CN115830006B - Abnormality detection method for improving hypersphere space learning based on neighbor contrast - Google Patents
Abnormality detection method for improving hypersphere space learning based on neighbor contrast Download PDFInfo
- Publication number
- CN115830006B CN115830006B CN202310053050.5A CN202310053050A CN115830006B CN 115830006 B CN115830006 B CN 115830006B CN 202310053050 A CN202310053050 A CN 202310053050A CN 115830006 B CN115830006 B CN 115830006B
- Authority
- CN
- China
- Prior art keywords
- feature
- loss
- neighbor
- representing
- neighbors
- 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
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 abnormality detection method for improving hypersphere space learning based on neighbor contrast, and belongs to the technical field of abnormality detection. The method comprises the following steps: s1, constructing a memory bank; s2, calculating the superball loss: searching 1-3 neighbors and 4-6 neighbors from a memory library by using the features extracted by the training images, calculating to obtain the loss close to the hypersphere and the loss far from the hypersphere, and obtaining the total hypersphere loss after adding and coupling; s3, calculating a brand new neighbor contrast loss; s4, calculating total loss by combining the hyper-sphere loss and the neighbor contrast loss, and training and updating the network; s5, calculating the abnormal score of the test sample, and obtaining a final abnormal detection result after the test is completed. Under the condition that only normal samples are used, the network can learn the clustering mode of the normal samples more fully by embedding the proposed neighbor contrast learning into the hypersphere space, so that the purpose of surface anomaly detection is achieved.
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 contrast, and belongs to the technical fields of contrast learning, anomaly detection and image processing.
Background
Detection of surface anomalies in products in industrial settings is critical to the development of industrial intelligence. Surface defect detection is a problem in locating abnormal areas in images, such as scratches and stains. However, in practical application, due to the low probability of abnormal samples and various abnormal forms, only a large number of normal samples are often used in the training process. Thus, a challenge in the task of surface anomaly detection is how to separate the normal and anomaly modes explicitly without training of the anomaly samples.
Currently, aiming at the pain point problem, a learner proposes to manually create a synthetic anomaly in a training stage, so as to train a segmentation network to explicitly segment a normal sample and an abnormal sample to learn the difference between the normal mode and the abnormal mode. This idea is good, but often the split network is overly suitable for synthetic anomalies and cannot generalize to real data. At the same time, there are also scholars who propose methods for reconstructing only normal images using a self-encoder in the training phase, which rely on the assumption that a self-encoder trained on normal samples can only successfully reconstruct normal regions, while failing for abnormal regions. However, the self-encoder, by virtue of its strong generalization capability, can be excessively generalized to abnormal samples, and can perform good reconstruction of even non-seen abnormal samples, resulting in failure of abnormality detection. For this problem, the learner increases the difference between the abnormal samples compared to the normal samples by introducing the normal feature memory module, but the performance tends to increase proportionally with the size of the memory module. Therefore, how to make the network learn the clustering pattern of normal samples sufficiently is the key to solve the pain point problem.
Disclosure of Invention
The invention aims to solve the problem that the normal property of an abnormal mode is overestimated due to insufficient knowledge of the normal mode in the surface abnormality detection algorithm in the prior art, and provides an abnormality 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 contrast 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 bank, inputting the next training image into the feature extraction network to extract features, sequentially searching nearest neighbors from newly extracted features for each feature in the memory bank, and then updating the features in the memory bank by calculating a moving average value until the training set is traversed;
s2, calculating the superball 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, forcing the features to be close to the 1 st, 2 nd and 3 rd neighbors and far from the 4 th, 5 th and 6 th neighbors respectively through the hyper-sphere loss, calculating to obtain a near hyper-sphere loss and a far hyper-sphere loss, and adding and coupling to obtain a total hyper-sphere loss;
s3, calculating brand new neighbor contrast loss: the features extracted in the step S2 are respectively formed into three opposite pairs with the 1 st, 2 nd and 3 rd neighbors in the memory library, and formed into 3 negative pairs with the 4 th, 5 th and 6 th neighbors, the similarity between the opposite pairs is maximized, the similarity between the negative pairs is minimized, and the neighbor contrast loss is calculated;
s4, calculating total loss by combining the hyper-sphere loss and the neighbor contrast loss, and training and updating the network;
s5, calculating abnormal scores of the test samples: in the test stage, inputting the images into a trained feature extraction network, searching a nearest neighbor from a memory library for each extracted feature, taking the Euclidean distance between the two features as the anomaly score of the feature, generating a pixel-level anomaly detection score map, taking the maximum value in the map as the image-level anomaly detection score, and completing the test to obtain a final anomaly detection result.
