CN117649552A - Image increment learning method based on contrast learning and active learning - Google Patents
Image increment learning method based on contrast learning and active learning Download PDFInfo
- Publication number
- CN117649552A CN117649552A CN202311617606.5A CN202311617606A CN117649552A CN 117649552 A CN117649552 A CN 117649552A CN 202311617606 A CN202311617606 A CN 202311617606A CN 117649552 A CN117649552 A CN 117649552A
- Authority
- CN
- China
- Prior art keywords
- learning
- image
- data
- model
- samples
- 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
- 238000000034 method Methods 0.000 title claims abstract description 56
- 238000000605 extraction Methods 0.000 claims description 29
- 238000012549 training Methods 0.000 claims description 28
- 238000005070 sampling Methods 0.000 claims description 23
- 239000013598 vector Substances 0.000 claims description 23
- 238000009826 distribution Methods 0.000 claims description 20
- 238000013434 data augmentation Methods 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 10
- 238000013528 artificial neural network Methods 0.000 claims description 7
- 230000006870 function Effects 0.000 claims description 6
- 230000003190 augmentative effect Effects 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 4
- 230000008859 change Effects 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 claims description 3
- 230000000644 propagated effect Effects 0.000 claims description 3
- 230000000007 visual effect Effects 0.000 claims description 3
- 230000015556 catabolic process Effects 0.000 claims description 2
- 238000006731 degradation reaction Methods 0.000 claims description 2
- 238000002372 labelling Methods 0.000 abstract 2
- 238000001514 detection method Methods 0.000 abstract 1
- 238000003709 image segmentation Methods 0.000 abstract 1
- 238000013459 approach Methods 0.000 description 7
- 238000002474 experimental method Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 238000004821 distillation Methods 0.000 description 3
- 238000013140 knowledge distillation Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000008685 targeting Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-supervised learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/091—Active learning
-
- 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
- G06V10/77—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
-
- 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
- G06V10/778—Active pattern-learning, e.g. online learning 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
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Multimedia (AREA)
- Biomedical Technology (AREA)
- General Engineering & Computer Science (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an image increment learning method based on contrast learning and active learning, which firstly utilizes the contrast learning method to learn robust image representation so as to enhance similarity and reduce difference between classes. And an active learning strategy is adopted, and a sample with the maximum gain of the current model is selected for marking, so that the marking cost is minimized and the performance of the model is improved. The invention combines contrast learning and active learning, thereby overcoming the limitation of the traditional image increment learning method. And generating robust image representation through contrast learning, so that the generalization capability of the model is improved. Meanwhile, the sample with the most information value is selected for labeling by utilizing an active learning strategy, so that the labeling cost can be effectively reduced and the learning efficiency of the model can be improved. By the method, efficient, accurate and stable model updating in the incremental learning task can be realized, and the method is suitable for various image related application fields such as image classification, target detection, image segmentation and the like.
Description
Technical Field
The invention relates to the technical field of computers, in particular to an image incremental learning method based on contrast learning and active learning.
Background
The background art relates to three major blocks: incremental learning, active learning, and contrast learning.
1) Incremental learning (Incremental Learning)
Incremental learning refers to updating a model by learning new data to accommodate new tasks or environments based on existing knowledge. In class increment learning, the data, model and algorithm angles can be roughly divided into three classes. Data-centric approaches focus on instance-based resolution of class increment problems, which can be further divided into data replay and data regularization. Model-centric approaches either regularize model parameters, prevent deviations, or extend network architecture to enhance representation capabilities. Finally, algorithm-centric approaches utilize knowledge distillation to prevent forgetting or correct deviations in the class delta model. Data-centric approaches can be divided into data replay and data regularization, where data replay can also be subdivided into direct original replay and generative replay. The method taking the model as the center can be divided into two types of dynamic network and parameter regularization, wherein the dynamic network can be divided into neuron expansion, central expansion and prompt expansion according to different modified structures. Algorithm-centric methods can be divided into knowledge distillation and model correction, where knowledge distillation can be divided into logarithmic probability distillation, feature distillation, and relational distillation. Model rectification can be classified into feature rectification, logarithmic probability rectification and weight rectification. The existing methods all need to find a balance between stability and plasticity, namely, when learning new knowledge and facing new tasks or new data, the performance of the model is ensured not to be significantly reduced or fluctuated.
