Reconstruction Error Research Articles

Unsupervised fabric defect detection based on multiscale image reconstruction and structural similarity assessment

AbstractFabric defect detection is a crucial aspect of the textile industry. Currently, deep learning methods have demonstrated exceptional performance in fabric defect detection tasks. However, their performance is greatly affected by the number of defect samples, which is a challenge to obtain during actual production. To address this issue, this article proposes an unsupervised anomaly detection method for fabric defects using image reconstruction networks. This method only requires defect‐free samples for training. During the training phase, the model compresses defect‐free samples to obtain a low‐dimensional manifold and reconstruct them. During the inference phase, the method assesses whether a sample is defective by calculating the reconstruction error between the input and output images, and locates the defect region by computing the difference in various patches. Furthermore, since fabric contains rich texture features, with high correlation between neighbouring pixels, a structure similarity index measure combined with mean absolute error is introduced to evaluate the reconstruction error, which enhances the model's representation ability for defect‐free samples. Additionally, considering the diverse texture backgrounds in fabric, a multiscale reconstruction module is designed to optimise the reconstruction effect. Experimental results demonstrate that compared with other related approaches, the proposed method achieves high accuracy (image‐based area under the curve (AUC) up to 98.2% and pixel‐based AUC up to 97.3%) on multiple datasets and has good generalisation ability for different fabric textures.

Anomaly detection method based on temporal spatial information enhancement

The proposed Anomaly Detection Method Based on Temporal Spatial Information Enhancement addresses the limitations of unsupervised techniques in detecting abnormal events beyond boundaries and limited samples. It incorporates a Serial Depth Separable Residual Block (Serial Block) as the backbone for predicting future frame. Additionally, a DenseReserve Subsample Module (DRSM) facilitates feature scale scaling, and a U-shaped Pyramid Attention Module (UPAM) guides feature fusion and enhances spatial details. During the prediction stage, reconstructed optical flow information aids in distinguishing abnormal and normal event features, with the abnormal score determined by a weighted fusion of optical flow reconstruction error and prediction error for future frames. Experimental results demonstrate the method’s outstanding performance, achieving area under the curve metrics of 99.7%, 92.1%, and 78.3% on UCSD Ped2, CUHK Avenue, and ShahanghaiTech datasets, respectively. This method offers significant advancements in detecting anomalous events in complex and multi-scene surveillance videos, with improved operational speed and reduced parameters.

MRFM: A timely detection method for DDoS attacks in IoT with multidimensional reconstruction and function mapping

To address the slow response time of existing detection modules to the Internet of Things (IoT) Distributed Denial of Service (DDoS) attacks, along with their low feature differentiation and poor detection performance, we propose MRFM, a timely detection method with multidimensional reconstruction and function mapping. Firstly, we employ a queue mechanism to capture and store incoming network traffic data within a predefined time frame. Subsequently, we introduce a multidimensional reconstruction neural network model, specifically designed to reconstruct quantitative features based on their respective indices by adjusting the loss function. This process is followed by the computation of multidimensional reconstruction errors and the transformation of vectors into mapping features, thereby augmenting the disparities among various types of traffic data and promoting the similarity within the same category of traffic data. Lastly, we extract frequency information from the qualitative feature matrix using information entropy calculations, enriching the feature profile of individual traffic instances. The experimental results on two benchmark datasets show that MRFM can effectively detect different types of DDoS attacks. Notably, MRFM consistently outperforms other existing methods, exhibiting an average metric improvement of up to 9.61 %.

Feasibility of Unobtrusively Estimating Blood Pressure Using Load Cells under the Legs of a Bed.

