基于稀疏编码的迁移学习及其在行人检测中的应用

行人检测是计算机视觉领域中的研究热点,其实质是一个二分类问题.目前基于统计的行人检测技术已取得了一定进展,但大都需要大量的训练数据.针对这一问题,提出了一种基于迁移学习的半监督行人分类方法:首先基于稀疏编码,从任意的未标记样本中,学习到一个紧凑、有效的特征表示;然后通过迁移学习,将学习到的特征表示方法迁移到行人分类中.在MIT行人数据库上的实验结果表明:该方法能有效地刻画出行人的特征,提高行人分类的性能,在标记样本少的情况下仍具有良好的分类效果,因此可应用于行人检测中.     

复杂地形下甘肃舟曲江顶崖滑坡自动提取方法

为精确获取甘肃省舟曲县南峪乡江顶崖在复杂地形下的滑坡区域分布,利用数字正射影像和数字高程模型准确获取滑坡区域,分析滑坡信息。在基于影像监督分类的基础上,叠加一定的影响因子约束分析进一步筛选滑坡区,利用数学形态学中的腐蚀、膨胀、开运算三种运算方式处理图像实现对滑坡区域自动识别提取。实验结果说明通过与数字正射影像叠加解译和在精度评价中生产者精度为81.25%,用户精度为92.85%验证能够较好地避免因复杂地形中丰富地物类型所带来的干扰,实现对复杂地形下舟曲县南峪乡江顶崖滑坡的自动识别提取,结果表明三种方法的结合可以为识别提取复杂地形下的滑坡区域提供借鉴。     

基于动变形控制法的软岩路基填筑材料回弹模量控制

根据路基路面变形协调原理,结合国内公路相关规范对路面弯沉的控制标准,获得软岩路基顶面动变形允许值;利用传递-反射矩阵方法和叠加原理推导了双轮胎振动荷载作用下弹性层状公路结构的动力响应解;采用Hankel数值逆变换技术计算出实例工程中典型软岩公路路基顶面的动变形,基于动变形控制法确定了不同软岩填筑区域对应填料的回弹模量临界值;根据现有公路规范规定的路基结构层厚度,提出满足动变形条件下的软岩填筑区域所需回弹模量控制要求,并以回弹模量为依据对软岩填料进行了分类.研究成果可为软岩填料分区优化填筑提供指导.     

BP_Adaboost算法的改进及在首轮融资时总票房分类预测中的应用

为获得改进的分类算法BP_Adaboost,利用思维进化算法(MEA)和列文伯格-马夸尔特算法(LM)结合改进的BP神经网络作为弱分类器,由改进的弱分类器集成得到MEA-LM-BP_Adaboost算法.提出了基于MEA-LM-BP_Adaboost算法的首轮融资时总票房分类预测方法,该方法包括变量选取及操作化处理、网络参数优化、MEA改进弱分类器、LM算法改进弱分类器、MEA-LM-BP_Adaboost算法的流程设计、待预测电影验证6个部分.选用2013～2018年的245部国产电影作为样本验证该预测方法和模型,测试集分类准确率可达73.3％.最后在模型准确率、稳定性、K折交叉验证3方面进行模型整体性能比较,结果表明本文提出的模型整体性能最好.     

基于强化局部特征的3D点云分类与分割网络

在点云的处理过程中,许多深度学习网络未能充分考虑局部点之间的复杂关系,导致大量空间几何信息丢失。针对该问题,提出了一个强化局部特征的网络,用于点云的目标分类和语义分割。该网络通过设计编码单元对点的多方向信息进行编码;通过注意力机制学习采样分组后形成局部点云的特征。同时提出了一种新的多维损失函数,结合使用交叉熵损失函数与中心损失函数作用于分类任务。在数据集ModelNet40和ScanNet上进行了大量实验,结果表明：该网络在点云的目标分类和语义分割任务上表现出较好的性能。     

Pierre Duhem’s good sense as a guide to theory choice

This paper examines Duhem’s concept of good sense as an attempt to support a non rule-governed account of rationality in theory choice. Faced with the underdetermination of theory by evidence thesis and the continuity thesis, Duhem tried to account for the ability of scientists to choose theories that continuously grow to a natural classification. I will examine the concept of good sense and the problems that stem from it. I will also present a recent attempt by David Stump to link good sense to virtue epistemology. I will argue that even though this approach can be useful for the better comprehension of the concept of good sense, there are some substantial differences between virtue epistemologists and Duhem. In the light of this reconstruction of good sense, I will propose a possible way to interpret the concept of good sense, which overcomes the noted problems and fits better with Duhem’s views on scientific method and motivation in developing the concept of good sense.     

