced situation | |
| Hsiao et al. | MTSbag: A Method to Solve Class Imbalance Problems | |
| CN117371423A (en) | Text classifier construction method and device under unbalanced condition |
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 |