Search

Due to speckle noise, changes in kidney shape and size between patients, and similar regions, segmenting kidneys in ultrasound images is challenging. To alleviate this challenge, we proposed a novel CNN model, namely multi-scale fusion network of structural features and detailed features (SDFNet), to segment kidneys accurately and robustly. Specifically, the architecture includes structure feature extraction network (S-Net), detail information extraction network (D-Net) and multi-scale fusion block (MCBlock), which are in charge of extracting structural features, capturing texture details and merging features, respectively. In S-Net, we designed a boundary detection (BD) module to obtain more complete kidney structural features. In addition, this paper also designed a step-by-step training mechanism to improve the generalization ability of the SDFNet. We validated the proposed method and compared the same kidney ultrasound dataset with several state-of-the-art methods using six quantitative indicators. The results demonstrate that the proposed approach achieves the best overall performance and outperforms all other approaches on kidney ultrasound image segmentation. It is worth noting that this paper quantitatively analyzes the loss function of segment kidney ultrasound images. This work is a good reference for future research.

This paper presents a novel multi-pooling architecture generated by combining the advantages of wavelet and max-pooling operations in convolutional neural networks (CNNs), focusing on semantic segmentation tasks. CNNs often use pooling to reduce the number of parameters, improve invariance to certain distortions, and enlarge the receptive field. However, pooling can cause information loss and thus is detrimental to further operations such as feature extraction and analysis. This problem is particularly critical for semantic segmentation, where each pixel of an image is assigned to a specific class to divide the image into disjoint regions of interest. To address this problem, pooling strategies based on wavelets-operations have been proposed with the promise to achieve a better trade-off between receptive field size and computational efficiency. Previous works have confirmed the superiority of wavelet pooling over the traditional one in semantic segmentation tasks. However, we have observed in our computational experiments that the expressive gains reported from the use of wavelet pooling in other segmentation tasks were not observed in the scope of aerial imagery due to imprecision in the segmentation of image details. The combination of wavelet pooling and max-pooling, a solution not yet reported in the literature, can address that issue. Such gap observed in the pooling area motivated the two proposals that are the main contributions of this paper: (a) A new multi-pooling strategy combining wavelet and traditional pooling in a new network structure suitable for aerial image segmentation tasks; (b) Two-stream architectures using the traditional max-pooling and wavelet pooling as streams. These proposals were implemented using the Segnet, a known architecture for semantic segmentation. The computational experiments, based on the IRRG images from the Potsdam and Vaihingen data sets, demonstrated that the proposed architectures surpassed the original Segnet architecture’s performance with results comparable to state-of-the-art approaches.

In this paper, we propose a Deep Belief Network (DBN) based approach for the classification of audio signals to improve work activity identification and remote surveillance of construction projects. The aim of the work is to obtain an accurate and flexible tool for consistently executing and managing the unmanned monitoring of construction sites by using distributed acoustic sensors. In this paper, ten classes of multiple construction equipment and tools, frequently and broadly used in construction sites, have been collected and examined to conduct and validate the proposed approach. The input provided to the DBN consists in the concatenation of several statistics evaluated by a set of spectral features, like MFCCs and mel-scaled spectrogram. The proposed architecture, along with the preprocessing and the feature extraction steps, has been described in details while the effectiveness of the proposed idea has been demonstrated by some numerical results, evaluated by using real-world recordings. The final overall accuracy on the test set is up to 98% and is a significantly improved performance compared to other state-of-the-are approaches. A practical and real-time application of the presented method has been also proposed in order to apply the classification scheme to sound data recorded in different environmental scenarios.

The increasing online reviews play an essential role in the e-commerce platform, which profoundly affects the purchase decisions of consumers. However, rampant dishonest sellers manipulate other buyers or robots to post deceptive reviews for profit. Recently, the detection of deceptive reviews has attracted general research attention, which mainly comprises two directions, traditional methods based on statistics and intelligent methods based on neural networks. These methods use a single feature or multiple features for classifier design. To make full use of different features for better feature representation of detecting deceptive reviews, this paper proposes a new feature fusion strategy and verifies its performance by comparing it with other feature fusion strategies. First, we utilize three independent models for feature extraction: the TextCNN, the Bidirectional Gated Recurrent Unit (GRU), and the Self-Attention are used to learn local semantic features, temporal semantic features, and weighted semantic features of reviews, respectively. Secondly, after obtaining different feature representations from the fully connected layers of these three models, we concatenate them together to form the final documental representation. Finally, we use a full connection layer and the sigmoid function to further learn and complete deceptive review detection. Experiments on three balanced and unbalanced in-domain small datasets (hotel, restaurant, doctor) and mixed-domain datasets show that our model is superior to baselines. Experiments on large-scale data with various imbalanced proportions verify the effectiveness of our method. We also analyze the results of different datasets from the perspective of part of speech to improve the model's interpretability.

