Search

Nowadays radar systems have been facing with the disordered electromagnetic spectrum access and utilization in shared spectrum environments with radio communication systems. Numerous waveform recognition methods have been studied with feature engineering and conventional machine learning (ML) for intelligent radar systems, but they are critically challenged by practical problems of scalability and reliability. Deep learning (DL) with the ability to automatically learn the representational features is leveraged to handle the aforementioned obstacles effectively. In this work, we proposed a high-accurate waveform recognition method for intelligent radar systems by developing a novel residual-attention multiscale-accumulation convolutional network (RamNet). By deliberately incorporating the residual connection and attention connection in selective-feature improvement blocks, RamNet can enrich high-impact features without vanishing gradient. Moreover, a structural multiscale-accumulation connection is deployed to improve feature utilization by gathering the high-relevant features at multiple signal resolutions. Experimental results from exhaustive simulation demonstrate that RamNet recognizes waveform robustly under rigorous channel impairments and presents superior performance compared to traditional ML and state-of-the-art DL models.

Objective: With the scenario of limited labeled dataset, this paper introduces a deep learning-based approach that leverages Diabetic Retinopathy (DR) severity recognition performance using fundus images combined with wide-field swept-source optical coherence tomography angiography (SS-OCTA). Methods: The proposed architecture comprises a backbone convolutional network associated with a Twofold Feature Augmentation mechanism, namely TFA-Net. The former includes multiple convolution blocks extracting representational features at various scales. The latter is constructed in a two-stage manner, i.e., the utilization of weight-sharing convolution kernels and the deployment of a Reverse Cross-Attention (RCA) stream. Results: The proposed model achieves a Quadratic Weighted Kappa rate of 90.2% on the small-sized internal KHUMC dataset. The robustness of the RCA stream is also evaluated by the single-modal Messidor dataset, of which the obtained mean Accuracy (94.8%) and Area Under Receiver Operating Characteristic (99.4%) outperform those of the state-of-the-arts significantly. Conclusion: Utilizing a network strongly regularized at feature space to learn the amalgamation of different modalities is of proven effectiveness. Thanks to the widespread availability of multi-modal retinal imaging for each diabetes patient nowadays, such approach can reduce the heavy reliance on large quantity of labeled visual data. Significance: Our TFA-Net is able to coordinate hybrid information of fundus photos and wide-field SS-OCTA for exhaustively exploiting DR-oriented biomarkers. Moreover, the embedded feature-wise augmentation scheme can enrich generalization ability efficiently despite learning from small-scale labeled data.

Nowadays, enhancing the living standard with smart healthcare via the Internet of Things is one of the most critical goals of smart cities, in which artificial intelligence plays as the core technology. Many smart services, deployed according to wearable sensor-based physical activity recognition, have been able to early detect unhealthy daily behaviors and further medical risks. Numerous approaches have studied shallow handcrafted features coupled with traditional machine learning (ML) techniques, which find it difficult to model real-world activities. In this work, by revealing deep features from deep convolutional neural networks (DCNNs) in fusion with conventional handcrafted features, we learn an intermediate fusion framework of human activity recognition (HAR). According to transforming the raw signal value to pixel intensity value, segmentation data acquired from a multisensor system are encoded to an activity image for deep model learning. Formulated by several novel residual triple convolutional blocks, the proposed DCNN allows extracting multiscale spatiotemporal signal-level and sensor-level correlations simultaneously from the activity image. In the fusion model, the hybrid feature merged from the handcrafted and deep features is learned by a multiclass support vector machine (SVM) classifier. Based on several experiments of performance evaluation, our fusion approach for activity recognition has achieved the accuracy over 96.0% on three public benchmark data sets, including Daily and Sport Activities, Daily Life Activities, and RealWorld. Furthermore, the method outperforms several state-of-the-art HAR approaches and demonstrates the superiority of the proposed intermediate fusion model in multisensor systems.

As perception-related applications are of great importance in industrial production and daily life nowadays, solutions for understanding given images semantically receive numerous attention from the literature. To this end, significant accomplishments have been reached for such pixel-wise segmentation problem thanks to novel manipulations of integrating global context into local details in convolutional neural networks. However, this strategy in the existing work did not exhaustively exploit middle-level features, which carry reasonable balance between fine-grained and semantic information. Therefore, this paper introduces a Cross-Attentional Bracket-shaped Convolutional Network (CAB-Net) to leverage their contribution to the tournament of constructing pixel-wise labeled map. In concrete, fine-to-coarse feature maps of interest from the backbone network are densely combined by an efficient fusion of channel-wisely and spatially attentional schemes in crossing manner, namely Cross-Attentional Fusion, to embed semantically rich features into finer patterns. Continuously, these newly decoded outputs repeat the same procedure round-by-round until shaping a final feature map having finest resolution for complete scene understanding. Consequently, the proposed CAB-Net achieves competitive mean Intersection of Union performance on PASCAL VOC 2012 (83.6% without MS-COCO pretraining), CamVid (76.4%) and Cityscapes (78.3%) datasets.

Hybrid excited (HE) machines have the synergies of high torque density and high efficiency of permanent magnet (PM) machines and excellent flux controllability of wound-field machines. The PMs and field coils (FCs) coexist in HE machines to provide excitations simultaneously, and can be coupled either in series or parallel. The topologies of HE machines are very diverse, which bring challenges in the selection and design of HE machines. Based on the partitioned stator (PS) platform, a pair of series and parallel HE machines having similar structures and PM usage are evaluated and compared to identify their merits and demerits. With the help of finite element analysis, the flux regulation, energy conversion capability, parasitic effects and irreversible demagnetization of the two kinds of the HE machines are comprehensively investigated and the guidance to the machine design is provided. A pair of prototypes are fabricated and tested to validate the analyses.

