INTERNATIONAL RESERVE MANAGEMENT: A DRIFT‐SWITCHING REFLECTED JUMP‐DIFFUSION MODEL

We study the cost of shocks, that is, jump risk, with respect to reserve management when the reserve process is formulated as a drift‐switching jump diffusion with a reflecting barrier at 0. Inspired by the Brownian drift switching model, our model results in a more realistic dynamic behavior of international reserves than the buffer stock model. The new model can capture both the jump behavior in reserve dynamics and the leptokurtic feature of the increment distribution which has a higher peak and two asymmetric heavier tails than the normal distribution. Through the selection of an initial distribution that reflects certain steady state behaviors, the reserve process becomes a regenerative process. This selection enables us to derive a closed‐form expression for the total expected discounted cost of managing reserves, thus helping us to numerically find management strategies that minimize costs. The numerical results show that shocks at the reserve level have a significant effect on reserve management strategies and that model misspecification can result in nonnegligible additional costs.     

采用航迹特征的智能机型识别

针对空战目标识别中机型自动识别比较困难的问题，提出了采用航迹特征的智能目标识别方法。利用卷积神经网络（Convolutional Neural Network，CNN）分层学习特征的能力，训练CNN算法模型自动地从航迹数据中学习有用的特征并分类。利用沿海实地采集的15个类别的飞机航迹数据，经一系列数据预处理后作为智能识别算法的训练和测试数据，在验证实验中描述了算法网络的相关配置,对比了CNN与其他分类器的识别结果。实验结果表明，CNN具有很好的识别性能。     

视频业务识别方法综述

视频用户体验的质量要求给网络传输带来了巨大挑战，未来的虚拟现实/增强现实(Virtual Reality/Augmented Reality，VR/AR) 视频业务则对交互时延提出了更高要求。视频业务识别是网络运营商及视频业务提供商对视频进行优先网络资源分配和自适应速率控制的基本前提。综述了业务识别领域的一般研究方法及其对视频业务的具体应用。首先，总结了每种业务识别方法的工作原理，对比了这些方法应用于视频业务时的优缺点及在准确率和召回率指标上的识别性能;然后，特别介绍了可以提升识别准确率和计算性能的特征选择算法;最后，阐述并分析了高吞吐量网络分类速度不够、无成熟视频业务识别平台和样本分布不均衡等视频业务识别面临的挑战，并针对现有业务识别方法存在的问题，给出了引入移动边缘计算、基于端口的方法串联其他识别方法等可能的解决方案。     

高分1号冬小麦解译面积核算方法研究

[目的]以高分1号(GF-1)融合2m卫星遥感影像为基础数据源,结合土地利用现状数据、高分多源遥感影像和地面样方等数据,对冬小麦分类提取中存在的面积误差问题进行研究和分析。[方法]文章以河南省永城市为研究区,在冬小麦提取结果聚类处理基础上,基于线性地物缓冲区数据,采用GIS空间运算实现线性地物面积扣除,接着分析了样方数据和土地利用现状数据再扣除零星地物面积比例上的差异,并采用样方零星地物平均扣除系数对全市各乡镇耕地与非耕地中冬小麦提取面积进行了相关统计和误差分析。[结果]永城市冬小麦最终解译面积11. 29万hm~2,其中线性地物和零星地物扣除面积分别为6 613. 08hm~2和3 875. 22hm~2,占研究区冬小麦解译面积的5. 86%和3. 32%,与统计上报数据相比,其处理前后误差由14. 12%降低至4. 41%,有效地提高冬小麦提取面积精度。[结论]误差来源分析与修正对冬小麦解译面积核算精度具有重要影响,该研究为县级区域尺度下冬小麦面积提取核算提供了思路和借鉴。     

