Search

Linear discriminant analysis (LDA) aims to find a low-dimensional space in which data points in the same class are to be close to each other while keeping data points from different classes apart. To improve the robustness of LDA to non-Gaussian distribution data, most existing discriminant analysis methods extend LDA by approximating the underlying manifold of data. However, these methods suffer from the following problems: 1) local affinity or reconstruction coefficients are learned on the basis of the relationships of all data pairs, which would lead to a sharp increase in the amount of computation and 2) they learn the manifold information in the original space, ignoring the interference of the noise and redundant features. Motivated by these challenges, this article represents a novel discriminant analysis model, called fast and adaptive locality discriminant analysis (FALDA), to improve the efficiency and robustness. First, with the anchor-based strategy, a bipartite graph of each class is constructed to characterize the local structure of data. Since the number of anchor points is far less than that of data points, learning of fuzzy membership relationships between data points and anchor points within each class can save training time. Second, a maximum entropy regularization is introduced to control the uniformity of the weights of graphs and avoid the trivial solution. Third, the above relationships are updated adaptively in the process of dimensionality reduction, which can suppress the interference of the noise and redundant features. Fourth, the whitening constraint is imposed on the projection matrix to remove the relevance between features and restrict the total scatter of data in the subspace. Last but not the least, data with complex distribution can be explicitly divided into sub-blocks according to the learned anchor points (or subclass center points). We test our proposed method on synthetic data, benchmark datasets, and imbalanced datasets. Promising experimental results demonstrate the success of this novel model.

Tensor-ring (TR) decomposition was recently studied and applied for low-rank tensor completion due to its powerful representation ability of high-order tensors. However, most of the existing TR-based methods tend to suffer from deterioration when the selected rank is larger than the true one. To address this issue, this article proposes a new low-rank sparse TR completion method by imposing the Frobenius norm regularization on its latent space. Specifically, we theoretically establish that the proposed method is capable of exploiting the low rankness and Kronecker-basis-representation (KBR)-based sparsity of the target tensor using the Frobenius norm of latent TR-cores. We optimize the proposed TR completion by block coordinate descent (BCD) algorithm and design a modified TR decomposition for the initialization of this algorithm. Extensive experimental results on synthetic data and visual data have demonstrated that the proposed method is able to achieve better results compared to the conventional TR-based completion methods and other state-of-the-art methods and, meanwhile, is quite robust even if the selected TR-rank increases.

Temporal point process is widely used for sequential data modeling. In this paper, we focus on the problem of modeling sequential event propagation in graph, such as retweeting by social network users, news transmitting between websites, etc. Given a collection of event propagation sequences, conventional point process model consider only the event history, i.e. embed event history into a vector, not the latent graph structure. We propose a Graph Biased Temporal Point Process (GBTPP) leveraging the structural information from graph representation learning, where the direct influence between nodes and indirect influence from event history is modeled respectively. Moreover, the learned node embedding vector is also integrated into the embedded event history as side information. Experiments on a synthetic dataset and two real-world datasets show the efficacy of our model compared to conventional methods and state-of-the-art., Comment: 9 pages, 6 figures, 2 tables

Handwritten Text Recognition has achieved an impressive performance in public benchmarks. However, due to the high inter- and intra-class variability between handwriting styles, such recognizers need to be trained using huge volumes of manually labeled training data. To alleviate this labor-consuming problem, synthetic data produced with TrueType fonts has been often used in the training loop to gain volume and augment the handwriting style variability. However, there is a significant style bias between synthetic and real data which hinders the improvement of recognition performance. To deal with such limitations, we propose a generative method for handwritten text-line images, which is conditioned on both visual appearance and textual content. Our method is able to produce long text-line samples with diverse handwriting styles. Once properly trained, our method can also be adapted to new target data by only accessing unlabeled text-line images to mimic handwritten styles and produce images with any textual content. Extensive experiments have been done on making use of the generated samples to boost Handwritten Text Recognition performance. Both qualitative and quantitative results demonstrate that the proposed approach outperforms the current state of the art., Comment: Accepted to TPAMI

Depth maps acquired with ToF cameras have a limited accuracy due to the high noise level and to the multi-path interference. Deep networks can be used for refining ToF depth, but their training requires real world acquisitions with ground truth, which is complex and expensive to collect. A possible workaround is to train networks on synthetic data, but the domain shift between the real and synthetic data reduces the performances. In this paper, we propose three approaches to perform unsupervised domain adaptation of a depth denoising network from synthetic to real data. These approaches are respectively acting at the input, at the feature and at the output level of the network. The first approach uses domain translation networks to transform labeled synthetic ToF data into a representation closer to real data, that is then used to train the denoiser. The second approach tries to align the network internal features related to synthetic and real data. The third approach uses an adversarial loss, implemented with a discriminator trained to recognize the ground truth statistic, to train the denoiser on unlabeled real data. Experimental results show that the considered approaches are able to outperform other state-of-the-art techniques and achieve superior denoising performances.

Augmented reality applications use object tracking to estimate the pose of a camera and to superimpose virtual content onto the observed object. Today, a number of tracking systems are available, ready to be used in industrial applications. However, such systems are hard to handle for a service maintenance engineer, due to obscure configuration procedures. In this article, we investigate options towards replacing the manual configuration process with a machine learning approach based on automatically synthesized data. We present an automated process of creating object tracker facilities exclusively from synthetic data. The data is highly enhanced to train a convolutional neural network, while still being able to receive reliable and robust results during real world applications only from simple RGB cameras. Comparison against related work using the LINEMOD dataset showed that we are able to outperform similar approaches. For our intended industrial applications with high accuracy demands, its performance is still lower than common object tracking methods with manual configuration. Yet, it can greatly support those as an add-on during initialization, due to its higher reliability.

Applications challenged by the joint problem of concept drift and class imbalance are attracting increasing research interest. This paper proposes a novel Reinforcement Online Active Learning Ensemble for Drifting Imbalanced data stream (ROALE-DI). The ensemble classifier has a long-term stable classifier and a dynamic classifier group which applies a reinforcement mechanism to increases the weight of the dynamic classifiers, which perform better on the minority class, and decreases the weight of the opposite. When the data stream is class imbalanced, the classifiers will lack the training samples of the minority class. To supply training samples, when creating a new classifier, the labeled instances buffer is used to provide instances of the minority class. Then, a hybrid labeling strategy that combines the uncertainty strategy and imbalance strategy is proposed to define whether to obtain the real label of an instance. An experimental evaluation compares the classification performance of the proposed method with semi-supervised and supervised algorithms on both real-world and synthetic data streams. The results show that the ROALE-DI achieves higher Area Under the ROC Curve (AUC) and accuracy values with even fewer real labels, and the labeling cost dynamically adjusts according to the concept drift and class imbalance ratio.