For person identification in non-interaction biometric systems, gait recognition has been recently encouraged in literature and industrial applications instead of face recognition. Although numerous advanced methods that learn object appearance by conventional machine learning models have been discussed in the last decade, most of them are strongly sensitive to scene background motion. In this research, we address the drawbacks of existing works by comprehensively studying gait information from 3D human skeleton data with a deep learning-based identifier. To capture the statistic gait information in the spatial dimension, we first extract the geometric gait features of joint distance and orientation. The dynamic gait information is then obtained by calculating the temporal description features with the mean and standard deviation of geometric features. Accordingly, the fully gait information of an individual is finally learned via a compact Deep Convolutional Neural Network which is explicitly designed with multiple stacks of asymmetric convolutional filters to fully gain the spatial correlation of in-frame body joints and the temporal relation of frame-wise posture at multiple scales. Based on the experimental results evaluated on four benchmark 3D gait datasets commonly used for person identification, including UPCV Gait, UPCV Gait K2, KS20 VisLab Multi-View Kinect Skeleton, and SDUGait, the proposed method presents the superior performance over that of several state-of-the-art approaches while maintaining a low computational capacity.

Recently, numerous methods have been introduced for three-dimensional (3-D) action recognition using handcrafted feature descriptors coupled traditional classifiers. However, they cannot learn high-level features of a whole skeleton sequence exhaustively. In this paper, a novel encoding technique—namely, pose feature to image (PoF2I), is introduced to transform the pose features of joint–joint distance and orientation to color pixels. By concatenating the features of all skeleton frames in a sequence, a color image is generated to depict spatial joint correlations and temporal pose dynamics of an action appearance. The strategy of end-to-end fine-tuning a pretrained deep convolutional neural network, which completely capture multiple high-level features at multiscale action representation, is implemented for learning recognition models. We further propose an efficient data augmentation mechanism for informative enrichment and overfitting prevention. The experimental results on six challenging 3-D action recognition datasets demonstrate that the proposed method outperforms state-of-the-art approaches.

This letter proposes a cost-efficient convolutional neural network (CNN) for robust automatic modulation classification (AMC) deployed for cognitive radio services of modern communication systems. The network architecture is designed with several specific convolutional blocks to concurrently learn the spatiotemporal signal correlations via different asymmetric convolution kernels. Additionally, these blocks are associated with skip connections to preserve more initially residual information at multi-scale feature maps and prevent the vanishing gradient problem. In the experiments, MCNet reaches the overall 24-modulation classification rate of 93.59% at 20 dB SNR on the well-known DeepSig dataset.

Recently, skeleton-based human action recognition has received more interest from industrial and research communities for many practical applications thanks to the popularity of depth sensors. A large number of conventional approaches, which have exploited handcrafted features with traditional classifiers, cannot learn high-level spatiotemporal features to precisely recognize complex human actions. In this paper, we introduce a novel encoding technique, namely Pose-Transition Feature to Image (PoT2I), to transform skeleton information to image-based representation for deep convolutional neural networks (CNNs). The spatial joint correlations and temporal pose dynamics of an action are exhaustively depicted by an encoded color image. For learning action models, we fine-tune end-to-end a pre-trained network to thoroughly capture multiple high-level features at multi-scale action representation. The proposed method is benchmarked on several challenging 3D action recognition datasets (e.g., UTKinect-Action3D, SBU-Kinect Interaction, and NTU RGB+D) with different parameter configurations for performance analysis. Outstanding experimental results with the highest accuracy of 90.33% on the most challenging NTU RGB+D dataset demonstrate that our action recognition method with PoT2I outperforms state-of-the-art approaches.

We introduce a variant of the timber grid shell, the timber Voronoi shell, whose surface is reticulated by Voronoi tessellation and whose edges are made of discrete dimensional timber. This work explores form-finding methods of the Voronoi shell as a compressive funicular shell. Two closed-form solutions to shape initialization are proposed. We develop methods for minimizing the deviation from coplanarity between timber members and adjacent surface normal in order to facilitate manufacturing. A well-defined fabrication process is important for making the physical structure consistent with the structural model. A 6-axis robot with a motor spindle is employed to prefabricate the timber so the in situ manual assembly becomes easier. A parametric model describes the joint details. We formulate the robotic toolpath as a closed-form function of the resultant mesh from form finding. Thus, a general-purpose programming language can directly implement the mesh optimization and manufacturing processes without CAD or CAM software. The physical implementations, including an exhibition pavilion, validated the approach.

Low-speed control of direct-drive permanent magnet synchronous motor (DD-PMSM) is a troublesome problem, partly since the motor is more sensitive to disturbance under low/ultra-low speed situations. To attenuate the low-frequency disturbances of the mechanical system and realize smooth low-speed control, this paper presents a control strategy using a two-degree-of-freedom (2-DOF) structure. The proposed method directly shapes the two closed-loop transfer functions of reference tracking and disturbance rejection with a 2-DOF scheme, which improves the disturbance attenuation while guarantees satisfactory tracking performance simultaneously. In addition, the system stability and robustness are also analyzed. The effectiveness of the proposed method and the advantages compared to other common-used control strategies are verified through simulations and experimental tests.

We present an integer program (IP) for urban design that (1) maximizes the floor area; (2) fills building volume with room templates; (3) encodes translational symmetry in urban layout; and (4) constructs economical urban routes. Regardless that integer programming is intensively studied in operational research (OR), its role in solving geometrical and topological problems in urban design was overlooked. Based on a regular grid, our 0–1 IP formulates the sunlight-gain rules, which give urban sites their shapes, especially for residential projects. With predefined plot templates, the gross floor area (volume) within a given site can be maximized under various sunlight requirements. Subsequently, the IP fills each building volume with 2D/3D room templates. Finally, an IP-based algorithm constructs routes that connect all plots and the site’s entrances to public transportation. Both the classical Steiner tree model and the latest coverage network model are extended to create reasonable routes. In addition, this work extends the concept of special ordered sets (SOS) to encode translational symmetry in urban layouts. Encoding layout symmetry can benefit from the solvers’ SOS2 mechanism in the Branch-and-Bound search algorithm. The results indicate that traditional decision making for cities could be partially automated by IP and an abundance of valid solutions are available for designers.

