Search

The dark channel prior (DCP)-based single image removal algorithm achieved excellent performance. However, due to the high complexity of the algorithm, it is difficult to satisfy the demands of real-time processing. In this article, we present a Graphics Processing Unit (GPU) accelerated parallel computing method for the real-time processing of high-definition video haze removal. First, based on the memory access pattern, we propose a simple but effective filter method called transposed filter combined with the fast local minimum filter algorithm and integral image algorithm. The proposed method successfully accelerates the parallel minimum filter algorithm and the parallel mean filter algorithm. Meanwhile, we adopt the inter-frame atmospheric light constraint to suppress the flicker noise in the video haze removal and simplify the estimation of atmospheric light. Experimental results show that our implementation can process the 1080p video sequence with 167 frames per second. Compared with single thread Central Processing Units (CPU) implementation, the speedup is up to 226&times, with asynchronous stream processing and qualified for the real-time high definition video haze removal.

This paper presents a mechanical control method for precise weeding based on deep learning. Deep convolutional neural network was used to identify and locate weeds. A special modular weeder was designed, which can be installed on the rear of a mobile platform. An inverted pyramid-shaped weeding tool equipped in the modular weeder can shovel out weeds without being contaminated by soil. The weed detection and control method was implemented on an embedded system with a high-speed graphics processing unit and integrated with the weeder. The experimental results showed that even if the speed of the mobile platform reaches 20 cm/s, the weeds can still be accurately detected and the position of the weeds can be located by the system. Moreover, the weeding mechanism can successfully shovel out the roots of the weeds. The proposed weeder has been tested in the field, and its performance and weed coverage have been verified to be precise for weeding.

While the cloth component in Unity engine has been used to represent the 3D cloth object for augmented reality (AR) and virtual reality (VR), it has several limitations in term of resolution and performance. The purpose of our research is to develop a stable cloth simulation based on a parallel algorithm. The method of a mass–spring system is applied to real-time cloth simulation with three types of springs. However, cloth simulation using the mass–spring system requires a small integration time-step to use a large stiffness coefficient. Furthermore, constraint enforcement is applied to obtain the stable behavior of the cloth model. To reduce the computational burden of constraint enforcement, the adaptive constraint activation and deactivation (ACAD) technique that includes the mass–spring system and constraint enforcement method is applied to prevent excessive elongation of the cloth. The proposed algorithm utilizes the graphics processing unit (GPU) parallel processing, and implements it in Compute Shader that executes in different pipelines to the rendering pipeline. In this paper, we investigate the performance and compare the behavior of the mass–spring system, constraint enforcement, and ACAD techniques using a GPU-based parallel method.

Recent advances in the development of large eddy simulation (LES) atmospheric models with corresponding atmospheric transport and dispersion (AT&amp, D) modeling capabilities have made it possible to simulate short, time-averaged, single realizations of pollutant dispersion at the spatial and temporal resolution necessary for common atmospheric dispersion needs, such as designing air sampling networks, assessing pollutant sensor system performance, and characterizing the impact of airborne materials on human health. The high computational burden required to form an ensemble of single-realization dispersion solutions using an LES and coupled AT&amp, D model has, until recently, limited its use to a few proof-of-concept studies. An example of an LES model that can meet the temporal and spatial resolution and computational requirements of these applications is the joint outdoor-indoor urban large eddy simulation (JOULES). A key enabling element within JOULES is the computationally efficient graphics processing unit (GPU)-based LES, which is on the order of 150 times faster than if the LES contaminant dispersion simulations were executed on a central processing unit (CPU) computing platform. JOULES is capable of resolving the turbulence components at a suitable scale for both open terrain and urban landscapes, e.g., owing to varying environmental conditions and a diverse building topology. In this paper, we describe the JOULES modeling system, prior efforts to validate the accuracy of its meteorological simulations, and current results from an evaluation that uses ensembles of dispersion solutions for unstable, neutral, and stable static stability conditions in an open terrain environment.

Plenoptic camera based system captures the light-field that can be exploited to estimate the 3D depth of the scene. This process generally consists of a significant number of recurrent operations, and thus requires high computation power. General purpose processor based system, due to its sequential architecture, consequently results in the problem of large execution time. A desktop graphics processing unit (GPU) can be employed to resolve this problem. However, it is an expensive solution with respect to power consumption and therefore cannot be used in mobile applications with low energy requirements. In this paper, we propose a modified plenoptic depth estimation algorithm that works on a single frame recorded by the camera and respective FPGA based hardware design. For this purpose, the algorithm is modified for parallelization and pipelining. In combination with efficient memory access, the results show good performance and lower power consumption compared to other systems.