超低分辨率人脸图像高速重建方法

文中采用软硬件相结合的方法,实现了一种基于学习的超低分辨率人脸图像的高速重建系统.硬件系统的工作频率为60MHz,采用了FPGA组成并行处理单元,采用多内存形成并行数据,高速实现了计算复杂度高的相似度计算、相似度排序,从而解决了基于学习的超低分辨率人脸图像重建的处理速度问题.文中实现了8×6,16×12,32×24三个级别的人脸图像尺寸的重建系统,重建倍率分别为4×4倍、8×8倍、16×16倍,取得了好的视觉效果和低的RMS误差率,在速度上与C相比,最大可达到7900多倍的加速比,在保证重建质量的同时在处理速度上也有显著的提高.     

Can we agree on an ecological classification of subterranean animals?

That the ties between any obligate subterranean species and the hypogean environment depend on the interplay of a species' own physiological characteristics and all of the ecological characteristics of the adjacent epigean habitat(s), including biotic factors, has been emphasized. The reasons why troglomorphy cannot be included within criteria for classifying cave dwellers have been demonstrated. After a review of historic classifications, standardized definitions of the most widely used terms are proposed. The most easily recognized are four categories which are terminological approximations of the classic Schiner‐Racovitza terminology: (1) troglobiont is a species or population, strictly bound to a hypogean habitat; (2) eutroglophile is an essentially epigean species, but able to maintain a permanent subterranean population; (3) subtroglophile is inclined perpetually or temporarily to inhabit a subterranean habitat but is bound to the surface for some biological functions (e.g. feeding); (4) trogloxene is a species only occurring sporadically underground.     

Classification,reconstructed phylogeny and geographic history of the Neotropical phorid flies of the genus Gymnophora (Diptera: Phoridae)

The Neotropical species of Gymnophora are revised to include 28 species, 17 of which are new. Two subgenera, Cerocratia (in a revised sense) and Gymnophora s.s. are recognized. Cerocratia includes Gymnophora triangularis sp. nov., G. acutangula Schmitz, G. latibrachia sp. nov., G. lacertosa sp. nov., G. chilensis Borgmeier, G. parachilensis sp. nov., G. aemula Borgmeier, G. forticornis Schmitz, and doubtfully G. colona Brues. G. adumbrata Borgmeier is considered as a junior synonym of G. aemula (syn. nov.). Within Gymnophora s.s. there are four major lineages, the G. commotria-group, the G. carina-group, the G. cymatoneura-group and the G. quartomollis-group. Of these, the G. carina-group and the G. quartomollis-group are known only from the Holarctic Region. The G. commotria-group consists of the G. forcipis-series, containing G. forcipis sp. nov., subuncata sp. nov. and uncata sp. nov., and G. damula-series, whose species are further organized into the G. strigula-subseries, containing G. strigula sp. nov. and G. unidentata sp. nov., plus the G. falciformis-subseries, containing G. alces sp. nov., damula sp. nov. and falciformis sp. nov. Also included in the G. commotria-group are eight residual species whose relationships cannot be deduced at this time. Three of these species, G. commotria Schmitz, penai sp. nov. and quadrata sp. nov. are more closely related to each other than to G. auricula sp. nov., brasiliensis Borgmeier, heteroneura Schmitz, inusitata sp. nov. and trispina sp. nov. The G. cymatoneura-group contains G. cymatoneura Enderlein and G. spiracularis Borgmeier, and the male of G. spiracularis is described for the first time. A key is given that permits identification of males of most species, but females cannot be identified with certainty. A partially sequenced classification of the genus indicates the relationships of most of the included species. The genus may have originated in the Neotropical Region, although the evidence is tenuous, and shows some similarity of distribution to Andean-centered plant groups.     

基于模糊粗糙集的层次分类增量特征选择

随着大数据时代的到来,数据的类标签数量急剧增加,对现有的分类任务带来了严峻的挑战。为了解决这个问题,人们通常将标签组织成层次结构,使用结构中所包含的信息来对任务进行学习。考虑样本的不断增加,使用模糊粗糙集信息熵设计了一种面向层次分类的增量特征选择算法。考虑兄弟策略,将现有的λ条件熵推广到了层次分类的情形,设计了一种非增量的层次分类特征选择算法,设计了λ增量条件熵,基于此设计了增量版本的特征选择算法。在实验中,采用了包括非增量版本在内的7种不同的特征选择算法在5个层次数据集上与增量算法进行比较,实验结果验证了2种算法的有效性,并且所设计的增量算法能在不影响性能的情况下加快特征选择的进程。