基于SSD的眨眼与打哈欠检测

提出了一种新的基于SSD网络的眨眼与打哈欠检测。首先,对原SSD架构进行四个方面的修改。第一,主干网用MobileNetV3替换了VGG,使得架构尺寸缩减三分之一。第二,重新设计了输出检测器,实现了目标检测与分类、目标定位与高宽比回归多任务的融合。第三,网络直接回归高宽比,比传统的从关键点离线计算高宽比的方法鲁棒性要高。第四,省去了非最大抑制层(NMS)中的重叠度(IOU)计算和阈值设定,使得单目标检测速度提高了40%。另外,在数据集YawDD上验证了所提出方法的有效性,眨眼与打哈欠的测全率为95.1%与100.0%,在单卡2080Ti上实现了41fps的高检测速度。     

基于Massive MIMO及频谱叠加的5G SA网络上行优化方法

大规模多输入多输出(Multiple-Input Multiple-Output, MIMO)可以有效提升5G SA网络的上行链路数据传输速率以及可靠性。针对5G SA网络上行链路速率和覆盖不均衡的情况,提出了基于大规模MIMO的分组算法,将发送信号矢量进行分组,组内采用最大似然检测,组外采用基于正交三角分解(QR分解)的干扰消除检测,并且结合5G频谱的叠加策略,在降低算法复杂度的同时,有效提升网络覆盖和速率。通过5G SA现网实测,通过MIMO降低分组数量能够提升分组检测性能,结合上行低频段频谱叠加策略能够有效提升5G SA网络上行覆盖30%,提升5G SA网络上行平均速率40%~80%,特别是弱覆盖边缘的网络速率,最高可达600%。     

“数字信号处理”课程的几何视角

“数字信号处理”的经典课程体系以“系统”为中心，以离散时间傅立叶变换、z变换和离散傅立叶变换为分析工具，以滤波器设计为目的，注重定理和公式推导及其性质分析，很难让学生建立一个形象的、完整统一的知识体系框架。本文从几何视角将 “系统”化和“代数”化的经典课程体系重构为面向“信号”的“几何”知识体系，从而提升学生对本质问题的“洞察力”，帮助学生建立系统的知识体系。     

Presentation attack detection based on two-stream vision transformers with self-attention fusion

Aiming at the performance degradation of the existing presentation attack detection methods due to the illumination variation, a two-stream vision transformers framework (TSViT) based on transfer learning in two complementary spaces is proposed in this paper. The face images of RGB color space and multi-scale retinex with color restoration (MSRCR) space are fed to TSViT to learn the distinguishing features of presentation attack detection. To effectively fuse features from two sources (RGB color space images and MSRCR images), a feature fusion method based on self-attention is built, which can effectively capture the complementarity of two features. Experiments and analysis on Oulu-NPU, CASIA-MFSD, and Replay-Attack databases show that it outperforms most existing methods in intra-database testing and achieves good generalization performance in cross-database testing.     

Simultaneous fractionation of multiple classes of polyphenols from honeybush tea using solid-phase extraction

The structural diversity of polyphenols and the inherent limitations of current extraction techniques pose a challenge to extract polyphenols using a simple and green method. Hence, in this study, a method was developed to simultaneously fractionate multiple classes of polyphenols by only varying ethanol-water solutions. Honeybush tea, which is rich in polyphenols, was selected as a model for this study. Solvent extraction followed by solid-phase extraction (SPE) was developed to obtain a polyphenol-rich fraction from six honeybush samples. Based on a gradient elution programme (10%, 30%, 50%, 70% and 90% (v/v) ethanol-water solution) of SPE, the Strata X cartridge showed a better recovery of most targeted polyphenols under 0.9 mL of the drying volume and 1 mL min⁻¹ of the dispensing speed. The elution programme for fractionating most polyphenols was as follows: single elution with 50% ethanol, followed by twice elution with 70% ethanol. The antioxidant capacity was used to analyse the differences among the polyphenol-rich fractions from six honeybush samples. Principal component analysis (PCA) revealed that unfermented C. genistoides (GG) has the greatest antioxidant capacity among the honeybush species studied. Additionally, mangiferin, isomangiferin and vicenin-2 were the main contributors to the antioxidant capacity in six honeybush fractions according to the correlation study.     

