Sparse Representation-based Classification Research Articles

Fault diagnosis of driving gear in a battery swapping system based on audio features and SRC-Adaboost

Abstract Because of electrification conditions, key components of battery swapping systems (BSSs) for electric heavy trucks are always damaged by electric erosion, which presents challenges to the safety and efficiency of high-intensity transportation. Due to the special working conditions of a BSS, the fault diagnosis of its driving gear encounters several issues, including reciprocation motion, low and fluctuating speed, and complicated noises. To solve these problems, audio features, including Mel-frequency cepstral coefficients and Gammatone cepstral coefficients, are extracted from the vibration signals. Then, these features are utilized to construct an original dictionary. Next, based on data augmentation and dictionary learning, a robust dictionary is generated from the original dictionary. Finally, with the robust dictionary, sparse representation-based classification is integrated into AdaBoost to achieve accurate fault diagnosis for the driving gear in BSS. The effectiveness of the fault diagnosis scheme is validated based on the monitoring data of the BSS, and the accuracy of fault diagnosis is 99.17%.

Fisher discrimination multiple kernel dictionary learning for robust identification of nonlinear features in machinery health monitoring

In sparse representation-based classification, Fisher discrimination dictionary learning (FDDL) has attracted widespread attention due to its advantages such as fewer parameters and good dictionary discriminability. However, due to its linear nature, it is difficult to handle nonlinear features. Kernel-based nonlinear transformation enables dictionary learning to capture nonlinear features embedded in samples, but traditional single-kernel methods have limited performance. In this paper, a Fisher discrimination multiple kernel dictionary learning (FDMKDL) method is proposed, in which Fisher discriminative criterion is imposed on both high-dimensional samples and coding coefficients. Specifically, we derive the kernel version of FDDL, i.e., Fisher discrimination kernel dictionary learning (FDKDL) to learn nonlinear features and promote dictionary discriminability. Meanwhile, to avoid the problem of poor adaptability and possible manual selection caused by single-kernel methods, we further derive a flexible multiple kernel learning (MKL) framework, which utilizes the complementary information of multiple kernel functions to adaptively obtain the optimal weights and synthesize a discriminative kernel space. Finally, FDMKDL combines FDKDL with this MKL framework to obtain a more discriminative dictionary. Experimental evaluations conducted on two datasets demonstrate the effectiveness and robustness of the proposed FDMKDL method in learning nonlinear features for machinery health monitoring when compared to several state-of-the-art methods.

Sparse representation-based motor imagery EEG classification towards asynchronous BCI systems

Sparse representation-based motor imagery EEG classification towards asynchronous BCI systems

Weight thresholded regularized robust coding for vision sensing-based hand gesture detection in collaborative robotics

Vision sensing-based hand gesture recognition is considered an important contemporary research problem of collaborative robotics in a human-robot coexisting environment. The problem becomes more complex in challenging environments, where a robot needs to correctly recognize a human gesture to perform navigation or some other designated job, even in the presence of poor illumination conditions, occlusion, etc. This work proposes a novel approach of successfully utilizing regularized robust coding (RRC)-based models to solve such hand gesture detection problems in real-life, challenging situations. The RRC model is used for robust regression of a signal or image and is known as an improvement over the classical sparse representation-based classification (SRC) model. In this work, we propose three novel variants of weight thresholding mechanism in conjunction with RRC (named WTRRC algorithms) which essentially employ more error-tolerant concepts of logistic functions for weight update, when the dictionary is formed by real-world, photometrically irregular hand gesture images. Extensive case studies in real-world environments (i) with poor illumination, and (ii) with both poor illumination and occlusions firmly establish superior performances of the proposed WTRRC variants compared to other state-of-the-art algorithms in such collaborative robotics problems.