The problem of representative selection amounts to sampling few informative exemplars from large datasets. Existing approaches to data selection often fall short of simultaneously handling non-linear data structures, sampling concise and non-redundant subsets, rejecting outliers, and yielding interpretable outcomes. This paper presents a novel representative selection approach, dubbed MOSAIC, for drawing descriptive sketches of arbitrary manifold structures. Resting upon a novel quadratic formulation, MOSAIC advances a multi-criteria selection approach that maximizes the global representation power of the sampled subset, ensures novelty of the samples by minimizing redundancy, and rejects disruptive information by effectively detecting outliers. Theoretical analyses shed light on geometrical characterization of the obtained sketch and reveal that the sampled representatives maximize a well-defined notion of data coverage in a transformed space. In addition, we present a highly scalable randomized implementation of the proposed algorithm shown to bring about substantial speedups. MOSAIC's superiority in achieving the desired characteristics of a representative subset all at once while exhibiting remarkable robustness to various outlier types is demonstrated via extensive experiments conducted on both real and synthetic data with comparisons to state-of-the-art algorithms.

As residential photovoltaic (PV) system installations continue to increase rapidly, utilities need to identify the locations of these new components to manage the unconventional two-way power flow and maintain sustainable management of distribution grids. But historical records are unreliable and constant re-assessment of active residential PV locations is resource intensive. To resolve these issues, we propose to model the solar detection problem in a machine learning set up based on labeled data, e.g., supervised learning. However, the challenge for most utilities is limited labels or labels on only one type of users. Therefore, we design new semi-supervised learning and one-class classification methods based on autoencoders, which greatly improve the nonlinear data representation of human behavior and solar behavior. The proposed methods have been tested and validated not only on synthetic data based on a publicly available dataset, but also on real-world data from utility partners. The numerical results show robust detection accuracy, laying down the foundation for managing distributed energy resources in distribution grids.

Monitoring air quality in cities as well as indoors has become an important topic in recent years. Being easily and densely deployable and rather low-cost, chemiresistive gas sensors indicate a feasible technology for this problem, especially if they are assembled as e-noses tracking multiple gases in one device. Nonetheless, the long-term stability of such sensors poses a severe problem for their measurement accuracy, which substantiates the need for drift compensation procedures as well as robust algorithms in the gas prediction process. In order to test such drift compensation methods and to generate synthetic data for algorithm training, a simulation model would be highly useful. In this paper, we present our investigations towards a simulation framework aiming at generating typical sensor responses of a graphene-based gas sensor with an emphasis on its drift behavior. Different drift models are studied, implemented and compared to real measurement data from a lab setup. The evaluation shows promising results when compared to experiments from a real chemiresistive sensor to which the model was parametrized. By using a second setup with fundamentally different heating characteristics we observed that recalibration is a necessary step to generalize to different sensing modes and to ensure the overall quality of the simulation data. Overall, these investigations provide a proper basis towards tackling the challenges in drift simulation.

Blind video quality assessment (BVQA) is of great importance for video-related applications, yet still challenging even in this deep learning era. The difficulty lies in the shortage of large-scale labeled data, thus making it hard to train a robust spatiotemporal encoder for BVQA. To relieve such difficulty, we first build a video dataset, which contains over 320K samples suffering from various compression and transmission artifacts. While manually annotating the dataset with subjective perception is much labor-intensive and time-consuming, we adopt reference-based VQA algorithms to weakly label the data automatically. We consider that single weak label is derived from single knowledge, which is deficient and incomplete for VQA. To alleviate the bias from single weak label (i.e., single knowledge) in the weakly labeled dataset, we propose HEterogeneous Knowledge Ensemble (HEKE) for spatiotemporal representation learning. Compared to learning from single knowledge, learning with HEKE is thought to achieve a lower infimum theoretically, and obtain richer representation. On the basis of the built dataset and the HEKE methodology, a feature encoder specific to BVQA is formed, and directly extract spatiotemporal representation from videos. Then, the video quality can be either acquired in a completely BVQA manner without ground truth, or via a finetuning-based regressor with labels. Extensive experiments on various VQA databases show that our BVQA model with the pretrained encoder achieves the state-of-the-art performance. More surprisingly, even trained on the synthetic data, our model still shows competitive performance on authentic databases. The data and source code will be available at https://github.com/Sissuire/BVQA-HEKE.

Information granulation and degranulation play a fundamental role in granular computing (GrC). Given a collection of information granules (referred to as reference information granules), the essence of the granulation process (encoding) is to represent each data (either numeric or granular) in terms of these reference information granules. The degranulation process (decoding) that realizes the reconstruction of original data is associated with a certain level of reconstruction error. An important issue is how to reduce the reconstruction error such that the data could be reconstructed more accurately. In this study, the granulation process is realized by involving fuzzy clustering. A novel neural network is leveraged in the consecutive degranulation process, which could help significantly reduce the reconstruction error. We show that the proposed degranulation architecture exhibits improved capabilities in reconstructing original data in comparison with other methods. A series of experiments with the use of synthetic data and publicly available datasets coming from the machine-learning repository demonstrates the superiority of the proposed method over some existing alternatives.

We present PALLAS, a practical method for gene regulatory network (GRN) inference from time series data, which employs penalized maximum likelihood and particle swarms for optimization. PALLAS is based on the Partially-Observable Boolean Dynamical System (POBDS) model and thus does not require ad-hoc binarization of the data. The penalty in the likelihood is a LASSO regularization term, which encourages the resulting network to be sparse. PALLAS is able to scale to networks of realistic size under no prior knowledge, by virtue of a novel continuous-discrete Fish School Search particle swarm algorithm for efficient simultaneous maximization of the penalized likelihood over the discrete space of networks and the continuous space of observational parameters. The performance of PALLAS is demonstrated by a comprehensive set of experiments using synthetic data generated from real and artificial networks, as well as real time series microarray and RNA-seq data, where it is compared to several other well-known methods for gene regulatory network inference. The results show that PALLAS can infer GRNs more accurately than other methods, while being capable of working directly on gene expression data, without need of ad-hoc binarization. PALLAS is a fully-fledged program, written in python, and available on GitHub (https://github.com/yukuntan92/PALLAS).

