CN116434037A - Multi-mode remote sensing target robust recognition method based on double-layer optimization learning - Google Patents
Multi-mode remote sensing target robust recognition method based on double-layer optimization learning Download PDFInfo
- Publication number
- CN116434037A CN116434037A CN202310434641.7A CN202310434641A CN116434037A CN 116434037 A CN116434037 A CN 116434037A CN 202310434641 A CN202310434641 A CN 202310434641A CN 116434037 A CN116434037 A CN 116434037A
- Authority
- CN
- China
- Prior art keywords
- network
- modulation
- double
- optimization
- identification
- 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
- 238000005457 optimization Methods 0.000 title claims abstract description 55
- 238000000034 method Methods 0.000 title claims abstract description 41
- 238000012549 training Methods 0.000 claims abstract description 27
- 230000006978 adaptation Effects 0.000 claims description 32
- 230000000694 effects Effects 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 230000006870 function Effects 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 2
- 238000000926 separation method Methods 0.000 claims description 2
- 238000012545 processing Methods 0.000 abstract description 3
- 210000002569 neuron Anatomy 0.000 description 12
- 238000013527 convolutional neural network Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 238000013256 Gubra-Amylin NASH model Methods 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
Images
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
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention belongs to the technical field of computer vision image processing, and provides a multi-mode remote sensing target robust identification method based on double-layer optimization learning. The method designs a training mode guiding model based on double-layer optimization to adaptively learn knowledge to be learned, and simultaneously utilizes a modulation network to selectively activate an identification network so as to help the identification network learn how to better extract the characteristic representation of an input image. The method and the system can fully utilize the potential of the existing data to deeply mine the model, help the model get rid of multi-mode interference, solve the problem of reduced accuracy of the model on multi-mode target data, and enable the model to obtain higher robustness under limited learning resources.
Description
Technical Field
The invention relates to the technical field of computer vision image processing, in particular to a multi-mode remote sensing target robust identification method based on double-layer optimization learning.
Background
The remote sensing image target identification is remote sensing image classification, which refers to the process of identifying and classifying the ground object target according to the spectrum characteristics, the space characteristics, the time phase characteristics and the like of the ground object in the remote sensing image. The method can provide auxiliary information for other applications such as target detection, and can also provide basic geographic information as a final result for the fields such as map drawing, rescue and relief work, military reconnaissance and the like. Early remote sensing image recognition was accomplished primarily through visual interpretation, relying primarily on the subjective consciousness of the researcher. In recent years, the recognition method mainly uses machine learning, features are extracted by using a convolutional neural network, and a large number of samples are input into a training network to search for the internal rules of the samples, so that an optimal model is obtained. A common network for classification is LeNet, alexNet, VGG, googLeNet, etc. In image net large-scale visual identification competition in 2015, he et al in document Deep Residual Learning for Image Recognition proposes a ResNet network, utilizes data preprocessing and a BN layer to solve gradient disappearance and gradient explosion problems, and simultaneously proposes a residual error structure to solve degradation problems in a deep network, so that the network structure is improved to a deeper level. Subsequently, zhang et al in the article "Shuffenet: an extremely efficient convolutional neural network for mobile devices" propose Shuffenet to increase the computational efficiency of the model using depth separable convolution. In addition to the traditional classification network, ian Goodhellow et al in paper Generative Adversarial Nets use the generator and discriminator for the same task to put forward a GAN model, which helps network training with the idea of oppositional gaming. Many other networks such as DCGAN, styleGAN, bigGAN, stackGAN, cycleGAN and the like are derived later on. With the deep research, ding et al in the literature RepVGG: making VGG-style ConvNets Great Again, equivalently replaced the trained model with a simpler model, and then tested, and the practicality of the model is improved by utilizing structural heavy parameterization. Liu et al in DARTS Differentiable Architecture Search uses gradient descent to search for architecture effectively based on continuous relaxation of architecture representation, solving the scalability problem of architecture search. However, most of research efforts including the above-described methods are directed to improving the classification accuracy or operation speed of the network, but do not pay attention to the distribution difference between the training data and the test data, so that the robustness is low when the model is applied to a remote sensing image having multi-modal information.