Hybrid permanent-magnet (PM) variable-flux memory (VFM) machines employ both high coercive force PM, i.e., the constant PM (CPM), and low coercive force PM, i.e., the variable PM (VPM). The high energy-product CPM guarantees the benefits of high torque density of the conventional PM machines, while the PM magnetization state of the VPM is flexibly regulated to match various operation requirements. In fact, the two kinds of PMs can be magnetically connected in either parallel or series, with which the machine characteristics are quite different. In this paper, based on the two-dimensional and three-dimensional finite-element analyses, a pair of VFM machines with parallel and series connections, respectively, are investigated with specific reference to identify their difference in end effects. The comparison results show that the series hybrid PM VFM machine suffers more serious end leakage flux. The prototype machines are manufactured and tested to validate the analyses and predictions.

In this study, a novel partitioned stator hybrid excited machine is proposed, in which the excitation sources including both permanent magnets (PMs) and field windings are allocated on the secondary stator that is separated from the original stator having armature windings. As a result, the available space is boosted to improve the torque density and flux regulation capability. Meanwhile, the rotor is simple without coil or PM. The PMs on the stator is placed on the slot openings between the two neighbour tooth tips, with which the uncontrolled voltage of the proposed machine is limited and it is favoured for safety-critical applications. The machine topology and operating principle are introduced in detail, and the basic electromagnetic performance is evaluated based on finite element method to validate the analysis.

The novel coronavirus 2019 (COVID-19) becomes recently a global pandemic as declared by the World Health Organization (WHO) in March 2020. COVID-19 rapidly spread and attacked people in more than 200 countries worldwide. The use of artificial intelligence (AI) techniques has become urgent to prevent the exacerbation of the astounding spread of this pernicious disease. This paper presents a novel rapid deep learning computer-aided Diagnosis (CAD) framework for simultaneously detecting and diagnosing the COVID-19 against different respiratory diseases such as pneumonia, atelectasis, cardiomegaly, infiltration, mass, pneumothorax, nodule, and effusion. To develop a useful patient triage system for detecting COVID-19 in early-stage, rapidly extract the visual diagnosis knowledge from the input chest X-ray (CXR) images is extremely required. The proposed CAD framework shows its capability to automatically detect and diagnose COVID-19 with a detection accuracy of 96.31% and a classification accuracy of 97.40%. Meanwhile, the real-time prediction speed of 0.0093 seconds is achieved for a single testing CXR image. To achieve a high accuracy of the knowledge extraction from the entire CXR images with a high prediction speed would represent a key for developing a comprehensive and useful smart patient triage system in hospitals and healthcare systems.

Deep learning (DL) nowadays activates the advancement of multiple research fields, from signal processing to computer vision and communication. Compared with traditional machine learning (ML) algorithms, some DL architectures such as convolutional neural network (CNN) allow learning feature automatically without expert knowledge in a specific domain. Inspired by the notable achievement of DL for general image classification, this work studies an efficient CNN-based approach for automatic modulation classification which is recognized as the fundamental task in several modern communication systems. Besides the architecture organized by several blocks of parallel asymmetric convolutional layers, the proposed network is specialized by two mechanisms: skip-connection for the prevention of vanishing gradient and reusable-feature depth-concatenation for an optimal feature utilization via accumulating the deep feature at multi-scale representation. According performance evaluation, our method gains the 24-modulation classification rate of 85.10% at +10 dB SNR on the most challenging and large-scale DeepSig: RadioML (version 2018.01A) dataset.

The Erdős-Rényi (ER) random graph G(n, p) analytically characterizes the behaviors in complex networks. However, attempts to fit real-world observations need more sophisticated structures (e.g., multilayer networks), rules (e.g., Achlioptas processes), and projections onto geometric, social, or geographic spaces. The p-adic number system offers a natural representation of hierarchical organization of complex networks. The p-adic random graph interprets n as the cardinality of a set of p-adic numbers. Constructing a vast space of hierarchical structures is equivalent for combining number sequences. Although the giant component is vital in dynamic evolution of networks, the structure of multiple big components is also essential. Fitting the sizes of the few largest components to empirical data was rarely demonstrated. The p-adic ultrametric enables the ER model to simulate multiple big components from the observations of genetic interaction networks, social networks, and epidemics. Community structures lead to multimodal distributions of the big component sizes in networks, which have important implications in intervention of spreading processes., Scientific Reports, 11 (1), ISSN:2045-2322

Modulation classification, recognized as the intermediate step between signal detection and demodulation, is widely deployed in several modern wireless communication systems. Although many approaches have been studied in the last decades for identifying the modulation format of an incoming signal, they often reveal the obstacle of learning radio characteristics for most traditional machine learning algorithms. To overcome this drawback, we propose an accurate modulation classification method by exploiting deep learning for being compatible with constellation diagram. Particularly, a convolutional neural network is developed for proficiently learning the most relevant radio characteristics of gray-scale constellation image. The deep network is specified by multiple processing blocks, where several grouped and asymmetric convolutional layers in each block are organized by a flow-in-flow structure for feature enrichment. These blocks are connected via skip-connection to prevent the vanishing gradient problem while effectively preserving the information identify throughout the network. Regarding several intensive simulations on the constellation image dataset of eight digital modulations, the proposed deep network achieves the remarkable classification accuracy of approximately 87% at 0 dB signal-to-noise ratio (SNR) under a multipath Rayleigh fading channel and further outperforms some state-of-the-art deep models of constellation-based modulation classification., Comment: Accepted for presentation at IEEE GLOBECOM 2020

Pseudo-random sequence is widely used in the field of information security. This design uses Verilog HDL to complete the construction of Logistic chaotic pseudo-random sequence generator with modular programming ideas. The test results show that the correlation performance and test statistics of the output sequence can meet the requirements of chaotic pseudo-sequence. The effect of applying it to image encryption also verifies that the sequence has good randomness.