In recent years, most of the studies have shown that the generalized iterated shrinkage thresholdings (GISTs) have become the commonly used first-order optimization algorithms in sparse learning problems. The nonconvex relaxations of the l₀-norm usually achieve better performance than the convex case (e.g., l₁-norm) since the former can achieve a nearly unbiased solver. To increase the calculation efficiency, this work further provides an accelerated GIST version, that is, AGIST, through the extrapolation-based acceleration technique, which can contribute to reduce the number of iterations when solving a family of nonconvex sparse learning problems. Besides, we present the algorithmic analysis, including both local and global convergence guarantees, as well as other intermediate results for the GIST and AGIST, denoted as (A)GIST, by virtue of the Kurdyka-Łojasiewica (KŁ) property and some milder assumptions. Numerical experiments on both synthetic data and real-world databases can demonstrate that the convergence results of objective function accord to the theoretical properties and nonconvex sparse learning methods can achieve superior performance over some convex ones.

Deep neural networks (DNNs) technique has achieved impressive performance on semantic segmentation,while its training process requires a large amount of pixel-wise labelled data. Domain adaptation,as a promising solution,can break the restriction by training the model on synthetic data,and generalising it in real-world data. However,there is still a lack of attention paid to the imbalance problems on semantic segmentation adaptation,including the imbalance problem between i) source and target data,ii) different classes. To solve these problems,a progressive hierarchical feature alignment method is proposed in this paper. To alleviate the data imbalance problem,the network is progressively trained by the data from multisource domains,so as to obtain domain-invariant features. To address the class imbalance problem,the features are aligned hierarchically across domains. According to the experimental results,our method shows the competitive adapted segmentation performance on three benchmark datasets.

Subspace clustering has been extensively studied from the hypothesis-and-test, algebraic, and spectral clustering-based perspectives. Most assume that only a single type/class of subspace is present. Generalizations to multiple types are non-trivial, plagued by challenges such as choice of types and numbers of models, sampling imbalance and parameter tuning. In many real world problems, data may not lie perfectly on a linear subspace and hand designed linear subspace models may not fit into these situations. In this work, we formulate the multi-type subspace clustering problem as one of learning non-linear subspace filters via deep multi-layer perceptrons (mlps). The response to the learnt subspace filters serve as the feature embedding that is clustering-friendly, i.e., points of the same clusters will be embedded closer together through the network. For inference, we apply K-means to the network output to cluster the data. Experiments are carried out on synthetic data and real world motion segmentation problems, producing state-of-the-art results.

Joint registration of point clouds obtained from multiple views is a key step of reconstruction for blades. However, due to the structural and surface characteristics of blades, some views do not meet the overlap constraints of registration, which results in significant initial errors of pose estimation. Thus, we propose a novel approach to recover the accuracy of poses estimation. The proposed method is robust to overlap extent of views through a stochastic framework. The approach formulates a variable-parameters graph optimization problem. Then a simulated annealing algorithm is used to solve the global optimal parameters. The candidate parameters in the simulated annealing process are obtained through the improved unscented Kalman filter, which reduces the initial errors and enhances the information matrices. The acceptance of the candidate parameters is determined by the optimization problem constrained by joint point correspondences and closed-loop consistency. And the parameters that can improve the registration accuracy are selected. We test our algorithm with simulated synthetic data and real data obtained by the robot measurement system. We compare the proposed algorithm with several state-of-the-art algorithms. The experimental results show that in the presence of significant initial errors, our method can estimate the poses more accurately and obtain better blade reconstructions.

We propose a real-time speaker-independent speech-to-facial animation system that predicts lip and jaw movements on a reference face for audio speech taken from any speaker. Our approach is motivated by two key observations; 1) Speaker-independent facial animation can be generated from phoneme labels, but to perform this automatically a speech recogniser is needed which, due to contextual look-ahead, introduces too much time lag. 2) Audio-driven speech animation can be performed in real-time but requires large, multi-speaker audio-visual speech datasets of which there are few. We adopt a novel three-stage training procedure that leverages the advantages of each approach. First we train a \emph{phoneme}-to-visual speech model from a large single-speaker audio-visual dataset. Next, we use this model to generate the synthetic visual component of a large multi-speaker audio dataset of which the video is not available. Finally, we learn an \emph{audio}-to-visual speech mapping using the synthetic visual features as the target. Furthermore, we increase the realism of the predicted facial animation by introducing two perceptually-based loss functions that aim to improve mouth closures and openings. The proposed method and loss functions are evaluated objectively using mean square error, global variance and a new metric that measures the extent of mouth opening. Subjective tests are carried out over the best performing systems. Results show that our approach produces audio-driven facial animation that is comparable to those produced from phoneme sequences and that improved mouth closures, particularly for bilabial closures, are achieved.

Artificial intelligence (AI)-enhanced automated fault diagnosis (AFD) has become increasingly popular for chiller fault diagnosis with promising classification performance. In practice, a sufficient number of fault samples are required by the AI methods in the training phase. However, faulty training samples are generally much more difficult to be collected than normal training samples. Data augmentation is introduced in these scenarios to enhance the training data set with synthetic data. In this study, a variational autoencoder-based conditional Wasserstein GAN with gradient penalty (CWGAN-GP-VAE) is proposed to diagnose various faults for chillers. A detailed comparative study has been conducted with real-world fault data samples to verify the effectiveness and robustness of the proposed methodology.

The conventional, L2-norm-based, regularization term in electromagnetic (EM) inversions implements smooth constraints on model complexity in the space domain, which can smoothen the boundaries of complex underground structures. To improve the resolution of 3-D frequency-domain airborne EM (AEM) inversions, we propose a new algorithm for sparse-regularized inversion based on the shearlet transform. Unlike traditional methods that invert the model parameters in the space domain, we first transform the 3-D resistivity model into the frequency domain and then invert the sparse coefficients using an L1-norm measure to ensure the sparseness of the solution. Finally, we transform the shearlet coefficients back to the space domain to update the model. The shearlet transform has inherent multiscale and multidirectional properties, making it capable of effectively extracting complex geometries such as curved boundaries. We adopt the finite-difference method and the iteratively reweighted least-squares scheme for our 3-D AEM modeling and inversions and apply the ``moving footprint'' technique to speed up the inversion. Tests using synthetic data show that sparse-regularized inversion based on the shearlet transform can obtain more-focused inversion results than conventional smoothness-constrained inversions based on the L2-norm. Tests using field survey data also reveal that the new method can achieve more realistic underground structures.

