To avoid the laborious annotation process for dense prediction tasks like semantic segmentation,unsupervised domain adaptation(UDA)methods have been proposed to leverage the abundant annotations from a source domain,s...To avoid the laborious annotation process for dense prediction tasks like semantic segmentation,unsupervised domain adaptation(UDA)methods have been proposed to leverage the abundant annotations from a source domain,such as virtual world(e.g.,3D games),and adapt models to the target domain(the real world)by narrowing the domain discrepancies.However,because of the large domain gap,directly aligning two distinct domains without considering the intermediates leads to inefficient alignment and inferior adaptation.To address this issue,we propose a novel learnable evolutionary Category Intermediates(CIs)guided UDA model named Leci,which enables the information transfer between the two domains via two processes,i.e.,Distilling and Blending.Starting from a random initialization,the CIs learn shared category-wise semantics automatically from two domains in the Distilling process.Then,the learned semantics in the CIs are sent back to blend the domain features through a residual attentive fusion(RAF)module,such that the categorywise features of both domains shift towards each other.As the CIs progressively and consistently learn from the varying feature distributions during training,they are evolutionary to guide the model to achieve category-wise feature alignment.Experiments on both GTA5 and SYNTHIA datasets demonstrate Leci's superiority over prior representative methods.展开更多
In Unsupervised Domain Adaptation(UDA)for person re-identification(re-ID),the primary challenge is reducing the distribution discrepancy between the source and target domains.This can be achieved by implicitly or expl...In Unsupervised Domain Adaptation(UDA)for person re-identification(re-ID),the primary challenge is reducing the distribution discrepancy between the source and target domains.This can be achieved by implicitly or explicitly constructing an appropriate intermediate domain to enhance recognition capability on the target domain.Implicit construction is difficult due to the absence of intermediate state supervision,making smooth knowledge transfer from the source to the target domain a challenge.To explicitly construct the most suitable intermediate domain for the model to gradually adapt to the feature distribution changes from the source to the target domain,we propose the Minimal Transfer Cost Framework(MTCF).MTCF considers all scenarios of the intermediate domain during the transfer process,ensuring smoother and more efficient domain alignment.Our framework mainly includes threemodules:Intermediate Domain Generator(IDG),Cross-domain Feature Constraint Module(CFCM),and Residual Channel Space Module(RCSM).First,the IDG Module is introduced to generate all possible intermediate domains,ensuring a smooth transition of knowledge fromthe source to the target domain.To reduce the cross-domain feature distribution discrepancy,we propose the CFCM Module,which quantifies the difficulty of knowledge transfer and ensures the diversity of intermediate domain features and their semantic relevance,achieving alignment between the source and target domains by incorporating mutual information and maximum mean discrepancy.We also design the RCSM,which utilizes attention mechanism to enhance the model’s focus on personnel features in low-resolution images,improving the accuracy and efficiency of person re-ID.Our proposed method outperforms existing technologies in all common UDA re-ID tasks and improves the Mean Average Precision(mAP)by 2.3%in the Market to Duke task compared to the state-of-the-art(SOTA)methods.展开更多
A common necessity for prior unsupervised domain adaptation methods that can improve the domain adaptation in unlabeled target domain dataset is access to source domain data-set and target domain dataset simultaneousl...A common necessity for prior unsupervised domain adaptation methods that can improve the domain adaptation in unlabeled target domain dataset is access to source domain data-set and target domain dataset simultaneously.However,data privacy makes it not always possible to access source domain dataset and target domain dataset in actual industrial equipment simulta-neously,especially for aviation component like Electro-Mechanical Actuator(EMA)whose dataset are often not shareable due to the data copyright and confidentiality.To address this problem,this paper proposes a source free unsupervised domain adaptation framework for EMA fault diagnosis.The proposed framework is a combination of feature network and classifier.Firstly,source domain datasets are only applied to train a source model.Secondly,the well-trained source model is trans-ferred to target domain and classifier is frozen based on source domain hypothesis.Thirdly,nearest centroid filtering is introduced to filter the reliable pseudo labels for unlabeled target domain data-set,and finally,supervised learning and pseudo label clustering are applied to fine-tune the trans-ferred model.In comparison with several traditional unsupervised domain adaptation methods,case studies based on low-and high-frequency monitoring signals on EMA indicate the effectiveness of the proposed method.展开更多
Sea fog detection with remote sensing images is a challenging task. Driven by the different image characteristics between fog and other types of clouds, such as textures and colors, it can be achieved by using image p...Sea fog detection with remote sensing images is a challenging task. Driven by the different image characteristics between fog and other types of clouds, such as textures and colors, it can be achieved by using image processing methods. Currently, most of the available methods are datadriven and relying on manual annotations. However, because few meteorological observations and buoys over the sea can be realized, obtaining visibility information to help the annotations is difficult. Considering the feasibility of obtaining abundant visible information over the land and the similarity between land fog and sea fog, we propose an unsupervised domain adaptation method to bridge the abundant labeled land fog data and the unlabeled sea fog data to realize the sea fog detection. We used a seeded region growing module to obtain pixel-level masks from roughlabels generated by the unsupervised domain adaptation model. Experimental results demonstrate that our proposed method achieves an accuracy of sea fog recognition up to 99.17%, which is nearly 3% higher than those vanilla methods.展开更多
:Cross-project defect prediction(CPDP)aims to predict the defects on target project by using a prediction model built on source projects.The main problem in CPDP is the huge distribution gap between the source project...:Cross-project defect prediction(CPDP)aims to predict the defects on target project by using a prediction model built on source projects.The main problem in CPDP is the huge distribution gap between the source project and the target project,which prevents the prediction model from performing well.Most existing methods overlook the class discrimination of the learned features.Seeking an effective transferable model from the source project to the target project for CPDP is challenging.In this paper,we propose an unsupervised domain adaptation based on the discriminative subspace learning(DSL)approach for CPDP.DSL treats the data from two projects as being from two domains and maps the data into a common feature space.It employs crossdomain alignment with discriminative information from different projects to reduce the distribution difference of the data between different projects and incorporates the class discriminative information.Specifically,DSL first utilizes subspace learning based domain adaptation to reduce the distribution gap of data between different projects.Then,it makes full use of the class label information of the source project and transfers the discrimination ability of the source project to the target project in the common space.Comprehensive experiments on five projects verify that DSL can build an effective prediction model and improve the performance over the related competing methods by at least 7.10%and 11.08%in terms of G-measure and AUC.展开更多