Molecular docking is widely used in computed drug discovery and biological target identification, but getting fast results can be tedious and often requires supercomputing solutions. AMIDE stands for AutoMated Inverse Docking Engine. It was initially developed in 2014 to perform inverse docking on High Performance Computing. AMIDE version 2 brings substantial speed-up improvement by using AutoDock-GPU and by pulling a total revision of programming workflow, leading to better performances, easier use, bug corrections, parallelization improvements and PC/HPC compatibility. In addition to inverse docking, AMIDE is now an optimized tool capable of high throughput inverse screening. For instance, AMIDE version 2 allows acceleration of the docking up to 12.4 times for 100 runs of AutoDock compared to version 1, without significant changes in docking poses. The reverse docking of a ligand on 87 proteins takes only 23 min on 1 GPU (Graphics Processing Unit), while version 1 required 300 cores to reach the same execution time. Moreover, we have shown an exponential acceleration of the computation time as a function of the number of GPUs used, allowing a significant reduction of the duration of the inverse docking process on large datasets.

Recently, ray tracing techniques have been highly adopted to produce high quality images and animations. In this paper, we present our design and implementation of a real-time ray-traced rendering engine. We achieved real-time capability for triangle primitives, based on the ray tracing techniques on GPGPU (general-purpose graphics processing unit) compute shaders. To accelerate the ray tracing engine, we used a set of acceleration techniques, including bounding volume hierarchy, its roped representation, joint up-sampling, and bilateral filtering. Our current implementation shows remarkable speed-ups, with acceptable error values. Experimental results shows 2.5–13.6 times acceleration, and less than 3% error values for the 95% confidence range. Our next step will be enhancing bilateral filter behaviors.

Methods of increasing complexity are currently being proposed for ultrasound (US) echographic signal processing. Graphics Processing Unit (GPU) resources allowing massive exploitation of parallel computing are ideal candidates for these tasks. Many high-performance US instruments, including open scanners like ULA-OP 256, have an architecture based only on Field-Programmable Gate Arrays (FPGAs) and/or Digital Signal Processors (DSPs). This paper proposes the implementation of the embedded NVIDIA Jetson Xavier AGX module on board ULA-OP 256. The system architecture was revised to allow the introduction of a new Peripheral Component Interconnect Express (PCIe) communication channel, while maintaining backward compatibility with all other embedded computing resources already on board. Moreover, the Input/Output (I/O) peripherals of the module make the ultrasound system independent, freeing the user from the need to use an external controlling PC.

The k-means problem is one of the most popular models in cluster analysis that minimizes the sum of the squared distances from clustered objects to the sought cluster centers (centroids). The simplicity of its algorithmic implementation encourages researchers to apply it in a variety of engineering and scientific branches. Nevertheless, the problem is proven to be NP-hard which makes exact algorithms inapplicable for large scale problems, and the simplest and most popular algorithms result in very poor values of the squared distances sum. If a problem must be solved within a limited time with the maximum accuracy, which would be difficult to improve using known methods without increasing computational costs, the variable neighborhood search (VNS) algorithms, which search in randomized neighborhoods formed by the application of greedy agglomerative procedures, are competitive. In this article, we investigate the influence of the most important parameter of such neighborhoods on the computational efficiency and propose a new VNS-based algorithm (solver), implemented on the graphics processing unit (GPU), which adjusts this parameter. Benchmarking on data sets composed of up to millions of objects demonstrates the advantage of the new algorithm in comparison with known local search algorithms, within a fixed time, allowing for online computation.