The distribution of physical features in the Earth's interior could be estimated by geophysical inversion from the acquired data at or above the surface. Inverse problems are generally considered as least-squares optimization issues in high-dimensional parameter space. Existing approaches are largely based on linear inversion methods, which are limited by the initial model. Nonlinear inversion methods, despite their significant ability in uncertainty quantification, still remain a formidable computational task. In this letter, a new gravity inversion approach is developed based on convolutional neural networks (CNNs). Although the training stage of this method is time-consuming, the actual prediction can be performed in only seconds. Thus, the high computational time of geophysical inversion can be considerably decreased once an appropriate network is constructed. The tests on synthetic data demonstrate that good results could be attained by applying this method to gravity data inversion compared with the least-squares regularization inversion and fully convolutional networks (FCNs).

Spatial information can play a supporting role in spectral unmixing. In this letter, we propose a dual branch autoencoder network to incorporate spatial-contextual information for spectral-spatial unmixing. The two branches leverage different architectures to efficiently extract spatial information and spectral information. In the first branch, we use fully connected layers to extract spectral information, where the neuron in each layer can capture all spectral features. In the second branch, 2-D convolution is adopted to exploit spatial features, which does not require hand-crafted assumptions compared with conventional methods. Then the extracted features are concatenated and propagated to generate the abundance and reconstruct the pixel. Moreover, to solve the drawbacks of the existing reconstruction functions, we propose a new function termed squared sine distance to improve the convergence quality of the proposed network. Experimental results reveal the effectiveness of our proposed method on both synthetic data and real-world data.

Fil: Pascual, Juan Pablo. Instutito Balseiro. Universidad Nacional de Cuyo. Río Negro, Argentina. Fil: Pascual, Juan Pablo. Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET). Argentina. Fil: Cogo, Jorge. Universidad Nacional de Río Negro. Centro Interdisciplinario de Telecomunicaciones, Electrónica, Computación y Ciencia Aplicada. Río Negro, Argentina. Fil: Cogo, Jorge. Instutito Balseiro. Universidad Nacional de Cuyo. Río Negro, Argentina. Fil: Collado Rosell, Arturo. Comisión Nacional de Energı́a Atómica (CNEA). Bariloche, Argentina. Fil: Collado Rosell, Arturo. Instutito Balseiro. Universidad Nacional de Cuyo. Río Negro, Argentina. Fil: Areta, Javier Alberto. Universidad Nacional de Río Negro. Centro Interdisciplinario de Telecomunicaciones, Electrónica, Computación y Ciencia Aplicada. Río Negro, Argentina. Fil: Areta, Javier Alberto. Instutito Balseiro. Universidad Nacional de Cuyo. Río Negro, Argentina. Fil: Areta, Javier Alberto. Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET). Argentina. In this article, we present a novel algorithm termed multipulse processing (MPP) for improving mean Doppler velocity estimation in weather radar applications. It can be used for both staggered pulse repetition time (PRT) and uniform-PRT sequences. Essentially, MPP consists of finding a particular zero of a functional composed of data autocorrelation estimates at multiple lags. To select the proper zero, an initial Doppler velocity estimate is required. Therefore, MPP can be considered as an estimation refinement stage. Its advantage lies in the fact that it uses the complete information contained in the radar signal autocorrelation. After a theoretical analysis, we compare the performance of MPP against other well-established methods of similar complexity and the Cramér–Rao lower bound, by means of Monte Carlo simulations using synthetic data. We show that the proposed estimator offers the lowest root-mean-square error (RMSE) at low signal-to-noise ratio (SNR) situations for a wide range of spectral widths. Finally, we evaluate the MPP algorithm performance using real data measured by the RMA Argentinian weather radar. The results of tests performed are consistent with those of Monte Carlo simulations and validate the proposed method. En este artículo, presentamos un algoritmo novedoso denominado procesamiento multipulso (MPP) para mejorar la estimación de la velocidad Doppler media en aplicaciones de radares meteorológicos. Se puede utilizar tanto para secuencias de tiempo de repetición de pulso (PRT) escalonadas como de PRT uniforme. Esencialmente, MPP consiste en encontrar un cero particular de un funcional compuesto por estimaciones de la autocorrelación de los datos en múltiples retardos. Para seleccionar el cero adecuado, se requiere una estimación inicial de la velocidad Doppler. Por lo tanto, MPP puede considerarse como una etapa de refinamiento de la estimación. Su ventaja radica en que utiliza la información completa contenida en la autocorrelación de la señal del radar. Después de un análisis teórico, comparamos el rendimiento de MPP con otros métodos bien establecidos de complejidad similar y con la cota Cramér-Rao, mediante simulaciones de Monte Carlo utilizando datos sintéticos. Mostramos que el estimador propuesto ofrece el error cuadrático medio (RMSE) más bajo en situaciones de relación señal-ruido (SNR) baja para una amplia gama de anchos espectrales. Finalmente, evaluamos el rendimiento del algoritmo MPP utilizando datos reales medidos por el radar meteorológico argentino RMA. Los resultados de las pruebas realizadas son consistentes con los de las simulaciones de Monte Carlo y validan el método propuesto.

Information extraction from synthetic aperture radar (SAR) images is heavily impaired by speckle noise, and hence, despeckling is a crucial preliminary step in scene analysis algorithms. The recent success of deep learning envisions a new generation of despeckling techniques that could outperform classical model-based methods. However, current deep learning approaches to despeckling require supervision for training, whereas clean SAR images are impossible to obtain. In the literature, this issue is tackled by resorting to either synthetically speckled optical images, which exhibit different properties with respect to true SAR images, or multitemporal SAR images, which are difficult to acquire or fuse accurately. In this article, inspired by recent works on blind-spot denoising networks, we propose a self-supervised Bayesian despeckling method. The proposed method is trained by employing only noisy SAR images and can, therefore, learn features of real SAR images rather than synthetic data. Experiments show that the performance of the proposed approach is very close to the supervised training approach on synthetic data and superior on real data in both quantitative and visual assessments.

