CN115830006A - Anomaly detection method for improving hypersphere space learning based on neighbor comparison - Google Patents
Anomaly detection method for improving hypersphere space learning based on neighbor comparison Download PDFInfo
- Publication number
- CN115830006A CN115830006A CN202310053050.5A CN202310053050A CN115830006A CN 115830006 A CN115830006 A CN 115830006A CN 202310053050 A CN202310053050 A CN 202310053050A CN 115830006 A CN115830006 A CN 115830006A
- Authority
- CN
- China
- Prior art keywords
- hypersphere
- loss
- feature
- neighbor
- features
- 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.)
- Granted
Links
Images
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
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to an anomaly detection method for improved hypersphere space learning based on neighbor comparison, and belongs to the technical field of anomaly detection. The method comprises the following steps: s1, constructing a memory library; s2, calculating the hypersphere loss: searching 1-3 neighbors and 4-6 neighbors from a memory library by using the characteristics extracted from the training images, calculating to obtain the close hyper-sphere loss and the far hyper-sphere loss, and adding and coupling to obtain the total loss of the hyper-sphere; s3, calculating brand new neighbor comparison loss; s4, calculating total loss by combining the hypersphere loss and neighbor comparison loss, and training an updating network; and S5, calculating the abnormal score of the test sample, and obtaining a final abnormal detection result after the test is finished. Under the condition of only using normal samples, the method enables the network to more fully learn the clustering pattern of the normal samples by embedding the proposed neighbor comparison learning into the hypersphere space so as to achieve the purpose of surface anomaly detection.
Description
Technical Field
The invention relates to an anomaly detection method, in particular to an anomaly detection method for improving hypersphere space learning based on neighbor comparison, and belongs to the technical field of comparison learning, anomaly detection and image processing.
Background
The detection of product surface anomalies in an industrial setting is crucial to the development of industrial intelligence. Surface defect detection is a problem in locating abnormal areas in images, such as scratches and smudges. However, in practical applications, because the probability of abnormal samples is low and the abnormal forms are various, there are often only a large number of normal samples in the training process. Therefore, a challenge of the task of detecting surface anomalies is how to clearly separate the normal mode from the anomalous mode without training of the anomalous samples.
Currently, in order to solve the pain problem, some researchers propose to artificially create synthetic anomalies in a training phase, so as to train a segmentation network to explicitly segment normal samples and abnormal samples to learn the difference between the normal mode and the abnormal mode. Although this idea is good, often the split network is overly suitable for synthetic anomalies and cannot be generalized to real data. Meanwhile, there are also studies on methods for reconstructing only normal images using an auto-encoder in the training stage, which rely on the assumption that an auto-encoder trained on normal samples can only successfully reconstruct normal regions, but fails for abnormal regions. However, the self-encoder, depending on its powerful generalization capability, may excessively generalize to abnormal samples, and may well reconstruct even the abnormal samples that are not seen, thereby resulting in failure of the abnormality detection. There is a scholarly on this problem to increase the difference between abnormal samples compared to normal samples by introducing normal feature storage blocks, but its performance tends to increase proportionally with the size of the storage block. Therefore, how to make the network sufficiently learn the clustering pattern of the normal samples is the key to solve the pain point problem.
Disclosure of Invention
The invention aims to overcome the problem that the normality of an abnormal mode is overestimated due to insufficient knowledge of the normal mode in a surface abnormal detection algorithm in the prior art, and provides an abnormal detection method for improving hypersphere space learning based on neighbor comparison.
The technical scheme adopted by the invention is as follows:
an anomaly detection method for improving hypersphere space learning based on neighbor comparison comprises the following steps:
s1, constructing a memory bank: inputting a training image into a feature extraction network to extract features, storing the extracted features into a memory library, inputting the next training image into the feature extraction network to extract features, sequentially searching nearest neighbors from the newly extracted features for each feature in the memory library, and then updating the features in the memory library by calculating a moving average value until the traversal of a training set is completed;
s2, calculating the hypersphere loss: inputting each image in the training set into a feature extraction network which is the same as that in the step S1 to extract features, searching 1-3 neighbors and 4-6 neighbors from a memory library by using the extracted features, respectively forcing the features to be close to the 1 st neighbors, the 2 nd neighbors and the 3 rd neighbors and to be far from the 4 th neighbors, the 5 th neighbors and the 6 th neighbors through hypersphere loss, calculating to obtain close hypersphere loss and far hypersphere loss, and adding and coupling to obtain total hypersphere loss;
s3, calculating the brand-new neighbor contrast loss: respectively forming three positive pairs of the features extracted in the step S2 and 1 st, 2 nd and 3 rd neighbors in a memory library, and forming 3 negative pairs of the features and the 4 th, 5 th and 6 th neighbors, maximizing the similarity between the positive pairs and minimizing the similarity between the negative pairs, and calculating the neighbor contrast loss;
s4, calculating total loss by combining the hypersphere loss and neighbor comparison loss, and training an updating network;
s5, calculating the abnormal score of the test sample: in the testing stage, the image is input into a trained feature extraction network, a nearest neighbor is searched from a memory bank for each extracted feature, the Euclidean distance between the two features is used as an abnormal score of the feature, a pixel-level abnormal detection score graph is generated, the maximum value in the graph is used as an image-level abnormal detection score, and the final abnormal detection result is obtained after the testing is finished.
