Search

In a high-dimensional sample space, the distribution characteristics of samples are complex. It is difficult to accurately describe the local distribution characteristics of samples directly in the original sample space. This paper proposes a multiple discriminant preserving support subspace radial basis function neural networks with graph similarity learning (DPSS-RBFNN) to describe the distribution characteristics of samples in the original high-dimensional space by multiple low-dimensional and simple discriminant preserving subspaces. DPSS-RBFNN includes the discriminant preserving support subspace (DPSS) learning module and the subspace distribution feature extraction (SDFE) module. In the DPSS learning module, the discriminativeness of each attribute and the joint discriminant between attributes are first considered to construct multiple subspaces. The discriminativeness of the features in these subspaces is not lower than that of the original samples. The graph similarity learning is used to measure the similarity of sample distributions between subspaces with different dimensions. Then multiple DPSSs are obtained. In SDFE module, the distribution characteristics of the samples are described by combining the local responses of the feature space extracted by the sub-RBFNN in each DPSS. The experimental results show that the proposed DPSS-RBFNN with few kernels can achieve higher accuracy in the recognition task than state-of-the-art algorithms. [ABSTRACT FROM AUTHOR]

Automated emotion recognition using brain electroencephalogram (EEG) signals is predominantly used for the accurate assessment of human actions as compared to facial expression or speech signals. Various signal processing methods have been used for extracting representative features from EEG signals for emotion recognition. However, the EEG signals are non-stationary and vary across the subjects as well as in different sessions of the same subject; hence it exhibits poor generalizability and low classification accuracy for an emotion classification of cross subjects. In this paper, EEG signals-based automated cross-subject emotion recognition framework is proposed using the Fourier-Bessel series expansion-based empirical wavelet transform (FBSE-EWT) method. This method is used to decompose the EEG signals from each channel into four sub-band signals. Manually ten channels are selected from the frontal lobe, from which entropy and energy features are extracted from each sub-band signal. The subject variability is reduced using an average moving filter method on each channel to obtain the smoothened feature vector of size 80. The three feature selection techniques, such as neighborhood component analysis (NCA), relief-F, and mRMR, are used to obtain an optimal feature vector. The machine learning models, such as artificial neural network (ANN), k -nearest neighborhood (k -NN) with two (fine and weighted) functions, and ensemble bagged tree classifiers are trained by the obtained feature vectors. The experiments are performed on two publicly accessible databases, named SJTU emotion EEG dataset (SEED) and dataset for emotion analysis using physiological signals (DEAP). The training and testing of the models have been performed using 10-fold cross-validation and leave-one-subject-out-cross-validation (LOSOCV). The proposed framework based on FBSE-EWT and NCA feature selection approach shows superior results for classifying human emotions compared to other state-of-art emotion classification models. [ABSTRACT FROM AUTHOR]

Automated emotion recognition using brain electroencephalogram (EEG) signals is predominantly used for the accurate assessment of human actions as compared to facial expression or speech signals. Various signal processing methods have been used for extracting representative features from EEG signals for emotion recognition. However, the EEG signals are non-stationary and vary across the subjects as well as in different sessions of the same subject; hence it exhibits poor generalizability and low classification accuracy for an emotion classification of cross subjects. In this paper, EEG signals-based automated cross-subject emotion recognition framework is proposed using the Fourier-Bessel series expansion-based empirical wavelet transform (FBSE-EWT) method. This method is used to decompose the EEG signals from each channel into four sub-band signals. Manually ten channels are selected from the frontal lobe, from which entropy and energy features are extracted from each sub-band signal. The subject variability is reduced using an average moving filter method on each channel to obtain the smoothened feature vector of size 80. The three feature selection techniques, such as neighborhood component analysis (NCA), relief-F, and mRMR, are used to obtain an optimal feature vector. The machine learning models, such as artificial neural network (ANN), k -nearest neighborhood (k -NN) with two (fine and weighted) functions, and ensemble bagged tree classifiers are trained by the obtained feature vectors. The experiments are performed on two publicly accessible databases, named SJTU emotion EEG dataset (SEED) and dataset for emotion analysis using physiological signals (DEAP). The training and testing of the models have been performed using 10-fold cross-validation and leave-one-subject-out-cross-validation (LOSOCV). The proposed framework based on FBSE-EWT and NCA feature selection approach shows superior results for classifying human emotions compared to other state-of-art emotion classification models. [ABSTRACT FROM AUTHOR]

