Unsupervised Adaptation Method Research Articles

Dual-polarization SAR rice growth model: A modeling approach for monitoring plant height by combining crop growth patterns with spatiotemporal SAR data

Plant height (PH) is a crucial indicator for monitoring rice growth. Synthetic aperture radar (SAR) has emerged as a valuable tool for monitoring rice growth due to its sensitivity to changes in vegetation. This paper introduces a dual-polarization SAR rice growth model for plant height (DPSRG-PH) that effectively combines crop growth patterns with PH models. To enhance the generalization ability of growth monitoring, we employ an unsupervised domain adaptive (DA) method. The experimental results demonstrate that the DPSRG-PH model considers the correlation between crop growth changes in the time dimension and the differences in crop growth at different phenological stages. Moreover, it can be effectively applied throughout the entire growth period, with the RMSE of most growth stages remaining below 10 cm. Furthermore, the modeling method proposed in this study effectively addresses the issues of overfitting caused by the number of sample points. This enhances the reasonability of the model parameters in evaluating the growth trend and holds practical implications for production.

A Long Time Span-Specific Emitter Identification Method Based on Unsupervised Domain Adaptation

Specific emitter identification (SEI) is a professional technology to recognize different emitters by measuring the unique features of received signals. It has been widely used in both civilian and military fields. Recently, many SEI methods based on deep learning have been proposed, most of which assume that the training set and testing set have the same data distribution. However, in reality, the testing set is generally used later than the training set and lacks labels. The long time span may change the signal transmission environment and fingerprint features. These changes result in considerable differences in data distribution between the training and testing sets, thereby affecting the recognition and prediction abilities of the model. Therefore, the existing works cannot achieve satisfactory results for a long time span SEI. To address this challenge and obtain stable fingerprints, we transform the long time span SEI problem into a domain adaptive problem and propose an unsupervised domain adaptive method called LTS-SEI. Noteworthily, LTS-SEI uses a multilayer convolutional feature extractor to learn feature knowledge and confronts a domain discriminator to generate domain-invariant shallow fingerprints. The classifier of LTS-SEI applies feature matching to source domain samples and target domain samples to achieve the domain alignment of deep fingerprints. The classifier further reduces the intraclass diversity of deep features to alleviate the misclassification problem of edge samples in the target domain. To confirm the effectiveness and reliability of LTS-SEI, we collect multiple sets of real satellite navigation signals using two antennas with 13 m- and 40 m-large apertures, respectively, and construct two available datasets. Numerous experiments demonstrate that LTS–SEI can considerably increase the recognition accuracy of the long time span SEI and is superior to the other existing methods.

Prompting and Tuning: A Two-Stage Unsupervised Domain Adaptive Person Re-identification Method on Vision Transformer Backbone

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.

SSVEP unsupervised adaptive feature recognition method based on self-similarity of same-frequency signals.

As an important human-computer interaction technology, steady-state visual evoked potential (SSVEP) plays a key role in the application of brain computer interface (BCI) systems by accurately decoding SSVEP signals. Currently, the majority SSVEP feature recognition methods use a static classifier. However, electroencephalogram (EEG) signals are non-stationary and time-varying. Hence, an adaptive classification method would be an alternative option to a static classifier for tracking the changes in EEG feature distribution, as its parameters can be re-estimated and updated with the input of new EEG data. In this study, an unsupervised adaptive classification algorithm is designed based on the self-similarity of same-frequency signals. The proposed classification algorithm saves the EEG data that has undergone feature recognition as a template signal in accordance with its estimated label, and the new testing signal is superimposed with the template signals at each stimulus frequency as the new test signals to be analyzed. With the continuous input of EEG data, the template signals are continuously updated. By comparing the classification accuracy of the original testing signal and the testing signal superimposed with the template signals, this study demonstrates the effectiveness of using the self-similarity of same-frequency signals in the adaptive classification algorithm. The experimental results also show that the longer the SSVEP-BCI system is used, the better the responses of users on SSVEP are, and the more significantly the adaptive classification algorithm performs in terms of feature recognition. The testing results of two public datasets show that the adaptive classification algorithm outperforms the static classification method in terms of feature recognition. The proposed adaptive classification algorithm can update the parameters with the input of new EEG data, which is of favorable impact for the accurate analysis of EEG data with time-varying characteristics.