2) Active Learning (Active Learning)
Active learning refers to a semi-supervised learning method aimed at training a machine learning model by minimizing the amount of data that needs to be labeled. In active learning, an algorithm selects among the data that needs to be marked in order to give the data to a human annotator for marking. In the process, the algorithm can try to select the data with the most information quantity for marking, so that the data quantity required to be marked can be reduced, and meanwhile, the performance of the machine learning model can be improved.
The following three query strategies can be broadly classified:
query strategy based on uncertainty sampling: an active learning based on uncertainty selects data samples with a high degree of offset uncertainty or upper level uncertainty, where offset uncertainty refers to the natural uncertainty of the data due to the impact of the data generation process, which itself is stochastic. Query strategy based on representativeness/diversity: sample batches representing unlabeled sets are selected based on a representative/diverse strategy and based on the intuition that the selected representative sample, once labeled, can serve as a surrogate for the entire data set. Clustering methods are widely used in representative-based strategies. One typical method is K-Means, which selects a centroid by iterative sampling in proportion to the squared distance of the point from the nearest previously selected centroid. Another widely adopted approach selects a collection of representative points based on a core set, which is a sub-sample of a dataset, that can act as a proxy for the entire set. And then sampling is repeated until the requirements are met. Comprehensive query strategies: the comprehensive query strategy aims at achieving a trade-off between uncertainty and representativeness/diversity in query selection.
3) Contrast learning (Contrastive Learning)
The contrast learning means that the model can learn the characteristic representation of the data by comparing different data samples, thereby improving the generalization capability and the robustness of the model. The core idea of contrast learning is to enable the model to distinguish between positive and negative samples, i.e. samples from the same distribution and samples from different distributions. The main challenge of contrast learning is how to construct an effective contrast loss function, i.e. how to define the similarity and difference between positive and negative samples. The existing contrast learning is usually used for self-supervision image learning, and the self-supervision image learning process using the contrast learning usually requires a great deal of calculation force, and is difficult to converge in some cases.
Disclosure of Invention
The invention aims to provide an image increment learning method based on contrast learning and active learning, which has the innovation point that a supervised contrast loss function is provided, the problem of insufficient generalization capability of common cross entropy loss is solved, and the problems that the contrast learning is difficult to converge and the training calculation force is high in requirement during self-supervision are also solved. The method also provides a sample selection strategy combining multiple methods, gives consideration to uncertainty of samples and improves diversity of sampled samples.
The specific technical scheme for realizing the aim of the invention is as follows:
an image increment learning method based on contrast learning and active learning comprises the following steps:
step one: determining tasks for incremental learning
Dividing tasks by adopting data sources, time or category factors; the method comprises the following steps: data is collected from different domains or scenes, and is divided according to the domains or scenes; the data is collected in time sequence and divided according to time intervals; the category of the data is dynamically changed and is divided according to the change of the category;
step two: module for determining model feature extraction and corresponding pre-training model thereof
The feature extraction module is a convolutional neural network, a graph neural network or a self-encoder for extracting useful feature representations from the image; the pre-training model is a model trained on a large-scale data set and is used for initializing parameters of a module for feature extraction, specifically, a ResNet pre-trained on an ImageNet is used as a module for feature extraction, and a fully connected neural network of a specified classification class is reconstructed to replace the fully connected neural network of the original pre-training model;
step three: build supervised contrast learning penalty and train
The training process is that iteration is carried out on the data of each task, each image is subjected to data augmentation at first when each iteration is carried out, random cutting, random overturning and random rotation are adopted, then the augmented image is input into a feature extraction module to obtain probability distribution of image categories, the probability distribution of the image categories and an image label are subjected to InfoNCE loss calculation, and back propagation and parameter updating are carried out; training of a task is completed after training is finished;
step four: sampling of active learning on trained samples for next incremental learning
The sampling process is carried out after the training of each task is finished, and each time of sampling, an algorithm combining maximum entropy, maximum variance and random sampling is adopted, and half of samples of each category of the original data set are selected to be used as a part of data of the next task;
step five: repeating the second, third and fourth steps to complete subsequent incremental learning
When the data of a new task arrives, combining the data with the sample obtained by the step one to serve as new training data, and repeating the step two, the step three and the step four to update the parameters of the module for feature extraction so as to adapt to new data distribution; this results in a series of different image classification tasks for each stage, with each subsequent model showing no significant degradation in the previous task.