In recent years, the number of people who endanger their lives has been increasing rapidly due to the mental burden of depression. The online social network (OSN) provides researchers with another perspective for detecting individuals suffering from depression. However, existing machine learning-based depression detection studies still leave relatively low classification performance, suggesting that there is significant improvement potential in their feature engineering. In this paper, we manually build and publish a large dataset on Sina Weibo (a leading OSN with the largest number of active users in the Chinese community), namely the Weibo User Depression Detection Dataset (WU3D). It includes more than 20,000 normal users and more than 10,000 depressed users, both of which are labeled and rechecked following the DSM-5 official medical and psychological depression document by professionals. Then, we conclude and propose ten statistical features by analyzing the user's text, social behavior, and posted pictures. In the meantime, text-based word features are extracted using the popular pretrained model XLNet. Moreover, we fuse these features from heterogeneous modalities and implement a multitask learning scheme to train our proposed deep neural network classification model, i.e. FusionNet (FN). The experimental results show that FN has excellent to recognize depressed users on the OSN, achieving the highest F1-Score of 0.9772 on the test set. Compared to existing studies, the proposed method has better classification performance and robustness for unbalanced training samples, as well as reasonable training and inference time. Our work provides a method to fuse multimodal information to detect individual-level depression and has reference significance for similar studies on other OSNs. [ABSTRACT FROM AUTHOR]

We present a method to detect the novelty of a research paper. Because novelty in scholarly literature also examines the larger research community, a network-based approach for extracting features is proposed. Two graphs are introduced, a macro-level graph, where authors and documents are used as nodes, and a micro-level graph, where keywords, topics, and words are used as nodes. After constructing the seed graph, papers are incrementally added while changes in the graph are recorded as the feature set of a paper. An autoencoder neural network is then used as the novelty detection model. The experimental results show that the commonly used text feature representations, TF-IDF and one-class SVM, are not suitable for detecting the novelty of a research paper. Among the constructed graphs, keyword-level graph features exhibit the best performance using regression analysis as the metric. We also combine the macro-level graph, micro-level graph, and all features and find that the combination of keywords, topics, and word features perform the best using regression and citation count analysis. Other factors that could affect the citation counts, impact, and audience, are also discussed. [ABSTRACT FROM AUTHOR]

In recent years, the study of linguistic steganalysis has been focused on altering network structure, such as replacing the basic neural units or increasing the model depth, which inevitably increases computational overhead and restricts further improvement in resource-constrained scenarios. In this paper, instead of relying on complex neural networks, we propose an alternative linguistic steganalysis framework based on hierarchical mutual learning to achieve higher detection accuracy with less inference time and model size. In the proposed method, networks with either identical or different structures are trained cooperatively to learn distinct text features from each other. To this end, in addition to the supervised learning loss function, we construct three mimicry loss functions at different feature extraction stages, which can integrate steganalytic features from various levels. Finally, we illustrate how the steganalysis framework can be extended from two networks to multiple networks. Four representative steganalysis networks with different structures are employed to verify the effectiveness of our framework. The experimental results show that the proposed framework can effectively assist networks with fewer resources to perform better in model size, inference time, and detection accuracy than state-of-the-art steganalysis algorithms. [ABSTRACT FROM AUTHOR]

Deep Neural Networks for image/video classification have obtained much success in various computer vision applications. Existing deep learning algorithms are widely used on RGB images or video data. Meanwhile, with the development of low-cost RGB-D sensors (such as Microsoft Kinect and Xtion Pro-Live), high-quality RGB-D data can be easily acquired and used to enhance computer vision algorithms [14]. It would be interesting to investigate how deep learning can be employed for extracting and fusing features from RGB-D data. In this paper, after briefly reviewing the basic concepts of RGB-D information and four prevalent deep learning models (i.e., Deep Belief Networks (DBNs), Stacked Denoising Auto-Encoders (SDAE), Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) Neural Networks), we conduct extensive experiments on five popular RGB-D datasets including three image datasets and two video datasets. We then present a detailed analysis about the comparison between the learned feature representations from the four deep learning models. In addition, a few suggestions on how to adjust hyper-parameters for learning deep neural networks are made in this paper. According to the extensive experimental results, we believe that this evaluation will provide insights and a deeper understanding of different deep learning algorithms for RGB-D feature extraction and fusion. [ABSTRACT FROM AUTHOR]