The ability to monitor blood pressure unobtrusively and continuously, even during sleep, may promote the prevention of cardiovascular diseases, enable the early detection of cardiovascular risk, and facilitate the timely administration of treatment. Publicly available data from forty participants containing synchronously recorded signals from four force sensors (load cells located under each leg of a bed) and continuous blood pressure waveforms were leveraged in this research. The focus of this study was on using a deep neural network with load-cell data as input composed of three recurrent layers to reconstruct blood pressure (BP) waveforms. Systolic (SBP) and diastolic (DBP) blood pressure values were estimated from the reconstructed BP waveform. The dataset was partitioned into training, validation, and testing sets, such that the data from a given participant were only used in a single set. The BP waveform reconstruction performance resulted in an R2 of 0.61 and a mean absolute error < 0.1 mmHg. The estimation of the mean SBP and DBP values was characterized by Bland-Altman-derived limits of agreement in intervals of [-11.99 to 15.52 mmHg] and [-7.95 to +3.46 mmHg], respectively. These results may enable the detection of abnormally large or small variations in blood pressure, which indicate cardiovascular health degradation. The apparent contrast between the small reconstruction error and the limit-of-agreement width owes to the fact that reconstruction errors manifest more prominently at the maxima and minima, which are relevant for SBP and DBP estimation. While the focus here was on SBD and DBP estimation, reconstructing the entire BP waveform enables the calculation of additional hemodynamic parameters.

Reconstruction error‐based fault detection of time series process data using generative adversarial auto‐encoders

AbstractFaults in time series process data are typically difficult to detect due to the complex temporal correlations of data samples. In this context, traditional unsupervised machine learning algorithms, such as principal component analysis, independent component analysis, and so forth, would yield only limited performance. Deep learning‐based methods have been employed in recent years to address these problems. Recently, generative adversarial networks have emerged as a promising generative modelling approach for learning data distributions. Inspired by the above, in this study, we present a novel reconstruction error‐based fault detection method for time series process data using generative adversarial auto‐encoder (GAAE) with Wasserstein loss, cycle consistency loss, and gradient penalty methods. The proposed method is designed to detect abnormal patterns in the time series data by training a GAAE to learn the underlying normal behaviour of the data. GAAEs help in effectively capturing the data's hidden distribution, and Wasserstein loss with gradient penalty is used to improve the accuracy of the latent space representation of the data, while the cycle consistency loss ensures consistency between the input and output data during the reconstruction process. Based on the extent of the reconstruction error metric of the GAAEs, we identify the potential faults in the process data stream. The proposed method is evaluated on two data process sets, namely, the Tennessee Eastman benchmark process dataset and the nuclear power flux real‐time dataset from a pressurized heavy water nuclear reactor, to validate the efficacy of the proposed approach.

Approximating the zero-norm penalized sparse signal recovery using a hierarchical Bayesian framework

Sparse Bayesian learning (SBL) with its self-regulation and uncertainty estimation features, has become a popular topic in the field of sparse signal recovery. However, it is a challenging problem to employ zero-norm penalty on signals in the hierarchical framework of SBL. In this paper, an SBL approach that approximates the zero-norm penalized spare signal recovery is proposed based on the idea that signals can be modeled using generalized Gaussian distribution (GGDs). It should be noted that the mean and variance of GGDs are related to its shape parameters, making the direct use of GGDs in Bayesian inference impractical. To solve this problem, we derives a proposition and further build a new hierarchical Bayesian framework (HrBayFw). The proposed HrBayFw is equivalent to assign ℓp-norm penalty to signals, and the parameter p can be designed to approximate 0. Thus, the zero-norm penalized spare signal recovery can be realized using SBL, and the use of the proposition enables tractable marginalization over all parameters. The main advantage of the proposed approach is that it achieves a lower reconstruction error than current Bayesian methods. Numerical simulations evidence that the proposed SBL approach achieves better accuracy performance in terms of normalized-mean-square-error comparing with other contemporary algorithms.

Dynamic Feature Extraction-Based Quadratic Discriminant Analysis for Industrial Process Fault Classification and Diagnosis.