And step three, constructing supervised contrast learning loss, comprising the following steps:
step 3.1: data augmentation of data for each task
The method of data augmentation is used, namely random cutting, random overturning and random rotation, and the appearance and the visual angle of the image are changed on the premise of not changing the semantic information of the image; the data augmentation is to enable the model to learn the same feature representation from different augmented images, thereby improving the robustness and invariance of the features;
step 3.2: inputting the amplified image into a feature extraction module to obtain probability distribution of image categories; calculating the comparison loss between the probability distribution of the image category and the image category label, and carrying out back propagation and parameter updating; the feature extraction contrast loss adopts InfoNCE loss, and the formula is as follows:
where q is the query feature vector, k + Is a positive sample feature vector, k i Is a negative sample feature vector, T is a temperature coefficient, and is used for controlling the distribution shape of the feature vector; the meaning of InfoNCE penalty is to give a query feature vector q to be matched with the positive sample feature vector k + Is larger relative to the other negative sample feature vectors, thereby making q and k + Is smaller, and q and other k i Is larger; infoNCE loss is seen as a multi-classification problem, with the goal of dividing q into k + The category in which it is located; after the InfoNCE loss is calculated, the parameters of the feature extraction module are back propagated and updated by using a gradient descent algorithm, so that the value of the loss function is reduced, and the performance of the feature extraction module is improved.
Step four, the improved active learning sampling algorithm
The active learning method mainly uses the following three methods:
the method comprises the following steps: maximum entropy. The method is based on information theory, and calculates the information entropy of the prediction probability of each sample, wherein the higher the information entropy is, the more uniform the prediction probability of the sample is, and the higher the uncertainty is. The advantage of this approach is that uncertainty in the prediction probability of the samples can be taken into account, which has the disadvantage that it may be too sensitive to classification problems.
The second method is as follows: random sampling method. This method is based on randomness, which randomly selects a proportion of samples from the trained samples as part of the data for the next task. The advantage of this method is that it is simple and easy to implement, and the disadvantage is that the uncertainty of the model is not utilized and some useless samples may be selected.
And a third method: variance maximum minimum method. This method is based on statistics, which calculates the variance of the prediction probability for each sample, the higher the variance, the more scattered the prediction probability for the sample, and the higher the uncertainty. The invention selects the samples with the maximum variance of half of the required samples and the samples with the minimum variance of half of the required samples, so as to ensure that the sampled method covers all the distributions as much as possible, and the actual variance of the sampled samples is extremely small due to the condition that the samples with the maximum variance are aggregated.
The invention provides an image increment learning method based on contrast learning and active learning. A large number of experiments are carried out on the Cifar-10 data set and the Cifar-100 data set, and the experimental results show that the method provided by the invention remarkably improves the accuracy of picture classification in the incremental learning process.
Drawings
FIG. 1 is a flow chart of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the following specific examples and drawings. The procedures, conditions, experimental methods, etc. for carrying out the present invention are common knowledge and common knowledge in the art, except for the following specific references, and the present invention is not particularly limited.
The invention relates to a model which is a supervision model based on contrast learning, comprising three modules: pre-training a model, comparing a loss function and sampling. The model training process of the invention is as follows: firstly, determining a proper task size and a proper model structure according to the characteristics of a data set; then, extracting data needing to be increased from the data set, and carrying out different data enhancement on each data as input of a pre-training model; then, calculating the contrast loss between the model output and the image label, and optimizing the parameters of the pre-training model, so that the distance between the feature vectors of different view angles of the same data is as small as possible, and the distance between the feature vectors of different view angles of different data is as large as possible; and finally, selecting a part of data from the data set by using a sampling method according to the trained model as the data of the subsequent incremental training so as to improve the generalization capability and adaptability of the model.