Identification of DNA N4-methylcytosine Sites via Multiview Kernel Sparse Representation Model

Identifying DNA N4-methylcytosine (4mC) sites is of great significance in biological research, such as chromatin structure, DNA stability, DNA-protein interaction and controlling gene expression. However, the traditional sequencing technology to identify 4mC sites is very time-consuming. In order to detect 4mC sites, we develop a multi-view learning method for achieving more effectively via merging multiple feature spaces. Furthermore, we think about whether the multi-view learning method can improve the across species classification ability by fusing data of multiple species. In our study, we propose a multi-view Laplacian kernel sparse representation-based classifier, called MvLapKSRC-HSIC. First, we make use of three feature extraction methods (PSTNP, NCP, DPP) to extract the DNA sequence features. MvLapKSRC-HSIC uses a kernel sparse representation-based classifier with graph regularization. In order to maintain the independence between various views, we add a multi-view regularization term constructed by Hilbert-Schmidt independence criterion (HSIC). In the experiments, MvLapKSRC-HSIC is applied on six datasets, so as to compare with other popular methods in single species and cross-species experiments. All experimental results show that MvLapKSRC-HSIC is superior to other outstanding methods on both single species and cross-species. Importantly, MvLapKSRC-HSIC can identify a series of potential DNA 4mC sites, which have not yet been experimentally evaluate on multiple species and merit further research.

A new method for predicting the prognosis of ischemic stroke based vascular structure features and lesion location features

ObjectiveDetermining the changes in the prognosis of the cerebral infarction area has an important guiding role in the selection of the treatment plan. The goal of this study is to propose a machine learning-based method that can predict the prognosis of stroke effectively and efficiently. Methods97 cases of stroke were analyzed retrospectively. Firstly, we extracted vascular structural features from computed tomography angiography (CTA) images and stroke location features from diffusion-weighted imaging (DWI) images to comprehensively characterize the lesions, respectively. Then, we performed sparse representation-based feature selection and classification to predict the prognosis of stroke based on the extracted features. Finally, we randomly divided the 97 cases into cross-validation set, independent testing set 1 and independent testing set 2 to validate the proposed model. Results464 vascular structure features and 116 positional features were extracted. After feature selection, 52 features were finally applied to build the classification model. The proposed model achieved promising prediction performance on the two independent testing sets, with the classification accuracies of 85.19% and 81.25%, respectively. ConclusionThe proposed machine learning approach can effectively mine and accurately quantify the features related to the prognosis, which include the vascular structural features and the stroke location features. In addition, the established prognostic prediction model based on these features has achieved interesting performances, which may provide valuable guidance for the clinical treatment of stroke.

Dictionary reduction in sparse representation-based classification of motor imagery EEG signals

Dictionary reduction in sparse representation-based classification of motor imagery EEG signals

18F-FDG PET and a classifier algorithm reveal a characteristic glucose metabolic pattern in adult patients with moyamoya disease and vascular cognitive impairment.

Vascular cognitive impairment (VCI) is a critical issue in moyamoya disease (MMD). However, the glucose metabolic pattern in these patients is still unknown. This study aimed to identify the metabolic signature of cognitive impairment in patients with MMD using 18F-2-fluoro-2-deoxy-D-glucose positron emission tomography (18F-FDG PET) and establish a classifier to identify VCI in patients with MMD. One hundred fifty-two patients with MMD who underwent brain 18F-FDG PET scans before surgery were enrolled and classified into nonvascular cognitive impairment (non-VCI, n = 52) and vascular cognitive impairment (VCI, n = 100) groups according to neuropsychological test results. Additionally, thirty-three health controls (HCs) were also enrolled. Compared to HCs, patients in the VCI group exhibited extensive hypometabolism in the bilateral frontal and cingulate regions and hypermetabolism in the bilateral cerebellum, while patients in the non-VCI group showed hypermetabolism only in the cerebellum and slight hypometabolism in the frontal and temporal regions. In addition, we found that the patients in the VCI group showed hypometabolism mainly in the left basal ganglia compared to those in the non-VCI group. The sparse representation-based classifier algorithm taking the SUVr of 116 Anatomical Automatic Labeling (AAL) areas as features distinguished patients in the VCI and non-VCI groups with an accuracy of 82.4%. This study demonstrated a characteristic metabolic pattern that can distinguish patients with MMD without VCI from those with VCI, namely, hypometabolic lesions in the left hemisphere played a more important role in cognitive decline in patients with MMD.

