基于改进MobileFaceNet的人脸识别方法北大核心

针对目前大多数人脸识别算法参数多、计算量大,难以部署到移动端和嵌入式设备中的问题,提出了一种基于改进MobileFaceNet的人脸识别方法。通过对MobileFaceNet模型结构的调整,将bottleneck模块优化为sandglass模块,改良深度卷积和逐点卷积的相对位置,适当增大sandglass模块的输出通道数,从而减少特征压缩时的信息丢失,增强人脸空间特征的提取。实验结果表明:改进后的方法在LFW测试数据集上准确率达99.15%,模型大小和计算量分别仅为原算法的61%和45%,验证了所提方法的有效性。     

一种基于FCOS神经网络的小建筑物目标检测方法北大核心

提出了一种基于FCOS神经网络的小建筑物目标检测算法,针对FCOS模型在特征提取阶段提取到的小建筑物目标特征较少问题,引入多尺度检测和可变形卷积方式,加强网络对小建筑物目标的特征提取能力,并通过改进后的SGE注意力机制降低特征图中的干扰噪声权重。改进后的网络可以提取到更多的小建筑物目标特征,对环境干扰噪声的鲁棒性更强。在自己搭建的数据集上进行了实验测试,结果表明,在相同环境下网络改进后建筑物的整体检测准确率提升了1.7%,其中对小建筑物目标提升了3.6%,减少了小建筑物目标漏检、误检的问题。     

MHD mixed convection of a Cu–water nanofluid flow through a channel with an open trapezoidal cavity and an elliptical obstacle

In the current work, numerical simulations are achieved to study the properties and the characteristics of fluid flow and heat transfer of (Cu–water) nanofluid under the magnetohydrodynamic effects in a horizontal rectangular canal with an open trapezoidal enclosure and an elliptical obstacle. The cavity lower wall is grooved and represents the heat source while the obstacle represents a stationary cold wall. On the other hand, the rest of the walls are considered adiabatic. The governing equations for this investigation are formulated, nondimensionalized, and then solved by Galerkin finite element approach. The numerical findings were examined across a wide range of Richardson number (0.1 ≤ Ri ≤ 10), Reynolds number (1 ≤ Re ≤ 125), Hartmann number (0 ≤ Ha ≤ 100), and volume fraction of nanofluid (0 ≤ φ ≤ 0.05). The current study's findings demonstrate that the flow strength increases inversely as the Reynolds number rises, which pushes the isotherms down to the lower part of the trapezoidal cavity. The Nu_avg rises as the Ri rise, the maximum Nu_avg = 10.345 at Ri = 10, Re = 50, ϕ = 0.05, and Ha = 0; however, it reduces with increasing Hartmann number. Also, it increase by increasing ϕ, at Ri = 10, the Nu_avg increased by 8.44% when the volume fraction of nanofluid increased from (ϕ = 0–0.05).     

基于条件信息卷积生成对抗网络的图像识别

传统的图像识别方法需要大量有标签样本进行训练，且模型训练难以达到稳定。针对这些问题，结合条件生成网络和信息最大化生成网络的结构优势建立了条件信息卷积生成网络（C-Info-DCGAN）。模型增加图像的类别信息和潜在信息作为输入数据，然后利用Q网络去更好地发挥类别信息和潜在信息对训练的引导作用，并且利用深度卷积网络来加强对图像特征的提取能力。实验结果表明，该方法能够加快模型训练收敛速度，并有效提高图像识别的准确率。     

基于并行多通道卷积长短时记忆网络的轴承寿命预测方法

在预测轴承剩余使用寿命时，数据间的时序特性是一个可以利用的重要隐藏信息。为了更好地提取具有时序信息的特征用于预测，提出了一种基于并行多通道卷积长短时记忆网络（PMCCNN-LSTM）的剩余使用寿命预测模型。该模型主要由两部分组成：前端为并行多通道卷积网络（PMCCNN），提取信号特征，挖掘数据的时序特性，并采用逐层训练和微调的方式提升参数的收敛性；后端为长短时记忆（LSTM）网络，基于特征进行剩余使用寿命预测，并采用加权平均的方法对预测结果进行平滑处理。在一个轴承加速寿命实验的公开数据集上使用留一法验证了该模型的准确性，实验结果表明：所提模型的平均误差与最大误差分别比传统的卷积神经网络（CNN）低23.38%和15.84%，比传统的LSTM低24.14%和19.01%，比卷积长短时记忆网络（CNN-LSTM）低30.32%和23.09%。     

Experimental Investigation of Heat Storage and Heat Transfer Rates during Melting of Nano-Enhanced Phase Change Materials (NePCM) in a Differentially-Heated Rectangular Cavity

