Search

Object Detection methods are an important subject in the implementations of Artificial Intelligent systems. Many attempts to build up a real-time object detection hardware/software in an SoC are presented in recent research. However, due to the demanding in both memory bandwidth and parallel computing resources, only few designs can fulfill the real-time requirement, which is important for applications or payloads such as Drones, UAVs, and autonomous vehicles. In this paper, issues in design of an SoC based object-detection will be discussed, and an on-going hardware design base on YOLO algorithm and ARC Platform will be presented. With the optimization in both numbers of Processing Element and Memory Bandwidth, An 176.3 GOPs CNN accelerator with 30 fps performance at 400MHz is presented. In addition to the Object-Detection engine, a ZCA image preprocessor and NMS postprocessing are also proposed to simplify the corresponding CNN model and enhance the real-time performance of Object-Detection. The emulation results and demonstration videos in the ARC platform will be presented. The post-layout simulation of current designverified the targeting real-time performance in a 28nm CMOS technology.

This work presents an object detection system based on Tiny YOLOv2 algorithm and ARC platform, and demonstrates on a Synopsys HAPS-80 FPGA.

After emergencies occur, evolutionary phenomena often occur. What brings great challenges to the emergency decision-making process is the complexity and uncertainty of the evolution process. In order to improve the completeness of information extraction in emergency decision-making, the "BiGRU-MultiAtt" model, which integrates two-way GRU and multi-channel self-attention mechanism, is used to extract the evolutionary relationship between emergencies. Compared with the mainstream "LSTM-ATT", this model will capture sentence-level semantic feature information more comprehensively, enhance the ability to learn urgent texts, and significantly reduce the overall complexity of the model. In addition, under the network model, a weight-based position embedding method is proposed to assist training, which further enhances the importance of text vocabulary position information and achieves the effect of enhancing training. In the evolutionary relationship extraction experiment, the accuracy and recall rate of the introduction of the multi-channel self-attention model method have been improved, and the F1 value has reached 90.41%.

This study proposes a control strategy for an energy storage system (ESS) based on the irradiance prediction. The energy output of photovoltaic (PV) systems is intermittent, which causes the power grid unstability and un reliability. It posts a great challenge to electric power industries. The development of the strategy is divided into two parts. First, a solar irradiance prediction model is proposed based on the Feed-forward Neural Networks (FNN), which uses the historical irradiance and satellite cloud images as the model inputs. The characteristic parameter are selected by the Principal Component Analysis (PCA). In order to improve the accuracy of prediction model, a hybrid method has also been proposed, which combines long-short-term prediction and can be used to predict the power generation of PV systems. Second, a control strategy for ESS has been developed, which considers the state of charge (SOC) of ESS, the microgrid (MG) power, and the change of the predicted power generation. The target of the control strategy is to reduce the grid power profile fluctuations which is interfered by the intermittent renewable energy generation, and thereby the strategy can improve the efficiency of the utilization of generated power, while reducing the operating costs and the energy loss caused by frequent power transmission in advance.

This paper reports the very first microfluidic circulatory platform mimicking physiological micro-environments in vitro to investigate cell interactions between different organs, and between organs and blood vessels. The platform consisting of a heart-mimicking pump, a device with blood-air barrier and a blood vessel-like micro-channel is demonstrated in this paper. Moreover, the platform can also include other well-developed organ-on-a-chip devices for advanced biomedical researches. In this paper, we exploit the developed platform to study endothelium responses and its interaction with epithelial cells at a blood-air barrier under biomimetic pulsatile flow perfusion.

Registration and fusion of magnetic resonance (MR) and ultrasound (US) images remain challenging tasks due to the great intensity variation between two modalities and the lack of ground truth spatial correspondence. In this paper, we proposed a mix-supervised registration network, driven by main vessel segmentation labels and modality independent neighborhood descriptor (MIND) similarity metric. The accurate alignment of the main vessel labels can prompt the registration network focus on the region of interest and MIND similarity metric constraints global deformation. Vessel label and similarity measures can promote each other and produce satisfactory registration results. This framework connects a Rigid-Net and a Bspline-Net to generate the transformation parameters of rigid and deformable registration in one shot. The well-trained network is able to register dynamic 3D US images to the preoperative MR image in real time, correcting the motion caused by a patient’s breathing. Experimental results show that the proposed method is effective for MR-US registration.