Smarter applications are making better use of the insights gleaned from data, having an impact on every industry and research discipline. At the core of this revolution lies the tools and the methods that are driving it, from processing the massive piles of data generated each day to learning from and taking useful action. Deep neural networks, along with advancements in classical ML and scalable general-purpose GPU computing, have become critical components of artificial intelligence, enabling many of these astounding breakthroughs and lowering the barrier to adoption. Python continues to be the most preferred language for scientific computing, data science, and machine learning, boosting both performance and productivity by enabling the use of low-level libraries and clean high-level APIs. This survey offers insight into the field of machine learning with Python, taking a tour through important topics to identify some of the core hardware and software paradigms that have enabled it. We cover widely-used libraries and concepts, collected together for holistic comparison, with the goal of educating the reader and driving the field of Python machine learning forward., Preprint of a manuscript accepted for publication in "Machine Learning with Python," a special issue of Information (ISSN 2078-2489)

An open accelerator (OpenACC)-aided graphics processing unit (GPU)-based finite difference time domain (FDTD) method is presented for the first time for the 3D evaluation of lightning radiated electromagnetic fields along a complex terrain with arbitrary topography. The OpenACC directive-based programming model is used to enhance the computational performance, and the results are compared with those obtained by using a CPU-based model. It is shown that OpenACC GPUs can provide very accurate results, and they are more than 20 times faster than CPUs. The presented results support the use of OpenACC not only in relation to lightning electromagnetics problems, but also to large-scale realistic electromagnetic compatibility (EMC) applications in which computation time efficiency is a critical factor.

In this study, a CUDA Fortran-based GPU-accelerated Laplace equation model was developed and applied to several cases. The Laplace equation is one of the equations that can physically analyze the groundwater flows, and is an equation that can provide analytical solutions. Such a numerical model requires a large amount of data to physically regenerate the flow with high accuracy, and requires computational time. These numerical models require a large amount of data to physically reproduce the flow with high accuracy and require computational time. As a way to shorten the computation time by applying CUDA technology, large-scale parallel computations were performed on the GPU, and a program was written to reduce the number of data transfers between the CPU and GPU. A GPU consists of many ALUs specialized in graphic processing, and can perform more concurrent computations than a CPU using multiple ALUs. The computation results of the GPU-accelerated model were compared with the analytical solution of the Laplace equation to verify the accuracy. The computation results of the GPU-accelerated Laplace equation model were in good agreement with the analytical solution. As the number of grids increased, the computational time of the GPU-accelerated model gradually reduced compared to the computational time of the CPU-based Laplace equation model. As a result, the computational time of the GPU-accelerated Laplace equation model was reduced by up to about 50 times.

This paper proposes a framework for the wireless sensor data acquisition using a team of Unmanned Aerial Vehicles (UAVs). Scattered over a terrain, the sensors detect information about their surroundings and can transmit this information wirelessly over a short range. With no access to a terrestrial or satellite communication network to relay the information to, UAVs are used to visit the sensors and collect the data. The proposed framework uses an iterative k-means algorithm to group the sensors into clusters and to identify Download Points (DPs) where the UAVs hover to download the data. A Single-Source–Shortest-Path algorithm (SSSP) is used to compute optimal paths between every pair of DPs with a constraint to reduce the number of turns. A genetic algorithm supplemented with a 2-opt local search heuristic is used to solve the multi-travelling salesperson problem and to find optimized tours for each UAVs. Finally, a collision avoidance strategy is implemented to guarantee collision-free trajectories. Concerned with the overall runtime of the framework, the SSSP algorithm is implemented in parallel on a graphics processing unit. The proposed framework is tested in simulation using three UAVs and realistic 3D maps with up to 100 sensors and runs in just 20.7 s, a 33.3× speed-up compared to a sequential execution on CPU. The results show that the proposed method is efficient at calculating optimized trajectories for the UAVs for data acquisition from wireless sensors. The results also show the significant advantage of the parallel implementation on GPU.

Fast reconstruction of power lines and corridors is a critical task in UAV (unmanned aerial vehicle)-based inspection of high-voltage transmission corridors. However, recent dense matching algorithms suffer the problem of low efficiency when processing large-scale high-resolution UAV images. This study proposes an efficient dense matching method for the 3D reconstruction of high-voltage transmission corridors with fine-scale power lines. First, an efficient random red-black checkerboard propagation is proposed, which utilizes the neighbor pixels with the most similar color to propagate plane parameters. To combine the pixel-wise view selection strategy adopted in Colmap with the efficient random red-black checkerboard propagation, the updating schedule for inferring visible probability is improved; second, strategies for decreasing the number of matching cost computations are proposed, which can reduce the unnecessary hypotheses for verification. The number of neighbor pixels necessary to propagate plane parameters is reduced with the increase of iterations, and the number of the combinations of depth and normal is reduced for the pixel with better matching cost in the plane refinement step; third, an efficient GPU (graphics processing unit)-based depth map fusion method is proposed, which employs a weight function based on the reprojection errors to fuse the depth map. Finally, experiments are conducted by using three UAV datasets, and the results indicate that the proposed method can maintain the completeness of power line reconstruction with high efficiency when compared to other PatchMatch-based methods. In addition, two benchmark datasets are used to verify that the proposed method can achieve a better F1 score, 4–7 times faster than Colmap.

