本期目录

2022年 29卷 5期 刊出日期：2022-10-31



Special Topic: Artificial Intelligence of Things

Design of high parallel CNN accelerator based on FPGA for AIoT

林志坚 高学伟 陈小培 祝志鹏 杜小勇 陈平平

中国邮电高校学报(英文), 2022, 29 (5). doi： 10.19682/j.cnki.1005-8885.2022.0026

摘要 ( 681 ) PDF (3802 KB)( 98 )

参考文献 | 相关文章 | 计量指标

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).

Vehicle-following system based on deep reinforcement learning in marine scene

张新 娄皓然 蒋励 肖前浩 蔡著文

中国邮电高校学报(英文), 2022, 29 (5). doi： 10.19682/j.cnki.1005-8885.2022.0025

摘要 ( 410 ) PDF (5117 KB)( 59 )

参考文献 | 相关文章 | 计量指标

In order to solve the problems that the feature data type are not rich enough in the data collection process about the vehicle-following task in marine scene which results in a long model convergence time and high training difficulty, a two-stage vehicle-following system was proposed. Firstly, semantic segmentation model predicts the number of pixels of the followed target, then the number of pixels of the followed target is mapped to the position feature. Secondly, deep reinforcement learning algorithm enables the control equipment to make decision action, to ensure that two moving objects remain within the safe distance. The experimental results show that the two-stage vehicle-following system has a 40% faster convergence rate than the model without position feature, and the following stability is significantly improved by adding the position feature.

Saliency guided self-attention network for pedestrian attribute recognition in surveillance scenarios

李娜 武阳阳 刘颖 李大湘 高嘉乐

中国邮电高校学报(英文), 2022, 29 (5). doi： 10.19682/j.cnki.1005-8885.2022.0007

摘要 ( 288 ) PDF (1938 KB)( 60 )

参考文献 | 相关文章 | 计量指标

Pedestrian attribute recognition is often considered as a multi-label image classification task. In order to make full use of attribute-related location information, a saliency guided sel-attention network ( SGSA-Net) was proposed to weakly supervise attribute localization， without annotations of attribute-related regions. Saliency priors were integrated into the spatial attention module ( SAM ). Meanwhile，channel-wise attention and spatial attention were introduced into the network. Moreover, a weighted binary cross-entropy loss ( WCEL) function was employed to handle the imbalance of training data. Extensive experiments on richly annotated pedestrian ( RAP) and pedestrian attribute ( PETA) datasets demonstrated that SGSA-Net outperformed other state-of-the-art methods.




RFID network planning based on improved brain storm optimization algorithm

林子涵 郑嘉利 谢孝德 冯敏瑜 何思怡

中国邮电高校学报(英文), 2022, 29 (5). doi： 10.19682/j.cnki.1005-8885.2022.0008

摘要 ( 300 ) PDF (1408 KB)( 78 )

参考文献 | 相关文章 | 计量指标

In order to improve the service quality of radio frequency identification (RFID) systems, multiple objectives should be comprehensively considered. An improved brain storm optimization algorithm GABSO, which incorporated adaptive learning operator and golden sine operator into the original brain storm optimization (BSO) algorithm, was proposed to solve the problem of RFID network planning (RNP). GABSO algorithm introduces learning operator and golden sine operator to achieve a balance between exploration and development. Based on GABSO algorithm, an optimization model is established to optimize the position of the reader. The GABSO algorithm was tested on the RFID model and dataset, and was compared with other methods. The GABSO algorithm's tag coverage was increased by 9.62% over the Cuckoo search (CS) algorithm, and 7.70% over BSO. The results show that the GABSO algorithm could be successfully applied to solve the problem of RNP.