In the anomaly detection method for improving hypersphere space learning based on neighbor contrast, the feature extraction network in the step S1 uses the resnet18, preferably features extracted from four layers before the resnet 18; the calculation formula of the moving average value is as follows:
wherein the method comprises the steps ofC i-1 A feature set in the memory bank representing the last state,C i representing the feature set in the updated memory bank,C i NN representation ofC i-1 The nearest neighbor feature set that matches among the newly extracted features,βindicating the super-parameters controlling the moving average,Nrepresenting the number of samples in the training set.
The calculation formula of the superball loss in the step S2 in the method is as follows:
wherein the method comprises the steps ofL att Representing a near (approach) hyper-sphere loss,Trepresenting the total number of features output by the feature extraction network,Krepresenting the number of neighbors to be approximated for each feature match,Drepresenting a predefined distance measure, here the euclidean distance,c t k representing the nearest neighbors to be approached for each feature match, here the 1 st, 2 nd, 3 rd neighbors respectively,p t representing each feature extracted by the feature extraction network. The above formula forces each featurep t Gradually approach toc t k Radius created for center isrIs a super ball of (2).
Wherein the method comprises the steps ofL rep Indicating a far (repeat) superball loss,Trepresenting the total number of features output by the feature extractor,Jrepresenting the number of neighbors to be moved away to which each feature matches,z t j representing that each feature matches a feature that is to be moved away from a neighboring feature, here the 4, 5, 6 neighbors, respectively, each feature is forced by the above formulap t Gradually get away fromz t j Radius created for center isrIs a super ball of (2).
The two losses above are coupled, and the total loss of the hyper sphere (hypersphere) is as follows:
this loss enables dense clustering of the normal features extracted by the network.
The neighbor contrast loss in step S3 in the above method is as follows:
wherein the method comprises the steps ofTRepresenting the total number of features output by the feature extraction network,Kthe number of the opposite faces is indicated,Jthe number of negative pairs is indicated as the number of pairs,c t k representation and eachThe features constitute a positive neighbor feature,z t j the representation and each feature constitute a neighbor feature of the negative pair,sim() Representing a cosine similarity function between the two vectors,representing a temperature parameter. This loss makes each feature extracted by the network more similar between its positive pair (1, 2, 3 neighbors) and less similar between the negative pair (4, 5, 6 neighbors), thus enabling a more dense clustering of normal features.
The total loss of the design combining the hyper-sphere loss and the neighbor contrast loss in step S4 is as follows:
The calculation formula of each feature anomaly score in step S5 is as follows:
wherein the method comprises the steps ofTRepresenting the total number of features output by the feature extraction network, here 8 x 8;p t representing each feature of the test image extracted by the feature extraction network;c i representing each feature in the memory bank;mrepresenting the total number of features in the memory bank.
It is a further object of the present invention to provide a storage device being a computer readable storage device having stored thereon a computer program for implementing the steps in the improved hyper-sphere space learning anomaly detection method based on neighbor contrast as described above.
The invention also provides an abnormality detection device for improving the hypersphere space learning based on the neighbor contrast, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor realizes the abnormality detection method for improving the hypersphere space learning based on the neighbor contrast when executing the program.