Recently, distributed optical fiber acoustic sensing (DAS) is regarded as a transformative technology in seismic exploration. However, both various complex background noise and weak desired signals significantly limit its practical application. To explore an effective denoising method for the vertical seismic profile (VSP) record received by DAS, we propose an improved residual encoder-decoder deep neural network (RED-Net) enhanced by deep iterative memory block (DMB) and channel aggregation block (CAB), called residual channel aggregation encoder-decoder network (RCEN). Here, DMB uses the weight accumulation theory to improve the feature extraction ability and achieve accurate noise elimination. Meanwhile, CAB, utilizing the multi-channel analysis architecture, enhances the weak signal retention performance. In addition, we leverage both synthetic data obtained by forward modeling and real DAS noise data to construct a sufficient training dataset with high authenticity, thereby meeting the requirement of network training. Both synthetic and field DAS-VSP data processing results demonstrate the advantage of RCEN compared to competing algorithms, including singular value decomposition (SVD), conventional RED-Net and feed-forward Denoising CNN (DnCNN).

Deep learning has been widely adopted in seismic inversion. One of the major obstacles when adopting deep learning in seismic inversion is the demand for labeled data sets. There are mainly two approaches to address this problem. One is to generate massive numbers of synthetic data and then transfer the trained model to real data. The other is to introduce theoretical constraints and reduce the parameter spaces of deep learning. In this letter, we propose a physics-constrained seismic impedance inversion method based on deep learning. Robinson convolution model is adopted to model the seismic forward process and provide theoretical constraints for the inversion process. Bilateral filtering is further combined to constrain the spatial continuity of the inversion results. The experimental results on both synthetic examples and real examples demonstrate that the proposed method can effectively improve the prediction accuracy and the spatial continuity of the inversion results.

Deep learning approaches require enough training samples to perform well, but it is a challenge to collect enough real training data and label them manually. In this letter, we propose a practical framework for automatically generating content-rich synthetic images with ground-truth annotations. By rendering 3-D CAD models, we generate two synthetic aircraft image data sets with wide distribution (Syn N and Syn U). For improving the quality of synthetic images, we propose a multiscale attention module which enhances the Cycle-Consistent Adversarial Network (CycleGAN) in spatial and channel dimensions. Then, we compare the synthetic images before and after translation qualitatively and quantitatively. Experiments on Northwestern Polytechnical University (NWPU) very high resolution (VHR)-10, University of Chinese Academy of Sciences, orientation robust object detection in aerial images (UCAS-AOD), and benchmark for object DetectIon in Optical Remote sensing images (DIOR) data sets demonstrate that synthetic data augmentation can improve the performance of aircraft detection in remote sensing images, especially when real data are insufficient. Synthetic data are available at: https://weix-liu.github.io/.

The attenuation of coherent noise plays a crucial role in reflection seismology but still poses some technical challenges. The multiple reflection-refraction (MRR) is one of the main coherent noises in land seismic surveys. The Cadzow filter can effectively attenuate incoherent noise. But it struggles in attenuating coherent noise. After keeping reflection events relatively horizontal, some researchers randomly rearrange the trace orders of the input data to realign coherent noise in an incoherent position so that the Cadzow filter can treat the realigned coherent noise as incoherent noise. Accordingly, a random-derangement-based FX Cadzow filter is proposed to attenuate MRRs in 2-D common-shot gathers based on the linear feature of MRR events and randomization of limited trace orders of the input data. In practice, after linear-move-out (LMO) with an estimated dip, the input data are divided into several small windows to obtain a group of relatively horizontal MRR events. Then, these windows are randomized and filtered one by one. Due to the limited trace orders of each window, there are many possible rearrangements after randomization. Some rearrangements may lead to undesirable filtering performance. To obtain the rearrangements that lead to good filtering performance, each window's trace orders are randomly deranged by resampling from the input orders with uniform distribution. A 2-D synthetic data experiment shows the influence of different rearrangements on the filtering performance. Both 2-D synthetic data and 2-D field data demonstrate that the proposed method outperforms the 2-D FK filter to attenuate both aliased and nonaliased MRRs.

The microseismic monitoring technique is widely applied to petroleum reservoirs to understand the process of hydraulic fracturing. Geophones continuously record the microseismic events triggered by fluid injection on the Earth’s surface or in monitoring wells. The microseismic event localization precision has a large impact on the performance of the technique. Deep learning has achieved significant progress in computer vision and natural language processing in recent years. We propose to use a deep convolutional neural network (CNN) to directly map the field records to their event locations. The biggest advantage of deep learning methods over conventional methods is that they can efficiently predict the characteristics of a huge amount of recorded data without human intervention. Thus, we use a CNN to predict the event location of field microseismic data, which were recorded during a hydraulic fracturing process of a shale gas play in Oklahoma, the United States. We use synthetic data with extracted field noise from the records to train the CNN. The synthetic training data allow us to produce the corresponding labels, and the extracted noise from the field data reduces the difference between the field and synthetic data. We use a correlation pre-processing step to avoid the need for event detection and picking of arrivals. We demonstrate that the proposed approach provides accurate microseismic event locations at a much faster speed than traditional imaging methods, such as time-reversal imaging. Comparison with an existing study on the same data is presented to evaluate the performance of the trained neural network.

We propose a blind nonlinear acoustic impedance inversion method. The seismic wavelet is first extracted through the Euclid deconvolution method from multichannel seismic data. Then, the acoustic impedance is inverted based on the exact nonlinear forward operator. The conventional Euclid deconvolution can theoretically estimate the reflectivity without special prior assumptions, but the method is extremely inefficient and unstable in the case of a large amount of data. We optimize the method and propose a frequency-domain algorithm to improve its efficiency. Conventional impedance inversions are almost implemented based on the linearized approximate formula, but the inversion errors will increase sharply when there is a strong reflection interface. We build the inversion objective function by the accurate nonlinear formula to improve accuracy. The total variation (TV) regularization and low-frequency components of well-logging curves are added to the objective function to make the inversion result have a block structure and converge to the absolute impedance. The nonlinear optimization problem is finally solved by the Levenberg-Marquardt (LM) algorithm. The results of synthetic data and field data verify that our method has high accuracy and good practicability.

