总访问量
今日访问
在线人数
To tackle the challenge of applying convolutional neural network (CNN) in field-programmable gate array (FPGA) due to its computational complexity, a high-performance CNN hardware accelerator based on Verilog hardware description language was designed, which utilizes a pipeline architecture with three parallel dimensions including input channels, output channels, and convolution kernels. Firstly, two multiply-and-accumulate (MAC) operations were packed into one digital signal processing (DSP) block of FPGA to double the computation rate of the CNN accelerator. Secondly, strategies of feature map block partitioning and special memory arrangement were proposed to optimize the total amount of off-chip access memory and reduce the pressure on FPGA bandwidth. Finally, an efficient computational array combining multiplicative-additive tree and Winograd fast convolution algorithm was designed to balance hardware resource consumption and computational performance. The high parallel CNN accelerator was deployed in ZU3EG of Alinx, using the YOLOv3-tiny algorithm as the test object. The average computing performance of the CNN accelerator is 127.5 giga operations per second (GOPS). The experimental results show that the hardware architecture effectively improves the computational power of CNN and provides better performance compared with other existing schemes in terms of power consumption and the efficiency of DSPs and block random access memory (BRAMs).
Mobile robots have been used for many industrial scenarios which can realize automated manufacturing process instead of human workers. To improve the quality of the optimal rapidly-exploring random tree ( RRT* ) for planning path in dynamic environment, a high-quality dynamic rapidly-exploring random tree ( HQD-RRT* ) algorithm is proposed in this paper, which generates a high-quality solution with optimal path length in dynamic environment. This method proceeds in two stages: initial path generation and path re-planning. Firstly, the initial path is generated by an improved smart rapidly-exploring random tree ( RRT* -SMART) algorithm, and the state tree information is stored as prior knowledge. During the process of path execution, a strategy of obstacle avoidance is proposed to avoid moving obstacles. The cost and smoothness of path are considered to re-plan the initial path to improve the path quality in this strategy. Compared with related work, a higher-quality path in dynamic
environment can be achieved in this paper. HQD-RRT* algorithm can obtain an optimal path with better stability. Simulations on the static and dynamic environment are conducted to clarify the efficiency of HQD-RRT* in avoiding unknown obstacles.
To deal with the secrecy issues and energy efficiency issues in the unmanned aerial vehicles ( UAVs) assisted communication systems, an UAV-enabled multi-hop mobile relay system is studied in an urban environment. Multiple rotary-wing UAVs with energy budget considerations are employed as relays to forward confidential information between two ground nodes in the presence of multiple passive eavesdroppers. The system secrecy energy efficiency ( SEE), defined by the ratio of minimum achievable secrecy rate ( SR) to total propulsion energy consumption (PEC), is maximized via jointly optimizing the trajectory and transmit power of each UAV relay. To solve the formulated non-convex fractional optimization problem subject to mobility, transmit power and information-causality constraints, an effective iterative algorithm is proposed by applying the updated-rate-assisted block coordinate decent method, successive convex approximation (SCA) technique and Dinkelbach method. Simulation
results demonstrate the effectiveness of the proposed joint trajectory design and power control scheme.
The task of multimodal sentiment classification aims to associate multimodal information, such as images and texts with appropriate sentiment polarities. There are various levels that can affect human sentiment in visual and textual modalities. However, most existing methods treat various levels of features independently without having effective method for feature fusion. In this paper, we propose a multi-level fusion classification (MFC) model to predict the sentiment polarity based on the fusing features from different levels by exploiting the dependency among them. The proposed architecture leverages convolutional neural networks ( CNNs) with multiple layers to extract levels of features in image and text modalities. Considering the dependencies within the low-level and high-level features, a bi-directional (Bi) recurrent neural network (RNN) is adopted to integrate the learned features from different layers in CNNs. In addition, a conflict detection module is incorporated to address the conflict between modalities. Experiments on the Flickr dataset demonstrate that the MFC method achieves comparable performance compared with strong baseline methods.
As the core technology of optical networks virtualization, virtual optical network embedding ( VONE) enables multiple virtual network requests to share substrate elastic optical network ( EON) resources simultaneously and hence has been applicated in edge computing scenarios. In this paper, we propose a reinforced virtual optical network embedding ( R-VONE ) algorithm based on deep reinforcement learning ( DRL) to optimize network embedding policies automatically. The network resource attributes are extracted as the environment state for model training, based on which DRL agent can deduce the node embedding probability. Experimental results indicate that R-VONE presents a significant advantage with lower blocking probability and higher resource utilization.
With the rapid development of Internet of thing (IoT) technology, it has become a challenge to deal with the increasing number and diverse requirements of IoT services. By combining burgeoning network function virtualization ( NFV) technology with cloud computing and mobile edge computing ( MEC), an NFV-enabled cloud-and-edge-collaborative IoT (CECIoT) architecture can efficiently provide flexible service for IoT traffic in the form of a service function chain (SFC) by jointly utilizing edge and cloud resources. In this promising architecture, a difficult issue is how to balance the consumption of resource and energy in SFC mapping. To overcome this challenge, an intelligent energy-and-resource-balanced SFC mapping scheme is designed in this paper. It takes the comprehensive deployment consumption as the optimization goal, and applies a deep Q-learning(DQL)-based SFC mapping (DQLBM) algorithm as well as an energy-based topology adjustment (EBTA) strategy to make efficient use of the limited network resources, while satisfying the delay requirement of users. Simulation results show that the proposed scheme can decrease service delay, as well as energy and resource consumption.