• Dense connection and spatial pyramid pooling based YOLO (DC-SPP-YOLO) is proposed. • The DC-SPP-YOLO is developed for ameliorating the object detection accuracy of YOLOv2 by employing the dense connection structure of convolutional layers and introducing an improved spatial pyramid pooling. • The PASCAL VOC standard datasets and the UA-DETRAC vehicle datasets are used to verify the effectiveness of the proposed method. • Results show that the proposed method not only outperforms the YOLOv2 method but gets the state-of-the-art performance on object detection tasks. Although the YOLOv2 method is extremely fast on object detection, its detection accuracy is restricted due to the low performance of its backbone network and the under-utilization of multi-scale region features. Therefore, a dense connection (DC) and spatial pyramid pooling (SPP) based YOLO (DC-SPP-YOLO) method is proposed in this paper for ameliorating the object detection accuracy of YOLOv2. Specifically, the backbone network of YOLOv2 adopts the dense connection of convolution layers, which strengthen the feature extraction and alleviate the vanishing-gradient problem. Moreover, an improved spatial pyramid pooling is introduced to pool and concatenate the multi-scale region features, so that the network learns the object features more comprehensively. The DC-SPP-YOLO model is established and trained based on a new loss function composed of MSE (mean square error) loss and cross-entropy loss. The experimental results indicate that the mAP (mean Average Precision) of DC-SPP-YOLO is higher than that of YOLOv2 on the PASCAL VOC datasets and the UA-DETRAC datasets. The effectiveness of DC-SPP-YOLO method proposed is demonstrated. [ABSTRACT FROM AUTHOR]

• A data-driven model is proposed based on multilevel features fusion to classify the variation trend of silicon content. • Multilevel features are extracted through variable analysis, statistical information, and stacked denoising autoencoder, respectively. • Silicon content trend labels are quantitative descripted by seven different primitives. • Both experimental simulation and industrial application verify the effectiveness and feasibility of the proposed model. The silicon content variation trend, which can reflect the quality of molten iron, provides significant information that can assist in ensuring the smooth operation of a blast furnace. This paper proposes a novel dynamic data-driven model for the online classification of the variation trend for the silicon content. Typically, a dynamic model for the silicon content variation trend primarily relies on process data feature extraction. First, a multilevel features fusion algorithm based on mutual information is developed to extract a rich and robust feature representation. Subsequently, the fused multilevel feature vectors and their corresponding trend labels are fed into a recurrent neural network model to capture the process dynamics and classify the variation trend. An experimental simulation and industrial application verified the effectiveness and feasibility of the proposed method. The classification results can provide guidance to ensure that the quality of molten iron is maintained within the desired range in the ironmaking process. [ABSTRACT FROM AUTHOR]

• An automatic audio-visual emotion recognition system based on CNN was proposed. • The audio-visual features were fused using a cascaded deep sparse autoencoder. • Signals were distributed to different cloud servers using multi-secret sharing scheme. • Edge computing was introduced to minimize the burden of transmission. This paper proposes an edge-cloud based privacy-preserved automatic emotion recognition system in which Internet of Things (IoT) devices capture facial image and speech signals from a user before distribution to different edge clouds using a secret sharing scheme, ensuring the privacy of the user. Simple signal preprocessing is performed in the edge clouds and then the signals are transmitted to the core cloud. In the core cloud, a Convolutional Neural Network (CNN)-based, pre-trained model is applied to extract deep-learned features from image and speech signals. Following the CNN model, two deep sparse auto-encoders fuse the signal features and introduce an extended non-linearity to the features. Finally, a support vector machine classifies the input in a corresponding emotion class. Several experiments were performed using the proposed system on two publicly available databases, namely, the RML database and eNTERFACE'05 database. The highest recognition accuracies by the proposed system on the two databases were 82.3% and 87.6%, respectively. These accuracies are higher than those by other state-of-the-art systems. [ABSTRACT FROM AUTHOR]

Recent developments on style transfer using convolution neural networks promote a series of computational art application, especially in generating various artistic styles images. However, the exaggerated portrait caricatures generation is still far from being perfect due to the fundamental challenges from the varied and extreme shape transformations of images. In this paper, we present a novel DeepExaggeration system to generate exaggerated portrait caricatures. First, the landmarks of portrait photos and exaggerated portrait caricatures are obtained for further geometric structural features extraction. Then, a generative model is designed to learn the mapping from the geometric structural features of portrait photos to those of exaggerated portrait caricature. After that, the exaggerated structural features can be generated by the designed generative model. Third, the generated exaggerated structural features are used for deforming the portrait photos to exaggerated portrait. Finally, we transfer the caricature style on the exaggerated portrait. Our method successfully relieves the difficulties from using convolution neural networks for geometric structural features extraction. Experimental results on exaggerated portrait caricatures synthesis demonstrate the effectiveness of our method. Moreover, we are the first to create the Magic Mirror Caricatures dataset by retrieving famous persons' portrait photos and exaggerated portrait caricatures from the Internet. [ABSTRACT FROM AUTHOR]