This paper introduces a novel method for enhancing fault classification and diagnosis in dynamic nonlinear processes. The method focuses on dynamic feature extraction within multivariate time series data and utilizes dynamic reconstruction errors to augment the feature set. A fault classification procedure is then developed, using the weighted maximum scatter difference (WMSD) dimensionality reduction criterion and quadratic discriminant analysis (QDA) classifier. This method addresses the challenge of high-dimensional, sample-limited fault classification, offering early diagnosis capabilities for online samples with smaller amplitudes than the training set. Validation is conducted using a cold rolling mill simulation model, with performance compared to classical methods like linear discriminant analysis (LDA) and kernel Fisher discriminant analysis (KFD). The results demonstrate the superiority of the proposed method for reliable industrial process monitoring and fault diagnosis.

Bridge progressive damage detection using unsupervised learning and self-attention mechanism

This paper presents a novel unsupervised method for detecting progressive damage in bridge structures. Traditional supervised learning approaches require a large number of damaged samples to train a high-performing detection model, which poses challenges when applied to normal in-service bridges. Since in-service bridges usually lack data from damage scenarios and existing methods fail to achieve real-time detection, unsupervised learning has gained popularity due to its ability to work without damaged samples and offer better real-time performance. In light of this, this paper proposes a novel unsupervised damage detection method that utilizes an unsupervised attention-convolutional auto-encoder to reconstruct real-time vibration signals from bridge. Two damage indicators are employed to evaluate the effectiveness of the reconstruction process, aiming to capture potential changes in the progressive damage of bridges. The results demonstrate that: the attention-convolutional auto-encoder achieves excellent reconstruction performance for vibration signals from intact structures, which reconstruction error converging to 0 and structural similarity approaching 1. However, the reconstruction performance is less satisfactory for vibration signals in damage scenarios, deteriorating further as the damage level increases. By analyzing the variation patterns of vibration signals in attention-convolutional auto-encoders, it was confirmed that the reliability of the proposed evaluation effects by the two different types of damage indicators. Finally, the trend of progressive damage was extracted through these indicators, showcasing the effectiveness of the method in detecting progressive damage. This approach facilitates the identification of potential progressive changes in the normal operation process of bridges, thus enhancing real-time damage detection capabilities and contributing to the knowledge of bridge monitoring.

Using Gaussian process for velocity reconstruction after coronary stenosis applicable in positron emission particle tracking: An in-silico study

Accurate velocity reconstruction is essential for assessing coronary artery disease. We propose a Gaussian process method to reconstruct the velocity profile using the sparse data of the positron emission particle tracking (PEPT) in a biological environment, which allows the measurement of tracer particle velocity to infer fluid velocity fields. We investigated the influence of tracer particle quantity and detection time interval on flow reconstruction accuracy. Three models were used to represent different levels of stenosis and anatomical complexity: a narrowed straight tube, an idealized coronary bifurcation with stenosis, and patient-specific coronary arteries with a stenotic left circumflex artery. Computational fluid dynamics (CFD), particle tracking, and the Gaussian process of kriging were employed to simulate and reconstruct the pulsatile flow field. The study examined the error and uncertainty in velocity profile reconstruction after stenosis by comparing particle-derived flow velocity with the CFD solution. Using 600 particles (15 batches of 40 particles) released in the main coronary artery, the time-averaged error in velocity reconstruction ranged from 13.4% (no occlusion) to 161% (70% occlusion) in patient-specific anatomy. The error in maximum cross-sectional velocity at peak flow was consistently below 10% in all cases. PEPT and kriging tended to overestimate area-averaged velocity in higher occlusion cases but accurately predicted maximum cross-sectional velocity, particularly at peak flow. Kriging was shown to be useful to estimate the maximum velocity after the stenosis in the absence of negative near-wall velocity.