Random mating mayfly algorithm for RFID network planning

谢孝德 郑嘉利 林子涵 何思怡 冯敏瑜

中国邮电高校学报(英文), 2022, 29 (5). doi： 10.19682/j.cnki.1005-8885.2022.0010

摘要 ( 351 ) PDF (4520 KB)( 69 )

参考文献 | 相关文章 | 计量指标

In order to improve robustness and efficiency of the radio frequency identification (RFID) network, a random mating mayfly algorithm (RMMA) was proposed. Firstly, RMMA introduced the mechanism of random mating into the mayfly algorithm (MA), which improved the population diversity and enhanced the exploration ability of the algorithm in the early stage, and find a better solution to the RFID nework planning (RNP) problem. Secondly, in RNP, tags are usually placed near the boundaries of the working space, so the minimum boundary mutation strategy was proposed to make sure the mayflies which beyond the boundary can keep the original search direction, as to enhance the ability of searching near the boundary. Lastly, in order to measure the performance of RMMA, the algorithm is then benchmarked on three well -known classic test functions, and the results are verified by a comparative study with particle swarm optimization (PSO), grey wolf optimization (GWO), and MA. The results show that the RMMA algorithm is able to provide very competitive results compared to these well-known meta-heuristics, RMMA is also applied to solve RNP problems. The performance evaluation shows that RMMA achieves higher coverage than the other three algorithms. When the number of readers is the same, RMMA can obtain lower interference and get a better load balance in each instance compared with other algorithms. RMMA can also solve RNP problem stably and efficiently when the number and position of tags change over time.

Artificial Intelligence

Automatic modulation classification based on AlexNet with data augmentation

张承畅 徐余 杨建鹏 李晓梦

中国邮电高校学报(英文), 2022, 29 (5). doi： 10.19682/j.cnki.1005-8885.2022.0009

摘要 ( 324 ) PDF (5760 KB)( 73 )

参考文献 | 相关文章 | 计量指标

Deep learning (DL) requires massive volume of data to train the network. Insufficient training data will cause serious overfitting problem and degrade the classification accuracy. In order to solve this problem, a method for automatic modulation classification ( AMC) using AlexNet with data augmentation was proposed. Three data augmentation methods is considered, i. e. , random erasing, CutMix, and rotation. Firstly, modulated signals are converted into constellation representations. And all constellation representations are divided into training dataset and test dataset. Then, training dataset are augmented by three methods. Secondly, the optimal value of execution probability for random erasing and CutMix are determined. Simulation results show that both of them perform optimally when execution probability is 0.5. Thirdly, the performance of three data augmentation methods are evaluated. Simulation results demonstrate that all augmentation methods can improve the classification accuracy. Rotation improves the classification accuracy by 13.04% when signal noise ratio (SNR) is 2 dB. Among three methods, rotation outperforms random erasing and CutMix when SNR is greater than - 6 dB. Finally, compared with other classification algorithms, random erasing, CutMix, and rotation used in this paper achieved the performance significantly improved. It is worth mentioning that the classification accuracy can reach 90.5% with SNR at 10 dB.

Spatiotemporal emotion recognition based on 3D time-frequency domain feature matrix

晁浩 连卫芳 刘永利

中国邮电高校学报(英文), 2022, 29 (5). doi： 10.19682/j.cnki.1005-8885.2022.0017

摘要 ( 354 ) PDF (4050 KB)( 54 )

参考文献 | 相关文章 | 计量指标

The research of emotion recognition based on electroencephalogram (EEG) signals often ignores the related information between the brain electrode channels and the contextual emotional information existing in EEG signals, which may contain important characteristics related to emotional states. Aiming at the above defects, a spatiotemporal emotion recognition method based on a 3-dimensional (3D) time-frequency domain feature matrix was proposed. Specifically, the extracted time-frequency domain EEG features are first expressed as a 3D matrix format according to the actual position of the cerebral cortex. Then, the input 3D matrix is processed successively by multivariate convolutional neural network (MVCNN) and long short-term memory (LSTM) to classify the emotional state. Spatiotemporal emotion recognition method is evaluated on the DEAP data set, and achieved accuracy of 87.58% and 88.50% on arousal and valence dimensions respectively in binary classification tasks, as well as obtained accuracy of 84.58% in four class classification tasks. The experimental results show that 3D matrix representation can represent emotional information more reasonably than two-dimensional (2D). In addition, MVCNN and LSTM can utilize the spatial information of the electrode channels and the temporal context information of the EEG signal respectively.