Referring to fig. 1, the present invention specifically includes:
step 1: the task for which incremental learning is required is determined. This may be divided according to factors such as the source of the data, time, category, etc. For example, if data is collected from different domains or scenarios, tasks may be split by domain or scenario. If the data is collected in a time sequence, the tasks may be separated by time intervals. If the category of data is dynamically changing, the tasks may be divided by the change in category. The invention uses a Cifar-10 data set and a Cifar-100 data set to divide the data sets into 5 and 10 increment tasks for increment learning demonstration, wherein each task of the Cifar-10 data set has 2 target classification categories, and each task of the Cifar-100 data set has 10 target classification categories.
Step 2: a module for determining model feature extraction and a corresponding pre-trained model thereof. The means for feature extraction may be a convolutional neural network, a graph neural network, a self-encoder, etc., for extracting useful feature representations from the image. The pre-trained model may be a model trained on a large-scale dataset for initializing parameters of a module for feature extraction or providing a better feature space. For example, a pre-trained ResNet on ImageNet can be used as a pre-training model as a module for feature extraction. The invention adopts a pretraining model of the rest Net18, discards all the full connection parts and only uses a convolution layer.
Step 3: data augmentation of data for each task
The method of data augmentation is used, namely random cutting, random overturning and random rotation, and the appearance and the visual angle of the image are changed on the premise of not changing the semantic information of the image; the data augmentation is to enable the model to learn the same feature representation from different augmented images, thereby improving the robustness and invariance of the features; inputting the amplified image into a feature extraction module to obtain probability distribution of image categories; calculating the comparison loss between the probability distribution of the image category and the image category label, and carrying out back propagation and parameter updating; the feature extraction contrast loss adopts InfoNCE loss, and the formula is as follows:
where q is a query feature vector, k + Is a positive sample feature vector, k i Is a negative sample feature vector, and T is a temperature coefficient for controlling the distribution shape of the feature vector. The meaning of InfoNCE penalty is to give a query feature vector q to be matched with the positive sample feature vector k + Is larger relative to the other negative sample feature vectors, thereby making q and k + Is smaller, and q and other k i Is larger. InfoNCE loss can be seen as a multi-classification problem, targeting q to k + The category in which they are located. After the InfoNCE loss is calculated, the parameters of the feature extraction module are back propagated and updated by using a gradient descent algorithm, so that the value of the loss function is reduced, and the performance of the feature extraction module is improved. The present embodiment inputs data for one task to train the model for the current task phase, uses the adam optimizer to perform gradient descent, and sets epoch to 50.
Step 4: the trained samples are actively learned for sampling for the next incremental learning. The active learning sampling is a sampling strategy based on active learning, and is used for selecting a part of samples from trained samples as a part of data of a next task so as to maintain the performance of a model on an old task and reduce the training difficulty of a new task. The sampling process is performed after the training of each task is finished, and each time the sample is sampled, an uncertainty score, such as information entropy, variance, maximum probability and the like, is calculated for each trained sample, and then a certain proportion of samples are selected from high to low according to the magnitude of the uncertainty score to be used as a part of data of the next task. This may enable the model to focus on samples that are more difficult for the model, thereby improving the learning efficiency of the model. In this implementation, sampling methods of maximum entropy, random sampling and maximum and minimum variance combination are adopted, and the total of 2000 and 500 samples are sampled for each category known by each task in the Cifar-10 data set and the Cifar-100 data set by the three methods.
Step 5: the three previous steps are repeated to complete the subsequent incremental learning. When the data of a new task arrives, the data can be combined with the sample obtained by the previous step to be used as new training data, and then the second step, the third step and the fourth step are repeated to update the parameters of the module for feature extraction so as to adapt to new data distribution. This may enable the model to continuously learn new knowledge from new data while retaining old knowledge that has been learned. Thus, a series of models for each stage of image classification task are obtained.
Examples
Experiments were performed on multiple data sets to evaluate the effectiveness of the proposed image delta approach based on contrast learning and active learning. In this experiment, cifar-10 was randomly divided into 5 groups of two by category, and Cifar-100 was randomly divided into 10 groups of 10 by category.