The problem of robustness of multi-mode remote sensing target identification aims at solving the problem of reduced accuracy when a model is transferred between different modes of a remote sensing image. The superior performance of most existing algorithms in the remote sensing field relies on an idealized assumption that the source data and the target data are considered to be independently and uniformly distributed, and the real problem of distribution differences between different modalities is ignored. Therefore, when the trained model is transferred to a new data set with a different modality in practical application, the test accuracy is often greatly reduced. This is caused by the fact that the image style difference between different modes of the remote sensing image is large, and the mode difference phenomenon is ubiquitous in the field of remote sensing image processing. How to ensure that a model trained on a source data set still keeps higher performance when being transferred to a target data set of other modes is what is needed to be researched for the problem of remote sensing task robustness.
Disclosure of Invention
Aiming at the problem that the remote sensing target recognition task is poor in robustness on multi-mode data, the invention designs a double-layer optimized training method, adjusts the remote sensing target recognition network by utilizing a modulation network, guides a model to mine self potential, and learns robust feature representation. In the method, a training mode based on double-layer optimization is designed to guide a model to adaptively learn knowledge to be learned, and meanwhile, a modulation network is utilized to selectively activate an identification network, so that the identification network is helped to learn how to better extract the characteristic representation of an input image. The method and the system can fully utilize the potential of the existing data to deeply mine the model, help the model get rid of multi-mode interference, solve the problem of reduced accuracy of the model on multi-mode target data, and enable the model to obtain higher robustness under limited learning resources.
The technical scheme of the invention is as follows: a multimode remote sensing target robust recognition method based on double-layer optimization learning is realized based on two parallel networks, wherein the parallel networks comprise a modulation network and a recognition network; training the identification network by using the modulation network as a gating function, and controlling the selective activation of the identification network; the output of each layer of the modulation network is respectively connected with different positions of the identification network and is used for guiding the parameter updating process of the identification network to be carried out towards the direction of learning the image content characteristics; the modulation network and the identification network are respectively updated in different stages of parallel network optimization through parallel network double-layer optimization, so that separation training of the modulation network and the identification network is realized;
during training, two data sets with different modes are used, namely a known data set D 1 And adaptation data set D 2 The method comprises the steps of carrying out a first treatment on the surface of the By means of a known dataset D 1 The recognition network is pre-trained separately for one time to obtain the recognition network about the known data set D 1 Optimal parameter θ P Recording this parameter for later use; before starting double-layer optimization, randomly initializing an identification network and a modulation network; the formal training process adopts a double-layer optimization method and is divided into an inner-layer adaptation stage and a modulation optimization stage; each time the double-layer optimization is started, namely before the inner layer adaptation stage is started, the optimal parameter theta is adopted P Reinitializing the identification network updated by the previous double-layer optimization stage to enable the parameter theta of the identification network 0 =θ P I.e. theta P As an initial parameter of the identification network in the double-layer optimization of the present round;
after the start of the adaptation phase of the inner layer,by adapting the dataset D 2 Training the adaptation process of the recognition network to the data sets of different modes, updating the recognition network, and not updating the modulation network in the whole inner layer adaptation stage;
where θ represents a parameter identifying the network,representing parameters of the modulation network, L c For network loss of the inner layer adaptation stage, alpha is the learning rate of the inner layer adaptation stage, the subscript n corresponds to the nth modulation optimization stage cycle, and the subscript i corresponds to the ith inner layer adaptation stage cycle;
after the inner layer adaptation phase is finished, acquiring the adaptation data set D which is learned by the identification network under the current modulation scheme 2 Optimal parameter θ * The method comprises the steps of carrying out a first treatment on the surface of the Starting the modulation optimization phase by the known data set D 1 Judging the effect of recognition network learning under the current modulation scheme; will know the dataset D 1 Loss L of input double-layer network in modulation optimization stage m The loss size is used to reflect the effect of the current modulation scheme; by loss L m Updating the modulation network and thus updating the modulation scheme while maintaining the identification network unchanged;
wherein L is m For network loss in the modulation optimization stage, beta is the learning rate in the modulation optimization stage;
when the one-time double-layer optimization process is completed, the modulation network corresponds to the updated modulation scheme when entering the next double-layer optimization process, and the modulation scheme is used for better screening out the parameters which are preferentially updated in the identification network, wherein the parameters which are preferentially updated are neurons.