In the anomaly detection method for improved hypersphere space learning based on neighbor comparison, the feature extraction network in the step S1 uses resnet18, preferably features extracted from the first four layers of resnet 18; the calculation formula of the moving average value is as follows:
whereinC i-1 The feature set in the memory pool representing the last state,C i representing the feature set in the updated memory base,C i NN to representC i-1 The matched nearest neighbor feature set in the newly extracted features,βrepresents a hyper-parameter that controls the moving average,Nrepresenting the number of samples in the training set.
In the method, the formula for calculating the hypersphere loss in the step S2 is as follows:
whereinL att Representing a near (irregular) hypersphere loss,Trepresents the total number of features output by the feature extraction network,Kindicating the number of neighbors to be approached for each feature match,Drepresenting a predefined distance measure, here euclidean distance,c t k the nearest neighbor to be approached, here the 1 st, 2 nd, 3 rd neighbor respectively,p t each feature extracted by the feature extraction network is represented. The above formula forces each feature top t Gradually approach toc t k The radius created for the center isrThe hypersphere of (1).
WhereinL rep Indicating a loss of distance (repel) hypersphere,Trepresents the total number of features output by the feature extractor,Jindicating the number of neighbors each feature matches to be far away,z t j indicating that each feature is matched to a feature that is far from its neighbors, here neighbors 4, 5, and 6, respectively, the above formula forces each feature top t Gradually, graduallyAway from toz t j A radius created for the center isrThe hypersphere of (1).
The two losses above are coupled, the total loss of the hypersphere (hypersphere) is as follows:
this loss allows dense clustering of the normal features extracted by the network.
The neighbor contrast loss described in step S3 in the above method is as follows:
whereinTRepresents the total number of features output by the feature extraction network,Kthe number of the opposite sides is shown,Jthe number of negative pairs is indicated and,c t k representing the nearest neighbor feature directly opposite each feature,z t j representing the nearest neighbor features that form a negative pair with each feature,sim() Representing a cosine similarity function between two vectors,
indicating a temperature parameter. This loss makes each feature extracted by the network more similar to its positive pair (neighbors 1, 2, 3) and less similar to its negative pair (neighbors 4, 5, 6), thus enabling more dense clustering of normal features.
The total loss in the combined hypersphere loss and neighbor contrast loss design in step S4 is as follows:
wherein
And
is a balance hyperparameter.
The calculation formula of each feature abnormality score in step S5 is as follows:
whereinTThe total number of features, here 8 x 8, representing the output of the feature extraction network;p t representing each feature of the test image extracted by the feature extraction network;c i representing each feature in the memory pool;mrepresenting the total number of features in the memory pool.
It is another object of the present invention to provide a storage device which is a computer readable storage device, said computer readable storage device having stored thereon a computer program for implementing the steps in the anomaly detection method for improved hypersphere spatial learning based on neighbor comparison as described above.
The invention also provides an anomaly detection device based on the improved hypersphere space learning of the neighbor comparison, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor executes the program to realize the anomaly detection method based on the improved hypersphere space learning of the neighbor comparison.
The invention has the beneficial effects that:
the invention uses the coupling hyper-sphere loss to realize the dense clustering of the normal features extracted by the network, and the close neighbor comparison loss design realizes the dense clustering of the normal features. Under the condition of only using the normal sample, the invention enables the network to more fully learn the clustering pattern of the normal sample by embedding the proposed neighbor comparison learning into the hypersphere space so as to achieve the purpose of surface anomaly detection, and overcomes the problem that the normality of the anomaly pattern is overestimated due to insufficient knowledge of the normal pattern in a surface anomaly detection algorithm.