Rolling bearing variable condition fault diagnosis method based on manifold embedding adaptive graph label propagation

Domain adaptation has shown good performance in fault diagnosis of bearing transferable working conditions. However, most of the existing cross-domain diagnosis models are carried out in linear space, which are subject to strong linear constraints and ignore the manifold correlation information between samples, which greatly limits the learning ability of the diagnosis model. Therefore, we propose a manifold embedding adaptive graph label propagation fault diagnosis method for rolling bearings under variable working conditions. The purpose is to reduce the cross-condition domain offset and mine the neglected nonlinear manifold information in the linear space to improve the adaptive ability of the diagnosis model. First, a similarity graph is constructed on the source domain of the known fault type and the target domain of the unknown information, and the cross-domain propagation of the label information is performed on the graph. Then, in this process, the clustering domain adaptation is performed simultaneously, the accuracy of label propagation is improved by reducing the cross-domain offset, and the graph is updated according to the new sample distribution. Finally, according to the known graph and target domain labels, the regression residuals of linear labels are introduced to relax the strong linear constraints in the original space, so that the linear label space approximates the manifold label space and obtains more accurate fault identification results. Experiments show that the new method, as an unsupervised domain adaptive method in transfer learning, has better fault recognition ability than similar transfer learning methods on Paderborn university datasets and own test rig datasets, and the convergence performance of the algorithm is excellent. The highest fault recognition accuracy can reach 100%.

Unsupervised model adaptation for multivariate calibration by domain adaptation-regularization based kernel partial least square

In chemometrics, calibration model adaptation is desired when training- and test-samples come from different distributions. Domain-invariant feature representation is currently a successful strategy to realize model adaptation and has received wide attention. The paper presents a nonlinear unsupervised model adaptation method termed as domain adaption regularization-based kernel partial least squares regression (DarKPLS). DarKPLS aims to minimize the source and target distributions in a low-dimensional latent space projected from the reproducing kernel Hilbert space (RKHS) generated with the labeled source data and unlabeled target data. Specially, the distributional means and variances between source and target latent variables are aligned in the RKHS. By extending existing domain invariant partial least square regression (di-PLS) with the projected maximum mean discrepancy (PMMD) to reduce the mean discrepancy in the RKHS further, DarKPLS could realize fine-grained domain alignment that further improves the adaptation performance. DarKPLS is applied to the γ-polyglutamic acid fermentation dataset, tobacco dataset and corn dataset, and it demonstrates improved prediction results in comparison with No adaptation partial least squares (PLS), null augmented regression (NAR), extended linear joint trained framework (ExtJT), scatter component analysis (SCA) and domain-invariant iterative partial least squares (DIPALS).

Novelty Detection of Leukocyte Image via Mean-Shifted Feature and Directly Optimized Subspace

Novelty detection of leukocyte images aims to learn effective data description from in-distribution leukocyte samples and detect out-of-distribution ones that deviate from the expected patterns. Recently, the methods of fine-tuning on the pre-trained model have demonstrated excellent performance in novelty detection. However, those methods are prone to collapse or feature deterioration due to the training data of limited categories and tiny differences of medical images such as leukocyte images. To alleviate the above issues, we propose an unsupervised feature adaptation method based on the Mean-Shifted feature and the directly optimized subspace for leukocyte novelty detection, named SubMSF. Specifically, we first use Mean-Shifted features to alleviate the catastrophic collapse when fine-tuning the pre-trained model via contrastive learning. Then, we combine the Mean-Shifted feature with the directly optimized subspace to recognize tiny differences between leukocytes. Compared with several state-of-the-art methods, the results on real-world leukocyte datasets demonstrate the effectiveness of the proposed SubMSF.