This paper presents a novel method for personal authentication of individuals using a deep learning approach. The reported method implements the fusion of two biometric attributes, namely fingernail plates and finger knuckles. Both these biometric traits have been extracted from the index, middle and ring fingers of a database of 890 (5 each from 178 persons) low-resolution hand dorsal images. AlexNet, a pre-trained deep learning model, has been tailored for feature extraction to suit the number of classes in the database. Deep learning features are extracted from both the traits using the customized Convolutional Neural Network (CNN). These features have been fused at the score-level and the rank-level for different combinations. Both these levels of fusion have been executed through well-established rules. Score-level fusion results show that verification can be done quite accurately with this multimodal framework. Results also depict that information fusion of these traits at the rank-level can be optimized to give Rank-1 identification rate of a perfect 100%. [ABSTRACT FROM AUTHOR]

This paper is concerned with neural networks-approximation based command filtering backstepping control for uncertain strict-feedback nonlinear systems with unknown disturbances. The “explosion of complexity” problem arising from the virtual controllers’ derivatives is resolved by utilizing the command filtering technique, and the shortcoming existing in dynamic surface method is properly overcome via an introduced error compensation mechanism (ECM). Moreover, the nonlinear functions of the underlying system are well approximated by exploiting neural networks-based framework. The developed strategy may cover two features with comparison of current achievements: 1) The filtering error can be eliminated in the light of the designed compensating signals; 2) The requirement of adaptive parameters is reduced to only one, which may enhance the control performance for realistic project implementation. At last, an application example in position tracking control of surface permanent magnet synchronous motor (SPMSM) is carried out to further verify the effectiveness and advantages of the theoretical result. [ABSTRACT FROM AUTHOR]

Hyperspectral images contain abundant spectral information, and three-dimensional (3D) feature extraction methods have been shown to be effective for classification. In this paper, we propose a hyperspectral image classification method that uses 3D multi-resolution wavelet convolutional network (3D MWCNNs) in which wavelets are first characterized by their time-frequency and multi-resolution. Then, the 3D-MWCNNs extract features from coarse to fine scales. In addition, 3D-MWCNNs work stably and effectively for approximation. In the conventional implementation of wavelets, empirical parameters must be determined in advance and the feature extraction process is not adaptive. Convolutional neural networks (CNNs) have strong adaptive learning capabilities and can extract features from low to high levels; however, they lack the theoretical underpinnings to perform multi-resolution approximation for filter learning. Therefore, by combining the CNNs framework with multi-resolution analysis theory, a model called 3D MWCNNs is proposed to extract the 3D features from different scales and different depths adaptively. 3D MWCNNs model is better at feature representation and approximation from 3D cube data; therefore, they capture the spatial and spectral features more discriminatively to improve the classification accuracy. Experimental results on three well-known hyperspectral images demonstrate that the proposed framework achieves considerably higher classification accuracy than do several state-of-the-art algorithms. [ABSTRACT FROM AUTHOR]

Since Drug-drug interactions (DDIs) can cause adverse effects when patients take two or more drugs and therefore increase health care costs, the extraction of DDIs is an important research area in patient safety. To improve the performance of Drug–drug interaction extraction (DDIE), we present a novel two-stage method in this paper. It first identifies the positive instances using a feature based binary classifier, and then a Long Short Term Memory (LSTM) based classifier is used to classify the positive instances into specific category. The experimental results show that the two-stage method has many advantages over one-stage ones, and among the factors related to LSTM, we find that the two layer bidirectional LSTM embedded with word, distance and Part-of-Speech obtains the highest F-score of 69.0%, which is state-of-the-art. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel approach based on deep learning for active classification of electrocardiogram (ECG) signals. To this end, we learn a suitable feature representation from the raw ECG data in an unsupervised way using stacked denoising autoencoders (SDAEs) with sparsity constraint. After this feature learning phase, we add a softmax regression layer on the top of the resulting hidden representation layer yielding the so-called deep neural network (DNN). During the interaction phase, we allow the expert at each iteration to label the most relevant and uncertain ECG beats in the test record, which are then used for updating the DNN weights. As ranking criteria, the method relies on the DNN posterior probabilities to associate confidence measures such as entropy and Breaking-Ties (BT) to each test beat in the ECG record under analysis. In the experiments, we validate the method on the well-known MIT-BIH arrhythmia database as well as two other databases called INCART, and SVDB, respectively. Furthermore, we follow the recommendations of the Association for the Advancement of Medical Instrumentation (AAMI) for class labeling and results presentation. The results obtained show that the newly proposed approach provides significant accuracy improvements with less expert interaction and faster online retraining compared to state-of-the-art methods. [ABSTRACT FROM AUTHOR]