By leveraging data from a fully labeled source domain,unsupervised domain adaptation(UDA)im-proves classification performance on an unlabeled target domain through explicit discrepancy minimization of data distributio...By leveraging data from a fully labeled source domain,unsupervised domain adaptation(UDA)im-proves classification performance on an unlabeled target domain through explicit discrepancy minimization of data distribution or adversarial learning.As an enhancement,category alignment is involved during adaptation to reinforce target feature discrimination by utilizing model prediction.However,there remain unexplored prob-lems about pseudo-label inaccuracy incurred by wrong category predictions on target domain,and distribution deviation caused by overfitting on source domain.In this paper,we propose a model-agnostic two-stage learning framework,which greatly reduces flawed model predictions using soft pseudo-label strategy and avoids overfitting on source domain with a curriculum learning strategy.Theoretically,it successfully decreases the combined risk in the upper bound of expected error on the target domain.In the first stage,we train a model with distribution alignment-based UDA method to obtain soft semantic label on target domain with rather high confidence.To avoid overfitting on source domain,in the second stage,we propose a curriculum learning strategy to adaptively control the weighting between losses from the two domains so that the focus of the training stage is gradually shifted from source distribution to target distribution with prediction confidence boosted on the target domain.Extensive experiments on two well-known benchmark datasets validate the universal effectiveness of our proposed framework on promoting the performance of the top-ranked UDA algorithms and demonstrate its consistent su-perior performance.展开更多
Labeled data scarcity of an interested domain is often a serious problem in machine learning.Leveraging the labeled data from other semantic-related yet co-variate shifted source domain to facilitate the interested do...Labeled data scarcity of an interested domain is often a serious problem in machine learning.Leveraging the labeled data from other semantic-related yet co-variate shifted source domain to facilitate the interested domain is a consensus.In order to solve the domain shift between domains and reduce the learning ambiguity,unsupervised domain adaptation(UDA)greatly promotes the transferability of model parameters.However,the dilemma of over-fitting(negative transfer)and under-fitting(under-adaptation)is always an overlooked challenge and potential risk.In this paper,we rethink the shallow learning paradigm and this intractable over/under-fitting problem,and propose a safer UDA model,coined as Bilateral Co-Transfer(BCT),which is essentially beyond previous well-known unilateral transfer.With bilateral co-transfer between domains,the risk of over/under-fitting is therefore largely reduced.Technically,the proposed BCT is a symmetrical structure,with joint distribution discrepancy(JDD)modeled for domain alignment and category discrimination.Specifically,a symmetrical bilateral transfer(SBT)loss between source and target domains is proposed under the philosophy of mutual checks and balances.First,each target sample is represented by source samples with low-rankness constraint in a common subspace,such that the most informative and transferable source data can be used to alleviate negative transfer.Second,each source sample is symmetrically and sparsely represented by target samples,such that the most reliable target samples can be exploited to tackle underadaptation.Experiments on various benchmarks show that our BCT outperforms many previous outstanding work.展开更多
Unsupervised transfer subspace learning is one of the challenging and important topics in domain adaptation,which aims to classify unlabeled target data by using source domain information.The traditional transfer subs...Unsupervised transfer subspace learning is one of the challenging and important topics in domain adaptation,which aims to classify unlabeled target data by using source domain information.The traditional transfer subspace learning methods often impose low-rank constraints,i.e.,trace norm,to preserve data structural information of different domains.However,trace norm is only the convex surrogate to approximate the ideal low-rank constraints and may make their solutions seriously deviate from the original optimums.In addition,the traditional methods directly use the strict labels of source domain,which is difficult to deal with label noise.To solve these problems,we propose a novel nonconvex and discriminative transfer subspace learning method named NDTSL by incorporating Schatten-norm and soft label matrix.Specifically,Schatten-norm can be imposed to approximate the low-rank constraints and obtain a better lowrank representation.Then,we design and adopt soft label matrix in source domain to learn a more flexible classifier and enhance the discriminative ability of target data.Besides,due to the nonconvexity of Schatten-norm,we design an efficient alternative algorithm IALM to solve it.Finally,experimental results on several public transfer tasks demonstrate the effectiveness of NDTSL compared with several state-of-the-art methods.展开更多
Unsupervised Domain Adaptation(UDA)intends to achieve excellent results by transferring knowledge from labeled source domains to unlabeled target domains in which the data or label distribution changes.Previous UDA me...Unsupervised Domain Adaptation(UDA)intends to achieve excellent results by transferring knowledge from labeled source domains to unlabeled target domains in which the data or label distribution changes.Previous UDA methods have acquired great success when labels in the source domain are pure.However,even the acquisition of scare clean labels in the source domain needs plenty of costs as well.In the presence of label noise in the source domain,the traditional UDA methods will be seriously degraded as they do not deal with the label noise.In this paper,we propose an approach named Robust Self-training with Label Refinement(RSLR)to address the above issue.RSLR adopts the self-training framework by maintaining a Labeling Network(LNet)on the source domain,which is used to provide confident pseudo-labels to target samples,and a Target-specific Network(TNet)trained by using the pseudo-labeled samples.To combat the effect of label noise,LNet progressively distinguishes and refines the mislabeled source samples.In combination with class rebalancing to combat the label distribution shift issue,RSLR achieves effective performance on extensive benchmark datasets.展开更多
This paper explores the Vision Transformer(ViT)backbone for Unsupervised Domain Adaptive(UDA)person Re-Identification(Re-ID).While some recent studies have validated ViT for supervised Re-ID,no study has yet to use Vi...This paper explores the Vision Transformer(ViT)backbone for Unsupervised Domain Adaptive(UDA)person Re-Identification(Re-ID).While some recent studies have validated ViT for supervised Re-ID,no study has yet to use ViT for UDA Re-ID.We observe that the ViT structure provides a unique advantage for UDA Re-ID,i.e.,it has a prompt(the learnable class token)at its bottom layer,that can be used to efficiently condition the deep model for the underlying domain.To utilize this advantage,we propose a novel two-stage UDA pipeline named Prompting And Tuning(PAT)which consists of a prompt learning stage and a subsequent fine-tuning stage.In the first stage,PAT roughly adapts the model from source to target domain by learning the prompts for two domains,while in the second stage,PAT fine-tunes the entire backbone for further adaption to increase the accuracy.Although these two stages both adopt the pseudo labels for training,we show that they have different data preferences.With these two preferences,prompt learning and fine-tuning integrated well with each other and jointly facilitated a competitive PAT method for UDA Re-ID.展开更多