Graphics processing units and video cards have seen a surge of usage in domains other than graphics computers, due to advances in hardware and software technologies, however, little uptake has been in the domain of systems engineering and real time control. This research article will demonstrate the use of video cards in multimodel adaptive robust control, using openGL and compute shaders. A software simulation will show the behavior of the adaptive robust multimodel control scheme as the target process is exposed to both parametric and structural disturbances and will show the viability of using graphics processing units in real time systems control.

This paper presents a near real-time automatic sub-pixel registration method of high-resolution panchromatic (PAN) and multispectral (MS) images using a graphics processing unit (GPU). In the first step, the method uses differential geo-registration to enable accurate geographic registration of PAN and MS images. Differential geo-registration normalizes PAN and MS images to the same direction and scale. There are also some residual misalignments due to the geometrical configuration of the acquisition instruments. These residual misalignments mean the PAN and MS images still have deviations after differential geo-registration. The second step is to use differential rectification with tiny facet primitive to eliminate possible residual misalignments. Differential rectification corrects the relative internal geometric distortion between PAN and MS images. The computational burden of these two steps is large, and traditional central processing unit (CPU) processing takes a long time. Due to the natural parallelism of the differential methods, these two steps are very suitable for mapping to a GPU for processing, to achieve near real-time processing while ensuring processing accuracy. This paper used GaoFen-6, GaoFen-7, ZiYuan3-02 and SuperView-1 satellite data to conduct an experiment. The experiment showed that our method’s processing accuracy is within 0.5 pixels. The automatic processing time of this method is about 2.5 s for 1 GB output data in the NVIDIA GeForce RTX 2080Ti, which can meet the near real-time processing requirements for most satellites. The method in this paper can quickly achieve high-precision registration of PAN and MS images. It is suitable for different scenes and different sensors. It is extremely robust to registration errors between PAN and MS.

History is often thought to be dull and boring – where large numbers of facts are memorized for passing exams. But the past informs the present and future, especially in delineating the context surrounding specific events that, in turn, help provide a deeper understanding of their causes and implications. Scientific progress (whether incremental or breakthroughs) is built upon prior work. Chronological examination of computational chemistry’s evolution reveals the existence of major “epochs” (e.g., transition from semi-empirical methods to first principles calculations), and the centrality of key ideas (e.g., Schrodinger equation and Born Oppenheimer approximation) in potentiating progress in the field. The longstanding question of whether computing power (both capacity and speed) or theoretical insights play a more important role in advancing computational chemistry was examined by taking into account the field’s development holistically. Specifically, availability of large amount of computing power at declining cost, and advent of graphics processing unit (GPU) powered parallel computing are enabling tools for solving hitherto intractable problems. On the other hand, this essay argues (using Born Oppenheimer approximation as an example) that theoretical insights’ role in unlocking problems through simple (but insightful) assumptions is often overlooked. Collectively, the essay should be useful as a primer for appreciating major development periods in computational chemistry, from which counterfactual questions illuminate the relative importance of theoretical insights and advances in computer science in moving the field forward.

The workloads of Convolutional Neural Networks (CNNs) exhibit a streaming nature that makes them attractive for reconfigurable architectures such as the Field-Programmable Gate Arrays (FPGAs), while their increased need for low-power and speed has established Application-Specific Integrated Circuit (ASIC)-based accelerators as alternative efficient solutions. During the last five years, the development of Hardware Description Language (HDL)-based CNN accelerators, either for FPGA or ASIC, has seen huge academic interest due to their high-performance and room for optimizations. Towards this direction, we propose a library-based framework, which extends TensorFlow, the well-established machine learning framework, and automatically generates high-throughput CNN inference engines for FPGAs and ASICs. The framework allows software developers to exploit the benefits of FPGA/ASIC acceleration without requiring any expertise on HDL development and low-level design. Moreover, it provides a set of optimization knobs concerning the model architecture and the inference engine generation, allowing the developer to tune the accelerator according to the requirements of the respective use case. Our framework is evaluated by optimizing the LeNet CNN model on the MNIST dataset, and implementing FPGA- and ASIC-based accelerators using the generated inference engine. The optimal FPGA-based accelerator on Zynq-7000 delivers 93% less memory footprint and 54% less Look-Up Table (LUT) utilization, and up to 10&times, speedup on the inference execution vs. different Graphics Processing Unit (GPU) and Central Processing Unit (CPU) implementations of the same model, in exchange for a negligible accuracy loss, i.e., 0.89%. For the same accuracy drop, the 45 nm standard-cell-based ASIC accelerator provides an implementation which operates at 520 MHz and occupies an area of 0.059 mm 2 , while the power consumption is &sim, 7.5 mW.