This article proposes a cognitive radar (CR) framework to improve the performance of the estimation and recognition for range-extended targets. To overcome the angular uncertainty and exploit the temporal correlation of target signatures, we introduce the static multimodel approach and the exponential correlation model to describe the dynamics of target signatures. The estimate of target signatures and the probability distribution of target hypotheses are updated recursively in response to the latest echo, ruled by the Bayesian theory. A waveform optimization module is integrated into the framework, which designs the waveform for the next emission according to the knowledge extracted from the echo history. A design criterion based on mutual information (MI) is derived, which implies the average estimation capability for all target hypotheses and the mean between-hypotheses divergence. The proposed method is examined using both synthetic data and electromagnetic (EM) simulation data of realistic targets.

Hyperspectral unmixing is an essential research topic for spectral data analysis due to the existence of mixed pixels. Recently, many methods based on sparse nonnegative matrix factorization (NMF) have been widely used for unmixing by incorporating similarity preserving. However, most of them conduct the similarity learning and unmixing by using two separate steps, which may lead to the case that the learned similarity matrix is not the optimal one for unmixing. Thus, the performance of unmixing would be influenced and become undesirable. To alleviate this problem, we propose an adaptive relationship preserving-based sparse NMF (ARP-NMF) for hyperspectral unmixing. Typically, we regard similarity learning and unmixing as an alternative optimization process. During this process, the learned similarity weights and unmixing results can be mutually improved. Thus, our proposed method can learn the optimal similarity weights for unmixing and obtain better generalization ability for different hyperspectral images than traditional methods. Moreover, by using the spectral and spatial local structure, our ARP-NMF method effectively preserves structure consistency between pixels and abundances. Experimental results both on the synthetic data and the real data reveal that our proposed method outperforms several representative sparse NMF-based methods.

Gravity and magnetic inversion are important methods for comprehensive quantitative interpretation of data obtained in, e.g., mineral, oil and gas, and geothermal exploration. At present, the 3-D joint inversion technology of gravity and magnetic data is facing challenges from large-scale data exploration applications. In this letter, a new algorithm for 3-D joint inversion of gravity and magnetic data with high accuracy and low computational cost is presented. We use the geometric trellis method to perform fast forward calculations and then introduce the sparse constraint and adaptive sensitivity matrix into the model constraint terms. The inexact structural resemblance method is then used to add the cross-gradient constraint penalty term to the objective function. Finally, an algorithm (DS-TGN) combining data-space (DS) and truncated Gauss-Newton (TGN) methods is used to solve the joint inversion objective function. Numerical experiments with synthetic data show that the proposed algorithm can significantly reduce the computational cost and obtain high accuracy density and magnetization models with structural resemblance and sharp boundaries. We also apply the DS-TGN algorithm to data obtained in the area of Greater Khingan in northwestern Heilongjiang, China. The underground density and magnetization distribution results provide a high-resolution geological model for the detection of skarn-type deposits.

A major challenges of deep learning (DL) is the necessity to collect huge amounts of training data. Often, the lack of a sufficiently large dataset discourages the use of DL in certain applications. Typically, acquiring the required amounts of data costs considerable time, material and effort. To mitigate this problem, the use of synthetic images combined with real data is a popular approach, widely adopted in the scientific community to effectively train various detectors. In this study, we examined the potential of synthetic data-based training in the field of intelligent transportation systems. Our focus is on camera-based traffic sign recognition applications for advanced driver assistance systems and autonomous driving. The proposed augmentation pipeline of synthetic datasets includes novel augmentation processes such as structured shadows and gaussian specular highlights. A well-known DL model was trained with different datasets to compare the performance of synthetic and real image-based trained models. Additionally, a new, detailed method to objectively compare these models is proposed. Synthetic images are generated using a semi-supervised errors-guide method which is also described. Our experiments showed that a synthetic image-based approach outperforms in most cases real image-based training when applied to cross-domain test datasets (+10% precision for GTSRB dataset) and consequently, the generalization of the model is improved decreasing the cost of acquiring images.

Seismic inversion is an indispensable part of earth exploration to precisely obtain the properties of subsurface media based on seismic data. However, the lack or inaccuracy of low-frequency (LF) information in seismic data constrains the correctness of the inversion. Traditional techniques encounter challenges in compensating the LF component of seismic data. Accessing the ability of deep learning to nonlinearly map inputs to expected outputs, we develop a neural network that can map post-stack data to broader band data, and then to impedance. We first propose an effective preprocessing scheme incorporating both well-logging and seismic data. Then, we extrapolate the LF information in the seismic data and invert the P-wave impedance with supervised and semi-supervised frameworks, respectively. In the synthetic data example, the coefficient of determination (R2) reaches 0.99 for LF extrapolation and 0.98 for impedance inversion. In the field data example, R2 is 0.826 between the inverted impedance and the real impedance of the validation well. Our experiments also reveal that the LF extrapolation improves the results of the impedance inversion.

In seismic exploration, seismic data usually contain anomalous amplitude noise whose high energy may affect the results of subsequent processing steps. In industry, this kind of noise is generally suppressed using the anomalous amplitude attenuation (AAA) method. The AAA method essentially suppresses abnormal amplitude noise using a median filter in the time-frequency domain. This makes its performance heavily dependent on the parameters, especially the window width of the median filter. Thus, we propose an improved anomalous amplitude attenuation (IAAA) method based on deep neural networks. The IAAA method contains two steps. In the first step, deep neural networks are used to detect the locations and the widths of noise regions. In the second step, the noise information (locations and widths) obtained at the previous step is exploited to apply the AAA method with more appropriate parameters to each noisy region. Compared with the conventional AAA method, the IAAA method can suppress the noise more effectively and preserve signals better. The experiments on both synthetic data and field data demonstrate that our method outperforms the conventional AAA method.