Drawings
FIG. 1 is a schematic diagram of a model structure in a training phase of the method of the present invention;
FIG. 2 is a schematic diagram of a model structure at a test stage of the method of the present invention;
FIG. 3 is a flow chart of the method for building a memory bank according to the present invention;
FIG. 4 is a flow chart of a training phase of the method of the present invention;
FIG. 5 is a flow chart of the testing phase of the method of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to specific embodiments and the accompanying drawings.
Example 1:
an anomaly detection method for improved hypersphere space learning based on neighbor comparison comprises the following steps:
s1, constructing a memory bank:
as shown in fig. 3, the first 3 × 256 normal samples from the training set are input into the feature extraction network to extract features, and the 8 × 512 features extracted through the first four layers of resnet18 (where the parameters of the first three layers are frozen) are flattened (rearranged into 64-dimensional 512 features) and stored in the memory bank. And then, sequentially taking out other samples of the training set, inputting the samples into the feature extraction network to extract features, flattening the extracted 8 x 512 features, and obtaining 64 features with the dimension of 512. For each 512-dimensional feature in the memory bank, a nearest neighbor is sequentially searched from the newly extracted 64 512-dimensional features, and then the features in the memory bank are updated by the following moving average formula:
whereinC i-1 The feature set in the memory pool representing the last state,C i representing the feature set in the updated memory base,C i NN to representC i-1 The matched nearest neighbor feature set in the newly extracted features,βrepresents a hyper-parameter that controls the moving average,Nrepresenting the number of samples in the training set.
S2, calculating the hypersphere loss:
as shown in fig. 1 and fig. 4, after the memory library is constructed, for each image 3 × 256 in the training set, the images are sequentially input into the feature extraction network (four layers before resnet 18) to obtain features 512 × 8. And sequentially searching the 1 st, 2 nd and 3 rd neighbors and the 4 th, 5 th and 6 th neighbors from the memory library for each feature with the dimension of 512. Then the feature is forced to approach the 1 st, 2 nd and 3 rd neighbors and to move away from the 4 th, 5 th and 6 th neighbors respectively through the hypersphere loss, and the formula is as follows:
whereinL att Representing the loss of a close (irregular) hypersphere,Tthe total number of features representing the output of the feature extraction network, here 8 x 8.KThe number of neighbors to be approached, here 3, representing each feature match;Drepresents a predefined distance metric, here euclidean distance;c t k the nearest neighbor to be approached, here the 1 st, 2 nd, 3 rd neighbor, respectively, representing each feature match;p t each feature extracted by the feature extraction network is represented. Forcing each feature as in the above formulap t Gradually approach toc t k The radius created for the center isrThe hypersphere of (1).
WhereinL rep Indicating a loss of distance (repel) hypersphere,Jrepresents the number of neighbors to be far away that each feature matches, here 3;z t j indicating that each feature is matched to a feature that is far away from the neighbor, here the 4 th, 5 th, and 6 th neighbors, respectively; the above formula forces each bitSign forp t Gradually move away fromz t j The radius created for the center isrThe hypersphere of (1).
Combining the above two losses allows each feature to be madep t Gradually approach toc t k The radius created for the center isrAnd gradually move away fromz t j The radius created for the center isrThe hypersphere of (1). Thus, the total loss of hypersphere (hypersphere) after coupling is as follows:
this loss allows dense clustering of the normal features extracted by the network.
S3, calculating the brand-new neighbor contrast loss:
and for each feature with dimension 512 extracted in S2, sequentially finding out the 1 st, 2 nd and 3 rd neighbors and the 4 th, 5 th and 6 th neighbors from the memory library. In addition to training using the post-coupling hypersphere loss described in S2, we also make this feature form three positive pairs with neighbors 1, 2, and 3 negative pairs with neighbors 4, 5, and 6, respectively, in order to maximize the similarity between positive pairs and minimize the similarity between negative pairs, using the following neighbor contrast loss:
whereinTRepresents the total number of features, here 8 x 8, output by the feature extractor;Krepresents the number of direct alignments, here 3;Jrepresents the number of negative pairs, also here 3;c t k representing the nearest neighbor feature directly opposite to each feature, here nearest 1, 2, 3;z t j the nearest neighbor features, here nearest 4, 5, 6, that form a negative pair with each feature;sim() Representing a cosine similarity function between two vectors,
indicating a temperature parameter. This loss makes each feature extracted by the network more similar to its positive pair (neighbors 1, 2, 3) and less similar to its negative pair (neighbors 4, 5, 6), thus enabling more dense clustering of normal features.