The modulation network comprises a plurality of convolution blocks which are connected in sequence, and a connecting module is connected among the rest convolution blocks except the initial convolution block in parallel; the identification network comprises a plurality of convolution blocks and classification heads which are connected in sequence; besides the initial convolution blocks, the front branches of all the other convolution blocks and the classification heads are transmitted to corresponding connection modules in the modulation network; each connection module is correspondingly transmitted to the cross multiplication operation among the rest convolution blocks and in front of the classification head.
For a network, different neurons have different degrees of importance for learning different knowledge. Some of the parameters of the neuron learning mainly affect the learning of the knowledge related to the content of the image, while other parameters of the neuron learning are more related to the style of the image. The invention aims to screen out neurons with lower importance for learning image content characteristics, and the neurons are preferably used for parameter updating when new knowledge needs to be learned, while neurons with higher importance for image content learning are prevented from being changed in subsequent training as much as possible. In this way, the guidance network learns new image style information without destroying the ability to learn the original content features. Through continuous training, when the network encounters sample data of different modes, the modulation network can control the recognition network to learn new mode knowledge by preferentially utilizing neurons which are less important for learning content characteristics, so that the capability of learning the content characteristics of an image which is acquired originally is not destroyed, and new mode information can be learned at the same time, and the robustness of the model is improved.
Specifically, the invention adopts two parallel networks as the modulation network and the identification network respectively, which are mutually independent and connected, wherein the output of each layer of the modulation network is respectively connected with different positions of the identification network, thereby realizing the selective activation of the identification network and guiding the parameter updating process of the identification network to be carried out towards the direction which is beneficial to learning the image content characteristics.
The invention has the beneficial effects that:
according to the invention, two parallel networks are designed to serve as a modulation network and an identification network respectively, a network training method based on double-layer optimization is provided, the modulation network is trained to selectively activate the identification network, the identification network is guided to update parameters by preferentially utilizing neurons with lower significance for learning content characteristics when facing sample data of different modes, the accuracy of the network is higher when the network is tested on multi-mode data, and the influence of multi-mode differences in a remote sensing image target identification task on results is improved.
Drawings
Fig. 1 is a schematic diagram of two parallel network architecture frameworks.
Fig. 2 is an overall training flow of a multi-modal remote sensing target robust recognition method based on double-layer optimization learning.
Detailed Description
The following describes the embodiments of the present invention further with reference to the drawings and technical schemes.
The whole training flow adopted by the invention is shown in figure 2, and mainly comprises an inner layer optimization stage for controlling the adaptation process of the identification network to the data of different modes and an outer layer optimization stage for controlling the updating of the modulation network. For convenience of description, these two optimization phases are hereinafter referred to as an inner layer adaptation phase and a modulation optimization phase, respectively. The training process requires the use of two different modalities of known data sets D 1 Adaptation data set D 2 Adapting data set D 2 An adaptation dataset belonging to a different modality than the known dataset is simulated. First, a known data set D with known sample labels is used before training begins 1 Pre-training the identification network once to obtain the identification network about the known data set D 1 Optimal parameter θ P . The theta is used before the inner layer adaptation stage begins in the subsequent training P Reinitializing the identification network, i.e. letting the identification network parameters theta 0 =θ P The modulation network is randomly initialized before the whole training is started.
When formally entering the double-layer optimization stage, the adaptive data set D is firstly utilized 2 Simulating adaptation of an identification network to data sets of different modalities, i.e. using adaptation data set D 2 The data of the identification network is updated, and the modulation network is not updated all the time in the whole inner layer adaptation stage so as to ensure that the identification network adapts to the dataSet D 2 Is performed under the same modulation scheme. Expressed by the formula:
where θ represents a parameter identifying the network,representing parameters of the modulation network, L c And α are the network loss and learning rate, respectively, of the inner layer adaptation phase.