Nonferrous metal price forecasting based on signal decomposition and ensemble learning

Nonferrous metals are indispensable raw materials for modern industry. The price forecasting of nonferrous metals is vital for business operators and investors. Based on the decomposition-integration framework, we propose a signal decomposition model combining variational mode decomposition (VMD) and an improved long-short time memory (LSTM) network. Using the MAE metric as a benchmark, the improved LSTM model (Mogrifier LSTM) obtained an average accuracy improvement of 5.99%. VMD is an efficient decomposition algorithm. However, it needs to set hyperparameters in advance. Unreasonable parameters will lead to poor decomposition results. Therefore, a method based on subseries complexity and reconstruction error (CAE) is proposed to reasonably decompose signals, improving 21.13% accuracy and reducing 37.56% computational overhead than other strategies. The structural model is introduced as a complement to the signal decomposition model, which learns different features by incorporating theoretical analyses into the choice of explanatory variables. The combining of two models achieves effective complementarity, obtaining an average accuracy improvement of 7.43%. Comparative tests on three datasets demonstrate the superiority of the proposed prediction framework. On the one hand, a reasonable decomposition strategy can play an essential role in the signal decomposition model. On the other hand, improving the prediction model and integrating different models is also an effective strategy to enhance accuracy.

A deep learning approach to the automatic detection of alignment errors in cryo-electron tomographic reconstructions

Electron tomography is an imaging technique that allows for the elucidation of three-dimensional structural information of biological specimens in a very general context, including cellular in situ observations. The approach starts by collecting a set of images at different projection directions by tilting the specimen stage inside the microscope. Therefore, a crucial preliminary step is to precisely define the acquisition geometry by aligning all the tilt images to a common reference. Errors introduced in this step will lead to the appearance of artifacts in the tomographic reconstruction, rendering them unsuitable for the sample study. Focusing on fiducial-based acquisition strategies, this work proposes a deep-learning algorithm to detect misalignment artifacts in tomographic reconstructions by analyzing the characteristics of these fiducial markers in the tomogram. In addition, we propose an algorithm designed to detect fiducial markers in the tomogram with which to feed the classification algorithm in case the alignment algorithm does not provide the location of the markers. This open-source software is available as part of the Xmipp software package inside of the Scipion framework, and also through the command-line in the standalone version of Xmipp.

Reconstructing Filaments around Galaxy Clusters from Spectroscopic Surveys using Machine Learning

Next-generation surveys such as the WEAVE Wide-Field Cluster Survey will soon give astronomers an unprecedented opportunity to study cosmic web structure and filamentary populations around clusters. Analysis of classical methods of extracting the cosmic web from simulated 2D projections has revealed significant incompleteness and contamination. In this note, we present the first results from a random forest trained and tested on the dark-matter simulation MDPL2. Our algorithm improves the precision of filament classification by 11% and decreases the structural reconstruction error by 43% compared to the previously published method.

Determination of quantity and volume of Carya cathayensis Sarg by line laser scanning combined with the point cloud fusion algorithm

Optical 3D measurement technology plays a vital role in diverse industries, particularly with the advancements in line laser scanning 3D imaging. In this paper, we propose a line laser scanning-based investigation for detecting Carya cathayensis Sarg. The Carya cathayensis Sarg specimens are scanned using a line laser to achieve three-dimensional reconstruction, enabling the calculation of their volume and quantity based on the acquired point cloud map. Through binocular acquisition and subsequent point cloud alignment and fusion, the error in the three-dimensional reconstruction is significantly reduced. The point cloud map facilitates the automatic identification of the number of scanned areas of Carya cathayensis Sarg areas and accurate volume calculations, with an error control of approximately 0.6% when compared to the actual volume. The application of this research in agriculture allows farmers to classify fruit sizes and optimize their selection, thus facilitating intelligent agricultural practices.

Hybrid treatment verification based on prompt gamma rays and fast neutrons: multivariate modelling for proton range determination

Robust and fast in vivo treatment verification is expected to increase the clinical efficacy of proton therapy. The combined detection of prompt gamma rays and neutrons has recently been proposed for this purpose and shown to increase the monitoring accuracy. However, the potential of this technique is not fully exploited yet since the proton range reconstruction relies only on a simple landmark of the particle production distributions. Here, we apply machine learning based feature selection and multivariate modelling to improve the range reconstruction accuracy of the system in an exemplary lung cancer case in silico. We show that the mean reconstruction error of this technique is reduced by 30%–50% to a root mean squared error per spot of 0.4, 1.0, and 1.9 mm for pencil beam scanning spot intensities of 108, 107, and 106 initial protons, respectively. The best model performance is reached when combining distribution features of both gamma rays and neutrons. This confirms the advantage of hybrid gamma/neutron imaging over a single-particle approach in the presented setup and increases the potential of this system to be applied clinically for proton therapy treatment verification.