With the development of polarimetric synthetic aperture radar (PolSAR), quantitative parameter inversion has been seen great progress, especially in the field of soil parameter inversion, which has achieved good results for applications. However, PolSAR data is also often many terabytes large. This huge amount of data also directly affects the efficiency of the inversion. Therefore, the efficiency of soil moisture and roughness inversion has become a problem in the application of this PolSAR technique. A parallel realization based on a graphics processing unit (GPU) for multiple inversion models of PolSAR data is proposed in this paper. This method utilizes the high-performance parallel computing capability of a GPU to optimize the realization of the surface inversion models for polarimetric SAR data. Three classical forward scattering models and their corresponding inversion algorithms are analyzed. They are different in terms of polarimetric data requirements, application situation, as well as inversion performance. Specifically, the inversion process of PolSAR data is mainly improved by the use of the high concurrent threads of GPU. According to the inversion process, various optimization strategies are applied, such as the parallel task allocation, and optimizations of instruction level, data storage, data transmission between CPU and GPU. The advantages of a GPU in processing computationally-intensive data are shown in the data experiments, where the efficiency of soil roughness and moisture inversion is increased by one or two orders of magnitude.

High-quality terrain rendering has been the focus of many visualization applications over recent decades. Many terrain rendering methods use the strategy of Level of Detail (LOD) to create adaptive terrain models, but the transition between different levels is usually not handled well, which may cause popping artefacts that seriously affect the reality of the terrain model. In recent years, many researchers have tried using modern Graphics Processing Unit (GPU) to complete heavy rendering tasks. By leveraging the great power of GPU, high quality terrain models with rich details can be rendered in real time, although the problem of popping artefacts still persists. In this study, we propose a real-time terrain rendering method with GPU tessellation that can effectively reduce the popping artefacts. Coupled with a view-dependent updating scheme, a multilevel terrain representation based on the flexible Dynamic Stitching Strip (DSS) is developed. During rendering, the main part of the terrain model is tessellated into appropriate levels using GPU tessellation. DSSs, generated in parallel, can seamlessly make the terrain transitions between different levels much smoother. Experiments demonstrate that the proposed method can meet the requirements of real-time rendering and achieve a better visual quality compared with other methods.

As an active microwave sensor, synthetic aperture radar (SAR) has the characteristic of all-day and all-weather earth observation, which has become one of the most important means for high-resolution earth observation and global resource management. Ship detection in SAR images is also playing an increasingly important role in ocean observation and disaster relief. Nowadays, both traditional feature extraction methods and deep learning (DL) methods almost focus on improving ship detection accuracy, and the detection speed is neglected. However, the speed of SAR ship detection is extraordinarily significant, especially in real-time maritime rescue and emergency military decision-making. In order to solve this problem, this paper proposes a novel approach for high-speed ship detection in SAR images based on a grid convolutional neural network (G-CNN). This method improves the detection speed by meshing the input image, inspired by the basic thought of you only look once (YOLO), and using depthwise separable convolution. G-CNN is a brand new network structure proposed by us and it is mainly composed of a backbone convolutional neural network (B-CNN) and a detection convolutional neural network (D-CNN). First, SAR images to be detected are divided into grid cells and each grid cell is responsible for detection of specific ships. Then, the whole image is input into B-CNN to extract features. Finally, ship detection is completed in D-CNN under three scales. We experimented on an open SAR Ship Detection Dataset (SSDD) used by many other scholars and then validated the migration ability of G-CNN on two SAR images from RadarSat-1 and Gaofen-3. The experimental results show that the detection speed of our proposed method is faster than the existing other methods, such as faster-regions convolutional neural network (Faster R-CNN), single shot multi-box detector (SSD), and YOLO, under the same hardware environment with NVIDIA GTX1080 graphics processing unit (GPU) and the detection accuracy is kept within an acceptable range. Our proposed G-CNN ship detection system has great application values in real-time maritime disaster rescue and emergency military strategy formulation.