Sparse Representation Classifier Guided Grassmann Reconstruction Metric Learning With Applications to Image Set Analysis

Modeling a sequence of video frames as a linear subspace on Grassmann manifold has recently become increasingly attractive in multiple computer vision applications. The success of such algorithms largely depends on a good distance measure, and learning an appropriate metric on Grassmann manifold remains a key challenge. Existing works address this by learning a discriminative mapping from the original Grassmann manifold to Hilbert space or a lower-dimensional, more discriminative Grassmann manifold. However, these approaches always highly rely on nearest neighbor matching of samples on the projected space, which is sensitive to noises and errors. Different from them, this paper proposes a Grassmann Reconstruction Metric Learning (GRML) algorithm guided by sparse representation-based classifier (SRC) for image set classification. SRC selects the coefficients associated with each class to reconstruct training samples, and then we employ it as a criterion to direct the design of a discriminant metric on Grassmann manifold. Specifically, GRML attempts to jointly maximize the inter-class reconstruction residual and minimize the intra-class reconstruction residual in the lower but more discriminative Grassmann manifold. To further explore the intrinsic geometry distance, we present a Grassmann Reconstruction Multiple Kernel Metric Learning (GRMKML) algorithm, which aims to jointly learn a metric and the corresponding kernel from a family of kernels for Grassmann manifold. Extensive experiments on eight benchmark datasets demonstrate that the proposed algorithms perform favorably against the state-of-the-art methods.

Stable and compact face recognition via unlabeled data driven sparse representation-based classification

Sparse representation-based classification (SRC) has attracted much attention by casting the recognition problem as simple linear regression problem. SRC methods, however, still is limited to enough labeled samples per category, insufficient use of unlabeled samples, and instability of representation. For tackling these problems, an unlabeled data driven inverse projection pseudo-full-space representation-based classification model is proposed with low-rank sparse constraints. The proposed model aims to mine the hidden semantic information and intrinsic structure information of all available data, which is suitable for few labeled samples and proportion imbalance between labeled samples and unlabeled samples problems in frontal face recognition. The mixed Gauss–Seidel and Jacobian ADMM algorithm is introduced to solve the model. The convergence, representation capability and stability of the model are analyzed. Experiments on three public datasets show that the proposed LR-S-PFSRC model achieves stable results, especially for proportion imbalance of samples.

A novel multi-branch wavelet neural network for sparse representation based object classification

Recent advances in acquisition and display technologies have led to an enormous amount of visual data, which requires appropriate storage and management tools. One of the fundamental needs is the design of efficient image classification and recognition solutions. In this paper, we propose a wavelet neural network approach for sparse representation-based object classification. The proposed approach aims to exploit the advantages of sparse coding, multi-scale wavelet representation as well as neural networks. More precisely, a wavelet transform is firstly applied to the image datasets. The generated approximation and detail wavelet subbands are then fed into a multi-branch neural network architecture. This architecture produces multiple sparse codes that are efficiently combined during the classification stage. Extensive experiments, carried out on various types of standard object datasets, have shown the efficiency of the proposed method compared to the existing sparse coding and deep learning-based methods.

A study of sparse representation-based classification for biometric verification based on both handcrafted and deep learning features

Biometric verification is generally considered a one-to-one matching task. In contrast, in this paper, we argue that the one-to-many competitive matching via sparse representation-based classification (SRC) can bring enhanced verification security and accuracy. SRC-based verification introduces non-target subjects to construct dynamic dictionary together with the client claimed and encodes the submitted feature. Owing to the sparsity constraint, a client can only be accepted when it defeats almost all non-target classes and wins a convincing sparsity-based matching score. This will make the verification more secure than those using one-to-one matching. However, intense competition may also lead to extremely inferior genuine scores when data degeneration occurs. Motivated by the latent benefits and concerns, we study SRC-based verification using two sparsity-based matching measures, three biometric modalities (i.e., face, palmprint, and ear) and their multimodal combinations based on both handcrafted and deep learning features. We finally approach a comprehensive study of SRC-based verification, including its methodology, characteristics, merits, challenges and the directions to resolve. Extensive experimental results demonstrate the superiority of SRC-based verification, especially when using multimodal fusion and advanced deep learning features. The concerns about its efficiency in large-scale user applications can be readily solved using a simple dictionary shrinkage strategy based on cluster analysis and random selection of non-target subjects.