利用改进特征金字塔模型的SAR图像多目标船舶检测

深度学习模型中的特征金字塔网络（Feature Pyramid Network，FPN）常被用作合成孔径雷达（Synthetic Aperture Radar，SAR）图像中多目标船舶的检测。针对复杂场景下多目标船舶检测问题，提出了一种基于改进锚点框的FPN模型。首先将特征金字塔模型嵌入传统的RPN（Region Proposal Network）并映射成新的特征空间用于目标检测，然后利用基于形状相似度距离（Shape Similar Distance,SSD）度量的Kmeans聚类算法优化FPN的初始锚点框，并使用SAR船舶数据集测试。实验结果表明，所提算法目标检测精确率达到98.62%，在复杂场景下与YOLO、Faster RCNN、FPN based on VGG/ResNet等模型进行对比，模型准确率提高，整体性能更好。     

利用频移动态特征的辐射源目标盲分选

利用辐射源目标的频移动态特征进行盲分选时，针对已有的分选方法不能很好解决非合作接收数据非周期、间断、交叠的问题，提出了一种改进的逻辑回归分选算法,将信号分选转化为轨迹分类的问题来处理。首先，借鉴多目标航迹起始模型中的M/N逻辑法的框架，设计新的数据关联门限，以建立目标起始曲线；其次，采用最小二乘拟合方法替代卡尔曼滤波，设计新的外推规则，建立载频变化预测曲线；分析新的接收数据与预测曲线的关联性，不断迭代回归或形成新的分支，最后获得目标分选结果。仿真与实测试验表明，利用逻辑回归分选算法，对不同运动目标搭载的同频辐射源可进行自动分选，4个仿真目标与2个真实目标的分选正确率达到100%。同时，该方法拓展了辐射源目标盲分选中的动态频移特征应用，为辐射源个体识别提供了新的支持。     

基于LBPH算法的人脸识别算法研究与设计

人脸识别技术是一种根据人脸的特征信息进行身份判定的技术,此技术经过几十年的发展,由弱人工智能向强人工智能转化,论文提出了一种Python语言利用LBPH算法提取人脸特征、实现人脸识别的方法。     

Vertical differentiation,product innovation,and dynamic competition

This paper presents an infinite horizon dynamic model in which two firms compete in a market vertically differentiated by the qualities of their products and consumers have heterogeneous preferences for quality. Given the product qualities offered, the firms engage in price competition that segments the market. In each period each firm can spend on product innovation that if successful increases the quality of its product. Three types of Markov perfect equilibria are identified. A running–coasting equilibrium exhibits increasing quality dominance with one firm undertaking innovation and the other coasting to free ride on the innovation by the first firm. The firm that coasts can have the larger dynamic payoff, so quality dominance does not imply payoff dominance. A second is a leap‐frog equilibrium in which the trailing firm undertakes innovation to leap into the lead. The trailing firm never innovates solely to narrow the gap with the leader, so catch up strategies are never used. In the third both firms undertake innovation, but if both have innovation successes, product differentiation remains the same and profits are reduced by the cost of innovation. The rivalry between Intel and AMD in microprocessors for personal computers provides a motivating example.     

Object detection based on visual memory: a feature learning and feature imagination process

ABSTRACT

Visual memory plays an important role for the human’s visual system to detect objects. The features of an object stored in the visual memory have much lower dimensions than the features contained within an image. We simulate the visual memory as a feature learning and feature imagination (FLFI) process to build an object detection algorithm. The method is constructed by a bottom-up feature learning and a top-down feature imagination. The proposed object detection method is tested using publicly available benchmark data sets, and the result indicates that it is fast and more robust.     

Organizational Design,Competition, and Financial Exchanges

We investigate the incentives for vertical and horizontal integration in the financial securities service industry. In a model with two exchanges and two central securities depositories (CSDs), we find that decentralized decisions might lead to privately and socially inferior industry equilibria with vertical integration of both CSDs with their respective exchanges. Allowing for horizontal integration of CSDs avoids privately inferior industry equilibria. However, we observe too little horizontal integration from the social perspective. We link our results to recent regulatory and institutional developments such as the emergence of multilateral trading facilities, over-the-counter regulation, and financial harmonization.