After the inner layer adaptation phase is completed, the data set D which is learned by the recognition network under the modulation scheme at the moment can be obtained 2 Optimal parameter θ * . The modulation optimization phase is then entered, the main function of which is to use the known data set D 1 To judge whether the network learning is good or bad under the current modulation scheme. Will know the dataset D 1 Inputting the whole model to obtain the loss L of the modulation optimization stage m The magnitude of the loss reflects how well the modulation scheme is doing at the time, and the modulation network is updated with the loss while the identification network remains unchanged. The purpose of this update is to allow the modulation network to learn a better modulation scheme and to better screen out neurons in the identification network that need to be updated preferentially (those neurons that are less important for learning image content characteristics). The corresponding formula is:
wherein L is m For network loss in the modulation optimization phase, β is the learning rate in the modulation optimization phase.
After the double-layer optimization process is finished, the modulation network corresponds to a better modulation scheme when entering the next optimization, so that the identification network can be better controlled to screen out neurons which are updated preferentially.
And continuously performing cyclic optimization according to the flow, and alternately updating the identification network and the modulation network to enable the network to learn to perform classification identification on the multi-mode data.
Through repeated optimization for many times in the training mode, the final model can have better capability of learning image content characteristics, and accuracy is improved when the model is tested on a multi-mode data set, so that the purpose of improving model robustness is achieved.
In summary, the multi-mode remote sensing target robust recognition method based on double-layer optimization learning is provided, so that the interference of modal differences on the accuracy of a model is effectively reduced, and the robustness of the model is improved.
Claims (2)
1. A multi-mode remote sensing target robust identification method based on double-layer optimization learning is characterized in that the method is realized based on two parallel networks, wherein the parallel networks comprise a modulation network and an identification network; training the identification network by using the modulation network as a gating function, and controlling the selective activation of the identification network; the output of each layer of the modulation network is respectively connected with different positions of the identification network and is used for guiding the parameter updating process of the identification network to be carried out towards the direction of learning the image content characteristics; the modulation network and the identification network are respectively updated in different stages of parallel network optimization through parallel network double-layer optimization, so that separation training of the modulation network and the identification network is realized;
during training, two data sets with different modes are used, namely a known data set D 1 And adaptation data set D 2 The method comprises the steps of carrying out a first treatment on the surface of the By means of a known dataset D 1 The recognition network is pre-trained separately for one time to obtain the recognition network about the known data set D 1 Optimal parameter θ P The method comprises the steps of carrying out a first treatment on the surface of the Before starting double-layer optimization, randomly initializing an identification network and a modulation network; the formal training process adopts a double-layer optimization method and is divided into an inner-layer adaptation stage and a modulation optimization stage; each time the double-layer optimization is started, namely before the inner layer adaptation stage is started, the optimal parameter theta is adopted P Reinitializing the identification network updated by the previous double-layer optimization stage to enable the parameter theta of the identification network 0 =θ P At θ P As a means ofIdentifying initial parameters of a network in the double-layer optimization of the round;
after the inner adaptation phase begins, the data set D is adapted 2 Training the adaptation process of the recognition network to the data sets of different modes, updating the recognition network, and not updating the modulation network in the whole inner layer adaptation stage;
where θ represents a parameter identifying the network,representing parameters of the modulation network, L c For network loss of the inner layer adaptation stage, alpha is the learning rate of the inner layer adaptation stage, the subscript n corresponds to the nth modulation optimization stage cycle, and the subscript i corresponds to the ith inner layer adaptation stage cycle;
after the inner layer adaptation phase is finished, acquiring the adaptation data set D which is learned by the identification network under the current modulation scheme 2 Optimal parameter θ * The method comprises the steps of carrying out a first treatment on the surface of the Starting the modulation optimization phase by the known data set D 1 Judging the effect of recognition network learning under the current modulation scheme; will know the dataset D 1 Loss L of input double-layer network in modulation optimization stage m The loss size is used to reflect the effect of the current modulation scheme; by loss L m Updating the modulation network and thus updating the modulation scheme while maintaining the identification network unchanged;
wherein L is m For network loss in the modulation optimization stage, beta is the learning rate in the modulation optimization stage;
when the one-time double-layer optimization process is completed, the modulation network corresponds to the updated modulation scheme when entering the next double-layer optimization process, and the updated modulation scheme is used for screening out the parameters updated in the identification network.