In this decade, although numerous conventional methods have been introduced for three-dimensional (3D) skeleton-based human action recognition, they have posed a primary limitation of learning a vulnerable recognition model from low-level handcrafted features. This paper proposes an effective deep convolutional neural network (CNN) with a dual-stream architecture to simultaneously learn the geometric-based static pose and dynamic motion features for high-performance action recognition. Each stream consists of several advanced blocks of regular and grouped convolutional layers, wherein various kernel sizes are configured to enrich representational features. Remarkably, the blocks in each stream are associated via skip-connection scheme to overcome the vanishing gradient problem, meanwhile, the blocks of two stream are jointly connected via a customized layer to partly share high-relevant knowledge gained during the model training process. In the experiments, the action recognition method is intensively evaluated on the NTU RGB+D dataset and its upgraded version with up to 120 action classes, where the proposed CNN achieves a competitive performance in terms of accuracy and complexity compared to several other deep models.

Coronavirus disease 2019 (COVID-19) is a novel harmful respiratory disease that has rapidly spread worldwide. At the end of 2019, COVID-19 emerged as a previously unknown respiratory disease in Wuhan, Hubei Province, China. The world health organization (WHO) declared the coronavirus outbreak a pandemic in the second week of March 2020. Simultaneous deep learning detection and classification of COVID-19 based on the full resolution of digital X-ray images is the key to efficiently assisting patients by enabling physicians to reach a fast and accurate diagnosis decision. In this paper, a simultaneous deep learning computer-aided diagnosis (CAD) system based on the YOLO predictor is proposed that can detect and diagnose COVID-19, differentiating it from eight other respiratory diseases: atelectasis, infiltration, pneumothorax, masses, effusion, pneumonia, cardiomegaly, and nodules. The proposed CAD system was assessed via five-fold tests for the multi-class prediction problem using two different databases of chest X-ray images: COVID-19 and ChestX-ray8. The proposed CAD system was trained with an annotated training set of 50,490 chest X-ray images. The regions on the entire X-ray images with lesions suspected of being due to COVID-19 were simultaneously detected and classified end-to-end via the proposed CAD predictor, achieving overall detection and classification accuracies of 96.31% and 97.40%, respectively. Most test images from patients with confirmed COVID-19 and other respiratory diseases were correctly predicted, achieving average intersection over union (IoU) greater than 90%. Applying deep learning regularizers of data balancing and augmentation improved the COVID-19 diagnostic performance by 6.64% and 12.17% in terms of the overall accuracy and the F1-score, respectively. It is feasible to achieve a diagnosis based on individual chest X-ray images with the proposed CAD system within 0.0093 s. Thus, the CAD system presented in this paper can make a prediction at the rate of 108 frames/s (FPS), which is close to real-time. The proposed deep learning CAD system can reliably differentiate COVID-19 from other respiratory diseases. The proposed deep learning model seems to be a reliable tool that can be used to practically assist health care systems, patients, and physicians.

With a rapid advancement of nanotechnology and the close integration of disciplines, research on nanozymes (nanomaterials with enzyme-like activities), is becoming an expeditiously developing field. In recent years, platinum group element-based (Pt, Pd, Ru, Rh, Ir, and Os) nanozymes developed successively, have not only promoted the research of nanozymes but also expanded the biomedical applications of nanomaterials. Generally speaking, platinum group element-based nanozymes process high catalytic efficiency, specific surface area, stability, and other physical/chemical properties, which benefit for their applications in biosensing, biological medicine, biomedical imaging, and environmental protection. This paper will introduce the research progress of platinum group element-based nanozymes including their synthesis, characterization, enzyme-like activities, stability, biocompatibility, toxicity, and applications for biological detection and clinical relevance. Our emphasis is put on unfolding the roles of platinum group element-based nanozymes in biomedical applications and how they overcome the limitations. Last but not least, trends and future perspectives of platinum group element-based nanozymes in biomedical applications are also provided.

Modulation classification, an intermediate step between signal detection and demodulation, is commonly deployed in many modern wireless communication systems. Although many approaches have been introduced in the last decades for identifying the modulation format of the incoming signal, they have the obstacle of mining radio characteristics for most traditional machine learning algorithms. To effectively handle this limitation, we propose an accurate modulation classification method by exploiting deep learning for being compatible with constellation diagram. A convolutional neural network (CNN), namely CRNet, is developed to proficiently learn the most relevant radio characteristics from transformed gray-scale constellation image by cross-residual connection, a novel structure for associating the intrinsic information between two processing flows specified by regular and grouped convolutional layers. Based on the experimental evaluation, CRNet achieves the classification rate of approximately 90% at +10 dB signal-to-noise ratio (SNR) under a multipath Rayleigh fading channel and further performs more accurately than some existing deep models for constellation-based modulation classification.

Diabetic Retinopathy (DR), the complication leading to vision loss, is generally graded according to the amalgamation of various structural factors in fundus photography such as number of microaneurysms, hemorrhages, vascular abnormalities, etc. To this end, Convolution Neural Network (CNN) with impressively representational power has been exhaustively utilized to address this problem. However, while existing multi-stream networks are costly, the conventional CNNs do not consider multiple levels of semantic context, which suffers from the loss of spatial correlations between the aforementioned DR-related signs. Therefore, this paper proposes a Densely Reversed Attention based CNN (DRAN) to leverage the learnable integration of channel-wise attention at multi-level features in a pretrained network for unambiguously involving spatial representations of important DR-oriented factors. Consequently, the proposed approach gains a quadratic weighted kappa of 85.6% on Kaggle DR detection dataset, which is competitive with the state-of-the-arts.