And S4, calculating total loss by combining the hypersphere loss and the neighbor comparison loss, training an updating network:
for the two loss functions obtained in S2 and S3, we combine them, and the total loss for the design is as follows:
wherein
And
is a balance over parameter and eventually updates the network with the total loss.
S5, calculating the abnormal score of the test sample in the test stage:
as shown in fig. 2 and 5, each 3 × 256 test image is input into the trained feature extraction network (the first four layers of resnet 18), features of 512 × 8 are extracted, then, for each feature with dimension 512, a nearest neighbor is sequentially searched from the memory library, and the euclidean distance between the two is used as the abnormal score of the feature. The calculation formula of each feature anomaly score is as follows:
whereinTRepresents the total number of features output by the feature extractor, here 8 x 8;p t representing each feature of the test image extracted by the feature extraction network;c i presentation memory libraryEach feature of (a);mrepresenting the total number of features in the memory bank, here 64.
Therefore, we obtain an 8 × 8 pixel level anomaly detection score map, which represents the pixel level anomaly detection result. We then upsample to the input image size, 256 × 256, and finally take the maximum value in the upsampled pixel level anomaly detection score map as the image level anomaly detection score. And (5) obtaining a final abnormal detection result after the test is finished.
Example 2:
a storage device that is a computer readable storage device having stored thereon a computer program for implementing the steps in the anomaly detection method for improved hypersphere space learning based on neighbor comparison as described in embodiment 1 above.
An anomaly detection device for improved hypersphere space learning based on neighbor comparison comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor executes the program to realize the anomaly detection method for improved hypersphere space learning based on neighbor comparison as described in embodiment 1 above.
The above description is only exemplary of the present invention and should not be taken as limiting the invention, as any modification, equivalent replacement, or improvement made within the spirit and scope of the present invention should be included in the present invention.
Claims (10)
1. An anomaly detection method for improved hypersphere space learning based on neighbor comparison is characterized by comprising the following steps:
s1, constructing a memory bank: inputting a training image into a feature extraction network to extract features, storing the extracted features into a memory library, inputting the next training image into the feature extraction network to extract features, sequentially searching nearest neighbors from the newly extracted features for each feature in the memory library, and then updating the features in the memory library by calculating a moving average value until the traversal of a training set is completed;
s2, calculating the hypersphere loss: inputting each image in the training set into a feature extraction network which is the same as that in the step S1 to extract features, searching 1-3 neighbors and 4-6 neighbors from a memory library by using the extracted features, respectively forcing the features to be close to the 1 st neighbors, the 2 nd neighbors and the 3 rd neighbors and to be far from the 4 th neighbors, the 5 th neighbors and the 6 th neighbors through hypersphere loss, calculating to obtain close hypersphere loss and far hypersphere loss, and adding and coupling to obtain total hypersphere loss;
s3, calculating the brand-new neighbor contrast loss: respectively forming three positive pairs of the features extracted in the step S2 and 1 st, 2 nd and 3 rd neighbors in a memory library, and forming 3 negative pairs of the features and the 4 th, 5 th and 6 th neighbors, maximizing the similarity between the positive pairs and minimizing the similarity between the negative pairs, and calculating the neighbor contrast loss;
s4, calculating total loss by combining the hypersphere loss and the neighbor comparison loss, and training an updating network;
s5, calculating the abnormal score of the test sample: in the testing stage, the image is input into a trained feature extraction network, a nearest neighbor is searched from a memory bank for each extracted feature, the Euclidean distance between the two features is used as an abnormal score of the feature, a pixel-level abnormal detection score graph is generated, the maximum value in the graph is used as an image-level abnormal detection score, and the final abnormal detection result is obtained after the testing is finished.
2. The anomaly detection method for improved hypersphere space learning based on neighbor comparison as claimed in claim 1, wherein said feature extraction network of S1 uses resnet18, and said moving average is calculated by the formula:
whereinC i-1 The feature set in the memory pool representing the last state,C i representing the feature set in the updated memory base,C i NN to representC i-1 The matched nearest neighbor feature set in the newly extracted features,βrepresents a hyper-parameter that controls the moving average,Nrepresenting the number of samples in the training set.
