基于锚框的高效单层特征目标检测

目标检测广泛应用在公共场合的智能监控、自动驾驶与计算机辅助诊断等领域。文章提出了单层特征目标检测替代复杂的特征金字塔结构,从而提升模型的推理速度和预测精度。在模型搭建过程中,瓶颈特征结构采用了单层空洞残差编码器,样本选择采用了统一匹配机制,并采用了任务对齐检测器。在COCO(Microsoft Common Objects in Context)数据集下,大量实验证明该方法的有效性,以Res Net50为基准,预测精度达到了38.2 m AP,比Retina Net的推理速度快1.4倍,精度提高2.3 m AP。该模型具有推理速度快、预测精度高等特点,可以应用在许多特定场景中。

基于改进K-Means的静脉特征学习与识别

针对传统的静脉识别中静脉特征的提取需要通过先验知识进行人工设计,并且特征的设计过程中需要对大量的参数进行调整,同时需要在后续的分类器设计中进行特殊的选择才能达到较好的识别效果等缺陷,提出了一种改进方法,对单层网络的特征学习结构中的K-Means方法进行针对性改进,并将其引入到静脉识别的静脉特征学习过程中,在分类器中采用SVM实现静脉分类。另外,引入SIFT特征结合改进词袋模型(SBOW)的传统特征学习和分类方法分别进行静脉识别,并将两者的识别结果进行对比,从而证明将基于单层网络特征学习方法引入静脉识别中的优越性和必要性。

Survey and establishment of Nanshimenzi Bed of Tiaojishan Formation in Qinglong,northern Hebei,China

In recent years,many fossil vertebrates,including feathered dinosaurs,have been discovered at Nanshimenzi village,Gangou Town,Qinglong County,Hebei Province,China.A geological section including the fossil-bearing strata was measured at Nanshimenzi,to determine the exact geological age and sedimentary characteristics of the section,and a new lithostratigraphic unit was named as the Nanshimenzi Bed.The Nanshimenzi Bed is about 56.6 m in thickness,mainly consisting of gray sandstones and siltstones and containing fossil vertebrates and several coal streaks.On the basis of survey and comparison,the Nanshimenzi Bed should be assigned to Tiaojishan Formation and probably to the upper part of the formation.

Growth and low-energy electron microscopy characteri- zation of monolayer hexagonal boron nitride on epitaxial cobalt

Low-energy electron microscopy （LEEM） has been used to study the structure, initial growth orientation, growth progression, and the number of layers of atomically thin hexagonal boron nitride （h-BN） films. The h-BN films are grown on heteroepitaxial Co using chemical vapor deposition （CVD） at low pressure. Our findings from LEEM studies include the growth of monolayer film having two, oppositely oriented, triangular BN domains commensurate with the Co lattice. The growth of h-BN appears to be self-limiting at a monolayer, with thicker domains only appearing in patches, presumably initiated between domain boundaries. Reflectivity measurements of the thicker h-BN films show oscillations resulting from the resonant electron transmission through quantized electronic states of the h-BN films, with the number of minima scaling up with the number of h-BN layers. First principles density functional theory （DFT） calculations show that the positions of oscillations are related to the electronic band structure of h-BN.

Chain formation in a monolayer of dipolar hard spheres under an external field

The phase behavior of a monolayer of dipolar hard spheres under an external field, which makes all dipoles of the monolayer orientate along its direction, is investigated. Using integral equation theory in the reference hypemetted chain （RHNC） approximation we calculate the correlation functions, which are used to obtain the response matrix of grand potential with respect to density fluctuations. The smallest eigenvalue of this response matrix determines the stability of the monolayer. When the smallest eigenvalue approaches zero, the monolayer becomes unstable and the corresponding eigenvector characterizes this instability. At dilute densities, with decreasing temperature the dipoles of the monolayer begin to form chains and simultaneously condensate. At medium and high densities, however, the dipoles of the monolayer have a stronger tendency to form dipolar chains with decreasing temperature and there is no condensation. The part of specific heat related to potential energy is investigated and found to increase sharply near the temperature of dipolar chain formation. This is in accordance with a sharp decrease in potential energy induced by the formation of a dipolar chain.