Blended acquisition can improve the acquisition efficiency, and thereby reduce the acquisition costs. It becomes an important acquisition method in seismic exploration. However, the blending noise imposes challenges for subsequent traditional seismic data processing procedures, and thus deblending algorithms are necessary to obtain deblended seismic data. Based on the assumption that the signal is coherent and the blending noise is randomized in a specific domain, traditional deblending methods using the L₂-norm measuring the data misfit can obtain separated gathers. However, the L₂-norm is improper when the appearing seismic noise bias the normal distribution, that is, abnormal noise appears. To attenuate the abnormal noise effects during iterative deblending, we proposed an accurate deblending algorithm based on a robust sparse Radon transform (RSRT). For the RSRT, the alternating split Bregman (ASB) algorithm is used for robust 2-D Radon model updating with an L₁-norm to measure data misfit in the mixed time-frequency domain and the sparsity constraint to the time-domain Radon model. Using the RSRT iteratively, the Radon model and the corresponding deblended data can be estimated robustly, accurately, and efficiently. Blended synthetic data with different levels of abnormal noise and with different trace intervals demonstrate the validity and flexibility of the proposed robust deblending method quantitatively with a high recovered SNR. Numerically blended field data further prove the effectiveness of the proposed method in attenuating the abnormal and blending noise.

Deep-learning-based methods have been successfully applied to seismic data random noise attenuation. Among them, the supervised deep-learning-based methods dominate the unsupervised ones. The supervised methods need accurate noise-free data as training labels. However, the field seismic data cannot meet this requirement. To circumvent it, some researchers utilized realistic-looking synthetic data or denoised results via conventional methods as labels. The former ones encounter the problem of weak generalization ability because it requires the same distribution of test and training data. The latter ones encounter the issue of insufficient denoising ability because its denoising ability is difficult to significantly exceed the conventional methods which were used to generate labels. To avoid preparing noise-free labels, we propose a novel deep learning framework for attenuating random noise of prestack seismic data in an unsupervised manner. The prestack seismic data, such as common-reflection-point (CRP) gathers and common-midpoint (CMP) gathers after normal moveout (NMO) correction, have high self-similarity. It is because their events are coherent in the time-space domain and approximately horizontal from shallow to deep layers. The generator convolutional neural network (GCN) first learns self-similar features before any learning. The useful signals are more self-similar than random noise, which is incoherent and randomly distributed. Therefore, the GCN extracts features of useful signals before random noise. We select the specific training iteration and adopt the early stopping strategy to suppress random noise. Both synthetic and field prestack seismic data examples demonstrate the validity of our methods.

This letter addresses the persymmetric adaptive detection problem of range-spread targets in compound Gaussian sea clutter. Based on the generalized likelihood ratio test (GLRT), Rao test, and Wald test, three novel compound Gaussian detectors are developed. Specifically, the sea clutter is modeled as compound Gaussian with inverse Gaussian distribution, and the detectors are derived by the two-step maximum a posterior (MAP) procedures. In the first step, we assume the clutter covariance matrix (CCM) and the inverse Gaussian texture are known and derive the proposed detectors' test statistics. In the second step, we use the persymmetric property and MAP criterion to estimate the CCM and inverse Gaussian texture parameters. Then the proposed detectors are proved to be constant false alarm rate (CFAR) detectors with respect to the real CCM. The simulation experiments are conducted by comparing the proposed detectors with their counterparts on both synthetic data and real sea clutter data. The numerical results indicate that the novel detectors exhibit better detection performance than their competitors.

Most previous studies focus on using complex deep neural networks to learn diverse features of massive synthetic data. In more realistic situations with a limited number of data pairs, complex networks not only have higher computational complexity, increasing training time and reducing inference speed, but also tend to over-fit a small amount of training data, thus having weak generalization capability. To address the aforementioned problem, we propose a lightweight architecture for real-time velocity inversion in realistic situations, the bilateral inversion network (BiInNet). BiInNet uses lightweight ResNet18, ShuffleNetV2, and modified MobileNetV2 as backbones, taking into account the inversion accuracy and inference speed. To reduce the redundant information in common shot gathers and focus on graphical property features which are strongly correlated with velocity, the intermediate results of velocity analysis, semblances, and interval velocity models are prepared as data pairs. Numerical experiments show that BiInNet can infer interval velocity models in real-time, with frames per second (FPS) up to 76.90 when ResNet18 is used as the backbone. Moreover, BiInNet achieves the best inversion accuracy on more realistic fold models, fault models, salt models, and noisy dataset (NFOMD) when adopting ShuffleNetV2 as the backbone, which illustrates that BiInNet can be applied to velocity inversion tasks of different geological structures and is robust to noise. Adopting transfer learning to fine-tune pretrained model, BiInNet is effectively applicable to velocity reversal models and field data, which further demonstrates the reliability of the proposed method and provides a practical velocity inversion scheme when the field data pairs are insufficient.

Surface nuclear magnetic resonance (sNMR) can directly and quantitatively detect groundwater, but its application in urban environments faces problems, such as low signal-to-noise ratios (SNRs) and difficulties in laying the coils. This study presents a rotational sNMR measurement scheme to accurately image a frozen urban lake. Through synthetic data experiments, we first demonstrate that sNMR data measured with six rotations can accurately invert underground water-bearing structures. Even when the environmental noise is high, this scheme can reflect the distribution of the water content in a frozen lake. Moreover, due to the small coil size, the inversion result is less affected by the underground resistivity. In field experiments, a large amount of high-quality sNMR data with average SNRs up to 12.8 dB were obtained from a high-noise environment using three reference coils. The 2-D distributions of the water content in the ice, water, and mud layers of the frozen lake were determined using the data measured from six rotations. The water content in the lake was found to be approximately equal to 1 m³/m³. Although there are still some problems with the measurements, such as inaccurate relaxation times and low resolutions in deep areas, further improvements in sNMR and the rotational detection scheme can facilitate the application of this approach to urban groundwater detection.

We propose a framework that estimates the inundation depth (maximum water level) and debris-flow-induced topographic deformation from remote sensing imagery by integrating deep learning and numerical simulation. A water and debris-flow simulator generates training data for various artificial disaster scenarios. We show that regression models based on Attention U-Net and LinkNet architectures trained on such synthetic data can predict the maximum water level and topographic deformation from a remote sensing-derived change detection map and a digital elevation model. The proposed framework has an inpainting capability, thus mitigating the false negatives that are inevitable in remote sensing image analysis. Our framework breaks limits of remote sensing and enables rapid estimation of inundation depth and topographic deformation, essential information for emergency response, including rescue and relief activities. We conduct experiments with both synthetic and real data for two disaster events that caused simultaneous flooding and debris flows and demonstrate the effectiveness of our approach quantitatively and qualitatively. Our code and data sets are available at https://github.com/nyokoya/dlsim.