Recently, deep learning (DL) is an innovative machine learning (ML) technique that has gained the outstanding achievements in computer vision and natural language processing. This work takes advantage of DL for effectively handling automatic modulation classification (AMC), which is the fundamental function of numerous cognitive radio-based and spectrum sensing-based applications in many modern communication systems. Concretely, a novel deep convolutional neural network (DCNN) is proposed for learning a classification model from a massive amount of modulated signals, in which the network architecture has several convolutional blocks specialized to simultaneously capture the temporal intra-signal correlations and the spatial inter-signal relations. To this end, each block comprises various convolutional layers of asymmetric convolution kernels, whose outputs are gathered via a concatenation layer. For the enrichment of multi-scale deep feature and the prevention of gradient vanishing problem, these blocks are associated by skip connections to take into account the useful residual information. In experiments, the proposed CNN-based AMC method achieves the overall 24-modulation classification rate of 88.22% at 10dB SNR on the well-known DeepSig dataset.

Recently, regarding several beneficial properties of depth camera, numerous 3D action recognition frameworks have studied high-level features by exploiting deep learning techniques, but nevertheless they cannot seize the meaningful characteristics of static human pose and dynamic action motion of a whole sequence. This paper introduces a deep network configured by two parallel streams of convolutional stacks for fully learning the deep intra-frame joint associations and inter-frame joint correlations, wherein the structure of each stream is learned from Inception-v3. In experiments, besides the compatibility verification with various backbone networks, the proposed approach achieves the state-of-theart performance in battle with several deep learning-based methods on the updated NTU RGB+D 120 dataset.

Background and Objective: The novel coronavirus 2019 (COVID-19) is a harmful lung disease that rapidly attacks people worldwide. At the end of 2019, COVID-19 was discovered as mysterious lung disease in Wuhan, Hubei province of China. World health organization (WHO) declared the coronavirus outbreak a pandemic in the second week of March 2020. Simultaneous deep learning detection and classification of COVID-19 from the entire digital X-ray images is the key to efficiently assist patients and physicians for a fast and accurate diagnosis.Methods: In this paper, a deep learning computer-aided diagnosis (CAD) based on the YOLO predictor is proposed to simultaneously detect and diagnose COVID-19 among the other eight lung diseases: Atelectasis, Infiltration, Pneumothorax, Mass, Effusion, Pneumonia, Cardiomegaly, and Nodule. The proposed CAD system is assessed via five-fold tests for multi-class prediction problem using two different databases of chest X-ray images: COVID-19 and ChestX-ray8. The proposed CAD system is trained using an annotated training set of 50,490 chest X-ray images.Results: The suspicious regions of COVID-19 from the entire X-ray images are simultaneously detected and classified end-to-end via the proposed CAD predictor achieving overall detection and classification accuracies of 96.31% and 97.40%, respectively. The most testing images of COVID-19 and other lunge diseases are correctly predicted achieving intersection over union (IoU) with their GTs greater than 90%. Applying deep learning regularizers of data balancing and augmentation improve the diagnostic performance by 6.64% and 12.17% in terms of overall accuracy and F1-score, respectively. Meanwhile, the proposed CAD system presents its feasibility to diagnose the individual chest X-ray image within 0.009 second. Thus, the presented CAD system could predict 108 frames/second (FPS) at the real-time of prediction.Conclusion: The proposed deep learning CAD system shows its capability and reliability to achieve promising COVID-19 diagnostic performance among all other lung diseases. The proposed deep learning model seems reliable to assist health care systems, patients, and physicians in their practical validations.

Modulation classification, an intermediate process between signal detection and demodulation in a physical layer, is now attracting more interest to the cognitive radio field, wherein the performance is powered by artificial intelligence algorithms. However, most existing conventional approaches pose the obstacle of effectively learning weakly discriminative modulation patterns. This paper proposes a robust modulation classification method by taking advantage of deep learning to capture the meaningful information of modulation signal at multi-scale feature representations. To this end, a novel architecture of convolutional neural network, namely Chain-Net, is developed with various asymmetric kernels organized in two processing flows and associated via depth-wise concatenation and element-wise addition for optimizing feature utilization. The network is evaluated on a big dataset of 14 challenging modulation formats, including analog and high-order digital techniques. The simulation results demonstrate that Chain-Net robustly classifies the modulation of radio signals suffering from a synthetic channel deterioration and further performs better than other deep networks., Comment: Accepted for presentation at IEEE GLOBECOM 2020

Numerous existing handcrafted feature-based and conventional machine learning-based approaches cannot seize the intensive correlations of skeleton structure in the spatiotemporal dimension. On another hand, some modern methods exploiting Long Short Term Memory (LSTM) to learn temporal action attribute, which lack an efficient scheme of revealing high-level informative features. To handle the aforementioned issues, this research introduces a novel hierarchical deep feature fusion model for 3D skeleton-based human action recognition, in which the deep information for modeling human appearance and action dynamic is gained by Convolutional Neural Networks (CNNs). The deep features of geometrical joint distance and orientation are extracted via a multi-stream CNN architecture to uncovering the hidden correlations in both the spatial and temporal dimensions. The experimental results on the NTU RGB+D dataset demonstrates the superiority of the proposed fusion model against several recently deep learning (DL)-based action recognition approaches.