高光谱成像用高帧频CMOS图像传感器北大核心

针对高帧频、全局曝光和光谱平坦等成像应用需求,设计了一款高光谱成像用CMOS图像传感器。其光敏元采用PN型光电二极管,读出电路采用5T像素结构。采用列读出电路以及高速多通道模拟信号并行读出的设计方案来获得低像素固定图像噪声(FPN)和非均匀性抑制。芯片采用ASMC 0.35μm三层金属两层多晶硅标准CMOS工艺流片,为了抑制光电二极管的光谱干涉效应,后续进行了光谱平坦化VAE特殊工艺,并对器件的光电性能进行了测试评估。电路测试结果符合理论设计预期,成像效果良好,像素具备积分可调和全局快门功能,最终实现的像素规模为512×256,像元尺寸为30μm×30μm,最大满阱电子为400 ke^(-),FPN小于0.2%,动态范围为72 dB,帧频为450 f/s,相邻10 nm波段范围内量子效率相差小于10%,可满足高光谱成像系统对CMOS成像器件的要求。     

基于双时间点检测的毫米波雷达人流量监测方法

针对现有基于视频监控的人流量统计方案成本高、算法复杂且不利于个人隐私保护的局限性，利用毫米波雷达体积小、成本低、分辨率高的特点，提出了一种基于双时间点检测的人流量监测方法。该方法先获取人体目标散射点位置和多普勒频移信息来构成点云数据，然后根据多普勒频移正负来判断人体的运动方向，并筛选具有高多普勒频移值的点云数据以降低干扰点对聚类结果的影响；在双时间点对特定区域内人员数量进行统计，并根据双时间点之间所获取的点云数据聚类结果对所统计人员数据进行修正。实验结果表明，该方法能够用匿名的方式以较高的正确率统计人员进出。     

Image sequence-based risk behavior detection of power operation inspection personnel

A novel image sequence-based risk behavior detection method to achieve high-precision risk behavior detection for power maintenance personnel is proposed in this paper. In this method, the original image sequence data is first separated from the foreground and background. Then, the free anchor frame detection method is used in the foreground image to detect the personnel and correct their direction. Finally, human posture nodes are extracted from each frame of the image sequence, which are then used to identify the abnormal behavior of the human. Simulation experiment results demonstrate that the proposed algorithm has significant advantages in terms of the accuracy of human posture node detection and risk behavior identification.     

Use of MALDI-TOF MS technology to evaluate adulteration of small ruminant milk with raw bovine milk

Detection of adulteration of small ruminant milk is very important for health and commercial reasons. New analytical and cost-effective methods need to be developed to detect new adulteration practices. In this work, we aimed to explore the ability of the MALDI-TOF mass spectrometry to detect bovine milk in caprine and ovine milk using samples from 18 dairy farms. Different levels of adulteration (0.5, 1, 5, 10, 20, 40, 60, and 80%) were analyzed during the lactation period of goat and sheep (in May, from 60 to 90 d in milk, and in August, from 150 to 180 d in milk). Two different ranges of peptide-protein spectra (500–4,000 Da; 4–20 kDa) were used to establish a calibration model for predicting the concentration of adulterant using partial least squares and generalized linear model with lasso regularization. The low molecular weight part of the spectra together with the generalized linear model with lasso regularization regression model appeared to have greater potential for our aim of detection of adulteration of small ruminants' milk. The subsequent prediction model was able to predict the concentration of bovine milk in caprine milk with a root mean square error of 11.4 and 17.0% in ovine milk. The results offer compelling evidence that MALDI-TOF can detect the adulteration of small ruminants' milk. However, the method is severely limited by (1) the complexity of the milk proteome resulting from the adulteration technique, (2) the potential degradation of thermolabile proteins, and (3) the genetic variability of tested samples. Additionally, the root mean square error of prediction based only on one individual sample adulteration series can drop down to 6.34% for quantification of adulterated caprine milk and 6.28% for adulterated ovine milk for the full set of concentrations or down to 2.33 and 4.00%, respectively, if we restrict only to low concentrations of adulteration (0, 0.5, 1, 5, 10%).     

基于改进残差特征蒸馏的轻量级超分辨率网络

基于深度学习的图像超分辨率算法通常采用递归的方式或参数共享的策略来减少网络参数，这将增加网络的深度，使得运行网络花费大量的时间，从而很难将模型部署到现实生活中。为了解决上述问题，本文设计一种轻量级超分辨率网络，对中间特征的关联性及重要性进行学习，且在重建部分结合高分辨率图像的特征信息。首先，引入层间注意力模块，通过考虑层与层之间的相关性，自适应地分配重要层次特征的权重。其次，使用增强重建模块提取高分辨率图像中更精细的特征信息，以此得到更加清晰的重建图片。通过大量的对比实验表明，本文设计的网络与其他轻量级模型相比，有更小的网络参数量，并且在重建精度和视觉效果上都有一定的提升。