Single-pixel imaging(SPI)is a prominent scattering media imaging technique that allows image transmission via one-dimensional detection under structured illumination,with applications spanning from long-range imaging ...Single-pixel imaging(SPI)is a prominent scattering media imaging technique that allows image transmission via one-dimensional detection under structured illumination,with applications spanning from long-range imaging to microscopy.Recent advancements leveraging deep learning(DL)have significantly improved SPI performance,especially at low compression ratios.However,most DL-based SPI methods proposed so far rely heavily on extensive labeled datasets for supervised training,which are often impractical in real-world scenarios.Here,we propose an unsupervised learningenabled label-free SPI method for resilient information transmission through unknown dynamic scattering media.Additionally,we introduce a physics-informed autoencoder framework to optimize encoding schemes,further enhancing image quality at low compression ratios.Simulation and experimental results demonstrate that high-efficiency data transmission with structural similarity exceeding 0.9 is achieved through challenging turbulent channels.Moreover,experiments demonstrate that in a 5 m underwater dynamic turbulent channel,USAF target imaging quality surpasses traditional methods by over 13 dB.The compressive encoded transmission of 720×720 resolution video exceeding 30 seconds with great fidelity is also successfully demonstrated.These preliminary results suggest that our proposed method opens up a new paradigm for resilient information transmission through unknown dynamic scattering media and holds potential for broader applications within many other scattering media imaging technologies.展开更多
Deep learning networks are increasingly exploited in the field of neuronal soma segmentation.However,annotating dataset is also an expensive and time-consuming task.Unsupervised domain adaptation is an effective metho...Deep learning networks are increasingly exploited in the field of neuronal soma segmentation.However,annotating dataset is also an expensive and time-consuming task.Unsupervised domain adaptation is an effective method to mitigate the problem,which is able to learn an adaptive segmentation model by transferring knowledge from a rich-labeled source domain.In this paper,we propose a multi-level distribution alignment-based unsupervised domain adaptation network(MDA-Net)for segmentation of 3D neuronal soma images.Distribution alignment is performed in both feature space and output space.In the feature space,features from different scales are adaptively fused to enhance the feature extraction capability for small target somata and con-strained to be domain invariant by adversarial adaptation strategy.In the output space,local discrepancy maps that can reveal the spatial structures of somata are constructed on the predicted segmentation results.Then thedistribution alignment is performed on the local discrepancies maps across domains to obtain a superior discrepancy map in the target domain,achieving refined segmentation performance of neuronal somata.Additionally,after a period of distribution align-ment procedure,a portion of target samples with high confident pseudo-labels are selected as training data,which assist in learning a more adaptive segmentation network.We verified the superiority of the proposed algorithm by comparing several domain adaptation networks on two 3D mouse brain neuronal somata datasets and one macaque brain neuronal soma dataset.展开更多
The purpose of unsupervised domain adaptation is to use the knowledge of the source domain whose data distribution is different from that of the target domain for promoting the learning task in the target domain.The k...The purpose of unsupervised domain adaptation is to use the knowledge of the source domain whose data distribution is different from that of the target domain for promoting the learning task in the target domain.The key bottleneck in unsupervised domain adaptation is how to obtain higher-level and more abstract feature representations between source and target domains which can bridge the chasm of domain discrepancy.Recently,deep learning methods based on autoencoder have achieved sound performance in representation learning,and many dual or serial autoencoderbased methods take different characteristics of data into consideration for improving the effectiveness of unsupervised domain adaptation.However,most existing methods of autoencoders just serially connect the features generated by different autoencoders,which pose challenges for the discriminative representation learning and fail to find the real cross-domain features.To address this problem,we propose a novel representation learning method based on an integrated autoencoders for unsupervised domain adaptation,called IAUDA.To capture the inter-and inner-domain features of the raw data,two different autoencoders,which are the marginalized autoencoder with maximum mean discrepancy(mAE)and convolutional autoencoder(CAE)respectively,are proposed to learn different feature representations.After higher-level features are obtained by these two different autoencoders,a sparse autoencoder is introduced to compact these inter-and inner-domain representations.In addition,a whitening layer is embedded for features processed before the mAE to reduce redundant features inside a local area.Experimental results demonstrate the effectiveness of our proposed method compared with several state-of-the-art baseline methods.展开更多
Unsupervised domain adaptation(UDA)has achieved great success in handling cross-domain machine learning applications.It typically benefits the model training of unlabeled target domain by leveraging knowledge from lab...Unsupervised domain adaptation(UDA)has achieved great success in handling cross-domain machine learning applications.It typically benefits the model training of unlabeled target domain by leveraging knowledge from labeled source domain.For this purpose,the minimization of the marginal distribution divergence and conditional distribution divergence between the source and the target domain is widely adopted in existing work.Nevertheless,for the sake of privacy preservation,the source domain is usually not provided with training data but trained predictor(e.g.,classifier).This incurs the above studies infeasible because the marginal and conditional distributions of the source domain are incalculable.To this end,this article proposes a source-free UDA which jointly models domain adaptation and sample transport learning,namely Sample Transport Domain Adaptation(STDA).Specifically,STDA constructs the pseudo source domain according to the aggregated decision boundaries of multiple source classifiers made on the target domain.Then,it refines the pseudo source domain by augmenting it through transporting those target samples with high confidence,and consequently generates labels for the target domain.We train the STDA model by performing domain adaptation with sample transport between the above steps in alternating manner,and eventually achieve knowledge adaptation to the target domain and attain confident labels for it.Finally,evaluation results have validated effectiveness and superiority of the proposed method.展开更多