The high demand for computational resources severely hinders the deployment of deep learning applications in resource-limited devices. In this work, we investigate the under-studied but practically important network efficiency problem and present a new, lightweight architecture for hand pose estimation. Our architecture is essentially a deeply-supervised pruned network in which less important layers and branches are removed to achieve a higher real-time inference target on resource-constrained devices without much accuracy compromise. We further make deployment optimization to facilitate the parallel execution capability of central processing units (CPUs). We conduct experiments on NYU and ICVL datasets and develop a demo1 using the RealSense camera. Experimental results show our lightweight network achieves an average running time of 32 ms (31.3 FPS, the original is 22.7 FPS) before deployment optimization. Meanwhile, the model is only about half parameters size of the original one with 11.9 mm mean joint error. After the further optimization with OpenVINO, the optimized model can run at 56 FPS on CPUs in contrast to 44 FPS running on a graphics processing unit (GPU) (Tensorflow) and it can achieve the real-time goal.

We are developing a software for energy loss simulation which is affected by jets in the nuclear matter described by relativistic hydrodynamics. Our program uses a Cartesian coordinate system in order to provide high spatial resolution for the analysis of jets propagation in nuclear matter. In this work, we use 7th order WENO numerical algorithm which is resistant to numerical oscillations and diffusions. For simulating energy losses in the bulk nuclear medium, we develop efficient hydrodynamic simulation program for parallel computing using Graphics Processing Unit (GPU) and Compute Unified Device Architecture (CUDA). It allows us to prepare event-by-event simulations in high computing precision in order to study jet modifications in the medium and event-by-event simulations of fluctuating initial conditions. In our simulation, we start the hydrodynamic simulation from generation initial condition based on the UrQMD model in order to simulate comparable nucleus-nucleus interaction in the RHIC and LHC energies. The main part of this simulation is the computation of hydrodynamic system evolution. We present obtained energy density distributions which can be compared to experimental results.

Welding frames with differing geometries is one of the most crucial stages in the production of high-end bicycles. This paper proposes a parallel algorithm and a mixed integer linear programming formulation for scheduling a two-machine robotic welding station. The time complexity of the introduced parallel method is O(log2n) on an n3-processor Exclusive Read Exclusive Write Parallel Random-Access Machine (EREW PRAM), where n is the problem size. The algorithm is designed to take advantage of modern graphics cards to significantly accelerate the computations. To present the benefits of the parallelization, the algorithm is compared to the state of art sequential method and a solver-based approach. Experimental results show an impressive speedup for larger problem instances—up to 314 on a single Graphics Processing Unit (GPU), compared to a single-threaded CPU execution of the sequential algorithm.

Smuggling of drugs and cigarettes in small inflatable boats across border rivers is a serious threat to the EU’s financial interests. Early detection of such threats is challenging due to difficult and changing environmental conditions. This study reports on the automatic detection of small inflatable boats and people in a rough wild terrain in the infrared thermal domain. Three acquisition campaigns were carried out during spring, summer, and fall under various weather conditions. Three deep learning algorithms, namely, YOLOv2, YOLOv3, and Faster R-CNN working with six different feature extraction neural networks were trained and evaluated in terms of performance and processing time. The best performance was achieved with Faster R-CNN with ResNet101, however, processing requires a long time and a powerful graphics processing unit.

This paper present contributions to the state-of-the art for graphics processing unit (GPU-based) embedded intelligence (EI) research for architectures and applications. This paper gives a comprehensive review and representative studies of the emerging and current paradigms for GPU-based EI with the focus on the architecture, technologies and applications: (1) First, the overview and classifications of GPU-based EI research are presented to give the full spectrum in this area that also serves as a concise summary of the scope of the paper; (2) Second, various architecture technologies for GPU-based deep learning techniques and applications are discussed in detail; and (3) Third, various architecture technologies for machine learning techniques and applications are discussed. This paper aims to give useful insights for the research area and motivate researchers towards the development of GPU-based EI for practical deployment and applications.