The emergence of convolutional neural network (CNN) has enabled facial expression recognition to accomplish significant outcomes nowadays. However, while existing multistream networks are subject to costly computation, the attention- embedded approaches do not involve multiple levels of semantic context in the predefined CNN. Based on the observation that emotions via a person's face are fusion of various muscular modalities, relying upon the outputs and corresponding attentional features of the deepest layer in the CNN is insufficient due to loss of informative details through multiple sub-sampling stages. Therefore, this paper introduces a CNN with densely backward attention to leverage the aggregation of channel-wise attention at multi-level features in a backbone network for reaching high recognition performance with cost-effective resource consumption. Particularly, cross-channel semantic information in high-level features are exploited densely to recalibrate finegrained details in low-level versions. Then, a step of multi-level aggregation is further executed for thorougly involving spatial representations of important facial modalities. As a consequence, the proposed approach gains highest mean class accuracy of 79.37% on RAF-DB, which is competitive with the state-of-the- arts.

The spread of illegal replication of digital image has seriously affected the protection of intellectual property rights. Nowadays, although watermarking is the most popular technology for copyright protection and ownership authentication, most of the current works cannot encounter numerous intentional attacks of watermark destruction. In this research, we introduce a novel framework of blind image watermarking that is able to learn attacking patterns effectively based on a deep convolutional encoder-decoder network. For details, a binary watermark image is hidden into selective wavelet blocks by the mean of an optimal encoding rule, wherein the quality image degradation is minimized over the Mean Square Error metric for a significant image imperceptibility enhancement. Then, the embedding maps, defined as the wavelet coefficient difference values, of various attacks simulated as digital image transformations of the watermarked image are revealed for training the deep learning-based watermark extraction model. Accordingly, the watermark information hidden in an attacked image can be precisely recovered from its embedding map by the trained model. From experimental results, the proposed watermark framework achieves a good performance trade-off between image imperceptibility and watermark robustness through an adjustable embedding strength. In addition, our approach is strongly defeats several state-of-the-art methods in terms of watermark robustness against many critical attacks.

Despite impressive achievements in image processing and artificial intelligence in the past decade, understanding video-based action remains a challenge. However, the intensive development of 3D computer vision in recent years has brought more potential research opportunities in pose-based action detection and recognition. Thanks to the advantages of depth camera devices like the Microsoft Kinect sensor, we developed an effective approach to in-depth analysis of indoor actions using skeleton information, in which skeleton-based feature extraction and topic model-based learning are two major contributions. Geometric features, i.e. joint distance, joint angle, and joint-plane distance are calculated in the spatio-temporal dimension. These features are merged into two types, called pose and transition features, and then are provided to codebook construction to convert sparse features into visual words by k-means clustering. An efficient hierarchical model is developed to describe the full correlation of feature - poselet - action based on Pachinko Allocation Model. This model has the potential to uncover more hidden poselets, which have been recognized as the valuable information and help to differentiate pose-sharing actions. The experimental results on several well-known datasets, such as MSR Action 3D, MSR Daily Activity 3D, Florence 3D Action, UTKinect-Action 3D, and NTU RGB+D Action Recognition, demonstrate the high recognition accuracy of the proposed method. Our method outperforms state-of-the-art methods in the field in most dataset benchmarks.

The rising popularity of data calls and the slowed global economy have posed a challenge to voice data networking—how to satisfy the growing user demand for VoIP calls under limited network resources. In a bandwidth-constrained network in particular, raising the bitrate for one call implies a lowered bitrate for another. Therefore, knowing whether it is worthwhile to raise one call's bitrate while other users might complain is crucial to the design of a user-centric rate control mechanism. To this end, previous work (Chen et al. 2012) has reported a log-like relationship between bitrate and user experience (i.e., QoE) in Skype calls. To show that the relationship extends to more general VoIP calls, we conduct a 60-participant user study via the Amazon Mechanical Turk crowdsourcing platform and reaffirm the log-like relationship between the call bitrate and user experience in widely used AMR-WB. The relationship gives rise to a simple and practical rate control scheme that exponentially quantizes the steps of rate change, therefore the name—exponential quantization (EQ). To support that EQ is effective in addressing the challenge, we show through a formal analysis that the resulting bandwidth allocation is optimal in both the overall QoE and the number of calls served. To relate EQ to existing rate control mechanisms, we show in a simulation study that the bitrates of calls administered by EQ converge over time and outperform those controlled by a (naïve) greedy mechanism and the mechanism implemented in Skype.

Visualization technology for network security situation adopts images to present the massive abstract data regarding network events. It reduces the workload of data analysis and benefits the manager to grasp the overall network status and trend. Secret information in the visual image requires confidentiality protection while transmitting. Comparing with some conventional methods realized by complicated encryptions such as DES and AES, we present a novel multilevel protection scheme based on visual cryptography (VC) with the beauty of decryption done only via the human eyes without using more computing devices. Essentially, a region incrementing VC scheme (RIVCS) is proposed in this paper dealing with the encoding of a secret situation image regarding network security. The secret image includes a number of regions, where each region is allocated with a certain secrecy level. Different secrecy levels can be decoded incrementally when different combinations of participants are gained. Firstly, we develop the model called the general AS (GAS) based RIVCS. Secondly, we design the algorithm for allocating secrecy levels. Thirdly, we construct the encoding matrices for sharing the secret pixels. Experimental results show that our method is more suitable to visualization data protection for network security situation with lower cost, higher reliability and richer application scenarios.

This paper presents a method for automatically constructing irregular floor plans. Given image patterns, the program produces a variety of layouts that satisfy users' geometric and topological requirements. The program extracts irregular regions from images via statistical region merging. It employs subgraph matching and simulated annealing to construct topologically feasible layouts. Methods for measuring similarities between the desired templates and irregular rooms are also reported. Applications using raster images and vector graphics are tested. Both applications generate feasible floor plans whose appearances resemble the input patterns. The results indicate that layout automation is valid without an orthogonal framework in the plane. Irregular layout synthesis allows architects to instantly implement visual preferences in early design stages.