After training and testing of the data sets, the accuracy performance of the final model on the three data sets was obtained and is shown in table 1. Compared with the Der model, the experimental result shows that the image increment learning method based on contrast learning and active learning can effectively avoid catastrophic forgetting and ensure considerable accuracy.
TABLE 1 comparison of the Performance of the present example with Der on the Cifar-10 dataset and the Cifar-100 dataset
| Cifar-10 | Cifar-100 | |
| Der | 71.99％ | 56.13％ |
| ours | 77.98％ | 57.39％ |
The protection of the present invention is not limited to the above embodiments. Variations and advantages that would occur to one skilled in the art are included in the invention without departing from the spirit and scope of the inventive concept, and the scope of the invention is defined by the appended claims.
Claims (3)
1. The image increment learning method based on contrast learning and active learning is characterized by comprising the following steps of:
step one: determining tasks for incremental learning
Dividing tasks by adopting data sources, time or category factors; the method comprises the following steps: data is collected from different domains or scenes, and is divided according to the domains or scenes; the data is collected in time sequence and divided according to time intervals; the category of the data is dynamically changed and is divided according to the change of the category;
step two: module for determining model feature extraction and corresponding pre-training model thereof
The feature extraction module is a convolutional neural network, a graph neural network or a self-encoder for extracting useful feature representations from the image; the pre-training model is a model trained on a large-scale data set and is used for initializing parameters of a module for feature extraction, specifically, a ResNet pre-trained on an ImageNet is used as a module for feature extraction, and a fully connected neural network of a specified classification class is reconstructed to replace the fully connected neural network of the original pre-training model;
step three: build supervised contrast learning penalty and train
The training process is that iteration is carried out on the data of each task, each image is subjected to data augmentation at first when each iteration is carried out, random cutting, random overturning and random rotation are adopted, then the augmented image is input into a feature extraction module to obtain probability distribution of image categories, the probability distribution of the image categories and an image label are subjected to InfoNCE loss calculation, and back propagation and parameter updating are carried out; training of a task is completed after training is finished;
step four: sampling of active learning on trained samples for next incremental learning
The sampling process is carried out after the training of each task is finished, and each time of sampling, an algorithm combining maximum entropy, maximum variance and random sampling is adopted, and half of samples of each category of the original data set are selected to be used as a part of data of the next task;
step five: repeating the second, third and fourth steps to complete subsequent incremental learning
When the data of a new task arrives, combining the data with the sample obtained by the step one to serve as new training data, and repeating the step two, the step three and the step four to update the parameters of the module for feature extraction so as to adapt to new data distribution; this results in a series of different image classification tasks for each stage, with each subsequent model showing no significant degradation in the previous task.
2. The image incremental learning method based on contrast learning and active learning according to claim 1, wherein the construction-supervised contrast learning loss in the step three includes the steps of:
step 2.1: data augmentation of data for each task
The method of data augmentation is used, namely random cutting, random overturning and random rotation, and the appearance and the visual angle of the image are changed on the premise of not changing the semantic information of the image; the data augmentation enables the model to learn the same characteristic representation from different augmented images, and the robustness and invariance of the characteristics are improved;
step 2.2: inputting the amplified image into a feature extraction module to obtain probability distribution of image categories; calculating the comparison loss between the probability distribution of the image category and the image category label, and carrying out back propagation and parameter updating; the feature extraction contrast loss adopts InfoNCE loss, and the formula is as follows:
where q is the query feature vector, k + Is a positive sample feature vector, k i Is a negative sample feature vector, T is a temperature coefficient, and is used for controlling the distribution shape of the feature vector; the meaning of InfoNCE penalty is to give a query feature vector q to be matched with the positive sample feature vector k + Is larger relative to the other negative sample feature vectors, thereby making q and k + Is smaller, and q and other k i Is larger; infoNCE loss is seen as a multi-classification problem, with the goal of dividing q into k + The category in which it is located; after the InfoNCE loss is calculated, the parameters of the feature extraction module are back propagated and updated by using a gradient descent algorithm, so that the value of the loss function is reduced, and the performance of the feature extraction module is improved.
