二值化身份感知图卷积神经网络北大核心CSCD

针对有限的内存资源导致图神经网络(graph neural network, GNN)无法完全加载属性图的问题,文中提出了二值化身份感知图卷积神经网络(binary identify-aware graph convolutional network, BID-GCN)。该网络通过在消息传递过程中递归地考虑节点的信息,为了获得一个给定的节点的嵌入,BID-GCN将提取以该节点为中心的自我网络,并进行多轮的异构消息传递,在自我网络的中心节点上应用与其他节点不同的参数。在消息传递过程中,对网络参数和输入节点特征进行二值化,并将原始的矩阵乘法修改为二值化以加速运算。通过理论分析和实验评估,BID-GCN可以减少网络参数和输入数据的平均约36倍的内存消耗,并加快引文网络上平均约49倍的推理速度,可以提供与全精度基线相当的性能,较好地解决内存资源有限的问题。     

Length matters: Scalable fast encrypted internet traffic service classification based on multiple protocol data unit length sequence with composite deep learning

As an essential function of encrypted Internet traffic analysis, encrypted traffic service classification can support both coarse-grained network service traffic management and security supervision. However, the traditional plaintext-based Deep Packet Inspection (DPI) method cannot be applied to such a classification. Moreover, machine learning-based existing methods encounter two problems during feature selection: complex feature overcost processing and Transport Layer Security (TLS) version discrepancy. In this paper, we consider differences between encryption network protocol stacks and propose a composite deep learning-based method in multiprotocol environments using a sliding multiple Protocol Data Unit (multiPDU) length sequence as features by fully utilizing the Markov property in a multiPDU length sequence and maintaining suitability with a TLS-1.3 environment. Control experiments show that both Length-Sensitive (LS) composite deep learning model using a capsule neural network and LS-long short time memory achieve satisfactory effectiveness in F1-score and performance. Owing to faster feature extraction, our method is suitable for actual network environments and superior to state-of-the-art methods.     

Constrained adversarial loss for generative adversarial network‐based faithful image restoration

Generative adversarial networks (GAN) have been successfully used in many image restoration tasks, including image denoising, super‐resolution, and compression artifact reduction. By fully exploiting its characteristics, state‐of‐the‐art image restoration techniques can be used to generate images with photorealistic details. However, there are many applications that require faithful rather than visually appealing image reconstruction, such as medical imaging, surveillance, and video coding. We found that previous GAN‐training methods that used a loss function in the form of a weighted sum of fidelity and adversarial loss fails to reduce fidelity loss. This results in non‐negligible degradation of the objective image quality, including peak signal‐to‐noise ratio. Our approach is to alternate between fidelity and adversarial loss in a way that the minimization of adversarial loss does not deteriorate the fidelity. Experimental results on compression‐artifact reduction and super‐resolution tasks show that the proposed method can perform faithful and photorealistic image restoration.     

2D Ruddlesden–Popper Perovskite Nanoplate Based Deep‐Blue Light‐Emitting Diodes for Light Communication

Ruddlesden–Popper perovskite, (PEA)₂PbBr₄ (PEA = C₈H₉NH₃), is a steady and inexpensive material with a broad bandgap and a narrow‐band emission. These features make it a potential candidate for deep‐blue light‐emitting diodes (LEDs). However, due to the weak exciton binding energy, LEDs based on the perovskite thin films usually possess a very low external quantum efficiency (EQE) of <0.03%. Here, for the first time, the construction of high‐performance deep‐blue LEDs based on 2D (PEA)₂PbBr₄ nanoplates (NPs) is demonstrated. The as‐fabricated (PEA)₂PbBr₄ NPs film shows a deep‐blue emission at 410 nm with excellent stability under ambient conditions. Impressively, LEDs based on the (PEA)₂PbBr₄ NPs film deliver a bright deep‐blue emission with a maximum luminance of 147.6 cd m^?2 and a high EQE up to 0.31%, which represents the most efficient and brightest perovskite LEDs operating at deep‐blue wavelengths. Furthermore, the LEDs retain over 80% of their efficiencies for over 1350 min under ≈60% relative humidity. The steady and bright deep‐blue LEDs can be used as an excitation light source to realize white light emission, which shows the potential for light communication. The work provides scope for developing perovskite into efficient and deep‐blue LEDs for low‐cost light source and light communication.     