3D/2D registration techniques can compensate for the deficiencies of X-ray angiography-based navigation in vascular interventional surgery, such as the lack of depth information and excessive use of contrast agents. In this study, we propose a novel Monte Carlo tree search-based 3D/2D vessel graph registration method. The registration problem is transferred to a tree search problem according to the topology of vessel centerlines. Then, the Monte Carlo tree search method is applied to find the optimal vessel matching associated with highest registration score. Experiments on uninitialized vessel data demonstrate that the proposed method can achieve the highest accuracy among four state-of-the-art methods. An average accuracy of 1.91 mm on clinical coronary artery data is obtained. For the independence of initial pose and robustness to noise, the proposed method can align 3D and 2D vessels without prior initialization in vascular interventional surgery.

In this paper, a microfluidic device capable of culturing cells in three-dimensional (3D) hydrogel under oxygen gradients is developed. The developed device is exploited to study the networking of endothelial cells, which plays an important role in vasculogenesis, in 3D matrix under different oxygen microenvironments. The vasculogenesis is an important process during normal development and disease progression in vivo. The device offers an in vitro model mimicking vasculogenesis process under physiological oxygen microenvironments in 3D matrix. The experimental results suggest that oxygen gradient plays a critical role in regulating endothelial cell networking in a 3D matrix. The device provides a powerful platform to study 3D cellular behaviors under oxygen gradients for various biomedical studies.

Convolutional Neural Networks (CNNs) have gained remarkable success in computer vision, which is mostly owe to their ability that enables learning rich image representations from large-scale annotated data. In the field of medical image analysis, large amounts of annotated data may be not always available. The number of acquired ground-truth data is sometimes insufficient to train the CNNs without overfitting and convergence issues from scratch. Hence application of the deep CNNs is a challenge in medical imaging domain. However, transfer learning techniques are shown to provide solutions for this challenge. In this paper, our target task is to implement diabetic retinopathy fundus image classification using CNNs based transfer learning. Experiments are performed on 1014 and 1200 fundus images from two publicly available DR1 and MESSIDOR datasets. In order to complete the target task, we carry out experiments using three different methods: 1) fine-tuning all network layers of each of different pre-trained CNN models; 2) fine-tuning a pre-trained CNN model in a layer-wise manner; 3) using pre-trained CNN models to extract features from fundus images, and then training support vector machines using these features. Experimental results show that convolutional neural networks based transfer learning can achieve better classification results in our task with small datasets (target domain), by taking advantage of knowledge learned from other related tasks with larger datasets (source domain). Transfer learning is a promising technique that promotes the use of deep CNNs in medical field with limited amounts of data.

Load modelling is an essential instrument to carry through the planning for modern power system with widely access of renewable energy generation (REG). And it is necessary to analyse the potential growth activity of REG for C02 reduction. Aiming at raising the scale of REG in final consumption segment, this paper presents the analysis of substitution effect among different kinds of energy. Within the strict constraint of energy balance and end-use efficiency by analogy to the efficiency function of Well-to-Wheel (WTW), this paper establishes a multi-objective optimized load model focusing on rebuilding an economic and environmental friendly structure of final energy consumption. Bi-level optimization is used to solve the model. 0bjective of the main-problem aims at system economy considering C02 emission, while the sub- problem with objective denoting to maximize the scale of alternative REG power consumption. Finally, to improve the REG utilization in the whole process from generation to consumption, a concept of flexible load curve combination and its relative capability of raising the renewable energy resource integration is proposed based on generalized energy storage.