3. The contrast learning and active learning based image incremental learning method of claim 1 wherein step four of the active learning sampling comprises:
maximum entropy: based on the information theory, calculating the information entropy of the prediction probability of each sample, wherein the higher the information entropy is, the more uniform the prediction probability of the sample is, and the higher the uncertainty is;
random sampling: randomly selecting a proportion of samples from the trained samples based on randomness as part of the data for the next task;
maximum and minimum variance: calculating the variance of the prediction probability of each sample based on statistics, wherein the higher the variance is, the more the prediction probability of the sample is dispersed and the higher the uncertainty is; and selecting the samples with the maximum variance of half of the required samples and the samples with the minimum variance of half of the required samples, so as to ensure that the sampled method covers all the distributions as much as possible, and the actual variances of the sampled samples are extremely small because the condition that the samples with the maximum variance are aggregated does not occur.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311617606.5A CN117649552A (en) | 2023-11-30 | 2023-11-30 | Image increment learning method based on contrast learning and active learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311617606.5A CN117649552A (en) | 2023-11-30 | 2023-11-30 | Image increment learning method based on contrast learning and active learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117649552A true CN117649552A (en) | 2024-03-05 |
Family
ID=90044545
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311617606.5A Pending CN117649552A (en) | 2023-11-30 | 2023-11-30 | Image increment learning method based on contrast learning and active learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117649552A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117909854A (en) * | 2024-03-20 | 2024-04-19 | 东北大学 | Zero sample composite fault diagnosis method based on multi-mode contrast embedding |
-
2023
- 2023-11-30 CN CN202311617606.5A patent/CN117649552A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117909854A (en) * | 2024-03-20 | 2024-04-19 | 东北大学 | Zero sample composite fault diagnosis method based on multi-mode contrast embedding |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111199242B (en) | Image increment learning method based on dynamic correction vector | |
| Zhang et al. | Improved deep hashing with soft pairwise similarity for multi-label image retrieval | |
| CN107480261B (en) | Fine-grained face image fast retrieval method based on deep learning | |
| CN113190699B (en) | Remote sensing image retrieval method and device based on category-level semantic hash | |
| CN108564129B (en) | Trajectory data classification method based on generation countermeasure network | |
| Liu et al. | Incdet: In defense of elastic weight consolidation for incremental object detection | |
| CN111125411B (en) | Large-scale image retrieval method for deep strong correlation hash learning | |
| CN112182221B (en) | Knowledge retrieval optimization method based on improved random forest | |
| Chang et al. | Automatic channel pruning via clustering and swarm intelligence optimization for CNN | |
| CN111599409B (en) | circRNA recognition method based on MapReduce parallelism | |
| CN111008224B (en) | Time sequence classification and retrieval method based on deep multitasking representation learning | |
| CN112819023A (en) | Sample set acquisition method and device, computer equipment and storage medium | |
| CN114444600A (en) | Small sample image classification method based on memory enhanced prototype network | |
| CN110598022B (en) | Image retrieval system and method based on robust deep hash network | |
| CN117649552A (en) | Image increment learning method based on contrast learning and active learning | |
| CN105320764A (en) | 3D model retrieval method and 3D model retrieval apparatus based on slow increment features | |
| Liu et al. | EACP: An effective automatic channel pruning for neural networks | |
| Zhang et al. | Multi-domain clustering pruning: Exploring space and frequency similarity based on GAN | |
| CN115358305A (en) | Incremental learning robustness improving method based on iterative generation of boundary samples | |
| CN114091650A (en) | Searching method and application of deep convolutional neural network architecture | |
| CN113032613A (en) | Three-dimensional model retrieval method based on interactive attention convolution neural network | |
| CN106569954A (en) | Method based on KL divergence for predicting multi-source software defects | |
| Li et al. | Deep metric learning with density adaptivity | |
| CN115329120A (en) | Weak label Hash image retrieval framework with knowledge graph embedded attention mechanism | |
| Qin et al. | Deep top similarity hashing with class-wise loss for multi-label image retrieval |
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 |