3. The method for detecting the anomaly of the improved hypersphere space learning based on the neighbor comparison as claimed in claim 1, wherein the formula for calculating the loss of the approaching hypersphere in S2 is as follows:
whereinL att Indicating that there is a loss of near super-spherical,Trepresents the total number of features output by the feature extraction network,Kindicating the number of neighbors to be approached for each feature match,Drepresenting a predefined distance measure, here euclidean distance,c t k the nearest neighbor to be approached, here the 1 st, 2 nd, 3 rd neighbor respectively,p t representing each feature extracted by the feature extraction network, forcing each feature to be as per the formula abovep t Gradually approach toc t k The radius created for the center isrThe hypersphere of (1).
4. The anomaly detection method for improved hypersphere space learning based on neighbor comparison as claimed in claim 1, wherein the calculation formula of far hypersphere loss in S2 is:
whereinL rep Indicating a loss away from the hypersphere,Trepresents the total number of features output by the feature extractor,Jindicating the number of neighbors each feature matches to be far away,z t j indicating that each feature is matched to a feature that is to be far from a neighbor, here a 4 th, 5 th, 6 th neighbor, respectively, as disclosed aboveEach feature is forcedp t Gradually move away fromz t j The radius created for the center isrThe hypersphere of (1).
5. The method for detecting the abnormal situation of the improved hypersphere space learning based on the neighbor comparison as claimed in claim 1, wherein the total hypersphere loss in S2 is calculated as follows:
6. the method for detecting the anomaly of the improved hypersphere spatial learning based on the neighbor comparison as claimed in claim 1, wherein the loss of the neighbor comparison in S3 is as follows:
where the total number of features output by the feature extraction network is represented,Kthe number of the opposite sides is shown,Jthe number of negative pairs is indicated,c t k representing the nearest neighbor feature directly opposite each feature,z t j representing the nearest neighbor features that form a negative pair with each feature,sim() Representing a cosine similarity function between two vectors,
indicating a temperature parameter.
7. The method for detecting the anomaly of the improved hypersphere spatial learning based on the neighbor comparison as claimed in claim 1, wherein the total loss of the hypersphere loss and neighbor comparison loss design in S4 is as follows:
wherein
And
is a balance hyperparameter.
8. The method for detecting the anomaly of the improved hypersphere space learning based on the neighbor comparison as claimed in claim 1, wherein the calculation formula of each feature anomaly score in the step S5 is as follows:
whereinTRepresenting the total number of features output by the feature extraction network;p t representing each feature of the test image extracted by the feature extraction network;c i representing each feature in the memory pool;mrepresenting the total number of features in the memory pool.
9. A storage device being a computer readable storage device, wherein said computer readable storage device has stored thereon a computer program for implementing the steps of the method for detecting anomalies based on improved hyperspectral space learning by neighbor comparison as claimed in any of claims 1 to 8.
10. An anomaly detection device for improved hypersphere space learning based on neighbor comparison, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the anomaly detection method for improved hypersphere space learning based on neighbor comparison as claimed in any one of claims 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310053050.5A CN115830006B (en) | 2023-02-03 | 2023-02-03 | Abnormality detection method for improving hypersphere space learning based on neighbor contrast |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310053050.5A CN115830006B (en) | 2023-02-03 | 2023-02-03 | Abnormality detection method for improving hypersphere space learning based on neighbor contrast |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115830006A true CN115830006A (en) | 2023-03-21 |
| CN115830006B CN115830006B (en) | 2023-05-30 |
Family
ID=85520736
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310053050.5A Active CN115830006B (en) | 2023-02-03 | 2023-02-03 | Abnormality detection method for improving hypersphere space learning based on neighbor contrast |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115830006B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117765363A (en) * | 2024-02-22 | 2024-03-26 | 山东省计算中心（国家超级计算济南中心） | Image anomaly detection method and system based on lightweight memory bank |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113994338A (en) * | 2019-07-30 | 2022-01-28 | 国立研究开发法人科学技术振兴机构 | Word representation device, word representation method, program, structure shape learning method, and structure design method for finite flow chart |
| CN114724043A (en) * | 2022-06-08 | 2022-07-08 | 南京理工大学 | Self-encoder anomaly detection method based on contrast learning |