Fault detection is a crucial task in seismic structure interpretation. Convolutional neural network (CNN)-based methods, in general, require large amount of labeled data for network training. One way to build the labeled data is to create synthetic seismic images with corresponding fault labels. However, it is hard to ensure that the synthetic data have the same fault feature distributions as the field data, which may lead to inaccurate and unreliable prediction results. Another way is to manually label the faults, which is time-consuming and subjective. In this letter, we propose that using knowledge distillation (KD) to improve the performance of fault detection by integrating the features from large number of synthetic samples and a small number of field samples. We distill knowledge from an ensemble of two teacher CNNs to train a student CNN (applied to final target) for seismic fault detection. In our work, one segmentation teacher CNN is trained on synthetic samples with known ground truth fault labels and another classification teacher CNN is trained on field samples with manually picked labels. Then, a classification student network is trained on samples generated by voting the results from two teacher models. The student CNN learns not only the general fault characteristics in the synthetic data but also the specific fault features of the target field data. Test on the field data shows that the student CNN highlights seismic fault more accurately with higher resolution than the teacher CNNs.

Traffic data collected from sensor networks often exhibit strong spatial correlations and recurrent temporal patterns. Learning these patterns and diagnosing anomalies in such spatiotemporal traffic data is critical to improving transportation systems and services. This paper proposes a dynamic framework to model spatiotemporal traffic data, with a particular application on diagnosing anomalies. Within the framework, we focus on characterizing the variation in system dynamics with a time-varying vector autoregressive model. We impose a low-rank tensor structure to model the collection of time-varying system matrices. As the temporal factor matrix captures the principal patterns/signatures across all time-varying system matrices, it is a useful tool to diagnose abnormal generative mechanisms and unexpected temporal patterns. We demonstrate the proposed tensor learning framework's effectiveness by experimenting with a synthetic data set and real-world spatiotemporal traffic speed data set. The results show the superiority of the proposed model in uncovering anomalous traffic network dynamics.

In many real-world settings, such as healthcare, machine learning models are trained and validated on one labeled domain and tested or deployed on another where feature distributions differ, i.e., there is covariate shift. When annotations are costly or prohibitive, an unsupervised domain adaptation (UDA) regime can be leveraged requiring only unlabeled samples in the target domain. Existing UDA methods are unable to factor in a model's predictive loss based on predictions in the target domain and therefore suboptimally leverage density ratios of only the input covariates in each domain. In this work we propose a model selection method for leveraging model predictions on a target domain without labels by exploiting the domain invariance of causal structure. We assume or learn a causal graph from the source domain, and select models that produce predicted distributions in the target domain that have the highest likelihood of fitting our causal graph. We thoroughly analyze our method under oracle knowledge using synthetic data. We then show on several real-world datasets, including several COVID-19 examples, that our method is able to improve on the state-of-the-art UDA algorithms for model selection.

Prediction of events such as part replacement and failure events plays a critical role in reliability engineering. Event stream data are commonly observed in manufacturing and teleservice systems. Designing predictive models for individual units based on such event streams is challenging and an under-explored problem. In this work, we propose a non-parametric prognostic framework for individualized event prediction based on the inhomogeneous Poisson processes with a multivariate Gaussian convolution process (MGCP) prior on the intensity functions. The MGCP prior on the intensity functions of the inhomogeneous Poisson processes maps data from similar historical units to the current unit under study which facilitates sharing of information and allows for analysis of flexible event patterns. To facilitate inference, we derive a variational inference scheme for learning and estimation of parameters in the resulting MGCP modulated Poisson process model. Experimental results are shown on both synthetic data as well as real-world data for fleet based event prediction.

The process of seismic multiple removal in oil seismic exploration is crucial for the imaging of underground structures with primary reflections. The inclusion of prediction and subtraction in the multiple removal method requires adaptive subtraction to remove the complex differences between the true and modeled multiples. The traditional adaptive subtraction method is generally expressed as a linear regression (LR) problem. In this article, we introduce U-net, a popular deep learning tool, to represent the complex differences between the true and modeled multiples in a nonlinear relationship. Thus, we present adaptive subtraction as a non-LR problem. The modeled multiples and full recorded seismic response with multiples and primaries are used as the input and labels to train U-net. The proposed U-net method is able to avoid over-fitting of the primaries due to the sufficient number of 2-D data windows for the training of U-net, as well as the network parameter regularization and the L1 norm minimization constraint on the primaries. The proposed U-net method attains 20.5 dB and 2.5 dB improvement in the signal-to-noise ratio (SNR) using the first synthetic data and second synthetic (Sigsbee2B) data set, respectively, compared with the traditional LR method, and a qualitative improvement for a real data set test.

Industrial prognosis refers to the prediction of failures of an industrial asset based on data collected by Internet of Things sensors. Prognostic models can experience the undesired effects of concept drift, namely, the presence of nonstationary phenomena that affects the data collected over time. Consequently, fault patterns learned from data become obsolete. To overcome this issue, contextual and operational changes must be detected and managed, triggering rapid model adaptation mechanisms. This article presents an adaptive learning approach based on a dendritic cell algorithm for drift detection and a deep neural network model that dynamically adapts to new operational conditions. A kernel density estimator with drift-based bandwidth is used to generate synthetic data for a faster adaptation, focusing on fine-tuning the lowest neural layers. Experimental results over a real-world industrial problem shed light on the outperforming behavior of the proposed approach when compared to other drift detectors and classification models.

In order to function in unstructured environments, robots need the ability to recognize unseen objects. We take a step in this direction by tackling the problem of segmenting unseen object instances in tabletop environments. However, the type of large-scale real-world dataset required for this task typically does not exist for most robotic settings, which motivates the use of synthetic data. Our proposed method, UOIS-Net, separately leverages synthetic RGB and synthetic depth for unseen object instance segmentation. UOIS-Net is comprised of two stages: first, it operates only on depth to produce object instance center votes in 2D or 3D and assembles them into rough initial masks. Secondly, these initial masks are refined using RGB. Surprisingly, our framework is able to learn from synthetic RGB-D data where the RGB is non-photorealistic. To train our method, we introduce a large-scale synthetic dataset of random objects on tabletops. We show that our method can produce sharp and accurate segmentation masks, outperforming state-of-the-art methods on unseen object instance segmentation. We also show that our method can segment unseen objects for robot grasping., Comment: Journal version of arXiv:1907.13236