Two-step Domain Adaptive Semantic Segmentation Algorithm of Driving Scene

Semantic segmentation of driving scenes is an important task in the field of automatic driving perception. The existing semantic segmentation based on deep learning method needs a lot of manpower cost to label data, and the changeable driving scenes will also lead to the performance degradation of semantic segmentation model in practical application. Therefore, in this paper, labeled computer-generated images are used as source domain data, unlabeled real driving scene images are used as target domain data, and unsupervised domain adaptive method is used to transfer knowledge when the target domain data is unlabeled. Based on the one-step domain adaptive semantic segmentation algorithm, a two-step domain adaptive semantic segmentation algorithm is proposed, and an image quality evaluation module is introduced to optimize the algorithm, so as to improve the performance index of cross-domain semantic segmentation tasks to some extent and reduce the dependence on data labeling and labeling cost.

S&GDA: An Unsupervised Domain Adaptive Semantic Segmentation Framework Considering Both Imaging Scene and Geometric Domain Shifts

Unsupervised domain adaptation uses labeled data from a source domain to help learn a target domain without any labeled data. Previous studies have not systematically analyzed the causes of remote sensing (RS) domain shifts, making it difficult to effectively model domain shifts caused by differences in geographic scene and platform imaging positions and attitudes. Therefore, this study conducts detailed analysis of the causes of domain shifts in RS images, and an unsupervised domain adaptive semantic segmentation framework, called S&GDA, that considers both imaging scene and geometric domain shifts is proposed. S&GDA comprised two modules: imaging scene simulation and imaging geometric simulation modules. The imaging scene simulation module is instrumental in mitigating domain shifts in geographical scenes due to variations in natural and human factors, thereby achieving cross-domain imaging scene consistency. Meanwhile, the imaging geometric simulation module allows for accurate simulation of domain shifts caused by changes in the position and attitude of a platform, ensuring cross-domain imaging geometry consistency. Note that none of these modules add additional parameters or computational complexity to the model as they only work on the input side of the data. Comprehensive experiments are conducted on the LoveDA and ISPRS datasets to evaluate S&GDA. Results indicate that S&GDA outperforms the state-of-the-art (SOTA) unsupervised domain adaptive semantic segmentation method by 3.12% of mIoU and can achieve 85% of the performance of the fully supervised method.

Study on Denoising Microseismic Signal Based on Autoencoder Convolutional Neural Networks

As a kind of dynamic real-time monitoring technology, microseismic monitoring technology has been widely used for rockburst warning. Due to the complexity of the actual monitoring environment, the monitoring signals often contain different types of noise, affecting the earning warning of rockburst. In this study, an Autoencoder Convolutional Neural Network denoising model based on deep learning has been proposed to denoising of the complex signals. The unsupervised adaptive training method is used to train the model, which only needs to set its initial parameters. The importance of an enhanced training dataset is illustrated by the comparison experiment. The results indicate that the training and verification shows well performance during training. The denoising efficiency of the proposed model is studied by the denoising of the synthetic noise-containing signals. Furthermore, the dataset from the water conveyance tunnel in the Hanjiang-to-Weihe River water diversion project (HJ-Project) in Shaanxi Province is taken as an engineering example to evolute the performance of the proposed model for practical project. The denoising performance of the model is analysed through the visual denoising results and evaluation index. The model can effectively denoise the complex noised signal which separate it into pure microseismic signal and noise signal, and improve the signal-to-noise ratio, which is benefit for arrive-time picking and source locating then improve the performance of early warning of rockburst.

Unsupervised domain adaptive tumor region recognition for Ki67 automated assisted quantification.