With the recent advent of deep learning in medical image processing, retinal blood vessel segmentation topic has been comprehensively handled by numerous research works. However, since the ratio between the number of vessel and background pixels is heavily imbalanced, many attempts utilized patches augmented from original fundus images along with fully convolutional networks for addressing such pixel-wise labeling problem, which significantly costs computational resources. In this paper, a method using Round-wise Features Aggregation on Bracket-shaped convolutional neural networks (RFA-BNet) is proposed to exclude the necessity of patches augmentation while efficiently handling the irregular and diverse representation of retinal vessels. Particularly, given raw fundus images, typical feature maps extracted from a pretrained backbone network are employed for a bracket-shaped decoder, wherein middle-scale features are continuously exploited round-by-round. Then, the decoded maps having highest resolution of each round are aggregated to enable the built model to flexibly learn various degrees of embedded semantic details while retaining proper annotations of thin and small vessels. Finally, the proposed approach showed its effectiveness in terms of sensitivity (0.7932), specificity (0.9741), accuracy (0.9511), and AUROC (0.9732) on DRIVE dataset.

Nowadays human activity recognition (HAR) plays an crucial role in the healthcare and wellness domains, for example, HAR contributes to context-aware systems like elder home assistance and care as a core technology. Despite promising performance in terms of recognition accuracy achieved by the advancement of machine learning for classification tasks, most of the existing HAR approaches, which adopt low-level handcrafted features, cannot completely deal with practical activities. Therefore, in this paper, we present an efficient wearable sensor based activity recognition method that allows encoding inertial data into color image data for learning highly discriminative features by convolutional neural networks (CNNs). The proposed data encoding technique converts tri-axial samples to color pixels and then arranges them for image-formed representation. Our method reaches the recognition accuracy of over 95% on two challenging activities datasets and further outperforms other deep learning-based HAR approaches.

We see a major hurdle, as well as a window of opportunity, in wearable system design - making the system soft and bendable for skin surface sensing. In this position statement, we present our design of an FPCB (flexible printed circuit board) based wearable system for muscle fatigue sensing, referred to as the Smart KT Tape.

Background Diabetic Retinopathy (DR) is considered a pathology of retinal vascular complications, which stays in the top causes of vision impairment and blindness. Therefore, precisely inspecting its progression enables the ophthalmologists to set up appropriate next-visit schedule and cost-effective treatment plans. In the literature, existing work only makes use of numerical attributes in Electronic Medical Records (EMR) for acquiring such kind of DR-oriented knowledge through conventional machine learning techniques, which require an exhaustive job of engineering most impactful risk factors. Objective In this paper, an approach of deep bimodal learning is introduced to leverage the performance of DR risk progression identification. Methods In particular, we further involve valuable clinical information of fundus photography in addition to the aforementioned systemic attributes. Accordingly, a Trilogy of Skip-connection Deep Networks, namely Tri-SDN, is proposed to exhaustively exploit underlying relationships between the baseline and follow-up information of the fundus images and EMR-based attributes. Besides that, we adopt Skip-Connection Blocks as basis components of the Tri-SDN for making the end-to-end flow of signals more efficient during feedforward and backpropagation processes. Results Through a 10-fold cross validation strategy on a private dataset of 96 diabetic mellitus patients, the proposed method attains superior performance over the conventional EMR-modality learning approach in terms of Accuracy (90.6%), Sensitivity (96.5%), Precision (88.7%), Specificity (82.1%), and Area Under Receiver Operating Characteristics (88.8%). Conclusions The experimental results show that the proposed Tri-SDN can combine features of different modalities (i.e., fundus images and EMR-based numerical risk factors) smoothly and effectively during training and testing processes, respectively. As a consequence, with impressive performance of DR risk progression recognition, the proposed approach is able to help the ophthalmologists properly decide follow-up schedule and subsequent treatment plans.

Recently, 3D action recognition has received more attention of research and industrial communities thanks to the popularity of depth sensors and the efficiency of skeleton estimation algorithms. Accordingly, a large number of methods have been studied by using either handcrafted features with traditional classifiers or recurrent neural networks. However, they cannot learn high-level spatial and temporal features of a whole skeleton sequence exhaustively. In this paper, we proposed a novel encoding technique to transform the pose features of joint-joint distance and joint-joint orientation to color pixels. By concatenating the features of all frames in a sequence, the spatial joint correlations and temporal pose dynamics of action appearance are depicted by a color image. For learning action models, we adopt the strategy of end-to-end fine-tuning a pre-trained deep convolutional neural networks to completely capture multiple high-level features at multi-scale action representation. The proposed method achieves the state-of-the-art performance on NTU RGB+D, the largest and most challenging 3D action recognition dataset, for both the cross-subject and cross-view evaluation protocols.

To build up a state-of-the-art semantic scene segmentation model, a balanced combination between coarsely and finely contextual details is required for eliminating class-wise ambiguities and reaching high accuracy of pixel-wise labeling, respectively. Accordingly, with deep learning integration, prior works have achieved impressive performance in general, but found difficulties in correctly labeling medium to small objects. For the purpose of overcoming such issue, this paper proposes a deep convolutional network with bracket-style decoder, namely B-Net, to leverage the utilization of features learned at middle layers in the backbone networks (encoder) for constructing a final prediction map of densely enhanced semantic information. In particular, every feature map of interest combines with its adjacent version of higher spatial resolution through lateral connection modules to produce finer outputs that repeat such routine round-by-round until retrieving the finest-resolution map for dense prediction. Consequently, benchmarking results on CamVid dataset showed the effectiveness of the proposed method with mean class-wise accuracy, pixel-wise accuracy, and mean union intersection of 76.2%, 87.1%, and 66.4%, respectively.