Hierarchical multi-label text classification (HMTC) has been gaining popularity in recent years thanks to its applicability to a plethora of real-world applications. The existing HMTC algorithms largely focus on the design of classifiers, such as the local, global, or a combination of them. However, very few studies have focused on hierarchical feature extraction and explore the association between the hierarchical labels and the text. In this paper, we propose a Label-based Attention for Hierarchical Multi-label Text Classification Neural Network (LA-HCN), where the novel label-based attention module is designed to hierarchically extract important information from the text based on the labels from different hierarchy levels. Besides, hierarchical information is shared across levels while preserving the hierarchical label-based information. Separate local and global document embeddings are obtained and used to facilitate the respective local and global classifications. In our experiments, LA-HCN outperforms other state-of-the-art neural network-based HMTC algorithms on four public HMTC datasets. The ablation study also demonstrates the effectiveness of the proposed label-based attention module as well as the novel local and global embeddings and classifications. By visualizing the learned attention (words), we find that LA-HCN is able to extract meaningful information corresponding to the different labels which provides explainability that may be helpful for the human analyst. National Research Foundation (NRF) This research/project is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG-100E-2019-031).

Objective: In this paper, a stacked autoencoder deep neural network is proposed to extract the QRS complex from raw ECG signals without any conventional feature extraction phase. Methods: A simple architecture has been deeply trained on many datasets to ensure the generalization of the network at inference. Results: The proposed method achieved a QRS detection accuracy of 99.6% using more than 1042000 beats which is competitive with all state-of-the-art QRS detectors. Moreover, the proposed method produced only 0.82% of Detection Error Rate using six unseen datasets containing more than 1470000 beats. Thus confirms the high performance of our method to detect QRSs. Conclusion: Stacked autoencoder neural networks are very effective in QRS detection. At inference, our algorithm processes 1042309 beats in less than 25.32 s. Thus, it is favorably comparable with state-of-the-art deep learning methods. Significance: The stacked autoencoder is an efficient tool for QRS detection, which could replace conventional systems to help practitioners make fast and accurate decisions.

A remarkable progress in sequential recommendation field lies on deep learning techniques, where deep learning was widely used to capture user preference from behavior records. However, researchers usually place emphasis on sequential change while ignore the correlation between user’s historical behaviors and candidate item’s characteristics (history-candidate association), which leads to the inaccurate matching between target users and candidate items. In this paper, we proposed a deep learning model to explore both sequential influence and history-candidate association in sequential recommendation, namely DeepAssociate. First, we considered the history-candidate association in user preference representation and obtained it by two steps, including sequential influence extraction and association feature extraction. Then, by defining a weighted objective function, we introduced an integrated framework which makes a combination of sequential and association features extraction and prediction module to enhance recommendation performance. Experimental results on four real-world datasets demonstrated that DeepAssociate model outperformed state-of-the-art methods on recommendation performance. Furthermore, a series of extensive experiments indicated the benefit of utilizing history-candidate association feature in reducing model complexity and accelerating model convergence.

Chronic obstructive pulmonary disease (COPD) is a chronic non-reversible lung disease. Other acute respiratory illness due to infections are termed as non-chronic. In general, the pulmonologist carries preliminary screening by accessing lung sounds. In this paper, we propose a methodology to automatically classify lung sounds associated with non-chronic and chronic categories. To accomplish the task, at first, the empirical mode decomposition (EMD) is applied to lung sound signals to obtain intrinsic mode functions (IMFs). Considering the availability of IMFs in entire dataset and by using a hybrid strategy, first four IMFs have been selected for feature extraction purpose. The IMFs are then further processed to construct two-dimensional (2D) and higher-dimensional (HD) phase space representation (PSR). The feature space includes the 95% confidence ellipse area from 2D-PSR and interquartile range (IQR), mean, median, standard deviation, skewness and kurtosis of Euclidian distances computed from HD-PSR. The process is carried out for the first four IMFs corresponding to the non-chronic and chronic categories of the lung sounds. Neighborhood component analysis is used to select best performing features. The computed and selected features depict a significant ability to discriminate the two categories of lung sound signals. To perform classification, we use ensemble classifiers. Simulation outcomes on ICBHI 2017 lung sound dataset show the ability of the proposed method in effectively classifying non-chronic and chronic lung sound signals. Ensemble of Bagged tree provides the highest classification accuracy of 97.14% over feature space constituted by 10D-PSR of fourth IMF.