Ki67 is a protein associated with tumor proliferation and metastasis in breast cancer and acts as an essential prognostic factor. Clinical work requires recognizing tumor regions on Ki67-stained whole-slide images (WSIs) before quantitation. Deep learning has the potential to provide assistance but largely relies on massive annotations and consumes a huge amount of time and energy. Hence, a novel tumor region recognition approach is proposed for more precise Ki67 quantification. An unsupervised domain adaptive method is proposed, which combines adversarial and self-training. The model trained on labeled hematoxylin and eosin (H&E) data and unlabeled Ki67 data can recognize tumor regions in Ki67 WSIs. Based on the UDA method, a Ki67 automated assisted quantification system is developed, which contains foreground segmentation, tumor region recognition, cell counting, and WSI-level score calculation. The proposed UDA method achieves high performance in tumor region recognition and Ki67 quantification. The AUC reached 0.9915, 0.9352, and 0.9689 on the validation set and internal and external test sets, respectively, substantially exceeding baseline (0.9334, 0.9167, 0.9408) and rivaling the fully supervised method (0.9950, 0.9284, 0.9652). The evaluation of automated quantification on 148 WSIs illustrated statistical agreement with pathological reports. The model trained by the proposed method is capable of accurately recognizing Ki67 tumor regions. The proposed UDA method can be readily extended to other types of immunohistochemical staining images. The results of automated assisted quantification are accurate and interpretable to provide assistance to both junior and senior pathologists in their interpretation.

Adaptation to CT Reconstruction Kernels by Enforcing Cross-Domain Feature Maps Consistency.

Deep learning methods provide significant assistance in analyzing coronavirus disease (COVID-19) in chest computed tomography (CT) images, including identification, severity assessment, and segmentation. Although the earlier developed methods address the lack of data and specific annotations, the current goal is to build a robust algorithm for clinical use, having a larger pool of available data. With the larger datasets, the domain shift problem arises, affecting the performance of methods on the unseen data. One of the critical sources of domain shift in CT images is the difference in reconstruction kernels used to generate images from the raw data (sinograms). In this paper, we show a decrease in the COVID-19 segmentation quality of the model trained on the smooth and tested on the sharp reconstruction kernels. Furthermore, we compare several domain adaptation approaches to tackle the problem, such as task-specific augmentation and unsupervised adversarial learning. Finally, we propose the unsupervised adaptation method, called F-Consistency, that outperforms the previous approaches. Our method exploits a set of unlabeled CT image pairs which differ only in reconstruction kernels within every pair. It enforces the similarity of the network’s hidden representations (feature maps) by minimizing the mean squared error (MSE) between paired feature maps. We show our method achieving a 0.64 Dice Score on the test dataset with unseen sharp kernels, compared to the 0.56 Dice Score of the baseline model. Moreover, F-Consistency scores 0.80 Dice Score between predictions on the paired images, which almost doubles the baseline score of 0.46 and surpasses the other methods. We also show F-Consistency to better generalize on the unseen kernels and without the presence of the COVID-19 lesions than the other methods trained on unlabeled data.

Fault Prediction of Railway Turnout Systems Based on Improved Sparse Auto Encoder and Gated Recurrent Unit Network

Railway turnout systems (RTS), as one of the most critical ground infrastructures of the high-speed rail, which directly impact the railway operation safety, are also susceptible to frequent faults. Hence, efficient fault prediction and intelligent maintenance decision support of RTS are significant to avoid unexpected outages and guarantee the operation safety of trains. A key challenge for fault prediction of RTS is to extract and select effective feature combinations that describe the degradation status of the system and then generate health indicators (HI) with high monotonicity, correlation, and robustness. In this paper, we focus on the jam fault caused by insufficient lubrication, which is one of the most frequent faults of RTS, and we present a novel unsupervised adaptive latent feature extraction method based on the improved sparse auto-encoder (ISAE). The locally weighted regression (LOESS) is performed to smooth the extracted latent features to overcome the influence of the noise. Improved auto-associative kernel regression (IAAKR) is utilized to fuse the feature combination and generate HI, which can intuitively describe the degradation status of RTS. Finally, the gated recurrent unit (GRU) network is introduced to enhance the correlation data process and construct the fault prediction models. The parameters of GRU are optimized by the improved particle swarm optimization (IPSO) algorithm. Extensive experiments are conducted on a real-world RTS degradation dataset collected from the Changsha Railway Station in China, which is related to the jam fault caused by the lubrication reduction, and the results show the proposed method achieves better prediction performance compared to state-of-the-art methods in terms of different evaluation metrics.