The beneficial effects of the invention are as follows:
the invention uses the coupling hyper-sphere loss to enable the normal characteristics extracted by the network to realize dense clustering, and the neighbor contrast loss design to realize the dense clustering of the normal characteristics. Under the condition that only normal samples are used, the network can learn the clustering mode of the normal samples more fully by embedding the proposed neighbor contrast learning into the hypersphere space so as to achieve the purpose of surface anomaly detection, and the problem that the normal of the abnormal mode is overestimated due to insufficient knowledge of the normal mode in the surface anomaly detection algorithm is solved.
Drawings
FIG. 1 is a schematic diagram of a training phase model structure of the method of the present invention;
FIG. 2 is a schematic diagram of a test phase model structure of the method of the present invention;
FIG. 3 is a flowchart of the method of the present invention for creating a memory bank;
FIG. 4 is a flow chart of the 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
The present invention will be further described in detail below with reference to specific embodiments and with reference to the accompanying drawings, in order to make the objects, technical solutions and advantages of the present invention more apparent.
Example 1:
an anomaly detection method for improving hypersphere space learning based on neighbor contrast comprises the following steps:
s1, constructing a memory bank:
as shown in fig. 3, the first 3×256×256 normal samples from the training set are input to the feature extraction network to extract features, and the features of 8×8×512 extracted by the first four layers of the resnet18 (in which the parameters of the first three layers are frozen) are flattened (rearranged into the features of 64 dimensions 512) and stored in the memory. And then sequentially taking out other samples of the training set, inputting the samples into a feature extraction network to extract features, and flattening the extracted features of 8 x 512 to obtain 64 features with the dimension of 512. For each feature of 512 dimensions in the memory bank, one nearest neighbor is found out from the newly extracted 64 features of 512 dimensions in turn, and then the features in the memory bank are updated by the following moving average formula:
wherein the method comprises the steps ofC i-1 A feature set in the memory bank representing the last state,C i representing the feature set in the updated memory bank,C i NN representation ofC i-1 The nearest neighbor feature set that matches among the newly extracted features,βindicating the super-parameters controlling the moving average,Nrepresenting the number of samples in the training set.
S2, calculating the superball loss:
as shown in fig. 1 and fig. 4, after the memory bank is constructed, for each image 3×256×256 in the training set, a feature extraction network (four layers before the resnet 18) is sequentially input to obtain 512×8×8 features. The 1 st, 2 nd, 3 rd neighbors and the 4 th, 5 th, 6 th neighbors are found from the memory pool in turn for each feature of dimension 512. Then the characteristic is respectively forced to approach the 1 st, 2 nd and 3 rd neighbors and to be far away from the 4 th, 5 th and 6 th neighbors through the hyper-sphere loss, and the formula is as follows:
wherein the method comprises the steps ofL att Representing a near (approach) hyper-sphere loss,Tthe total number of features representing the output of the feature extraction network is here 8 x 8.KThe number of neighbors to be approximated, here 3, representing each feature match;Drepresenting a predefined distance measure, here the euclidean distance;c t k nearest neighbors to be approached, here 1 st, 2 nd, 3 rd neighbors, respectively, representing each feature match;p t representing each feature extracted by the feature extraction network. The above formula forces each featurep t Gradually approach toc t k Radius created for center isrIs a super ball of (2).
Wherein the method comprises the steps ofL rep Indicating a far (repeat) superball loss,Jrepresenting the number of neighbors to be moved away, here 3, that each feature matches;z t j indicating that each feature matches a feature that is to be far from a neighbor, here the 4 th, 5 th, 6 th neighbors, respectively; the above formula forces each featurep t Gradually get away fromz t j Radius created for center isrIs a super ball of (2).
Combining the above two losses can then make each featurep t Gradually approach toc t k Radius created for center isrIs far away from the ballz t j Radius created for center isrIs a super ball of (2). Thus, the total loss of post-coupling hyper sphere (hypersphere) is as follows:
this loss enables dense clustering of the normal features extracted by the network.