Hybrid representation based on dictionaries for hyperspectral image classification

In real-world applications, the hyperspectral image (HSI) classification faces two critical problems: one is the Hughes phenomenon and the other is the large spatial and spectral variabilities among the pixels of the HSI, especially for the case where only few labeled pixels per class are available. To address these problems, we propose a united framework for the hybrid representation model based on two dictionaries in this paper, where the input pixel (or its feature) is explicitly decomposed into the class-specific component and the variation component via the Bayesian approach. Based on the proposed framework, two hybrid representation-based classification approaches, named as the sparse and dense hybrid representation-based classification and the hybrid sparse representation-based classification, are proposed based on different priors on the separate components over the associated dictionaries. Moreover, the techniques for designing the associated dictionaries and estimating the hyperparameters are also present so the practicality of the proposed approaches is further improved. Experimental results show that the proposed two approaches outperform the conventional classifiers and are robust to the spatial and spectral variabilities, based on four real hyperspectral datasets.

Group Class Residual ℓ1-Minimization on Random Projection Sparse Representation Classifier for Face Recognition

Sparse Representation-based Classification (SRC) has been seen to be a reliable Face Recognition technique. The ℓ1 Bayesian based on the Lasso algorithm has proven to be most effective in class identification and computation complexity. In this paper, we revisit classification algorithm and then recommend the group-based classification. The proposed modified algorithm, which is called as Group Class Residual Sparse Representation-based Classification (GCR-SRC), extends the coherency of the test sample to the whole training samples of the identified class rather than only to the nearest one of the training samples. Our method is based on the nearest coherency between a test sample and the identified training samples. To reduce the dimension of the training samples, we choose random projection for feature extraction. This method is selected to reduce the computational cost without increasing the algorithm’s complexity. From the simulation result, the reduction factor (ρ) 64 can achieve a maximum recognition rate about 10% higher than the SRC original using the downscaling method. Our proposed method’s feasibility and effectiveness are tested on four popular face databases, namely AT&T, Yale B, Georgia Tech, and AR Dataset. GCR-SRC and GCR-RP-SRC achieved up to 4% more accurate than SRC random projection with class-specific residuals. The experiment results show that the face recognition technology based on random projection and group-class-based not only reduces the dimension of the face data but also increases the recognition accuracy, indicating that it is a feasible method for face recognition.

Mental State Assessment in College English Teaching Courses Based on Deep Learning

College English teaching aims at students with different foundations and characteristics. Also, it is necessary to grasp the mental state of students in time during the teaching process. Based on the bimodal information of facial expressions and speech of classroom students, this paper designs a mental state assessment method based on deep learning. For bimodal emotion recognition of facial expressions and speech, a feature fusion method based on sparse canonical correlation analysis (SCCA) is proposed in this paper. First, the emotional features of the facial expression and speech are extracted, respectively. Then, SCCA is used to fuse the emotional features of the two modalities. Finally, the sparse representation-based classification (SRC) is used as the classifier for emotional prediction. Based on the prediction results, the mental state of different students can be grasped, so as to adjust the teaching strategy in a targeted manner. Experiments are carried out based on public datasets. First, the proposed method achieves the average classification accuracy of 92.4%, which is higher than those from the present methods for comparison. Second, under the condition of noise corruption, the proposed method keeps the superior robustness over the comparison methods. The results show that the proposed bimodal emotion recognition method based on SCCA and SRC can achieve higher recognition rates than some present methods.

Dictionary learning-based damage detection under varying environmental conditions using only vibration responses of numerical model and real intact State: Verification on an experimental offshore jacket model

Monitoring structural damage is critical for preserving the service life of engineering systems. In varying operational environments, the working loads are changing all the time and they are typically unknown; in such environments, access to damage data is difficult and sometimes even impossible, and generally, intact data of the system is available. From this standpoint, this study aims to propose a novel vibration-based method for damage detection of real systems using Dictionary Learning (DL) based on a FE model and real intact state under different uncertainties such as varying working loads. A Frequency Domain Sparse Representation-based Classification (FDSRC) model is designed via DL to learn the features and damage detection of real systems. The main contribution of this study is the training process of the proposed FDSRC using the raw frequency data of a real intact system and various states of the related Finite Element (FE) model under simple working loads (one working load), which is then tested using the raw frequency data of the various states of the real system under complex working loads (varying working loads), to make more realistic assumptions and simulate the operational conditions. The raw frequency data is generated using the Frequency Domain Decomposition (FDD) method. The performance of the proposed method is validated using real measurements from an experimental offshore jacket model. The results of the study indicate that in the case of changing the working load, the proposed method can achieve higher accuracy than other comparative methods. As a result, the proposed method is robust for structural damage monitoring under different uncertainties.

