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。     

激光在线测厚振动分析与精度优化

激光测厚具有安全可靠、测量精度高、测量范围大等优点,广泛应用于纸张、电池极片等薄膜类材料厚度的在线测量。带材宽幅方向扫描测厚时由于扫描架往复运动会产生机械振动,影响在线测厚精度。针对该问题,以锂离子电池极片厚度测量为例,使用双激光差动式测厚平台对电池极片和铜箔分别进行厚度测量,然后对测厚数据进行频谱分析,探究其振动规律的相似性,并基于频谱分析结果采用滑动带阻滤波方式对测厚数据进行处理,滤波后极片和铜箔的厚度极差分别降低了33.4%和73.8%,有效过滤了机械振动导致的测量误差,可满足极片和铜箔厚度测量的精度要求。     

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。     

Polynomial filtering for nonlinear stochastic systems with state‐ and disturbance‐dependent noises

This article is concerned with the polynomial filtering problem for a class of nonlinear stochastic systems governed by the Itô differential equation. The system under investigation involves polynomial nonlinearities, unknown‐but‐bounded disturbances, and state‐ and disturbance‐dependent noises ((x,d)‐dependent noises for short). By expanding the polynomial nonlinear functions in Taylor series around the state estimate, a new polynomial filter design method is developed with hope to reduce the conservatism of the existing results. In virtue of stochastic analysis and inequality technique, sufficient conditions in terms of parameter‐dependent linear matrix inequalities (PDLMIs) are derived to guarantee that the estimation error system is input‐to‐state stable in probability. Moreover, the desired polynomial matrix can be obtained by solving the PDLMIs via the sum‐of‐squares approach. The effectiveness and applicability of the proposed method are illustrated by two numerical examples with one concerning the permanent magnet synchronous motor.     

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%。     

Learning Association Rules and Tracking the Changing Concepts on Webpages: An Effective Pornographic Websites Filtering Approach

We applied the decision tree algorithm to learn association rules between webpage’s category (pornographic or normal) and the critical features. Based on these rules, we proposed an efficient method of filtering pornographic webpages with the following major advantages: 1) a weighted window-based technique was proposed to estimate for the condition of concept drift for the keywords found recently in pornographic webpages; 2) checking only contexts of webpages without scanning pictures; 3) an incremental learning mechanism was designed to incrementally update the pornographic keyword database.     

相似系数阈值滤波的数据驱动控制束偏移

高斯束偏移不仅具有接近波动方程偏移的精度，而且具有Kirchhoff偏移灵活、高效的特点。然而当实际地震采集数据中含有较强噪声时，易产生偏移假象而影响成像质量。为此，在传统高斯束偏移的基础上，根据有效信号和干扰信号在τ-p域中的相干性差异，发展了一种基于相似系数阈值滤波的数据驱动控制束偏移方法。采用数据驱动策略，在高斯束偏移成像过程中，先计算τ-p道集的相似系数，再通过设定相似系数阈值控制干扰信号，从而降低偏移剖面中的随机噪声；控制束偏移可以直接提取角度域共成像点道集，无需复杂的角度映射变换且具有更高信噪比。模型测试及实际资料处理结果表明：对于低信噪比数据，控制束偏移剖面的信噪比明显高于常规高斯束偏移，但会损失相对振幅的可靠性；尽管控制束偏移在τ-p道集的滤波过程增加了一定的计算量，但总体与常规高斯束偏移方法的计算效率相当；相似系数阈值参数选取十分关键，阈值较小时偏移噪声较强，但过大阈值也可能压制部分有效信息或产生偏移假象，选取合适的阈值参数才能得到较理想的偏移剖面。     

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.     

New version of continuous–discrete cubature Kalman filtering for nonlinear continuous–discrete systems

For nonlinear continuous–discrete systems, this paper elaborates a new accurate implementation of continuous–discrete cubature Kalman filter (CD-CKF). As the main contribution of this work, the new Kalman prediction stage begins by integrating the nonlinear continuous model for all the cubature sample vectors; the prior estimate state and covariance prediction are based on the weighted statistics of these integrated cubature sample vectors and the Gauss–Legendre approximation scheme. The new square root form CD-CKF is also derived and accurately implemented by combining with the modified variable stepsize NIRK. As the advantages of proposed approach, the complicated and error-prone processes of solving covariance differential equation or calculating derivatives are avoided, while the positive semi-definiteness of prior error covariance are numerically guaranteed. Simulations of traffic control scenarios further confirm the new approach’s superior filtering performance in both reliability and accuracy.