S3, calculating brand new neighbor contrast loss:
and (3) sequentially finding the 1 st, 2 nd and 3 rd neighbors and the 4 th, 5 th and 6 th neighbors from the memory bank for each feature with the dimension of 512 extracted in the step (S2). In addition to training using post-coupling hyper-sphere loss as described in S2, we have this feature formed three pairs with 1, 2, 3 neighbors, and 3 negative pairs with 4, 5, 6 neighbors, respectively, aimed at maximizing similarity between pairs, minimizing similarity between negative pairs, using the following neighbor contrast loss:
wherein the method comprises the steps ofTRepresenting the total number of features output by the feature extractor, here 8 x 8;Krepresents the number of facing pairs, here 3;Jrepresents the number of negative pairs, here also 3;c t k representing nearest neighbor features which are opposite to each feature, namely 1 st, 2 nd and 3 rd nearest neighbors;z t j representing the nearest neighbor features that form a negative pair with each feature, here the 4 th, 5 th, 6 th nearest neighbor;sim() Representing a cosine similarity function between the two vectors,representing a temperature parameter. This loss makes each feature extracted by the network more similar between its positive pair (1, 2, 3 neighbors) and less similar between the negative pair (4, 5, 6 neighbors), thus enabling a more dense clustering of normal features.
S4, calculating total loss by combining the hyper-sphere loss and the neighbor contrast loss, and training an updating network:
for the two loss functions obtained in S2 and S3, we combine them, the total loss of the design is as follows:
wherein the method comprises the steps ofAnd->Is a balanced superparameter, and eventually the network is updated with total losses.
S5, calculating abnormal scores of the test samples in the test stage:
as shown in fig. 2 and fig. 5, each 3×256 test image is input into a trained feature extraction network (four layers before the resnet 18), 512×8×8 features are extracted, then a nearest neighbor is sequentially searched from a memory library for each feature with dimension of 512, and the euclidean distance between the two features is regarded as an anomaly score of the feature. The calculation formula of each feature anomaly score is as follows:
wherein the method comprises the steps ofTRepresenting 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 representing each feature in the memory bank;mrepresenting the total number of features in the memory bank, here 64.
Thus, we obtain an 8 x 8 pixel level anomaly detection score plot, which represents the pixel level anomaly detection result. We then upsample to the input image size 256 by 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) finishing the test to obtain a final abnormality detection result.
Example 2:
a storage device which is a computer readable storage device having stored thereon a computer program for implementing the steps in the method of anomaly detection for improved hypersphere space learning based on neighbor contrast as described in embodiment 1 above.
An abnormality detection apparatus for improved hypersphere space learning based on neighbor contrast, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, which when executed implements the abnormality detection method for improved hypersphere space learning based on neighbor contrast as described in embodiment 1 above.
The above description is only of the preferred embodiments of the present invention and is not intended to limit the invention, but any modifications, equivalents, and improvements made within the spirit and principle of the present invention should be included in the scope of the present invention.