This research proposed a pneumatic multi-pipe deep lunar soil drilling device under the background of manned lunar-landing and used coupled simulation by EDEM/Fluent for the drilled process, pneumatic drilling principle has been proved feasible by the simulation result. The drilled device this research proposed can collect lunar soil five meters under the surface, and ensure that the bedding information of the collected lunar soil is complete.

Landing-impact between lander and lunar regolith is the decisive factor to soft landing and relates directly to the safety of detection equipment. In this paper, the landing-impact process and the interaction between lander and regolith is researched using the discrete element simulation. Firstly, the landing-impact model is established based on discrete element theory and the mechanical properties of lunar regolith. Many simulations in different working conditions are conducted to study the vertical impact. Secondly, to study the process of landing-impact, dynamic response, stress distribution and motion in a certain condition are studied and the phenomenon of landing impact is explained. Finally, the influence of velocity and mass on the landing-impact is summarized through contrastive analysis. The research in this paper provides reference for the design of lander buffering mechanism and the establishment of the theoretical model.

Real time online data truly are high frequency data as part of the Big Data problem from the Internet Of Things. High frequency data, particularly in finance, tend to be more fat-tailed in distribution than lower frequency data. Fat tailed distribution simulation of random variables like high frequency data has been brought into attention as a better phenomenon in in-sample fit and out-of-sample prediction. The existing approaches include parameter estimation by Least Square Regression (LSR) and Maximum Likelihood Estimation (MLE), and exponential generalized autoregressive conditional heteroscedastic (EGARCH) model simulation based on Stochastic Volatility Jump-Diffusion process. However, there is still room for accuracy of fat tailed distribution simulation. We propose using Ito's drift-diffusion equation as the kernel function for a univariate stochastic process without introducing extra jump coefficients in EGARCH or performing LSR or MLE between x t and x t−1 . The mean reverting rate and stochastic volatility are conditional variables and automatically become correlated. Their estimation can be solved by applying Least Square Estimation on the kernel function given the mean value as unconditional variable. The proposed method is proved to be a fat tailed distribution simulation of original data with smaller mean squared error. The paper compares the simulation of the stochastic process of Google finance's online real time stock price by the proposed method to that by LSR, MLE and EGARCH. The proposed method simulates a fat tailed distribution of original data with lower mean squared error and yet being outlier-prone in a more reasonable way.

In this work, a cascade classifier is trained to detect multi-view face samples. Comparing with most of face detection system which use different classifier to classify frontal face and profile face, our system has advantage in detection speed. The proposed face detector extracts the Haar-like feature from the training samples and train a cascade classifier by using Adaboost learing algorithm. Different from the existing algorithms, our detection system only contains a cascade classifier model. Our preliminary experiments demonstrate that our cascade classifier can achieve similiar accuracy and 60% higher speed detection than the multi-view face detection system which consist of two sparate cascade classifiers.

Retinal vessel delineation is a hot research topic owing to its importance in a lot of clinic application. Several methods have been proposed in the past decades. Here we will present a new supervised method for retinal vessel segmentation. The method is designed to explore the complex relationship between retinal images and their corresponding vessel label maps. Specifically, in order to build a model describing the direct transformation from retinal image to vessel map, we introduce a deep convolutional neural network (abbreviation as CNN), which has strong enough induction ability. For the purpose of constructing the whole vessel probability map, we also design a synthesis method. Our method shows better performance on DRIVE dataset than state-of-the-art of reported approaches in the light of sensitivity (abbreviation as Se), specificity (abbreviation as Sp) and accuracy (abbreviation as Acc). Our proposed method has great potential to be applied in existing computer-assisted diagnostic system of ophthalmologic diseases. Meanwhile, the method may offer a novel, general computing framework for segmentation in other fields.