Unsupervised domain adaptation(UDA),which aims to use knowledge from a label-rich source domain to help learn unlabeled target domain,has recently attracted much attention.UDA methods mainly concentrate on source clas...Unsupervised domain adaptation(UDA),which aims to use knowledge from a label-rich source domain to help learn unlabeled target domain,has recently attracted much attention.UDA methods mainly concentrate on source classification and distribution alignment between domains to expect the correct target prediction.While in this paper,we attempt to learn the target prediction end to end directly,and develop a Self-corrected unsupervised domain adaptation(SCUDA)method with probabilistic label correction.SCUDA adopts a probabilistic label corrector to learn and correct the target labels directly.Specifically,besides model parameters,those target pseudo-labels are also updated in learning and corrected by the anchor-variable,which preserves the class candidates for samples.Experiments on real datasets show the competitiveness of SCUDA.展开更多
Deep neural networks have been successfully applied to numerous machine learning tasks because of their impressive feature abstraction capabilities.However,conventional deep networks assume that the training and test ...Deep neural networks have been successfully applied to numerous machine learning tasks because of their impressive feature abstraction capabilities.However,conventional deep networks assume that the training and test data are sampled from the same distribution,and this assumption is often violated in real-world scenarios.To address the domain shift or data bias problems,we introduce layer-wise domain correction(LDC),a new unsupervised domain adaptation algorithm which adapts an existing deep network through additive correction layers spaced throughout the network.Through the additive layers,the representations of source and target domains can be perfectly aligned.The corrections that are trained via maximum mean discrepancy,adapt to the target domain while increasing the representational capacity of the network.LDC requires no target labels,achieves state-of-the-art performance across several adaptation benchmarks,and requires significantly less training time than existing adaptation methods.展开更多
Scene segmentation is widely used in autonomous driving for environmental perception.Semantic scene segmentation has gained considerable attention owing to its rich semantic information.It assigns labels to the pixels...Scene segmentation is widely used in autonomous driving for environmental perception.Semantic scene segmentation has gained considerable attention owing to its rich semantic information.It assigns labels to the pixels in an image,thereby enabling automatic image labeling.Current approaches are based mainly on convolutional neural networks(CNN),however,they rely on numerous labels.Therefore,the use of a small amount of labeled data to achieve semantic segmentation has become increasingly important.In this study,we developed a domain adaptation framework based on optimal transport(OT)and an attention mechanism to address this issue.Specifically,we first generated the output space via a CNN owing to its superior of feature representation.Second,we utilized OT to achieve a more robust alignment of the source and target domains in the output space,where the OT plan defined a well attention mechanism to improve the adaptation of the model.In particular,the OT reduced the number of network parameters and made the network more interpretable.Third,to better describe the multiscale properties of the features,we constructed a multiscale segmentation network to perform domain adaptation.Finally,to verify the performance of the proposed method,we conducted an experiment to compare the proposed method with three benchmark and four SOTA methods using three scene datasets.The mean intersection-over-union(mIOU)was significantly improved,and visualization results under multiple domain adaptation scenarios also show that the proposed method performed better than semantic segmentation methods.展开更多
Osteoporotic Vertebral Fracture(OVFs)is a common lumbar spine disorder that severely affects the health of patients.With a clear bone blocks boundary,CT images have gained obvious advantages in OVFs diagnosis.Compared...Osteoporotic Vertebral Fracture(OVFs)is a common lumbar spine disorder that severely affects the health of patients.With a clear bone blocks boundary,CT images have gained obvious advantages in OVFs diagnosis.Compared with CT images,X-rays are faster and more inexpensive but often leads to misdiagnosis and miss-diagnosis because of the overlapping shadows.Considering how to transfer CT imaging advantages to achieve OVFs classification in X-rays is meaningful.For this purpose,we propose a multi-modal semantic consistency network which could do well X-ray OVFs classification by transferring CT semantic consistency features.Different from existing methods,we introduce a feature-level mix-up module to get the domain soft labels which helps the network reduce the domain offsets between CT and X-ray.In the meanwhile,the network uses a self-rotation pretext task on both CT and X-ray domains to enhance learning the high-level semantic invariant features.We employ five evaluation metrics to compare the proposed method with the state-of-the-art methods.The final results show that our method improves the best value of AUC from 86.32 to 92.16%.The results indicate that multi-modal semantic consistency method could use CT imaging features to improve osteoporotic vertebral fracture classification in X-rays effectively.展开更多
Unsupervised domain adaptation enables neural networks to transfer from a labeled source domain to an unlabeled target domain by learning domain-invariant representations.Recent approaches achieve this by directly mat...Unsupervised domain adaptation enables neural networks to transfer from a labeled source domain to an unlabeled target domain by learning domain-invariant representations.Recent approaches achieve this by directly matching the marginal distributions of these two domains.Most of them,however,ignore exploration of the dynamic trade-off between domain alignment and semantic discrimination learning,thus rendering them susceptible to the problems of negative transfer and outlier samples.To address these issues,we introduce the dynamic parameterized learning framework.First,by exploring domain-level semantic knowledge,the dynamic alignment parameter is proposed,to adaptively adjust the optimization steps of domain alignment and semantic discrimination learning.Besides,for obtaining semantic-discriminative and domain-invariant representations,we propose to align training trajectories on both source and target domains.Comprehensive experiments are conducted to validate the effectiveness of the proposed methods,and extensive comparisons are conducted on seven datasets of three visual tasks to demonstrate their practicability.展开更多
Although neural approaches have yielded state-of-the-art results in the sentence matching task,their perfor-mance inevitably drops dramatically when applied to unseen domains.To tackle this cross-domain challenge,we a...Although neural approaches have yielded state-of-the-art results in the sentence matching task,their perfor-mance inevitably drops dramatically when applied to unseen domains.To tackle this cross-domain challenge,we address unsupervised domain adaptation on sentence matching,in which the goal is to have good performance on a target domain with only unlabeled target domain data as well as labeled source domain data.Specifically,we propose to perform self-su-pervised tasks to achieve it.Different from previous unsupervised domain adaptation methods,self-supervision can not on-ly flexibly suit the characteristics of sentence matching with a special design,but also be much easier to optimize.When training,each self-supervised task is performed on both domains simultaneously in an easy-to-hard curriculum,which gradually brings the two domains closer together along the direction relevant to the task.As a result,the classifier trained on the source domain is able to generalize to the unlabeled target domain.In total,we present three types of self-super-vised tasks and the results demonstrate their superiority.In addition,we further study the performance of different usages of self-supervised tasks,which would inspire how to effectively utilize self-supervision for cross-domain scenarios.展开更多
基金Australian Research Council Project(FL-170100117).