In this work, an experimental study of melting heat transfer of nano-enhanced phase change materials(NePCM) in a differentially-heated rectangular cavity was performed. Two height-to-width aspect ratios of the cavity, i.e., 0.9 and 1.5, were investigated. The model Ne PCM samples were prepared by dispersing graphene nanoplatelets(GNP) into 1-tetradecanol, having a nominal melting point of 37℃, at loadings up to 3 wt.%. The viscosity was found to have a more than 10-fold increase at the highest loading of GNP. During the melting experiments, the wall superheat at the heating boundary was set to be 10℃ or 30℃. It was shown that with increasing the loading of GNP, both the heat storage and heat transfer rates during melting decelerate to some extent, at all geometrical and thermal configurations. This suggested that the use of NePCM in such cavity may not be able to enhance the heat storage rate due to the dramatic growth in viscosity, which deteriorates significantly natural convective heat transfer during melting to overweigh the enhanced heat conduction by only a decent increase in thermal conductivity. This also suggested that the numerically predicted melting accelerations and heat transfer enhancements, as a result of the increased thermal conductivity, in the literature are likely overestimated because the negative effects due to viscosity growth are underestimated.     

太赫兹可调谐滤波器设计

为了实现太赫兹信号的可调谐滤波，设计了一种基于柔性材料的太赫兹可调谐滤波器。通过扭曲特氟龙（Teflon）波导形成环型谐振器，实现了160～200 GHz频段的带阻滤波功能。改变谐振腔长可实现自由频谱范围（FSR）和滤波频点的调谐，实验测试了自由频谱范围在1.9 GHz和2.8 GHz间切换以及相应的滤波特性。研究表明，谐振腔长一定时，改变弯曲半径可实现滤波阻带抑制度的调节，柔性材料太赫兹环型谐振器可用于可调谐滤波，且具有较高的自由度。     

高温气冷堆舱室抗商用飞机撞击的耦合数值分析

本文建立了大型商用飞机撞击典型高温气冷堆核电站反应堆舱室的非线性有限元模型，计算中混凝土舱室直接采用工程用钢筋混凝土的损伤塑性本构模型，飞机结构采用Johnson-Cook本构模型。对飞机高速撞击高温气冷堆核电站反应堆舱室非线性撞击过程进行模拟计算，得出正面和侧面撞击条件下的撞击载荷曲线、撞击位移云图、反应堆舱室混凝土破坏情况等结果。评估表明，反应堆舱室结构在撞击条件下的整体损伤微小，可为保护内部关键设备提供重要的屏障功能。     

Super-Resolution Reconstruction of Images Based on Microarray Camera

In the field of images and imaging, super-resolution (SR) reconstruction of images is a technique that converts one or more low-resolution (LR) images into a highresolution (HR) image. The classical two types of SR methods are mainly based on applying a single image or multiple images captured by a single camera. Microarray camera has the characteristics of small size, multi views, and the possibility of applying to portable devices. It has become a research hotspot in image processing. In this paper, we propose a SR reconstruction of images based on a microarray camera for sharpening and registration processing of array images. The array images are interpolated to obtain a HR image initially followed by a convolution neural network (CNN) procedure for enhancement. The convolution layers of our convolution neural network are 3×3 or 1×1 layers, of which the 1×1 layers are used to improve the network performance particularly. A bottleneck structure is applied to reduce the parameter numbers of the nonlinear mapping and to improve the nonlinear capability of the whole network. Finally, we use a 3×3 deconvolution layer to significantly reduce the number of parameters compared to the deconvolution layer of FSRCNN-s. The experiments show that the proposed method can not only ameliorate effectively the texture quality of the target image based on the array images information, but also further enhance the quality of the initial high resolution image by the improved CNN.     

The microstructure of buccal cavity and alimentary canal of Siganus rivulatus: Scanning electron microscope study

The microstructure of the oral cavity and alimentary canal of herbivorous fish Siganus rivulatus collected from the Red Sea were investigated by using scanning electron microscope (SEM). The results showed that S. rivulatus has three types of teeth, tri‐cusped, bi‐cusped, and papilliform. A taste bud (Type I) was recorded in the oropharyngeal cavity. Characteristic styles of microridges on the cell's surface inside the buccal cavity were recorded. Also, the distribution of the mucous cells in the lining of the mouth cavity, alimentary canal was observed. Mucosal folds along the distinct parts of alimentary canal, showed characteristic pattern which was complex in the intestinal mucosa. The results concluded that there are characteristic microstructures according to feeding habitat compared with other bony fishes.