Deep learning-based anomaly detection for individual drone vehicles performing swarm missions

This study explores methods for the detection and identification of anomalous instances in individual drone vehicles when performing missions in a swarm formation. Conventional anomaly detection (AD) in unmanned aerial vehicle (UAV) clusters typically involves manual inspection, which is time and resource-inefficient, followed by machine learning techniques. In this study, a novel machine-learning-based framework was proposed for the automatic detection of anomalous individual drones within a swarm and the rapid identification of faulty channels. Considering the imbalance between normal and abnormal states when using real flight data, semi-supervised models were selected. Four models with one-dimensional (1D) convolutions were trained on normal data. These models were based on a variational autoencoder and three popular AD-specific models (AnoGAN, GANomaly, and Skip-GANomaly), and their performances considering several metrics were compared. Data preprocessing was performed according to various scaling methods, and the hyperparameters that affect the training process were determined through Bayesian optimization. After training, AD was performed using a two-step process. First, detection through binary classification was performed by generating a reconstruction of the testing data and thresholding the reconstruction error. After detection, if the data were determined to be abnormal, each of the 16 channels was ranked in terms of its probability of being the source of the anomaly. The proposed scheme for detecting anomalies was tested and verified using real-world flight data, and analysis of the results revealed the major types of faults and identified specific abnormal channels. This has implications on future research of implementing necessary responses to abnormal readings to maintain UAV autonomy.

A likelihood based method for compressive signal recovery under Gaussian and saturation noise

Practical sensing systems have a limited dynamic range, which gives rise to saturation noise in signal acquisition if the signal attains values outside this range at certain locations. Despite rapid advances in compressed sensing theory and systems over the years, the effect of saturation on compressed sensing measurements, especially in the presence of additive Gaussian noise, has often been overlooked in many existing approaches. Some methods for compressive recovery under saturation noise simply discard saturated measurements, while others impose some ad hoc hard constraints to ensure consistency with the known saturation threshold. In this paper, we propose a novel maximum likelihood based estimator for compressive reconstruction in the presence of both Gaussian and saturation noise. Our proposed method ensures probabilistic consistency between the estimated signal and the saturated measurements and helps account for both the saturation effect and the potentially unbounded Gaussian noise in the measurements. We also derive upper bounds on the reconstruction error for our estimator, and argue that they follow intuitive trends. We analyze an important curvature term in these bounds, and show that it is superior to methods such as saturation rejection, indicating better stability and robustness. Furthermore, we present extensive simulation results to demonstrate the effectiveness of our method over several other existing methods in reconstructing synthetic signals and images from compressive measurements with saturation and Gaussian noise. We also show simulation results in audio signal de-clipping.

Learning joint surface reconstruction and segmentation, from brain images to cortical surface parcellation.

Reconstructing and segmenting cortical surfaces from MRI is essential to a wide range of brain analyses. However, most approaches follow a multi-step slow process, such as a sequential spherical inflation and registration, which requires considerable computation times. To overcome the limitations arising from these multi-steps, we propose SegRecon, an integrated end-to-end deep learning method to jointly reconstruct and segment cortical surfaces directly from an MRI volume in one single step. We train a volume-based neural network to predict, for each voxel, the signed distances to multiple nested surfaces and their corresponding spherical representation in atlas space. This is, for instance, useful for jointly reconstructing and segmenting the white-to-gray-matter interface and the gray-matter-to-CSF (pial) surface. We evaluate the performance of our surface reconstruction and segmentation method with a comprehensive set of experiments on the MindBoggle, ABIDE and OASIS datasets. Our reconstruction error is found to be less than 0.52 mm and 0.97 mm in terms of average Hausdorff distance to the FreeSurfer generated surfaces. Likewise, the parcellation results show over 4% improvements in average Dice with respect to FreeSurfer, in addition to an observed drastic speed-up from hours to seconds of computation on a standard desktop station.