Research on emotional space for movie and TV drama videos

李玉杰 张晶晶 蒋伟 王春晓

中国邮电高校学报(英文), 2022, 29 (5). doi： 10.19682/j.cnki.1005-8885.2022.0006

摘要 ( 380 ) PDF (2372 KB)( 55 )

参考文献 | 相关文章 | 计量指标

Emotional space refers to a multi-dimensional emotional model that describes a group of subjective feelings or emotions. Since the existing discrete emotional space is mainly aimed at human’s primary emotions, it cannot describe the complex emotions evoked when watching movies. In order to solve this problem, an emotional fusion space for videos was constructed by selecting movies and TV dramas with rich emotional semantics as the research objects. Firstly, emotional words based on movie and TV drama videos are acquired and analyzed by using subjective evaluation and semantic analysis methods. Then, the emotional word vectors obtained from the above analysis are fused, reduced dimension by t-distributed stochastic neighbor embedding (t-SNE) algorithm, and clustered by bisecting K-means clustering algorithm to get a discrete emotional space for movie and TV drama videos. This emotional fusion space can obtain different categories by changing the value of the emotion classification number without re-labeling and calculation.




TSR: algorithm of image hole-filling based on three-step repairing

李富澄 邓军勇 朱筠 罗佳莹 任含

中国邮电高校学报(英文), 2022, 29 (5). doi： 10.19682/j.cnki.1005-8885.2022.0005

摘要 ( 286 ) PDF (6258 KB)( 57 )

参考文献 | 相关文章 | 计量指标

In order to solve the hole-filling mismatch problem in virtual view synthesis, a three-step repairing (TSR) algorithm was proposed. Firstly, the image with marked holes is decomposed by the non-subsampled shear wave transform ( NSST), which will generate high-/ low-frequency sub-images with different resolutions. Then the improved Criminisi algorithm was used to repair the texture information in the high-frequency sub-images, while the improved curvature driven diffusion (CDD) algorithm was used to repair the low-frequency sub-images with the image structure information. Finally, the repaired parts of high-frequency and low-frequency sub-images are synthesized to obtain the final image through inverse NSST. Experiments show that the peak signal-to-noise ratio (PSNR) of the TSR algorithm is improved by an average of 2 - 3 dB and 1 - 2 dB compared with the Criminisi algorithm and the nearest neighbor interpolation (NNI) algorithm, respectively.

FS-LSTM: sales forecasting in e-commerce on feature selection

张函 井音吉 赵永利

中国邮电高校学报(英文), 2022, 29 (5). doi： 10.19682/j.cnki.1005-8885.2022.0018

摘要 ( 377 ) PDF (904 KB)( 96 )

参考文献 | 相关文章 | 计量指标




There are many studies on sales forecasting in e-commerce, most of which focus on how to forecast sales volume with related e-commerce operation data. In this paper, a deep learning method named FS-LSTM was proposed, which combines long short-term memory (LSTM) and feature selection mechanism to forecast the sales volume. The indicators with most contributions by the extreme gradient boosting (XGBoost) model are selected as the input features of LSTM model. FS-LSTM method can get less mean average error (MAE) and mean squared error (MSE) in the forecasting of e-commerce sales volume, comparing with the LSTM model without feature selection. The results show that the FS-LSTM can improve the performance of original LSTM for forecasting the sales volume.




Others

Compact multilayer dual-band bandpass filter design using stepped impedance resonators

韩宇南 梁珮云 林洋 程春悦 姚雨辰 戴琳 金铭

中国邮电高校学报(英文), 2022, 29 (5). doi： 10.19682/j.cnki.1005-8885.2022.0016

摘要 ( 289 ) PDF (2676 KB)( 58 )

参考文献 | 相关文章 | 计量指标

Compact dual-band bandpass filter (BPF) for the 5th generation mobile communication technology (5G) radio frequency (RF) front-end applications was presented based on multilayer stepped impedance resonators (SIRs). The multilayer dual-band SIR BPF can achieve high selectivity and four transmission zeros (TZs) near the passband edges by the quarter-wavelength tri-section SIRs. The multilayer dual-band SIR BPF is fabricated on a 3-layer FR-4 substrate with a compact dimension of 5.5 mm ×5.0 mm ×1.2 mm. The measured two passbands of the multilayer dual-band SIR BPF are 3.3 GHz -3.5 GHz and 4.8 GHz -5.0 GHz with insertion loss (IL) less than 2 dB respectively. Both measured and simulated results suggest that it is a possible candidate for the application of 5G RF front-end at sub-6 GHz frequency band.