文摘To avoid the laborious annotation process for dense prediction tasks like semantic segmentation,unsupervised domain adaptation(UDA)methods have been proposed to leverage the abundant annotations from a source domain,such as virtual world(e.g.,3D games),and adapt models to the target domain(the real world)by narrowing the domain discrepancies.However,because of the large domain gap,directly aligning two distinct domains without considering the intermediates leads to inefficient alignment and inferior adaptation.To address this issue,we propose a novel learnable evolutionary Category Intermediates(CIs)guided UDA model named Leci,which enables the information transfer between the two domains via two processes,i.e.,Distilling and Blending.Starting from a random initialization,the CIs learn shared category-wise semantics automatically from two domains in the Distilling process.Then,the learned semantics in the CIs are sent back to blend the domain features through a residual attentive fusion(RAF)module,such that the categorywise features of both domains shift towards each other.As the CIs progressively and consistently learn from the varying feature distributions during training,they are evolutionary to guide the model to achieve category-wise feature alignment.Experiments on both GTA5 and SYNTHIA datasets demonstrate Leci's superiority over prior representative methods.
文摘In Unsupervised Domain Adaptation(UDA)for person re-identification(re-ID),the primary challenge is reducing the distribution discrepancy between the source and target domains.This can be achieved by implicitly or explicitly constructing an appropriate intermediate domain to enhance recognition capability on the target domain.Implicit construction is difficult due to the absence of intermediate state supervision,making smooth knowledge transfer from the source to the target domain a challenge.To explicitly construct the most suitable intermediate domain for the model to gradually adapt to the feature distribution changes from the source to the target domain,we propose the Minimal Transfer Cost Framework(MTCF).MTCF considers all scenarios of the intermediate domain during the transfer process,ensuring smoother and more efficient domain alignment.Our framework mainly includes threemodules:Intermediate Domain Generator(IDG),Cross-domain Feature Constraint Module(CFCM),and Residual Channel Space Module(RCSM).First,the IDG Module is introduced to generate all possible intermediate domains,ensuring a smooth transition of knowledge fromthe source to the target domain.To reduce the cross-domain feature distribution discrepancy,we propose the CFCM Module,which quantifies the difficulty of knowledge transfer and ensures the diversity of intermediate domain features and their semantic relevance,achieving alignment between the source and target domains by incorporating mutual information and maximum mean discrepancy.We also design the RCSM,which utilizes attention mechanism to enhance the model’s focus on personnel features in low-resolution images,improving the accuracy and efficiency of person re-ID.Our proposed method outperforms existing technologies in all common UDA re-ID tasks and improves the Mean Average Precision(mAP)by 2.3%in the Market to Duke task compared to the state-of-the-art(SOTA)methods.
基金supported by the National Natural Science Foundation of China(No.52075349)the Aeronautical Science Foundation of China(No.201905019001)the China Scholarship Council(No.202106240078).
文摘A common necessity for prior unsupervised domain adaptation methods that can improve the domain adaptation in unlabeled target domain dataset is access to source domain data-set and target domain dataset simultaneously.However,data privacy makes it not always possible to access source domain dataset and target domain dataset in actual industrial equipment simulta-neously,especially for aviation component like Electro-Mechanical Actuator(EMA)whose dataset are often not shareable due to the data copyright and confidentiality.To address this problem,this paper proposes a source free unsupervised domain adaptation framework for EMA fault diagnosis.The proposed framework is a combination of feature network and classifier.Firstly,source domain datasets are only applied to train a source model.Secondly,the well-trained source model is trans-ferred to target domain and classifier is frozen based on source domain hypothesis.Thirdly,nearest centroid filtering is introduced to filter the reliable pseudo labels for unlabeled target domain data-set,and finally,supervised learning and pseudo label clustering are applied to fine-tune the trans-ferred model.In comparison with several traditional unsupervised domain adaptation methods,case studies based on low-and high-frequency monitoring signals on EMA indicate the effectiveness of the proposed method.
基金supported in part by the Ministry of Education-China Mobile Communication Corp(MoE-CMCC)Artificial Intelligence Project,China(No.MCM20190701)。
文摘Sea fog detection with remote sensing images is a challenging task. Driven by the different image characteristics between fog and other types of clouds, such as textures and colors, it can be achieved by using image processing methods. Currently, most of the available methods are datadriven and relying on manual annotations. However, because few meteorological observations and buoys over the sea can be realized, obtaining visibility information to help the annotations is difficult. Considering the feasibility of obtaining abundant visible information over the land and the similarity between land fog and sea fog, we propose an unsupervised domain adaptation method to bridge the abundant labeled land fog data and the unlabeled sea fog data to realize the sea fog detection. We used a seeded region growing module to obtain pixel-level masks from roughlabels generated by the unsupervised domain adaptation model. Experimental results demonstrate that our proposed method achieves an accuracy of sea fog recognition up to 99.17%, which is nearly 3% higher than those vanilla methods.
基金This paper was supported by the National Natural Science Foundation of China(61772286,61802208,and 61876089)China Postdoctoral Science Foundation Grant 2019M651923Natural Science Foundation of Jiangsu Province of China(BK0191381).
文摘:Cross-project defect prediction(CPDP)aims to predict the defects on target project by using a prediction model built on source projects.The main problem in CPDP is the huge distribution gap between the source project and the target project,which prevents the prediction model from performing well.Most existing methods overlook the class discrimination of the learned features.Seeking an effective transferable model from the source project to the target project for CPDP is challenging.In this paper,we propose an unsupervised domain adaptation based on the discriminative subspace learning(DSL)approach for CPDP.DSL treats the data from two projects as being from two domains and maps the data into a common feature space.It employs crossdomain alignment with discriminative information from different projects to reduce the distribution difference of the data between different projects and incorporates the class discriminative information.Specifically,DSL first utilizes subspace learning based domain adaptation to reduce the distribution gap of data between different projects.Then,it makes full use of the class label information of the source project and transfers the discrimination ability of the source project to the target project in the common space.Comprehensive experiments on five projects verify that DSL can build an effective prediction model and improve the performance over the related competing methods by at least 7.10%and 11.08%in terms of G-measure and AUC.
