改进HSR-FCN的服装图像识别分类算法研究

目前网络上的服装图像数量增长迅猛，对于大量服装图像实现智能分类的需求日益增加。将基于区域的全卷积网络（Region-Based Fully Convolutional Networks，R-FCN）引入到服装图像识别中，针对服装图像分类中网络训练时间长、形变服装图像识别率低的问题，提出一种新颖的改进框架HSR-FCN。新框架将R-FCN中的区域建议网络和HyperNet网络相融合，改变图片特征学习方式，使得HSR-FCN可以在更短的训练时间内达到更高的准确率。在模型中引入了空间转换网络，对输入服装图像和特征图进行了空间变换及对齐，加强了对多角度服装和形变服装的特征学习。实验结果表明，改进后的HSR-FCN模型有效地加强了对形变服装图像的学习，且在训练时间更短的情况下，比原来的网络模型R-FCN平均准确率提高了大约3个百分点，达到96.69%。     

A new perspective of co-doping and Nd segregation effect on proton stability and transportation in Y and Nd co-doped BaCeO3

Bulk and surface properties of proton stability and transportation in Y and Nd co-doped BaCeO₃ (BCYN), especially the effect of Nd segregation, were investigated by first-principles calculations. Since the structure of doped BaCeO₃ at the operating temperature of proton-conducting has been unclear for a long time, we have summarized the latest experimental results and calculated the structure of the asymmetric BCYN for the first time. The results show that compared with Y, Nd doping promotes oxygen vacancy formation, however reduces proton stability. Our calculation can also provide a possible explanation for the formation of space charge layer at the grain boundary of doped BaCeO₃ in experiment. Unlike the stable Y in BCYN, Nd is calculated to be easily segregated, which can facilitate both proton hydration and proton transportation near the surface. Moreover, Nd segregation at the grain boundary is predicted to be beneficial for proton transportation between grains.     

移动边缘计算环境下的动态资源分配策略

在通讯设备爆炸式增长的时代,移动边缘计算作为5G通讯技术的核心技术之一,对其进行合理的资源分配显得尤为重要。移动边缘计算的思想是把云计算中心下沉到基站部署(边缘云),使云计算中心更加靠近用户,以快速解决计算资源分配问题。但是,相对于大型的云计算中心,边缘云的计算资源有限,传统的虚拟机分配方式不足以灵活应对边缘云的计算资源分配问题。为解决此问题,提出一种根据用户综合需求变化的动态计算资源和频谱分配算法(DRFAA),采用"分治"策略,并将资源模拟成"流体"资源进行分配,以寻求较大的吞吐量和较低的传输时延。实验仿真结果显示,动态计算资源和频谱分配算法可以有效地降低用户与边缘云之间的传输时延,也可以提高边缘云的吞吐量。     

一种基于改进NMS的牛脸检测方法

针对牲畜牛身份认证的多牛脸检测场景，本文给出一种基于改进Faster R-CNN的牛脸检测方法。使用Inception v2替换ZF网络作为Faster R-CNN的基础网络，模型精度得到显著提升；针对多牛检测场景对NMS（Non-Maximum Suppression）进行相应优化，使模型的召回率得到显著提升。通过和其他目标检测模型对比实验，本文的改进模型在精确率和召回率上均优于其他模型。     

Combating Biofilm Associated Infection In Vivo: Integration of Quorum Sensing Inhibition and Photodynamic Treatment based on Multidrug Delivered Hollow Carbon Nitride Sphere

Recently, quorum sensing (QS) inhibitors (QSIs) have been combined with antibiotics to enhance antibiofilm efficacy in vitro and in vivo. However, targeting QS signals alone is not enough to prevent bacterial infections. Drug resistance and recurrence of biofilms makes it difficult to eradicate. Herein, photodynamic therapy (PDT) is selected to unite QSIs and antibiotics. A synergistically antibiofilm system, which combines QSIs, antibiotics, and PDT based on hollow carbon nitride spheres (HCNSs) is envisaged. First, HCNS provides the multidrug delivering ability, enabling QSIs and antibiotics to be released in sequence. Subsequently, multistage releases sensitize bacteria effectively, potentiating the chemotherapeutic effects of the antibiotics. Finally, the integration of QSIs and PDT not only minimizes the possibility of drug resistance, but also overcomes the problem of limited mass and extension of PDT. Even after 48 h of incubation, the bacterial biofilm is obviously inhibited. And its biofilm disperse efficiency exceeds 48% (compared with QSI‐potentiated chemotherapy group) and 40% (compared with PDT group). Besides, the inhibition of the QS system influences phenotypes related to virulence factor production and surface hydrophobicity, which weaken biofilm invasion and formation. Eventually, this system is applied to disperse bacterial biofilm in vivo. Overall, PDT and QS modulation are devoted to eradicate drug resistance and recurrence of the biofilm.     

基于时频域的卷积神经网络运动想象脑电信号识别方法

针对目前运动想象脑电（EEG）信号识别率较低的问题，考虑到脑电信号蕴含着丰富的时频信息，提出一种基于时频域的卷积神经网络（CNN）运动想象脑电信号识别方法。首先，利用短时傅里叶变换（STFT）对脑电信号的相关频带进行预处理，并将多个电极的时频图组合构造出一种二维时频图；然后，针对二维时频图的时频特性，通过一维卷积的方法设计了一种新颖的CNN结构；最后，通过支持向量机（SVM）对CNN提取的特征进行分类。基于BCI数据集的实验结果表明，所提方法的平均识别率为86.5%，优于其他传统运动想象脑电信号识别方法；同时将该方法应用在智能轮椅上，验证了其有效性。     

Screening of frozen-thawed conditions for keeping nutritive compositions and physicochemical characteristics of goat milk

Frozen milk can help producers overcome the seasonality of goat milk production, low goat production and short lactation periods, and avoid discarding milk during some special periods. We investigated effects of combination between freezing (cryogenic refrigerator of ?16 to ?20°C or ultra-cryogenic refrigerator of ?76 to ?80°C) and thawing (homeothermy of 20 to 25°C or refrigeration of 2 to 4°C) on nutritive compositions and physicochemical characteristics of raw goat milk during storage period (80 d). Compared with fresh goat milk, the frozen-thawed milk decreased contents of fat, protein, and lactose, as well as surface tension and stability coefficient, whereas increased effective diameter and polydispersity index. The average values of color values (L*, a*, and b*) in 4 group samples changed from 83.01 to 82.25, ?1.40 to ?1.54, 3.51 to 3.81, respectively, and the ΔE of most samples did not exceed 2. In contrast to the other 3 frozen-thawed treatments, goat milk treated with ultra-cryogenic freezing-homeothermic thawing (UFHT) possessed higher fat (5.20 g/100 g), smaller effective particle diameter (0.32 µm), and the lowest polydispersity index value (0.26). The color and confocal laser scanning microscopy images of UFHT were similar to those of fresh goat milk, illustrating UFHT was the optimal approach to maintain the natural quality of goat milk. Our finding provides a theoretical basis for producers to freeze surplus milk.