2. The multi-mode remote sensing target robust recognition method based on double-layer optimization learning according to claim 1, wherein the modulation network comprises a plurality of convolution blocks which are connected in sequence, and a connecting module is connected among all the other convolution blocks except the initial convolution block in parallel; the identification network comprises a plurality of convolution blocks and classification heads which are connected in sequence; besides the initial convolution blocks, the front branches of all the other convolution blocks and the classification heads are transmitted to corresponding connection modules in the modulation network; each connection module is correspondingly transmitted to the cross multiplication operation among the rest convolution blocks and in front of the classification head.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310434641.7A CN116434037B (en) | 2023-04-21 | 2023-04-21 | Multi-mode remote sensing target robust recognition method based on double-layer optimization learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310434641.7A CN116434037B (en) | 2023-04-21 | 2023-04-21 | Multi-mode remote sensing target robust recognition method based on double-layer optimization learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN116434037A true CN116434037A (en) | 2023-07-14 |
CN116434037B CN116434037B (en) | 2023-09-22 |
Family
ID=87079513
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310434641.7A Active CN116434037B (en) | 2023-04-21 | 2023-04-21 | Multi-mode remote sensing target robust recognition method based on double-layer optimization learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116434037B (en) |
Citations (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20090070281A1 (en) * | 2007-06-01 | 2009-03-12 | Solomon Research Llc | System for hybridized efficient genetic algorithms to solve bi-objective optimization problems with application to network computing |
WO2020240477A1 (en) * | 2019-05-31 | 2020-12-03 | Thales Canada Inc. | Method and processing device for training a neural network |
CN112784818A (en) * | 2021-03-03 | 2021-05-11 | 电子科技大学 | Identification method based on grouping type active learning on optical remote sensing image |
CN113065520A (en) * | 2021-04-25 | 2021-07-02 | 江南大学 | Multi-modal data-oriented remote sensing image classification method |
CN113095249A (en) * | 2021-04-19 | 2021-07-09 | 大连理工大学 | Robust multi-mode remote sensing image target detection method |
CN113505792A (en) * | 2021-06-30 | 2021-10-15 | 中国海洋大学 | Multi-scale semantic segmentation method and model for unbalanced remote sensing image |
JP6980958B1 (en) * | 2021-06-23 | 2021-12-15 | 中国科学院西北生態環境資源研究院 | Rural area classification garbage identification method based on deep learning |
CN114067177A (en) * | 2021-11-18 | 2022-02-18 | 中国人民解放军国防科技大学 | Remote sensing image classification network robustness improving method based on self-supervision learning |
WO2022052367A1 (en) * | 2020-09-10 | 2022-03-17 | 中国科学院深圳先进技术研究院 | Neural network optimization method for remote sensing image classification, and terminal and storage medium |
KR20220073582A (en) * | 2020-11-26 | 2022-06-03 | 주식회사 딥핑소스 | Method for training and testing obfuscation network capable of performing distinct concealing processes for distinct parts of original medical image, and training device and testing device using the same |
US20220374630A1 (en) * | 2021-05-11 | 2022-11-24 | Guangxi Academy of Science | Person re-identification system and method integrating multi-scale gan and label learning |
US20230011635A1 (en) * | 2021-07-09 | 2023-01-12 | Viettel Group | Method of face expression recognition |
-
2023
- 2023-04-21 CN CN202310434641.7A patent/CN116434037B/en active Active
Patent Citations (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20090070281A1 (en) * | 2007-06-01 | 2009-03-12 | Solomon Research Llc | System for hybridized efficient genetic algorithms to solve bi-objective optimization problems with application to network computing |
WO2020240477A1 (en) * | 2019-05-31 | 2020-12-03 | Thales Canada Inc. | Method and processing device for training a neural network |
WO2022052367A1 (en) * | 2020-09-10 | 2022-03-17 | 中国科学院深圳先进技术研究院 | Neural network optimization method for remote sensing image classification, and terminal and storage medium |
KR20220073582A (en) * | 2020-11-26 | 2022-06-03 | 주식회사 딥핑소스 | Method for training and testing obfuscation network capable of performing distinct concealing processes for distinct parts of original medical image, and training device and testing device using the same |