基金the 111 Project(No.BP0719010)the Project of the Science and Technology Commission of Shanghai Municipality(No.18DZ2270700)。
文摘By leveraging data from a fully labeled source domain,unsupervised domain adaptation(UDA)im-proves classification performance on an unlabeled target domain through explicit discrepancy minimization of data distribution or adversarial learning.As an enhancement,category alignment is involved during adaptation to reinforce target feature discrimination by utilizing model prediction.However,there remain unexplored prob-lems about pseudo-label inaccuracy incurred by wrong category predictions on target domain,and distribution deviation caused by overfitting on source domain.In this paper,we propose a model-agnostic two-stage learning framework,which greatly reduces flawed model predictions using soft pseudo-label strategy and avoids overfitting on source domain with a curriculum learning strategy.Theoretically,it successfully decreases the combined risk in the upper bound of expected error on the target domain.In the first stage,we train a model with distribution alignment-based UDA method to obtain soft semantic label on target domain with rather high confidence.To avoid overfitting on source domain,in the second stage,we propose a curriculum learning strategy to adaptively control the weighting between losses from the two domains so that the focus of the training stage is gradually shifted from source distribution to target distribution with prediction confidence boosted on the target domain.Extensive experiments on two well-known benchmark datasets validate the universal effectiveness of our proposed framework on promoting the performance of the top-ranked UDA algorithms and demonstrate its consistent su-perior performance.
基金supported by National Key R&D Program of China(2021YFB3100800)National Natural Science Foundation of China(62271090)+1 种基金Chongqing Natural Science Fund(cstc2021jcyjjqX0023)supported by Huawei computational power of Chongqing Artificial Intelligence Innovation Center.
文摘Labeled data scarcity of an interested domain is often a serious problem in machine learning.Leveraging the labeled data from other semantic-related yet co-variate shifted source domain to facilitate the interested domain is a consensus.In order to solve the domain shift between domains and reduce the learning ambiguity,unsupervised domain adaptation(UDA)greatly promotes the transferability of model parameters.However,the dilemma of over-fitting(negative transfer)and under-fitting(under-adaptation)is always an overlooked challenge and potential risk.In this paper,we rethink the shallow learning paradigm and this intractable over/under-fitting problem,and propose a safer UDA model,coined as Bilateral Co-Transfer(BCT),which is essentially beyond previous well-known unilateral transfer.With bilateral co-transfer between domains,the risk of over/under-fitting is therefore largely reduced.Technically,the proposed BCT is a symmetrical structure,with joint distribution discrepancy(JDD)modeled for domain alignment and category discrimination.Specifically,a symmetrical bilateral transfer(SBT)loss between source and target domains is proposed under the philosophy of mutual checks and balances.First,each target sample is represented by source samples with low-rankness constraint in a common subspace,such that the most informative and transferable source data can be used to alleviate negative transfer.Second,each source sample is symmetrically and sparsely represented by target samples,such that the most reliable target samples can be exploited to tackle underadaptation.Experiments on various benchmarks show that our BCT outperforms many previous outstanding work.
基金supported by the National Natural Science Foundation of China(Grant No.61922087)the Huxiang Young Talents Program of Hunan Province(2021RC3070).
文摘Unsupervised transfer subspace learning is one of the challenging and important topics in domain adaptation,which aims to classify unlabeled target data by using source domain information.The traditional transfer subspace learning methods often impose low-rank constraints,i.e.,trace norm,to preserve data structural information of different domains.However,trace norm is only the convex surrogate to approximate the ideal low-rank constraints and may make their solutions seriously deviate from the original optimums.In addition,the traditional methods directly use the strict labels of source domain,which is difficult to deal with label noise.To solve these problems,we propose a novel nonconvex and discriminative transfer subspace learning method named NDTSL by incorporating Schatten-norm and soft label matrix.Specifically,Schatten-norm can be imposed to approximate the low-rank constraints and obtain a better lowrank representation.Then,we design and adopt soft label matrix in source domain to learn a more flexible classifier and enhance the discriminative ability of target data.Besides,due to the nonconvexity of Schatten-norm,we design an efficient alternative algorithm IALM to solve it.Finally,experimental results on several public transfer tasks demonstrate the effectiveness of NDTSL compared with several state-of-the-art methods.
基金supported by the National Key R&D Program of China(2022ZD0114801)the National Natural Science Foundation of China(Grant No.61906089)the Jiangsu Province Basic Research Program(BK20190408).
文摘Unsupervised Domain Adaptation(UDA)intends to achieve excellent results by transferring knowledge from labeled source domains to unlabeled target domains in which the data or label distribution changes.Previous UDA methods have acquired great success when labels in the source domain are pure.However,even the acquisition of scare clean labels in the source domain needs plenty of costs as well.In the presence of label noise in the source domain,the traditional UDA methods will be seriously degraded as they do not deal with the label noise.In this paper,we propose an approach named Robust Self-training with Label Refinement(RSLR)to address the above issue.RSLR adopts the self-training framework by maintaining a Labeling Network(LNet)on the source domain,which is used to provide confident pseudo-labels to target samples,and a Target-specific Network(TNet)trained by using the pseudo-labeled samples.To combat the effect of label noise,LNet progressively distinguishes and refines the mislabeled source samples.In combination with class rebalancing to combat the label distribution shift issue,RSLR achieves effective performance on extensive benchmark datasets.
基金This work was supported by the National Key Research and Development Program of China in the 13th Five-Year(No.2016YFB0801301)in the 14th Five-Year(Nos.2021YFFO602103,2021YFF0602102,and 20210Y1702).
文摘This paper explores the Vision Transformer(ViT)backbone for Unsupervised Domain Adaptive(UDA)person Re-Identification(Re-ID).While some recent studies have validated ViT for supervised Re-ID,no study has yet to use ViT for UDA Re-ID.We observe that the ViT structure provides a unique advantage for UDA Re-ID,i.e.,it has a prompt(the learnable class token)at its bottom layer,that can be used to efficiently condition the deep model for the underlying domain.To utilize this advantage,we propose a novel two-stage UDA pipeline named Prompting And Tuning(PAT)which consists of a prompt learning stage and a subsequent fine-tuning stage.In the first stage,PAT roughly adapts the model from source to target domain by learning the prompts for two domains,while in the second stage,PAT fine-tunes the entire backbone for further adaption to increase the accuracy.Although these two stages both adopt the pseudo labels for training,we show that they have different data preferences.With these two preferences,prompt learning and fine-tuning integrated well with each other and jointly facilitated a competitive PAT method for UDA Re-ID.
基金supported by the Natural Science Foundation of China Project(No.62031011).