Offline Signature verification is a biometric method with important applications in financial, legal and administrative procedures. The verification process includes comparing the extracted features of a questioned signature with those of genuine signatures of a certain individual. There are many challenges in designing offline signature verification as dynamic temporal features of signatures are not available. Deep Convolutional Neural Networks (DCNNs) have the great capability of extracting features from signature images. Despite the important advantages of these networks, they are unable to recognize the spatial properties of each feature in a signature. In addition, max-pooling layers usually eliminate some features that are crucial for forgery detection. In this paper, we propose a novel signature verification model with a combination of a CNN and Capsule Neural Networks (CapsNet) in order to capture spatial properties of signature features, improve the feature extraction phase, and reduce the complexity of the network. Moreover, we designed a new training mechanism in which a single network is trained simultaneously by two images at the same level so that the training parameters are reduced by half. Such mechanism does not require two separate networks for learning the features. Finally, a composite backbone architecture is presented with the hybrid of the proposed CNN-CapsNet models which we name CBCapsNet. The evaluation results demonstrate that our proposed model can improve accuracy and outperform prevalent signature verification methods in the community.

Fake news is the information or stories that are intentionally created to deceive or mislead the readers. In recent times, Fake news detection has attracted the attention of researchers and practitioners due to its many-fold benefits, including bringing in preventive measures to tackle the dissemination of misinformation that could otherwise disturb the social fabrics. Social media in recent times are heavily loaded with multimedia news and information. People prefer online news reading and find it more informative and convenient if they have access to multimedia content in the forms of text, images, audio, and videos. In early studies, researchers have proposed several fake news detection mechanisms that mostly utilize the textual features and not proper to learn multimodal (textual + visual) shared representation. To overcome these limitations, in this paper, we propose a multimodal fake news detection framework with appropriate multimodal feature fusion that leverages information from text and image and tries to maximize the correlation between them to get the efficient multimodal shared representation. We empirically show that text, when combined with the image, can improve the performance of the model. The model detects the post once it is introduced into the network in an early stage. At the early stage of a news post’s introduction into the network, the model takes the text and image of the post as input and decides whether this is fake or genuine. Since this model only analyzes news contents, It does not require any prior information regarding the user and network details. This framework has four different sub-modules viz. Attention Based Stacked Bidirectional Long Short Term Memory (ABS-BiLSTM) for textual feature representation, Attention Based Multilevel Convolutional Neural Network–Recurrent Neural Network (ABM-CNN–RNN) for visual feature extraction, multimodal Factorized Bilinear Pooling (MFB) for feature fusion and finally Multi-Layer Perceptron (MLP) for the classification. We perform experiments on two publicly available datasets, viz. Twitter and Weibo. Evaluation results show the efficacy of our proposed approach that performs significantly better compared to the state-of-the-art models. It shows to outperform the current state-of-the-art by approximately 10 points for the Twitter dataset. In contrast, the Weibo dataset achieves an overall better performance with balanced F1-scores between fake and real classes. Furthermore, the complexity of our proposed model is significantly lower than the state-of-the-art.

The Electrocardiogram (ECG) arrhythmia classification has become an interesting research area for researchers and developers as it plays a vital role in early prevention and diagnosis of cardiovascular diseases. In ECG signal classification, the feature extraction and selection processes are critical steps. Thus, in this paper, different ECG signal descriptors based on one-dimensional local binary pattern (LBP), wavelet, higher-order statistical (HOS), and morphological information are introduced for feature extraction. For feature selection and classification processes, a new hybrid ECG arrhythmia classification approach called MRFO-SVM that combines a metaheuristic algorithm termed Manta ray foraging optimization (MRFO) with support vector machine (SVM) is proposed to automatically determine the relevance features of LBP, HOS, wavelet and magnitude values. In MRFO-SVM approach, the MRFO is utilized to optimize the parameters of SVM and to select the significant features subset that provides the best classification performance, meanwhile SVM is used for classification purposes. The proposed MRFO-SVM approach is trained on the MIT-BIH Arrhythmia database containing four abnormal and one normal heartbeats. The experimental results of ECG arrhythmia classification using the proposed MRFO-SVM revealed with evidence its superiority with overall classification accuracy of 98.26% over seven well-known metaheuristic algorithms.