CN112784818A (en) * | 2021-03-03 | 2021-05-11 | 电子科技大学 | Identification method based on grouping type active learning on optical remote sensing image |
CN113095249A (en) * | 2021-04-19 | 2021-07-09 | 大连理工大学 | Robust multi-mode remote sensing image target detection method |
CN113065520A (en) * | 2021-04-25 | 2021-07-02 | 江南大学 | Multi-modal data-oriented remote sensing image classification method |
US20220374630A1 (en) * | 2021-05-11 | 2022-11-24 | Guangxi Academy of Science | Person re-identification system and method integrating multi-scale gan and label learning |
JP6980958B1 (en) * | 2021-06-23 | 2021-12-15 | 中国科学院西北生態環境資源研究院 | Rural area classification garbage identification method based on deep learning |
CN113505792A (en) * | 2021-06-30 | 2021-10-15 | 中国海洋大学 | Multi-scale semantic segmentation method and model for unbalanced remote sensing image |
US20230011635A1 (en) * | 2021-07-09 | 2023-01-12 | Viettel Group | Method of face expression recognition |
CN114067177A (en) * | 2021-11-18 | 2022-02-18 | 中国人民解放军国防科技大学 | Remote sensing image classification network robustness improving method based on self-supervision learning |
Non-Patent Citations (3)
Title |
---|
WENDA ZHAO 等: "Diversity Consistency Learning for Remote-Sensing Object Recognition With Limited Labels", 《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING ( VOLUME: 60)》 * |
张军军;万广通;张洪群;李山山;冯旭祥;: "迁移学习下高分快视数据道路快速提取", 中国图象图形学报, no. 07 * |
曲景影;孙显;高鑫;: "基于CNN模型的高分辨率遥感图像目标识别", 国外电子测量技术, no. 08 * |
Also Published As
Publication number | Publication date |
---|---|
CN116434037B (en) | 2023-09-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112163465B (en) | Fine-grained image classification method, fine-grained image classification system, computer equipment and storage medium | |
CN108664924B (en) | Multi-label object identification method based on convolutional neural network | |
CN112750140B (en) | Information mining-based disguised target image segmentation method | |
Dong et al. | Research and discussion on image recognition and classification algorithm based on deep learning | |
CN108170736A (en) | A kind of document based on cycle attention mechanism quickly scans qualitative method | |
CN109002755B (en) | Age estimation model construction method and estimation method based on face image | |
CN107463954B (en) | A kind of template matching recognition methods obscuring different spectrogram picture | |
CN111339935B (en) | Optical remote sensing picture classification method based on interpretable CNN image classification model | |
CN113806746A (en) | Malicious code detection method based on improved CNN network | |
CN110287985B (en) | Depth neural network image identification method based on variable topology structure with variation particle swarm optimization | |
Kembuan et al. | Convolutional neural network (CNN) for image classification of indonesia sign language using tensorflow | |
CN112232395B (en) | Semi-supervised image classification method for generating countermeasure network based on joint training | |
CN114048843A (en) | Small sample learning network based on selective feature migration | |
CN112487933B (en) | Radar waveform identification method and system based on automatic deep learning | |
CN116434037B (en) | Multi-mode remote sensing target robust recognition method based on double-layer optimization learning | |
Kiani et al. | Classification of Persian handwritten digits using spiking neural networks | |
CN113449601B (en) | Pedestrian re-recognition model training and recognition method and device based on progressive smooth loss | |
CN114067155A (en) | Image classification method, device, product and storage medium based on meta learning | |
Yao et al. | Regional attention reinforcement learning for rapid object detection | |
Yang et al. | NAM Net: Meta-network with normalization-based attention for few-shot learning | |
CN113077009A (en) | Tunnel surrounding rock lithology identification method based on migration learning model | |
CN112686228B (en) | Pedestrian attribute identification method and device, electronic equipment and storage medium | |
Eghbali et al. | Deep Convolutional Neural Network (CNN) for Large-Scale Images Classification | |
Tan et al. | Chinese Traditional Visual Cultural Symbols Recognition Based on Convolutional Neural Network | |
CN117892841B (en) | Self-distillation method and system based on progressive association 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 |