文摘Single-pixel imaging(SPI)is a prominent scattering media imaging technique that allows image transmission via one-dimensional detection under structured illumination,with applications spanning from long-range imaging to microscopy.Recent advancements leveraging deep learning(DL)have significantly improved SPI performance,especially at low compression ratios.However,most DL-based SPI methods proposed so far rely heavily on extensive labeled datasets for supervised training,which are often impractical in real-world scenarios.Here,we propose an unsupervised learningenabled label-free SPI method for resilient information transmission through unknown dynamic scattering media.Additionally,we introduce a physics-informed autoencoder framework to optimize encoding schemes,further enhancing image quality at low compression ratios.Simulation and experimental results demonstrate that high-efficiency data transmission with structural similarity exceeding 0.9 is achieved through challenging turbulent channels.Moreover,experiments demonstrate that in a 5 m underwater dynamic turbulent channel,USAF target imaging quality surpasses traditional methods by over 13 dB.The compressive encoded transmission of 720×720 resolution video exceeding 30 seconds with great fidelity is also successfully demonstrated.These preliminary results suggest that our proposed method opens up a new paradigm for resilient information transmission through unknown dynamic scattering media and holds potential for broader applications within many other scattering media imaging technologies.
基金supported by the Fund of Key Laboratory of Biomedical Engineering of Hainan Province(No.BME20240001)the STI2030-Major Projects(No.2021ZD0200104)the National Natural Science Foundations of China under Grant 61771437.
文摘Deep learning networks are increasingly exploited in the field of neuronal soma segmentation.However,annotating dataset is also an expensive and time-consuming task.Unsupervised domain adaptation is an effective method to mitigate the problem,which is able to learn an adaptive segmentation model by transferring knowledge from a rich-labeled source domain.In this paper,we propose a multi-level distribution alignment-based unsupervised domain adaptation network(MDA-Net)for segmentation of 3D neuronal soma images.Distribution alignment is performed in both feature space and output space.In the feature space,features from different scales are adaptively fused to enhance the feature extraction capability for small target somata and con-strained to be domain invariant by adversarial adaptation strategy.In the output space,local discrepancy maps that can reveal the spatial structures of somata are constructed on the predicted segmentation results.Then thedistribution alignment is performed on the local discrepancies maps across domains to obtain a superior discrepancy map in the target domain,achieving refined segmentation performance of neuronal somata.Additionally,after a period of distribution align-ment procedure,a portion of target samples with high confident pseudo-labels are selected as training data,which assist in learning a more adaptive segmentation network.We verified the superiority of the proposed algorithm by comparing several domain adaptation networks on two 3D mouse brain neuronal somata datasets and one macaque brain neuronal soma dataset.
基金supported by the National Natural Science Foundation of China(Grant Nos.61906060,62076217,62120106008)the Yangzhou University Interdisciplinary Research Foundation for Animal Husbandry Discipline of Targeted Support(yzuxk202015)+1 种基金the Opening Foundation of Key Laboratory of Huizhou Architecture in Anhui Province(HPJZ-2020-02)the Open Project Program of Joint International Research Laboratory of Agriculture and AgriProduct Safety(JILAR-KF202104).
文摘The purpose of unsupervised domain adaptation is to use the knowledge of the source domain whose data distribution is different from that of the target domain for promoting the learning task in the target domain.The key bottleneck in unsupervised domain adaptation is how to obtain higher-level and more abstract feature representations between source and target domains which can bridge the chasm of domain discrepancy.Recently,deep learning methods based on autoencoder have achieved sound performance in representation learning,and many dual or serial autoencoderbased methods take different characteristics of data into consideration for improving the effectiveness of unsupervised domain adaptation.However,most existing methods of autoencoders just serially connect the features generated by different autoencoders,which pose challenges for the discriminative representation learning and fail to find the real cross-domain features.To address this problem,we propose a novel representation learning method based on an integrated autoencoders for unsupervised domain adaptation,called IAUDA.To capture the inter-and inner-domain features of the raw data,two different autoencoders,which are the marginalized autoencoder with maximum mean discrepancy(mAE)and convolutional autoencoder(CAE)respectively,are proposed to learn different feature representations.After higher-level features are obtained by these two different autoencoders,a sparse autoencoder is introduced to compact these inter-and inner-domain representations.In addition,a whitening layer is embedded for features processed before the mAE to reduce redundant features inside a local area.Experimental results demonstrate the effectiveness of our proposed method compared with several state-of-the-art baseline methods.
基金This work was partially supported by the National Natural Science Foundation of China under Grant Nos.61702273 and 62076062the Natural Science Foundation of Jinangsu Province of China under Grant No.BK20170956+1 种基金the Open Projects Program of National Laboratory of Pattern Recognition under Grant No.20200007was also sponsored by Qing Lan Project.
文摘Unsupervised domain adaptation(UDA)has achieved great success in handling cross-domain machine learning applications.It typically benefits the model training of unlabeled target domain by leveraging knowledge from labeled source domain.For this purpose,the minimization of the marginal distribution divergence and conditional distribution divergence between the source and the target domain is widely adopted in existing work.Nevertheless,for the sake of privacy preservation,the source domain is usually not provided with training data but trained predictor(e.g.,classifier).This incurs the above studies infeasible because the marginal and conditional distributions of the source domain are incalculable.To this end,this article proposes a source-free UDA which jointly models domain adaptation and sample transport learning,namely Sample Transport Domain Adaptation(STDA).Specifically,STDA constructs the pseudo source domain according to the aggregated decision boundaries of multiple source classifiers made on the target domain.Then,it refines the pseudo source domain by augmenting it through transporting those target samples with high confidence,and consequently generates labels for the target domain.We train the STDA model by performing domain adaptation with sample transport between the above steps in alternating manner,and eventually achieve knowledge adaptation to the target domain and attain confident labels for it.Finally,evaluation results have validated effectiveness and superiority of the proposed method.
基金This work was supported by the National Natural Science Foundation of China(Grant Nos.61876091,61772284)the China Postdoctoral Science Foundation(2019M651918)the Open Foundation of MIIT Key Laboratory of Pattern Analysis and Machine Intelligence.
文摘Unsupervised domain adaptation(UDA),which aims to use knowledge from a label-rich source domain to help learn unlabeled target domain,has recently attracted much attention.UDA methods mainly concentrate on source classification and distribution alignment between domains to expect the correct target prediction.While in this paper,we attempt to learn the target prediction end to end directly,and develop a Self-corrected unsupervised domain adaptation(SCUDA)method with probabilistic label correction.SCUDA adopts a probabilistic label corrector to learn and correct the target labels directly.Specifically,besides model parameters,those target pseudo-labels are also updated in learning and corrected by the anchor-variable,which preserves the class candidates for samples.Experiments on real datasets show the competitiveness of SCUDA.