Hearing impaired individuals can easily overcome the barriers in communicating with other members of the society via computer technology. In this study, the recognition of dynamic words in Turkish Sign Language (TSL) with two hands was studied using the Leap Motion Controller (LMC) device. 50 dynamic words were determined considering the similarities and differences among themselves, and a dataset was created using 4 signers. For the system proposed in this paper, a comprehensive feature extraction process is executed. Applying feature selection algorithms and PCA, LDA and PCA+LDA dimension reduction methods to this dataset, new datasets with less dimension were obtained. For the first time, ELM architectures were operated as a classifier in a sign language recognition system. Recognition performance was tested with 5 different ELM networks and 2 classical classifiers and the results were compared. In addition, comparisons of classical and ELM based classifiers were presented. The 10-fold cross validation method was used to test the validity of the proposed system and the accuracy of the results obtained. Based on the results obtained by a comprehensive analysis, it was observed that the ML-KELM classifier maintains its performance rate and gives the highest performance rate. At the same time, it has been observed that ML-KELM classifier has a stable structure, which offers less user intervention.

Parkinson’s Disease (PD) is a progressive neurological disorder that affects the functioning of the brain. As PD progresses, there arises a problem of maintaining coordination between brain and different parts of the body. Subjects have the problem of performing daily activities. Its early diagnosis is important to improve the quality of life of Parkinson’s patients. Most of Parkinson’s patients are likely to have voice disorders in the early stages. Therefore, the analysis of voice recordings with machine learning-based models can improve the diagnosis process. Voice signals are non-linear, non-stationary signals which exhibit oscillatory behavior. In this paper, a hybrid approach is proposed to extract features from resonance-based and time-frequency based information. A combination of Resonance based Sparse Signal Decomposition (RSSD) + Time-Frequency (T-F) algorithm is proposed. Two datasets are collected for the study. The first dataset D1 is a public dataset consisting of 16 PD + 21 Healthy Controls (HC) and the second dataset D2 is collected from 20 HC to explore the effect of diversity in the diagnosis process. To explore the potential of deep learning techniques in diagnosing speech impairments in PD patients, the potential of Convolution Neural Network (CNN) is also explored. To evaluate the method, we conducted several experiments with state-of-the-art feature extraction and classification techniques. The proposed hybrid approach distinguished PD and HC with a validation accuracy of 99.37%. We have also explored the effect of diversity in the data collection process and it is observed that diversity in the socio-cultural group plays very important role to put the system in the clinical practice.

The Extreme learning machine (ELM) designed by Huang et al. is proved to be a fast and good classifier over a decade, but existing ELM is non-deterministic in nature as well as kernel dependent and needs attention to optimize the selection of kernels. In ELM feature space is obtained with the help of single kernel function. The choice of kernel depends on perceptiveness of classification problem. So a generalized framework with deterministic nature along with optimized kernel is ought to be designed that can be applied to large domain of real world heterogeneous pattern classification problems. This paper presents a deterministic extreme learning machine for neural network with feedforward architecture which is formulated with multiple kernel learning. We further enhance this approach by incorporating Gray level co-occurrence matrix (GLCM) for multi-modal feature extraction. Two formulation of kernel extreme learning machine are introduced, with target kernel function as a linear combination of different base kernels. The first one is based on deterministic multiple kernel learning while the second one uses deterministic multiple kernel learning along with GLCM for extracting the invariant feature vectors. The performance of proposed algorithms are analyzed on pattern recognition problem for face recognition by changing the number of training set, types of kernel used and coefficients used for combining base kernels. The superior recognition rate is achieved for prominent multi-class face databases, when compared with contemporary methods that prove the efficacy of proposed algorithms.

Accurate segmentation of the liver and tumors in computed tomography (CT) images is critical for intelligent computer-aided diagnosis (CAD). The commonly used segmentation methods based on fully convolutional networks (FCN) only take a single image into consideration but do not make good use of sequence information. In this paper, two more feasible sequence segmentation strategies than 3D U-net which can utilize inter-slice and intra-slice features simultaneously at the lower hardware and time cost are studied to improve the segmentation result. U-net serves as the backbone model of segmentation and Bi-directional convolutional long short-term memory (Bi-ConvLSTM) is chosen to extract and fuse the inter-slice feature. Strategy A corrects the pre-segmented results of U-net in the fusion of sequence information as a post-processing, where Mod-1, Mod-2 and Mod-3 models are built to compare the effects of width, depth, and residual structure on the modified model of sequence segmentation. Strategy B directly integrates the fusion of sequence information into the feature extraction of U-net, and then an end-to-end model called W-net is built based on it. The experiment results show that both strategies improve the liver and tumor segmentation performance in various aspects. The results based on strategy A are closer to the ground truth with less misdiagnose region: Mod-1 achieves better accuracy on liver contour segmentation because of the largest model width; Mod-2 can obtain more accurate tumor contour since the greatest depth of feature extraction process; and Mod-3 is at the average segmentation performance. Therefore, strategy A is recommended in the application of surgery planning of tumors. Strategy B achieves better space coincidence degree and less training time cost, which is more suitable for the early screening of liver cancer.