Unsupervised Domain Adaptive Salient Object Detection through Uncertainty-Aware Pseudo-Label Learning

Recent advances in deep learning significantly boost the performance of salient object detection (SOD) at the expense of labeling larger-scale per-pixel annotations. To relieve the burden of labor-intensive labeling, deep unsupervised SOD methods have been proposed to exploit noisy labels generated by handcrafted saliency methods. However, it is still difficult to learn accurate saliency details from rough noisy labels. In this paper, we propose to learn saliency from synthetic but clean labels, which naturally has higher pixel-labeling quality without the effort of manual annotations. Specifically, we first construct a novel synthetic SOD dataset by a simple copy-paste strategy. Considering the large appearance differences between the synthetic and real-world scenarios, directly training with synthetic data will lead to performance degradation on real-world scenarios. To mitigate this problem, we propose a novel unsupervised domain adaptive SOD method to adapt between these two domains by uncertainty-aware self-training. Experimental results show that our proposed method outperforms the existing state-of-the-art deep unsupervised SOD methods on several benchmark datasets, and is even comparable to fully-supervised ones.

MetaPan: Unsupervised Adaptation With Meta-Learning for Multispectral Pansharpening

Multispectral (MS) pansharpening aims to improve the spatial resolution of MS images (MSI) using the spatial details of panchromatic (PAN) images. Due to the gap of prior knowledge between the simulated data and real-world cases, unsupervised learning-based approaches have grown increasing interest. However, some key hyper-parameters, such as the initial weights of the networks, are set manually, which significantly impacts the fusion performance. To tackle this problem, we propose a novel unsupervised adaptation method with meta-learning for MS pansharpening (MetaPan), in which the meta-learning aims to automatically learn the initial parameters of a three-stream fusion network (TSFNet) for unsupervised adaptation learning (UAL). Specifically, the TSFNet consists of a PAN stream, an MS stream, and a fusion stream, where the fusion stream implicitly leverages domain-specific knowledge of input image pairs while the other two streams explicitly inject spatial details and spectral information into the fusion stream. The MetaPan consists of a pre-training stage, a meta-learning stage, and a UAL stage. At the pre-training stage, the TSFNet is trained with the supervision of simulated ground truth such that it is universal for all image pairs. Then, the process of meta-learning optimizes for an internal representation of network parameters that can adapt to a specific image pair with UAL through only a few steps. Finally, the learned internal representation is fine-tuned to a real-world image pair (a test image pair) with UAL. Experiments on two datasets show that our method performs better than state-of-the-art methods in both quantitative metrics and visual appearance.

Transferable network with Siamese architecture for anomaly detection in hyperspectral images

The purpose of hyperspectral anomaly detection is to distinguish abnormal objects from the surrounding background. In actual scenes, however, the complexity of ground objects, the high-dimensionality of data and the non-linear correlation of bands have high requirements for the generalizability, feature extraction ability and nonlinear expression ability of anomaly detection algorithms. In order to address the above problems, we propose a transferable network with Siamese architecture for hyperspectral image anomaly detection (TSN-HAD). The contribution of TSN-HAD is three-fold. First, we address the anomaly detection problem through similarity metric evaluation. The similarity scores are estimated with the proposed Siamese network, which is able to identify pixels that deviate from the neighboring pixels. Second, a spectral-angle-based contrastive constraint is proposed to compel the network to extract more discriminative features, which can enhance the generalizability of the network effectively. Third, to enhance the transfer capability of the network, an unsupervised adaptive clustering method is employed to fine-tune the network with pseudo-labels learned from the test image without any prior knowledge. Comparisons with the state-of-the-art methods on multiple datasets demonstrate the excellence of the proposed method.