Magnetic logic devices have advantages of non-volatility, radiation hardness, scalability down to the sub-5-nm range, and three-dimensional (3D) integration capability. Despite these advantages, nanomagnetic applications for information processing are limited today. The main stumbling block is the inadequately high energy required to switch information states in the spin-based nanodevices. Recently, the spin-transfer torque (STT) effect has been introduced as a promising solution [1-2]. In STT magnetic tunneling junctions (MTJs), using a spin-polarized electric current to switch magnetic states has the potential to solve the energy problem. However, the switching current density on the order of 1 MA/cm2 in the current STT-MTJ devices, with the smallest reported to date characteristic cross-sectional size on the order of 10 nm, still remains inadequately high for enabling a wide range of information processing applications [3-4]. For the technology to be competitive in the near future, it is critical to show that it could be favorably scaled into the sub-10-nm range. On a positive note, due to a new physics in this previously poorly explored size range, nanomagnetic devices may display promising characteristics that can make them superior to their semiconductor counterparts. In this size range, the underlying spin physics is due to the surface as much as it is due the volume. As a result, the effective thermal reservoir that absorbs spin excitations is substantially reduced, which in turn leads to a reduced spin damping and consequently to a reduced switching current density. Hence, the current study to understand the junction size dependence of the spin switching current is timely. This paper presents the key results of this study.

Fundus image mosaic can expand the visual field of view and provide auxiliary diagnosis for fundus diseases. Due to the low contrast and uneven grayscale of fundus images, traditional methods based on gray characteristic are no longer applicable. We propose a fundus image mosaic algorithm based on Scale-invariant feature transform (SIFT), along with a quantitative evaluation criterion. We detect feature points in Gaussian scale space and generate feature descriptors with scale invariance for all feature points. The matching feature point pairs are obtained by evaluating the similarity of their feature descriptors and outliers are eliminated using Random sample consensus (RANSAC) algorithm employed the projective transformation model. After that, the transformation matrix can be calculated based on the obtained matching point pairs. Finally, the multi-view fundus images are fused based on the transformation matrix. To the best of our knowledge, there is no quantitative evaluation reported for fundus image mosaic. Here we propose a criterion based on manual marker to evaluate the mosaic results. The experimental results show that the average accuracy is 2.27±0.5 pixels. Our proposed method has great potential to be applied in existing computer-assisted diagnostic system of ophthalmologic diseases.

The purpose of retinal image registration is to establish the coherent correspondences between the multi-model retinal image for applying into the ophthalmological surgery. Vessel landmarks detection in retinal image is the vital step in the retinal image registration. In this paper, a novel approach is proposed, firstly, a deep learning technology is used to vessel segmentation to generate the probability map of the retinal image, which is more reliable for optimizing the feature detection in retinal image. Secondly, we detect the landmarks using the multi-scale Hessian response on the probability map of the retinal image. Compared to the traditional methods, the results show that our method enable a majority of the bifurcation points, crossover points and curvature extreme points to be detected out simultaneously. Moreover, the impact of image noise and pathology can be reduced significantly.

The demand for data service is increasing dramatically and wireless systems with high throughput and the capability to serve a large number of User Equipments (UEs) are desired. The massive MIMO system is considered as one of the most promising systems for the fifth generation of mobile telecommunication technology. Compared to conventional wireless systems, the spectrum efficiency of massive MIMO can be increased by an order of magnitude as tens of UEs can be served on the same time-frequency resource by exploring spatial multiplexing. Though extensively studied, massive MIMO still faces tough practical issues, such as excessive cost for fronthaul, channel estimation, beamforming matrix calculation within the latency requirement, signal quality at the deep fading zones or cell edge, etc. This paper presents a repeater-enhanced massive MIMO system with improved channel quality and reduced fronthaul cost, where Amplify-and-Forward Repeaters (AFRs) are employed without additional baseband processing at the base station or UEs. The disadvantages of AFRs, such as extra delay, increased delay spread and compromised channel reciprocity, are discussed and the corresponding solutions are suggested. Performances of different beamforming methods are analyzed and verified by simulations.