To realize deep learning techniques, a type of deep neural network (DNN) called a convolutional neural networks (CNN) is among the most widely used models aimed at image recognition applications. However, there is growing demand for light-weight and low-power neural network accelerators, not only for inference but also for training process. In this paper, we propose a training accelerator that provides low power and compact chip size targeted for mobile and edge computing applications. It accelerates to achieve the real-time processing of both inference and training using concurrent floating-point data paths. The proposed accelerator can be externally controlled and employs resource sharing and an integrated convolution-pooling block to achieve low area and low energy consumption. We implemented the proposed training accelerator in an FPGA (Field Programmable Gate Array) and evaluated its training performance using an MNIST CNN example in comparison with a PC with GPU (Graphics Processing Unit). While both methods achieved a similar training accuracy of 95.1%, the proposed accelerator, when implemented in a silicon chip, reduced the energy consumption by 480 times compared to the counterpart. Additionally, when implemented on an FPGA, an energy reduction of over 4.5 times was achieved compared to the existing FPGA training accelerator for the MNIST dataset. Therefore, the proposed accelerator is more suitable for deployment in mobile/edge nodes compared to the existing software and hardware accelerators.

Gravity data have been playing an important role in marine exploration and research. However, obtaining gravity data over an extensive marine area is expensive and inefficient. In reality, marine gravity anomalies are usually calculated from satellite altimetry data. Over the years, numerous methods have been presented for achieving this purpose, most of which are time-consuming due to the integral calculation over a global region and the singularity problem. This paper proposes a fast method for the calculation of marine gravity anomalies. The proposed method introduces a novel scheme to solve the singularity problem and implements the parallel technique based on a graphics processing unit (GPU) for fast calculation. The details for the implementation of the proposed method are described, and it is tested using the geoid height undulation from the Earth Gravitational Model 2008 (EGM2008). The accuracy of the presented method is evaluated by comparing it with marine shipboard gravity data. Its efficiency is demonstrated through comparison with the conventional sequential method. The tests demonstrate that the proposed method can be employed for accurately calculating marine gravity anomalies and provides an advantage on computational efficiency.

Magnetic Resonance Imaging-based prostate segmentation is an essential task for adaptive radiotherapy and for radiomics studies whose purpose is to identify associations between imaging features and patient outcomes. Because manual delineation is a time-consuming task, we present three deep-learning (DL) approaches, namely UNet, efficient neural network (ENet), and efficient residual factorized convNet (ERFNet), whose aim is to tackle the fully-automated, real-time, and 3D delineation process of the prostate gland on T2-weighted MRI. While UNet is used in many biomedical image delineation applications, ENet and ERFNet are mainly applied in self-driving cars to compensate for limited hardware availability while still achieving accurate segmentation. We apply these models to a limited set of 85 manual prostate segmentations using the k-fold validation strategy and the Tversky loss function and we compare their results. We find that ENet and UNet are more accurate than ERFNet, with ENet much faster than UNet. Specifically, ENet obtains a dice similarity coefficient of 90.89% and a segmentation time of about 6 s using central processing unit (CPU) hardware to simulate real clinical conditions where graphics processing unit (GPU) is not always available. In conclusion, ENet could be efficiently applied for prostate delineation even in small image training datasets with potential benefit for patient management personalization.

As outlined in the 3Gpp Release 16, 5G satellite access is important for 5G network development in the future. A terrestrial-satellite network integrated with 5G has the characteristics of low delay, high bandwidth, and ubiquitous coverage. A few researchers have proposed integrated schemes for such a network, however, these schemes do not consider the possibility of achieving optimization of the delay characteristic by changing the computing mode of the 5G satellite network. We propose a 5G satellite edge computing framework (5GsatEC), which aims to reduce delay and expand network coverage. This framework consists of embedded hardware platforms and edge computing microservices in satellites. To increase the flexibility of the framework in complex scenarios, we unify the resource management of the central processing unit (CPU), graphics processing unit (GPU), and field-programmable gate array (FPGA), we divide the services into three types: system services, basic services, and user services. In order to verify the performance of the framework, we carried out a series of experiments. The results show that 5GsatEC has a broader coverage than the ground 5G network. The results also show that 5GsatEC has lower delay, a lower packet loss rate, and lower bandwidth consumption than the 5G satellite network.