Exploiting reliable pseudo-labels for unsupervised domain adaptive person re-identification

Person re-identification, getting impressive performance under the single-domain setting, often suffers huge performance drop when deploying to the unseen target domain owing to domain gap. Current research mainly focuses on unsupervised domain adaptation to alleviate the domain gap, and the methods by clustering the target-domain samples have achieved significant results. However, some inaccurate pseudo-labels, i.e., noisy pseudo-labels, may be generated on clustering, which will seriously affect the performance of the model. In order to solve the above problem, we propose a novel unsupervised domain adaptive person re-identification method by exploiting reliable pseudo-labels (RPL) from two aspects, i.e., adaptive dynamic clustering (ADC) and cross-camera similarity evaluation (CCSE). Specifically, firstly, for the methods based on the density-based clustering algorithm, we propose the adaptive dynamic clustering which calculates the clustering radius adaptively and dynamically to obtain more reasonable clustering results in the iterative optimization of the model. Next, for noisy pseudo-labels caused by small inter-class variations under the same camera, we propose the cross-camera similarity evaluation to filter out these noises to further improve the discrimination of the model. Extensive experiments on three publicly available large-scale datasets show that the proposed method can achieve state-of-the-art performance on unsupervised domain adaptation person re-identification.

Unsupervised Subcategory Domain Adaptive Network for 3D Object Detection in LiDAR

Three-dimensional object detection based on the LiDAR point cloud plays an important role in autonomous driving. The point cloud distribution of the object varies greatly at different distances, observation angles, and occlusion levels. Besides, different types of LiDARs have different settings of projection angles, thus producing an entirely different point cloud distribution. Pre-trained models on the dataset with annotations may degrade on other datasets. In this paper, we propose a method for object detection using an unsupervised adaptive network, which does not require additional annotation data of the target domain. Our object detection adaptive network consists of a general object detection network, a global feature adaptation network, and a special subcategory instance adaptation network. We divide the source domain data into different subcategories and use a multi-label discriminator to assign labels dynamically to the target domain data. We evaluated our approach on the KITTI object benchmark and proved that the proposed unsupervised adaptive method could achieve a remarkable improvement in the adaptation capabilities.

DCN-Based unsupervised domain adaptive person re-identification method

DCN-Based unsupervised domain adaptive person re-identification method

A Word Embedding-Based Method for Unsupervised Adaptation of Cooking Recipes

Studying food recipes is indispensable to understand the science of cooking. An essential problem in food computing is the adaptation of recipes to user needs and preferences. The main difficulty when adapting recipes is in determining ingredients relations, which are compound and hard to interpret. Word embedding models can catch the semantics of food items in a recipe, helping to understand how ingredients are combined and substituted. In this work, we propose an unsupervised method for adapting ingredient recipes to user preferences. To learn food representations and relations, we create and apply a specific-domain word embedding model. In contrast to previous works, we not only use the list of ingredients to train the model but also the cooking instructions. We enrich the ingredient data by mapping them to a nutrition database to guide the adaptation and find ingredient substitutes. We performed three different kinds of recipe adaptation based on nutrition preferences, adapting to similar ingredients, and vegetarian and vegan diet restrictions. With a 95% of confidence, our method can obtain quality adapted recipes without a previous knowledge extraction on the recipe adaptation domain. Our results confirm the potential of using a specific-domain semantic model to tackle the recipe adaptation task.