基金supported by the National Key R&D Program of China(No.2016YFB1200203)the National Natural Science Foundation of China(Nos.41427806 and 61273233)
文摘Deep neural networks have been successfully applied to numerous machine learning tasks because of their impressive feature abstraction capabilities.However,conventional deep networks assume that the training and test data are sampled from the same distribution,and this assumption is often violated in real-world scenarios.To address the domain shift or data bias problems,we introduce layer-wise domain correction(LDC),a new unsupervised domain adaptation algorithm which adapts an existing deep network through additive correction layers spaced throughout the network.Through the additive layers,the representations of source and target domains can be perfectly aligned.The corrections that are trained via maximum mean discrepancy,adapt to the target domain while increasing the representational capacity of the network.LDC requires no target labels,achieves state-of-the-art performance across several adaptation benchmarks,and requires significantly less training time than existing adaptation methods.
基金supported by the National Natural Science Foundation of China(11971296)National Key R&D Program of China(2021YFA1003004).
文摘Scene segmentation is widely used in autonomous driving for environmental perception.Semantic scene segmentation has gained considerable attention owing to its rich semantic information.It assigns labels to the pixels in an image,thereby enabling automatic image labeling.Current approaches are based mainly on convolutional neural networks(CNN),however,they rely on numerous labels.Therefore,the use of a small amount of labeled data to achieve semantic segmentation has become increasingly important.In this study,we developed a domain adaptation framework based on optimal transport(OT)and an attention mechanism to address this issue.Specifically,we first generated the output space via a CNN owing to its superior of feature representation.Second,we utilized OT to achieve a more robust alignment of the source and target domains in the output space,where the OT plan defined a well attention mechanism to improve the adaptation of the model.In particular,the OT reduced the number of network parameters and made the network more interpretable.Third,to better describe the multiscale properties of the features,we constructed a multiscale segmentation network to perform domain adaptation.Finally,to verify the performance of the proposed method,we conducted an experiment to compare the proposed method with three benchmark and four SOTA methods using three scene datasets.The mean intersection-over-union(mIOU)was significantly improved,and visualization results under multiple domain adaptation scenarios also show that the proposed method performed better than semantic segmentation methods.
基金National Natural Science Foundation of China(U21A20390)National Key Research and Development Program of China(2018YFC2001302)+2 种基金Development Project of Jilin Province of China(nos.20200801033GH,20200403172SF,YDZJ202101ZYTS128)Jilin Provincial Key Laboratory of Big Data Intelligent Computing(no.20180622002JC)The Fundamental Research Funds for the Central University,JLU.
文摘Osteoporotic Vertebral Fracture(OVFs)is a common lumbar spine disorder that severely affects the health of patients.With a clear bone blocks boundary,CT images have gained obvious advantages in OVFs diagnosis.Compared with CT images,X-rays are faster and more inexpensive but often leads to misdiagnosis and miss-diagnosis because of the overlapping shadows.Considering how to transfer CT imaging advantages to achieve OVFs classification in X-rays is meaningful.For this purpose,we propose a multi-modal semantic consistency network which could do well X-ray OVFs classification by transferring CT semantic consistency features.Different from existing methods,we introduce a feature-level mix-up module to get the domain soft labels which helps the network reduce the domain offsets between CT and X-ray.In the meanwhile,the network uses a self-rotation pretext task on both CT and X-ray domains to enhance learning the high-level semantic invariant features.We employ five evaluation metrics to compare the proposed method with the state-of-the-art methods.The final results show that our method improves the best value of AUC from 86.32 to 92.16%.The results indicate that multi-modal semantic consistency method could use CT imaging features to improve osteoporotic vertebral fracture classification in X-rays effectively.
基金supported by the National Natural Science Foundation of China(No.61932009)the Starry Night Science Fund of Zhejiang University Shanghai Institute for Advanced Study,China。
文摘Unsupervised domain adaptation enables neural networks to transfer from a labeled source domain to an unlabeled target domain by learning domain-invariant representations.Recent approaches achieve this by directly matching the marginal distributions of these two domains.Most of them,however,ignore exploration of the dynamic trade-off between domain alignment and semantic discrimination learning,thus rendering them susceptible to the problems of negative transfer and outlier samples.To address these issues,we introduce the dynamic parameterized learning framework.First,by exploring domain-level semantic knowledge,the dynamic alignment parameter is proposed,to adaptively adjust the optimization steps of domain alignment and semantic discrimination learning.Besides,for obtaining semantic-discriminative and domain-invariant representations,we propose to align training trajectories on both source and target domains.Comprehensive experiments are conducted to validate the effectiveness of the proposed methods,and extensive comparisons are conducted on seven datasets of three visual tasks to demonstrate their practicability.
基金supported by the National Natural Science Foundation of China under Grant Nos.61922085 and 61976211the National Key Research and Development Program of China under Grant No.2020AAA0106400+2 种基金the Key Research Program of the Chinese Academy of Sciences under Grant No.ZDBS-SSW-JSC006the Independent Research Project of the National Laboratory of Pattern Recognition under Grant No.Z-2018013the Youth Innovation Promotion Association of Chinese Academy of Sciences under Grant No.2020138.
文摘Although neural approaches have yielded state-of-the-art results in the sentence matching task,their perfor-mance inevitably drops dramatically when applied to unseen domains.To tackle this cross-domain challenge,we address unsupervised domain adaptation on sentence matching,in which the goal is to have good performance on a target domain with only unlabeled target domain data as well as labeled source domain data.Specifically,we propose to perform self-su-pervised tasks to achieve it.Different from previous unsupervised domain adaptation methods,self-supervision can not on-ly flexibly suit the characteristics of sentence matching with a special design,but also be much easier to optimize.When training,each self-supervised task is performed on both domains simultaneously in an easy-to-hard curriculum,which gradually brings the two domains closer together along the direction relevant to the task.As a result,the classifier trained on the source domain is able to generalize to the unlabeled target domain.In total,we present three types of self-super-vised tasks and the results demonstrate their superiority.In addition,we further study the performance of different usages of self-supervised tasks,which would inspire how to effectively utilize self-supervision for cross-domain scenarios.