Cosine 2DPCA With Weighted Projection Maximization.

Two-dimensional principal component analysis (2DPCA), known to be very sensitive to outliers, employs the square F -norm as the distance metric and only satisfies the optimal objective of maximizing projection variance. However, the objective of minimizing reconstruction errors for all samples is not optimized as much as possible. To handle the problem, a novel cosine objective function is first presented for maximizing weighted projection, in which the 2-norm of vectors with an adjustable power parameter is employed as the distance metric. Not only the objective with the maximum projection distance is accomplished in the cosine objective function, but also the objective with the minimum sum of reconstruction errors is also optimized indirectly. Then, the cosine 2DPCA (Cos-2DPCA) method is proposed, and the greedy iterative algorithm to solve Cos-2DPCA is also developed. The convergence and correlation of solutions are proved theoretically and discussed in detail. Finally, the series of experiments are carried out on the artificial dataset and eight standard datasets, respectively. The results demonstrate that the performances of Cos-2DPCA are significantly improved on the reconstruction, correlation, complexity, and classification, and it outperforms most of the existing robust 2DPCA methods.

Efficient Feature Reconstruction via l 2,1-Norm Regularization for Few-Shot Classification

Few-shot classification, which aims to recognize novel classes with the aid of learned knowledge and a few labeled samples, remains challenging and draws emerging concerns in computer vision. Metric-based methods incorporate the idea of metric learning into tackling the few-shot classification problem, i.e., the class membership of query samples can be determined in the latent space. Several recent methods propose to infer such memberships by feature reconstruction, which, however either exploit attention mechanisms or closed-form solutions. In this paper, we revisit the essence of feature reconstruction applied to few-shot classification and introduce a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">l</i> <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2,1</sub> -norm regularization to constraint it. As with the intention of attention mechanisms, we attempt to guide feature reconstruction towards focusing more on semantically rich target regions and diminishing the contribution of profitless features. We thereby obtain more discriminative query reconstructions for each class and perform classification based on these reconstruction errors. We formulate the proposal as a constrained optimization problem and design a simple and efficient iterative method as a solution. Furthermore, we provide the analysis on the convergence of proposed iterative method in theory. We conduct few-shot classification on two fine-grained and two general datasets. Extensive experimental results reveal that our method achieves new state-of-the-art performance.

Locally-Enriched Cross-Reconstruction for Few-Shot Fine-Grained Image Classification

Few-shot fine-grained image classification has attracted considerable attention in recent years for its realistic setting to imitate how humans conduct recognition tasks. Metric-based few-shot classifiers have achieved high accuracies. However, their metric function usually requires two arguments of vectors, while transforming or reshaping three-dimensional feature maps to vectors can result in loss of spatial information. Image reconstruction is thus involved to retain more appearance details: the test images are reconstructed by different classes and then classified to the one with the smallest reconstruction error. However, discriminative local information, vital to distinguish sub-categories in fine-grained images with high similarities, is not well elaborated when only the base features from a usual embedding module are adopted for reconstruction. Hence, we propose the novel local content-enriched cross-reconstruction network (LCCRN) for few-shot fine-grained classification. In LCCRN, we design two new modules: the local content-enriched module (LCEM) to learn the discriminative local features, and the cross-reconstruction module (CRM) to fully engage the local features with the appearance details obtained from a separate embedding module. The classification score is calculated based on the weighted sum of reconstruction errors of the cross-reconstruction tasks, with weights learnt from the training process. Extensive experiments on four fine-grained datasets showcase the superior classification performance of LCCRN compared with the state-of-the-art few-shot classification methods. Codes are available at: https://github.com/lutsong/LCCRN.