Claims (6)
1. An anomaly detection method for improving hypersphere space learning based on neighbor contrast 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 bank, inputting the next training image into the feature extraction network to extract features, sequentially searching nearest neighbors from newly extracted features for each feature in the memory bank, and then updating the features in the memory bank by calculating a moving average value until the training set is traversed;
s2, calculating the superball loss: inputting each image in the training set into a feature extraction network which is the same as that in the step S1, extracting features, searching 1-3 neighbors and 4-6 neighbors from a memory bank by using the extracted features, searching 1 st, 2 nd, 3 rd neighbors and 4 th, 5 th and 6 th neighbors from the memory bank in sequence for each feature, and then forcing the feature to be close to the 1 st, 2 nd and 3 rd neighbors and far from the 4 th, 5 th and 6 th neighbors respectively through hyper-sphere loss, wherein the formula is as follows:
wherein the method comprises the steps ofL att Indicating that the loss of the ball is approaching,Trepresenting the total number of features output by the feature extraction network,Kthe number of neighbors to be approximated, here 3, representing each feature match;Drepresenting a predefined distance measure, here the Euclidean distanceSeparating;
c t k nearest neighbors to be approached, here 1 st, 2 nd, 3 rd neighbors, respectively, representing each feature match;p t representing each feature extracted by the feature extraction network, each feature is forced by the above formulap t Gradually approach toc t k Radius created for center isrIs a super ball of (2);
wherein the method comprises the steps ofL rep Indicating a loss away from the oversphere,Jrepresenting the number of neighbors to be moved away, here 3, that each feature matches;z t j indicating that each feature matches a feature that is to be far from a neighbor, here the 4 th, 5 th, 6 th neighbors, respectively; the above formula forces each featurep t Gradually get away fromz t j Radius created for center isrIs a super ball of (2);
combining the above two losses can then make each featurep t Gradually approach toc t k Radius created for center isrIs far away from the ballz t j Radius created for center isrThus, the total loss of the nanospheres after coupling is as follows:
wherein the method comprises the steps ofL att Indicating that the loss of the ball is approaching,L rep indicating loss away from the superball;L hs indicating total loss of superball , The loss enables the normal features extracted by the network to realize dense clustering;
s3, calculating brand new neighbor contrast loss: and (3) respectively forming three opposite pairs with the 1 st, 2 nd and 3 rd neighbors and forming 3 negative pairs with the 4 th, 5 th and 6 th neighbors in the memory library, maximizing the similarity between the opposite pairs, minimizing the similarity between the negative pairs, and calculating neighbor contrast loss, wherein the neighbor contrast loss is as follows:
wherein the method comprises the steps ofTRepresenting the total number of features output by the feature extractor;Krepresents the number of facing pairs, here 3;Jrepresents the number of negative pairs, here also 3;c t k representing nearest neighbor features which are opposite to each feature, namely 1 st, 2 nd and 3 rd nearest neighbors;z t j representing the nearest neighbor features that form a negative pair with each feature, here the 4 th, 5 th, 6 th nearest neighbor;sim() Representing a cosine similarity function between the two vectors,representing temperature parameters, the loss makes each feature extracted by the network more similar to the positive pair and less similar to the negative pair, so that dense clustering of normal features is realized;
s4, calculating total loss by combining the hyper-sphere loss and the neighbor contrast loss, and training and updating the network;
s5, calculating abnormal scores of the test samples: in the test stage, inputting the images into a trained feature extraction network, searching a nearest neighbor from a memory library for each extracted feature, taking the Euclidean distance between the two features as the anomaly score of the feature, generating a pixel-level anomaly detection score map, taking the maximum value in the map as the image-level anomaly detection score, and completing the test to obtain a final anomaly detection result.
2. The anomaly detection method for improving hypersphere space learning based on neighbor contrast according to claim 1, wherein the feature extraction network of S1 uses a resnet18, and the calculation formula of the moving average value is:
wherein the method comprises the steps ofC i-1 A feature set in the memory bank representing the last state,C i representing the feature set in the updated memory bank,C i NN representation ofC i-1 The nearest neighbor feature set that matches among the newly extracted features,βindicating the super-parameters controlling the moving average,Nrepresenting the number of samples in the training set.
3. The anomaly detection method for improving hypersphere space learning based on neighbor contrast according to claim 1, wherein the total loss of the design combining the hypersphere loss and the neighbor contrast loss in S4 is as follows:
4. The anomaly detection method for improving hypersphere space learning based on neighbor contrast according to claim 1, wherein the calculation formula of each characteristic anomaly score in step S5 is as follows:
wherein the method comprises the steps ofTRepresenting 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 bank;mrepresenting the total number of features in the memory bank.