The millimeter Wave (mmWave) is considered as a promising technology of the future fifth-generation wireless systems due to the currently underutilized multi-GHz spectrum surrounding the 60GHz carrier frequency, particularly with the recent advances in low cost sub-terahertz semiconductor circuitry. Since the mmWave suffers high attenuation during transmission, the repeater is a key technique to enable mmWave systems with seamless coverage. This paper considers the following practical design issues of Amplified-and-Forward Repeater (AFR) enhanced mmWave systems. First, the distribution of excess delay through multiple AFR hops is derived, which helps the design of cyclic-prefix length to avoid inter-symbol interference in orthogonal frequency division multiplexing systems. In addition, we show that the AFR decreases the channel coherence bandwidth and then propose a novel AFR design with a finite impulse response filter based channel equalizer to address this problem. Simulation results show that the proposed method is able to significantly increase channel coherence bandwidth and effectively reduce bit error rate.

Zero-Forcing (ZF) has been considered as one of the potential practical precoding and detection method for massive MIMO systems. One of the most important advantages of massive MIMO is the capability of supporting a large number of users in the same time-frequency resource, which requires much larger dimensions of matrix inversion for ZF than conventional multi-user MIMO systems. In this case, Neumann Series (NS) has been considered for the Matrix Inversion Approximation (MIA), because of its suitability for massive MIMO systems and its advantages in hardware implementation. The performance-complexity trade-off and the hardware implementation of NS-based MIA in massive MIMO systems have been discussed. In this paper, we analyze the effects of the ratio of the number of massive MIMO antennas to the number of users on the performance of NS-based MIA. In addition, we derive the approximation error estimation formulas for different practical numbers of terms of NS-based MIA. These results could offer useful guidelines for practical massive MIMO systems., accepted to conference; Proc. IEEE ICC 2015

Massive Multiple-Input Multiple-Output (MIMO) systems, dense Small-Cells (SCs), and full duplexing are three candidates for next-generation wireless systems. The cooperation of the three techniques could offer more benefits, e.g., SC in-band wireless backhaul in massive MIMO systems. In this paper, three strategies of SC in-band wireless backhaul in massive MIMO systems are introduced and compared. Simulation results show that SC in-band wireless backhaul has the potential to improve the throughput for massive MIMO systems, and applying full-duplexing techniques at SCs could provide greater gain.

In this paper, we proposed an intelligent campus equipment ubiquitous-management system. This RFID-based intelligent ubiquitous-management system use RFID to monitor the status of equipment in real time, and use wire or wireless network to send real-time situation to display on manager's PC or PDA. In addition, the RFID-based intelligent management system will also synchronize with database to record and reserve message. Furthermore, the status will display not only to a single manager but also a number of managers. In order to increase efficiency between Graphical User Interface (GUI) and Database, the system adopts Sql Dependency Object of ADO.NET so that the database changed any situation (eg. Add, Modify, and Delete) could be known immediately and synchronized with manager's PC or PDA. In order to obtain a feasible solution to this NP-Complete problem efficiently, we apply genetic algorithm to enhance the efficiency of finding optimum solution for equipment utilization. We could assign constraints into this system, and it will post back the optimum solution simultaneously on the screen. As a consequence, we compare the simulated annealing based approach (SA) proposed in [4][6] with our genetic algorithm based approach (GA) for maximizing the equipment utilization. Experimental result shows experimentally that our GA approach achieves an average 79.66% improvement in equipment utilization in an acceptable run time.

In order to study the influence of stator ventilation channel cutting length on its eddy current loss, fluid flow inside ventilation groove and stator temperature, the 1000MW hydro-generator is taken as an example in this paper. Its three-dimension transient electromagnetic field model under double poles and coupled field model of fluid and temperature under one slot pitch are created. The eddy current loss of ventilation channel and the influence of cutting length on that are obtained by the transient electromagnetic field. The distribution of fluid velocity inside stator radial ventilation groove and stator temperature is obtained by the coupled field. The variation of radial fluid velocity and stator temperature is obtained by quantitative analysis when different cutting length is applied in ventilation channel. Results show that the appropriate cutting length of the ventilation channel can be conductive to the fluid flow inside the ventilation groove and lower stator temperature. The conclusion can provide reference for the optimization design of ventilation system of large hydro-generator.