Magnetic Resonance-Electrical Properties Tomography (MR-EPT) is a method to reconstruct the electrical properties (EPs) of bio-tissues from the measured radiofrequency (RF) field in Magnetic Resonance Imaging (MRI). Current MR-EPT approaches reconstruct the EP profile by solving a second-order partial differential wave equation problem, which is sensitive to noise and can induce large reconstruction artefacts near tissue boundaries and areas with inaccurate field measurements. In this paper, a novel MR-EPT approach is proposed, which is based on a direct solution to Maxwell&rsquo, s curl equations. The distribution of EPs is calculated by iteratively fitting the RF field calculated by the finite-difference-time-domain (FDTD) technique to the measured values. To solve the time-consuming problem of the iterative fitting, a graphics processing unit (GPU) is used to accelerate the FDTD technique to process the field calculation kernel. The new EPT method was evaluated by investigating a numerical head phantom, and it was found that EP values can be accurately calculated and were relatively insensitive to simulated RF field errors. The feasibility of the proposed method was further validated using phantom-based experimental data collected from a 9.4 Tesla (T) Magnetic Resonance Imaging (MRI) system. The results also indicated that more accurate EP values could be reconstructed from the new method compared with conventional methods. Moreover, even in the absence of phase information of the RF field, the proposed approach is still capable of offering robust EPT solutions, thus having improved practicality for potential clinical applications.

Object detection, one of the most fundamental and challenging problems in computer vision. Nowadays some dedicated embedded systems have emerged as a powerful strategy for deliver high processing capabilities including the NVIDIA Jetson family. The aim of the present work is the recognition of objects in complex rural areas through an embedded system, as well as the verification of accuracy and processing time. For this purpose, a low power embedded Graphics Processing Unit (Jetson Nano) has been selected, which allows multiple neural networks to be run in simultaneous and a computer vision algorithm to be applied for image recognition. As well, the performance of these deep learning neural networks such as ssd-mobilenet v1 and v2, pednet, multiped and ssd-inception v2 has been tested. Moreover, it was found that the accuracy and processing time were in some cases improved when all the models suggested in the research were applied. The pednet network model provides a high performance in pedestrian recognition, however, the sdd-mobilenet v2 and ssd-inception v2 models are better at detecting other objects such as vehicles in complex scenarios.

Parallel reservoir simulation is an important approach to solving real-time reservoir management problems. Recently, there is a new trend of using a graphics processing unit (GPU) to parallelize the reservoir simulations. Current GPU-aided reservoir simulations focus on compute unified device architecture (CUDA). Nevertheless, CUDA is not functionally portable across devices and incurs high amount of code. Meanwhile, domain decomposition is not well used for GPU-based reservoir simulations. In order to address the problems, we propose a parallel method with OpenACC to accelerate serial code and reduce the time and effort during porting an application to GPU. Furthermore, the GPU-aided domain decomposition is developed to accelerate the efficiency of reservoir simulation. The experimental results indicate that (1) the proposed GPU-aided approach can outperform the CPU-based one up to about two times, meanwhile with the help of OpenACC, the workload of the transplant code was reduced significantly by about 22 percent of the source code, (2) the domain decomposition method can further improve the execution efficiency up to 1.7&times, The proposed parallel reservoir simulation method is a efficient tool to accelerate reservoir simulation.

Numerous imaging techniques measure data that are mathematically wrapped to the finite interval [−π, π], corresponding to the principle value domain of the arctangent function. A wide range of reconstruction algorithms has been developed to obtain the true, unwrapped phase by adding an integral multiple of 2π to each point of the wrapped grid. However, the phase unwrapping procedure is hampered by the presence of noise, phase vortices or insufficiently sampled digital data. Unfortunately, reliable phase unwrapping algorithms are generally computationally intensive and their design often requires multiple iterations to reach convergence, leading to high execution times. In this paper, we present a high-speed phase unwrapping algorithm that is robust against noise and phase residues. By executing the parallel implementation of a single-step Fourier-based phase unwrapping algorithm on the graphics processing unit of a standard graphics card, we were able to reduce the total processing time of the phase unwrapping algorithm to &lt, 5 ms when executed on a 640 × 480-pixel input map containing an arbitrarily high density of phase jumps. In addition, we expand upon this technique by inserting the obtained solution as a preconditioner in the conjugate gradient technique. This way, phase maps that contain regions of low-quality or invalid data can be unwrapped iteratively through weighting of local phase quality.