5. A storage device, which is a computer readable storage device, wherein a computer program is stored on the computer readable storage device for implementing the steps in the method for anomaly detection based on improved hyper-sphere space learning of neighbor contrast according to any one of claims 1-4.
6. An abnormality detection apparatus for improved hypersphere space learning based on neighbor contrast comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the abnormality detection method for improved hypersphere space learning based on neighbor contrast as claimed in any one of claims 1 to 4 when executing the program.
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 CN115830006A (en) | 2023-03-21 |
CN115830006B true 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) |
Families Citing this family (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 (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114724043A (en) * | 2022-06-08 | 2022-07-08 | 南京理工大学 | Self-encoder anomaly detection method based on contrast learning |
Family Cites Families (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP4006918A4 (en) * | 2019-07-30 | 2023-07-26 | Japan Science and Technology Agency | Finite flow pattern word representation device, word representation method, and program, structure shape learning method, and structure designing method |
WO2021247962A1 (en) * | 2020-06-05 | 2021-12-09 | Google Llc | Classifying out-of-distribution data using a contrastive loss |
US20220326118A1 (en) * | 2021-04-09 | 2022-10-13 | Wolfgang Richter | System and a method to monitor moving parts in machines |
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 |
CN115424621B (en) * | 2022-10-31 | 2023-04-07 | 北京科技大学 | Training method of voiceprint recognition model, voiceprint recognition method, device and equipment |
-
2023
- 2023-02-03 CN CN202310053050.5A patent/CN115830006B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114724043A (en) * | 2022-06-08 | 2022-07-08 | 南京理工大学 | Self-encoder anomaly detection method based on contrast learning |
Also Published As
Publication number | Publication date |
---|---|
CN115830006A (en) | 2023-03-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Min et al. | A two-stream mutual attention network for semi-supervised biomedical segmentation with noisy labels | |
CN111414968A (en) | Multi-mode remote sensing image matching method based on convolutional neural network characteristic diagram | |
CN115830006B (en) | Abnormality detection method for improving hypersphere space learning based on neighbor contrast | |
CN115063373A (en) | Social network image tampering positioning method based on multi-scale feature intelligent perception | |
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 | |
CN111242125A (en) | Natural scene image text detection method, storage medium and terminal device | |
CN114926892A (en) | Fundus image matching method and system based on deep learning and readable medium | |
CN117635991A (en) | Optimization method for image feature point matching based on deep learning | |
CN116645369A (en) | Anomaly detection method based on twin self-encoder and two-way information depth supervision | |
CN116778164A (en) | Semantic segmentation method for improving deep V < 3+ > network based on multi-scale structure | |
CN113222016B (en) | Change detection method and device based on cross enhancement of high-level and low-level features | |
CN116227586A (en) | Meta learning fault diagnosis method and device based on depth residual error shrinkage prototype network | |
CN112149804B (en) | Novel convolutional neural network model and application | |
Cui et al. | Autocorrelation Aware Aggregation Network for Salient Object Detection of Strip Steel Surface Defects | |
CN115880553A (en) | Multi-scale change target retrieval method based on space-time modeling | |
CN115049901A (en) | Small target detection method and device based on feature map weighted attention fusion | |
CN115131244A (en) | Single image rain removing method and system based on counterstudy | |
CN114529455A (en) | Task decoupling-based parameter image super-resolution method and system | |
Lu et al. | A Nov Al Global-Local Adversarial Network for Unsupervised Cross-Domain Road Detection | |
CN111797732A (en) | Video motion identification anti-attack method insensitive to sampling | |
Zeng et al. | Comparison between the traditional and deep learning algorithms on image matching | |
Pan et al. | Learning Hierarchical Graph Representation for Image Manipulation Detection | |
CN117095447B (en) | Cross-domain face recognition method and device, computer equipment and storage medium | |
Fan et al. | Ship target detection algorithm based on improved deep learning |
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 |