Unlabeled data driven cost-sensitive inverse projection sparse representation-based classification with 1/2 regularization

Sparse representation-based classification (SRC) has been widely used because it just relies on simple linear regression ideas to do classification, and it does not need learning. However, the performance of SRC is limited by needing sufficient labeled samples per class and the sensitivity to class imbalance. For tackling these problems, an improved SRC model is constructed in this paper. For alleviating the problem of insufficient labeled samples, an unlabeled data-driven inverse projection sparse representation-based classification model is constructed to achieve effective and stable representation and recognition results. The L1/2 and S1/2 regularizations are introduced to capture the sparsity of 1-D and 2-D, and to make the model have good statistical properties. Moreover, the cost-sensitive strategy is integrated into the model’s classification criteria to improve the imbalance of class distribution adaptively, especially for multiclass imbalanced data. A solver utilizing the mixed Gauss-Seidel and Jacobian ADMM algorithm is developed to obtain the approximate solution. Experiments on common public test databases show that the proposed model achieves competitive results compared with the latest published results and some deep-learning algorithms.

Regularization on Augmented Data to Diversify Sparse Representation for Robust Image Classification.

Image classification is a fundamental component in modern computer vision systems, where sparse representation-based classification has drawn a lot of attention due to its robustness. However, on the optimization of sparse learning systems, regularization and data augmentation are both powerful, but currently isolated. We believe that regularization and data augmentation can cooperate to generate a breakthrough in robust image classification. In this article, we propose a novel framework, regularization on augmented data (READ), which creates diversification in the data using the generic augmentation techniques to implement robust sparse representation-based image classification. When the training data are augmented, READ applies a distinct regularizer, l1 or l2 , in particular, on the augmented training data apart from the original data, so that regularization and data augmentation are utilized and enhanced synchronously. We introduce an elaborate theoretical analysis on how to optimize the sparse representation by both l1 -norm and l2 -norm with the generic data augmentation and demonstrate its performance in extensive experiments. The results obtained on several facial and object datasets show that READ outperforms many state-of-the-art methods when using deep features.

Face Recognition Based on SRC Combined with Sparse Embedding Dimensionality Reduction

<p>Sparse representation-based classification (SRC) method has achieved good recognition results and shown strong robustness for face recognition, especially when the face image is affected by illumination variations, expression changes and occlusion. SRC method simply uses the training set as a dictionary to encode test samples. However, the high-dimensional training face data usually contain a large amount of redundant information, which will increase the complexity of this method. Therefore, the image dimensionality reduction procedure is separately performed by most of the existing methods before SRC is launched, but this may not be able to make full use of the discriminative information of the training samples. In this paper, based on the efficient SRC method, a sparse embedding dimensionality reduction strategy is combined with to achieve a face recognition method. For the proposed method, a projection matrix is used to project high-dimensional data into a low-dimensional space. At the same time, a discriminative coefficient constraint term in the objective function is introduced to reduce the classification residual of the sample through the distance relationship between all coefficients. Then the label information of the sample is used to iteratively update the projection matrix and coefficient representation. Finally, the test samples are projected into the low-dimensional space for classification. A large number of experimental results on three widely used face datasets show that the proposed method improves the discrimination of face images in low-dimensional space and can achieve better face recognition results.</p> <p> </p>

Application of Deep Learning in Civil Engineering Management.

Construction safety issues are of great significance in civil engineering management. In this paper, the entry point is the recognition of workers wearing helmets during the construction process, and the recognition performance is improved by combining deep learning and traditional classifiers to achieve intelligent recognition of construction safety clothing. In the specific process, the deep residual networks (ResNet) and sparse representation-based classification (SRC) are used as basic classifiers to classify samples with unknown categories. The results of the two decisions are fused and the reliability of the fused decision is determined. Afterwards, the reliable test samples are added to the original training samples to update the classifier, so as to obtain more reliable recognition results. The proposed method is tested and verified with actual measured data. The experimental results show the effectiveness and robustness of the proposed method.