基于FMCW雷达的双流融合神经网络手势识别方法

针对传统光学摄像头和无线技术的手势识别方法受光照环境影响和空间纵向、横向特征不全的问题,该文提出一种基于调频连续波（Frequency Modulated Continuous Wave,FMCW）雷达信号的双流融合神经网络（Two-Stream Fusion Neural Network,TS-FNN）手势识别方法.首先,利用二维快速傅立叶变换（Fast Fourier Transform,FFT）求取中频信号的频谱,估计手势的距离和速度,并利用多重信号分类（Multiple Signal Classification,MUSIC）方法计算角度.其次,利用这三维参数在时间上的累积,将一个手势动作映射为32帧距离-速度矩阵图和角度时间图.最后,建立TS-FNN进行手势特征提取和特征融合.实验结果表明,TS-FNN方法与传统卷积神经网络相比,手势的平均识别准确率提升了约5%.     

基于主成分分析与深度神经网络的快速噪声水平估计算法

鉴于从噪声图像分解获得的原生图块集合的协方差矩阵前若干个特征值（按照升序排序）与图像噪声水平值具有强相关性,提出了一种基于主成分分析和深度神经网络的快速噪声水平估计算法.该算法首先选用原生图块集合协方差矩阵前若干个特征值构成刻画图像噪声水平高低的特征矢量,然后在大量有代表性且已标定噪声水平值的噪声图像集合上利用深度神经网络训练预测模型以实现将特征矢量直接映射为噪声水平值,最后为获得更高的预测准确性,采用粗精预测模型相结合的两步预测方式实现.实验表明：文中算法在各个噪声级别上都具有稳定的预测准确性,且执行效率非常高,作为降噪算法的前置预处理模块具有更好的综合优势.     

非刚性稠密匹配大位移运动光流估计

光流场是目标检测,无人机定位等众多计算机视觉任务的重要基础.本文针对非刚性大位移运动等困难运动类型图像序列光流计算的准确性与鲁棒性问题,提出一种基于非刚性稠密匹配的TV-L1（Total Variational with L1 norm,TV-L1）大位移光流计算方法.首先,使用非刚性稠密块匹配计算图像序列初始最近邻域场,其次根据图像相邻块区域的相似性消除初始最近邻域场中的非一致性区域以得到准确的图像最近邻域场.然后,在图像金字塔分层计算框架下,将图像最近邻域场引入基于非局部约束的TV-L1光流估计模型,通过Quadratic Pseudo-Boolean Optimization（QPBO）融合算法在金字塔分层图像光流计算时对TV-L1模型光流估计进行大位移运动补偿.最后,采用标准测试图像序列对本文方法和当前代表性的变分方法LDOF（Large Displacement Optical Flow,LDOF）、Classic+NL、NNF（Nearest Neighbor Fields,NNF）以及深度学习方法FlowNet2.0进行对比分析.实验结果表明,本文方法能有效提高非刚性运动、大位移运动以及运动遮挡等困难运动类型光流估计的精度与鲁棒性.     

PBN坩埚生长碲锌镉单晶的性能与缺陷研究

采用垂直布里奇曼法,利用热解氮化硼(PBN)坩埚生长碲锌镉单晶(CdZnTe),并对晶体进行定向、切割与磨抛制成(111)面碲锌镉晶片。采用傅里叶红外透过光谱仪、红外透射显微镜、金相显微镜、X射线形貌仪、X射线衍射仪和X射线双晶衍射仪等仪器系统地检测和研究了碲锌镉晶体的性能与缺陷。研究发现:相比于石英坩埚,采用PBN坩埚生长的碲锌镉晶体单晶率和出片率均大幅增加,孪晶、小角晶界和位错密度明显减少。     

基于VGG网络的双波段图像融合方法

针对红外与可见光图像中物体信息具有各自优点的情况,提出了基于VGG网络的红外与可见光图像融合方法来提高对夜间或复杂背景情况下的物体检测识别能力。首先将图像分别输入到一个经过训练得到的VGG网络中,经过不同的卷积层提取各自的特征图;然后将特征图经过ZCA白化处理,去除冗余信息;再通过归一化处理,将特征图的维度降到二维,并通过双三次插值法将其缩放到与源图像尺寸一致;最后通过加权取平均得到融合后的图像。实验结果表明,本文的方法在第四和第五层卷积得到的融合结果优于前三层的融合结果。同时,本文融合方法与其他3种融合方法相比视觉效果较好,在标准差、平均梯度、相关系数、熵值等评价指标上分别平均提升了12.79 %、11.04 %、9.94 %和2.54 %,并且在融合时间上保持在1秒以内。这说明该方法融合效果较好,速度较快,能够较多地保留红外与可见光图像信息和较好地提升目标的显著性。