Thousands of equipments of the wireless network and the mobile devices are widely used and the demand of dynamic spectrum resources is significantly growing. How to maximize the social utility in modern technique becomes an important issue. In this paper, we propose a genetic algorithm based approach to allocate the power spectrum dynamically. The objective is to maximize the sum of individual Shannon utilities with the background interference and cross talk consideration. Compared to the approach in [4], experimental result shows that our approach obtains a better result within an acceptable amount of run time.

In this paper, a two-phase study was accomplished to explore the error identification and classification in terms of designing a self-check and correction circuit for crosstalk timing errors. A signature fault model (SFM) was firstly derived from an error space and a logical signature model (LSM) to recovery crosstalk timing errors such as glitch, delay, and speedup. Second, a transistor-level error detector, called crosstalk- error self-repairer (CESR), was designed to simplify the proposed SFM model. The proposed circuit has the capability of online error detection, correction, and error tolerant to obtain more reliable on-chip communication. In addition, the experimental results had been conducted thoroughly to demonstrate the effectiveness of our proposed work.

Based on the current situation which the bucket wheel excavator systems can't been arranged according to the previous design scheme at Heidaigou Open-pit Coal Mine, three feasible alternatives to arrange bucket wheel excavator systems are put forward. Then their reliability is calculated, and the optimal alternative is recommended according to the calculated reliability. Finally the economic benefit is estimated if the recommend alternative were put in practice.

Objective: To preliminarily investigate the correlation between coronary plaque type and acute coronary syndrome (ACS) using 320-slice volume computed tomography (CT). Methods: Fifty-seven patients who were suspected of having ACS and underwent CT within 30 days of onset were included in this retrospective analysis. All patients underwent 320-slice volume CT coronary image examination and were divided into four groups according to the coronary plaque components: Type I, fully calcified plaque; Type II, mixed plaque (calcification > 50%); Type III, mixed plaque (non-calcification > 50%); Type IV, non-calcified plaque. Results: Coronary plaques were found in 53/57 patients (93%). Of these, 16 (30.2%) had type I, nine (17.0%) had type II, 15 (28.3%) had type III and 13 (24.5%) had type IV plaques, respectively. The presence of Type II-IV non-calcified plaques was significantly associated with ACS (17/53, 32.1%, p 0.05). Conclusion: 320-slice volume CT coronary image can clearly depict the components of coronary plaque. Mixed plaques with non-calcified component > 50% are significantly associated with ACS.

Humorous language cannot stand by its own since language makes sense only in the position of discourse, in the situational and cultural context. Discourse Analysis has an interpretive power to count for humor.

In modern knowledge society, venture funds are playing important roles in SMEs financing and there are many risks in venture funds' investment. Can venture funds manage the risks with knowledge management? That is the topic the paper will study. The paper first discusses the definitions of venture funds and knowledge, and then analyzes the need for knowledge management in venture funds to manage risks. Finally, the paper takes market risk management as an example to study ways to manage risks with knowledge management. With the application of knowledge management in venture funds, SMEs will acquire more funds and venture funds will acquire more rewards.

The design of measurement and control system of the radon detection instrument, which used for geological prospecting and environmental monitoring, is introduced in this paper. This paper includes the present research situation of the radon detection instrument, the software and hardware design, the focus is the data processing software design which based on the fourth language Delphi. The data processing software includes the data serial communication, the data acceptance, the data display, the data storage, the curve drawing, the database management and so on.

consider the start-up time limits, start-up sequences, recovery paths and operation constraints, etc. Ref.[12] and [13] apply genetic algorithm to system partitioning with the unit start-up time limits in consideration, but they restore only a part of predetermined units and load buses and do not consider the balance of the recovery times of the subsystems. This paper considers the units’ start characteristics and the system state, partitions the system according to the black-start power sources with the aim of restoring most units and substations in the shortest time. The partitioning results obtained include all the units and substations that could be start within the time limit, the subsystems with approximately same restoration times and their recovery sequences. The proposed parallel restoration strategy can effectively reduce the time consumption of the black-start restoration process.