This paper highlights one of the most promising research directions for automatic modulation recognition algorithms, although it does not provide a final solution. We study the design of a high-precision classifier for recognizing PSK, QAM and DVB-S2 APSK modulation signals. First, an efficient pattern recognition model that includes three main modules for feature extraction, feature optimization and classification is presented. The feature extraction module extracts the most useful combinations of up to six high-order cumulants that embed sixth-order moments and uses logarithmic function properties to improve the distribution curve of the six-order cumulants. To the best of our knowledge, this is the first time that these combinations and the improved feature criteria have been applied in this area. The optimizer module selects optimal features via principal component analysis (PCA). Then, in the classifier module, we study two important supervised neural network classifiers (i.e., multilayer perceptron (MLP)- and radial basis function (RBF)-based classifiers). Through an experiment, we determine the best classifier for recognizing the considered modulations. Then, we propose an RBF-PCA combined recognition system in which an optimization module is added to enhance the overall classifier performance. This module optimizes the classifier performance by searching for the best subset of features to use as the classifier input. The simulation results illustrate that the RBF-PCA classifier combination achieves high recognition accuracy even at a low signal-to-noise ratio (SNR) and with limited training samples.

Temporal feature extraction is an important issue in video-based action recognition. Optical flow is a popular method to extract temporal feature, which produces excellent performance thanks to its capacity of capturing pixel-level correlation information between consecutive frames. However, such a pixel-level correlation is extracted at the cost of high computational complexity and large storage resource. In this paper, we propose a novel temporal feature extraction method, named Attentive Correlated Temporal Feature (ACTF), by exploring inter-frame correlation within a certain region. The proposed ACTF exploits both bilinear and linear correlation between successive frames on the regional level. Our method has the advantage of achieving performance comparable to or better than optical flow-based methods while avoiding the introduction of optical flow. Experimental results demonstrate our proposed method achieves the state-of-the-art performances of 96.3% on UCF101 and 76.3% on HMDB51 benchmark datasets., 24 pages (exclude reference), 7 figures, 4 tables

Current state-of-the-art myoelectric interfaces employ traditional pattern recognition (PR) algorithms to decode the Electromyogram (EMG) signals into hand movements for controlling artificial limbs. Recently, deep learning (DL) models have also been exploited for EMG feature learning/extraction. Models like Convolutional Neural Networks (CNN), which capture the spatial correlations, and Long Short-Term Memory (LSTM), which capture the non-linear temporal dynamics of EMG time-series data, have been shown to outperform traditional EMG PR systems. Nevertheless, the large number of model parameters, long training times, and large amounts of data required to train these DL models remain limiting factors that may hinder their translation into clinically viable prostheses. Consequently, rather than applying DL directly, this paper leverages concepts derived from these models to build upon our proposed concept of a Fusion of Time Domain Descriptors (FTDD). The FTDD are augmented with Range Spatial Filtering (RSF) to capture the spatial correlations and combined into an LSTM-style framework. This process, denoted as Recurrent Spatial-Temporal Fusion (RSTF), can be applied in combination with any traditional feature extraction method to exploit temporal and spatial correlations, with the potential for bi-directional applications. The advantages of the proposed RSTF method include (1) the memory concept, capturing long- and short-term spatial and temporal dependencies of the EMG signals, (2) significantly improved performance outperforming other state-of-the-art models and (3) the simplicity and the fairly low computational costs for feature extraction. Results are bench-marked against several feature extraction methods, proving the power of the RSTF using data from 82 subjects from five EMG databases with varying recording characteristics. The proposed method significantly outperforms all other methods tested for EMG pattern recognition, including a deep LSTM and other CNN methods previously reported in the literature and at a fracture of the computational cost. On the most challenging dataset, improvements of as much as 15% were found.