| US20220326118A1 (en) * | 2021-04-09 | 2022-10-13 | Wolfgang Richter | System and a method to monitor moving parts in machines |
| CN115190999A (en) * | 2020-06-05 | 2022-10-14 | 谷歌有限责任公司 | Classifying data outside of a distribution using contrast loss |
| CN115205592A (en) * | 2022-07-15 | 2022-10-18 | 东北大学 | Multi-mode data based rebalance long-tail image data classification method |
| CN115424621A (en) * | 2022-10-31 | 2022-12-02 | 北京科技大学 | Training method of voiceprint recognition model, voiceprint recognition method, device and equipment |
| CN115526234A (en) * | 2022-08-30 | 2022-12-27 | 长沙理工大学 | Cross-domain model training and log anomaly detection method and device based on transfer learning |
-
2023
- 2023-02-03 CN CN202310053050.5A patent/CN115830006B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113994338A (en) * | 2019-07-30 | 2022-01-28 | 国立研究开发法人科学技术振兴机构 | Word representation device, word representation method, program, structure shape learning method, and structure design method for finite flow chart |
| CN115190999A (en) * | 2020-06-05 | 2022-10-14 | 谷歌有限责任公司 | Classifying data outside of a distribution using contrast loss |
| US20220326118A1 (en) * | 2021-04-09 | 2022-10-13 | Wolfgang Richter | System and a method to monitor moving parts in machines |
| CN114724043A (en) * | 2022-06-08 | 2022-07-08 | 南京理工大学 | Self-encoder anomaly detection method based on contrast learning |
| CN115205592A (en) * | 2022-07-15 | 2022-10-18 | 东北大学 | Multi-mode data based rebalance long-tail image data classification method |
| CN115526234A (en) * | 2022-08-30 | 2022-12-27 | 长沙理工大学 | Cross-domain model training and log anomaly detection method and device based on transfer learning |
| CN115424621A (en) * | 2022-10-31 | 2022-12-02 | 北京科技大学 | Training method of voiceprint recognition model, voiceprint recognition method, device and equipment |
Non-Patent Citations (4)
| Title |
|---|
| GUO JUN ZHANG: "a hypersphere method for plant leaves classification" * |
| 叶剑华: "基于变分模态分解和多模型融合的用户级综合能源系统超短期负荷预测" * |
| 孙启玉: "河长制信息化关键技术研究及平台应用" * |
| 范洪光: "超球冠上的模式分类及图像识别技术研究" * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117765363A (en) * | 2024-02-22 | 2024-03-26 | 山东省计算中心（国家超级计算济南中心） | Image anomaly detection method and system based on lightweight memory bank |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115830006B (en) | 2023-05-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108230278B (en) | Image raindrop removing method based on generation countermeasure network | |
| CN111476249B (en) | Construction method of multi-scale large-receptive-field convolutional neural network | |
| CN111612051A (en) | Weak supervision target detection method based on graph convolution neural network | |
| CN111125397B (en) | Cloth image retrieval method based on convolutional neural network | |
| CN115830006A (en) | Anomaly detection method for improving hypersphere space learning based on neighbor comparison | |
| CN112766170B (en) | Self-adaptive segmentation detection method and device based on cluster unmanned aerial vehicle image | |
| CN110674326A (en) | Neural network structure retrieval method based on polynomial distribution learning | |
| CN113807340B (en) | Attention mechanism-based irregular natural scene text recognition method | |
| CN115035418A (en) | Remote sensing image semantic segmentation method and system based on improved deep LabV3+ network | |
| CN116089883B (en) | Training method for improving classification degree of new and old categories in existing category increment learning | |
| CN116416497A (en) | Bearing fault diagnosis system and method | |
| CN111104924A (en) | Processing algorithm for effectively identifying low-resolution commodity image | |
| CN114998703A (en) | Remote sensing image change detection method based on high-resolution convolutional neural network | |
| Cui et al. | Autocorrelation Aware Aggregation Network for Salient Object Detection of Strip Steel Surface Defects | |
| CN112949748A (en) | Dynamic network anomaly detection algorithm model based on graph neural network | |
| CN111241326B (en) | Image visual relationship indication positioning method based on attention pyramid graph network | |
| CN115861306B (en) | Industrial product abnormality detection method based on self-supervision jigsaw module | |
| CN116778164A (en) | Semantic segmentation method for improving deep V < 3+ > network based on multi-scale structure | |
| CN116227586A (en) | Meta learning fault diagnosis method and device based on depth residual error shrinkage prototype network | |
| CN115641449A (en) | Target tracking method for robot vision | |
| CN111354019B (en) | Visual tracking failure detection system based on neural network and training method thereof | |
| CN113222016A (en) | Change detection method and device based on cross enhancement of high-level and low-level features | |
| CN109840552B (en) | Dynamic image classification method | |
| CN113807366A (en) | Point cloud key point extraction method based on deep learning | |
| CN112288629A (en) | Super-resolution method of cross-scale graph neural network |
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 |