The memory machine concept is recently extended to switched-flux (SF) machines, forming a series of SF hybrid magnet memory machines (SF-HMMMs). They exhibit the merits of good demagnetization withstand capability and effective flux adjustability. Nevertheless, the torque density is inevitably compromised due to the geometric conflict within the stationary part. Meanwhile, the inactivated dc coil, excluding online magnetization transients, results in system redundancy. Hence, in this paper, a new hybrid-excited (HE) concept is developed and implemented in a partitioned-stator SF-HPMM (PS-SF-HMMM). Thereby, the distinct synergies of a dual-magnet memory machine and an HE machine are achieved. With slight compensated field excitation, the torque can be improved at low-speed operation. Meanwhile, the high-speed constant-power region can be further extended without sacrificing high efficiency. The stator/rotor pole numbers are optimized first. The operating mechanism and optimal stepwise HE implementation over a whole operating envelop are then addressed. In addition, a comparison between PS-SF-HMMM with hybrid-excitation and its pure HE counterpart is established. Finally, both the finite-element simulation and experiments are carried out to verify the utility of the proposed HE concept.

Dancing to the rhythm in music comes natural for most of us. This however is a little far-fetched for the hearing impaired. Not being able to hear the music, the hearing impaired rely on visual aid and techniques such as mind counting to dance. To ease the learning process and alleviate the cognitive load in a dance performance, we propose DanceVibe, a wearable device that replays the beats in music via vibrations. In a 35-volunteer user study conducted over 3-month time, we find the system adds to the visual aid in the learning process and is effective enhancing dance performance. The system is particularly useful enabling on-stage performance without the need to memorize and mind count the beats. A word of caution before using DanceVibe and DanceVibe only on stage is that it does require practice and familiarity to the concept of rhythm.

Dihedral angles are good descriptors of the numerous conformations visited by large, flexible systems, but their analysis requires directional statistics. A single package including the various multivariate statistical methods for angular data that accounts for the distinct topology of such data does not exist. Here, we present a lightweight standalone, operating-system independent package called Clustangles to fill this gap. Clustangles will be useful in analyzing the ever-increasing number of structures in the Protein Data Bank and clustering the copious conformations from increasingly long molecular dynamics simulations.

The effective end-to-end transport of delay-sensitive voice data has long been a problem in multimedia networking. One of the major issues is determining the sending rate of real-time VoIP streams such that the user experience is maximized per unit network resource consumed. A particularly interesting complication that remains to be addressed is that the available bandwidth is often dynamic. Thus, it is unclear whether a marginal increase warrants better user experience. If a user naively tunes the sending rate to the optimum at any given opportunity, the user experience could fluctuate. To investigate the effects of magnitude and frequency of rate changes on user experience, we recruited 127 human participants to systematically score emulated Skype calls with different combinations of rate changes, including varying magnitude and frequency of rate changes. Results show that 1) the rate change frequency affects the user experience on a logarithmic scale, echoing Weber-Fechner's Law [1], 2) the effect of rate change magnitude depends on how users perceive the quality difference, and 3) this study derives a closed-form model of user perception for rate changes for Skype calls.

Switched reluctance machines (SRMs) have been comprehensively researched, and they play an important role in the modern industrial field. The topology of SRMs is quite simple and only one set of windings is employed. Consequently, the winding configuration has an essential impact on performance. In this paper, a new winding configuration for SRMs with only half-teeth wound (HT-SRM) is proposed, which is totally different from the conventional SRMs with full-teeth wound (FT-SRM). First, the basic topology and operating principle of HT-SRMs are presented by a three-phase 12/8 prototype. Then, based on the 2-D finite element analysis, an optimization of the original 12/8 HT-SRM is conducted and a series of static characteristics of three SRMs, namely, a conventional 12/8 FT-SRM, an original 12/8 HT-SRM with the same iron laminations, and an optimal HT-SRM with the optimized dimensions, are compared. Finally, three 12/8 SRM prototypes are fabricated and the experimental results verify the theoretical analysis and the corresponding finite element analysis-based predictions.

This paper presents a novel background subtraction method that is flexible for various background scenarios. The method includes automated-directional masking (ADM) algorithm for adaptive background modeling and historical intensity pattern reference (HIPaR) algorithm for foreground segmentation. By selecting an appropriate mask in a set based on directional feature, ADM updates background smoothly and precisely following a boundary-based strategy with an intensity correction rule. In order to segment foreground, HIPaR refers intensity patterns of previous backgrounds and input frames and then compares their mean difference with a checking threshold to make foreground decision. Experimental results prove that our proposed ADM-HIPaR outperforms other state-of-the-art methods in terms of foreground detection accuracy.

This paper improves a remarkable background estimation algorithm, namely Neighbor-based Intensity Correction (NIC) which is used in the background subtraction technique for foreground detection. The algorithm has an efficient intensity correction scheme to update the current background based on calculating the standard deviation of two windows captured from the background and the input frame in which the windows are constructed by a squared-structure binary mask. Although the NIC algorithm achieved the comparative results with existing approaches on the foreground detection accuracy and processing speed, its performance in the multi-modal background including high-speed motion and camera jitter should be improved. In the original algorithm, we recognize that the shape of a binary mask further affects the updating performance besides the window size which was already analyzed. Various shapes are therefore recommended for the multi-modal background adaptation. Moreover, an adaptive threshold identified by referring several previous Otsu thresholds to cope with the high-speed motion challenge is proposed. Experimental results on some standard datasets such CAVIAR 2004, AVSS2007, PETS 2009, and CDNET 2014, demonstrate that the foreground detection accuracy is significantly boosted with 2.6--6.7% of the F-measure metric.

Program slicing is a program analysis and understanding of technology. Sequence fault localization refers to the use of specific methods for faults in the program. Currently, the research program fault positioning is more and more people's attention and gets some results which is the more mainstream software fault localization method. Program slicing technique currently used to locate the fault procedures, which primarily to take advantage of dynamic slicing technique. Based on the full analysis of the advantages and disadvantages on the basis of previous work, we propose a flexible slicing rule and give a new method based on the slicing rule.