改进AOD-Net的道路交通图像去雾算法

针对现有图像去雾算法在处理道路交通图像时无法同时兼顾去雾效果和实时性的问题,本文以快速一体化网络去雾算法(AOD-Net)为基础进行改进.首先,在AOD-Net中添加SE通道注意力,以自适应的方式分配通道权重,关注重要特征;其次,引入金字塔池化模块,扩大网络的感受野,并融合不同尺度特征,更好地捕捉图像信息;最后,使用复合损失函数同时关注图像像素信息和结构纹理信息.实验结果表明,改进后的AOD-Net算法对道路交通图像去雾后的峰值信噪比提升了2.52 dB,结构相似度达到了91.2%,算法复杂度和去雾耗时略微增加,但仍满足实时要求.     

基于深度神经网络的行人头部检测

行人检测已成为安防、智能视频监控、景区人流量统计所依赖的核心技术,最新目标检测方法包括快速的区域卷积神经网络Fast RCNN、单发多重检测器 SSD、部分形变模型DPM等,皆为对行人整体的检测。在大场景下,行人姿态各异,物体间遮挡频繁,只有通过对行人身体部分位置建模,抓住人的局部特征,才能实现准确的定位。利用Faster RCNN深度网络原型,针对行人头部建立检测模型,同时提取行人不同方向的头部特征,并加入空间金字塔池化层,保证检测速率,有效解决大场景下行人的部分遮挡问题,同时清晰地显示人群大致流动方向,相比普通的人头估计,更有利于人流量统计。     

Headquarter-centered common sourcing management through order coordination and consolidation

It is common for subsidiaries of a group company to use the same types of components for producing similar products. Different subsidiary companies may well procure such components from the same suppliers. This paper studies two sourcing management models. One is the Subsidiary-Autonomous Sourcing Management (SD-ASM) where subsidiaries manage their inventories and place purchasing orders independent of each other. The other is the Headquarter-centered Common Sourcing Management (HQ-CSM) where purchasing orders of subsidiaries are processed centrally through some kind of headquarter coordination. In the SD-ASM model, each subsidiary places replenishment orders at a time interval corresponding to their economic order quantity (EOQ). In the HQ-CSM model, two purchasing order management policies are examined. One is the Order Coordination policy in which common replenishment epochs or time periods are proposed by the headquarter and the subsidiaries are encouraged to coordinate the timing of their orders based on the common replenishment epochs. The other is the Order Consolidation policy in which the subsidiaries combine the quantity of their orders and the headquarter places a combined order with the supplier. In the Order Coordination policy, classic RAND heuristic is used to find the best common replenishment epoch and the best replenishment timing of each subsidiary. In the Order Consolidation policy, the optimal order quantity of the combined order is obtained from a mathematical model. The combined order is then allocated to the subsidiaries according to a proportional allocation rule. A series of numerical studies is conducted to compare the costs of the SD-ASM and HQ-CSM policies. The results show that HQ-CSM outperforms SD-ASM in terms of cost and robustness against demand uncertainties. This achievement is largely due to the economies of process (synergistic ordering process), the economies of scale (large order quantity with price discount) and risk pooling effect (transshipments). The results also reveal that the Order Consolidation policy with a combined order always performs better than the Order Coordination policy with common replenishment epochs especially in face of high demand uncertainties and high service level in the global market.     

基于改进U-Net网络的腺体细胞图像分割算法

针对腺体图像在自动分割过程中由于多尺度目标和信息丢失影响导致准确率降低的问题,文中采用了一种引入注意力模块的全卷积神经网络模型。该模型遵循编码器-解码器结构,在编码网络中用空洞残差卷积层代替原有的普通卷积层,并添加空洞金字塔池;再在解码网络中加入注意力模块,使模型输出高分辨率特征图,提高对多尺度目标的分割精度。实验结果表明,提出的网络模型参数少分割精度高,对腺体图像的平均分割精度高达89.7%,具有较好的鲁棒性。     

煤层气藏边界类型、成藏主控因素及富集区预测

如何定义煤层气藏、如何界定气藏边界是认识煤层气藏的关键,要评价煤层气富集区首先要找出控制其成藏的主要因素。为此,结合我国含煤盆地的典型煤层气藏特征,围绕煤层气藏边界类型、成藏主控因素进行了探讨,并预测了沁水、鄂尔多斯、准噶尔盆地的煤层气富集区。结论认为：①煤层气藏的边界是煤层气藏划分的前提,煤层气藏主要有5种地质边界类型：水动力边界、风氧化带边界、物性边界、断层边界和岩性边界;②煤层气藏的形成经历了煤层气的生成和吸附、煤层的吸附能力增加和煤层气的解吸-扩散和保存阶段;③保存阶段是成藏的关键,区域构造演化、水动力条件和封闭条件是主要的成藏控制因素。进而在沁水盆地、鄂尔多斯盆地和准噶尔盆地优选出9个煤层气有利富集区：晋城、阳城、安泽、阳泉、大宁-吉县、韩城、吴堡、准南、阜康。     

基于二阶混合注意力的行人重识别

基于注意力机制的行人重识别方法更多利用图像中 一阶信息,忽略了特征中二阶信息 ,不能挖掘特征图之间的相关性和细粒度信息。提出一种基于二阶混合注意力的行人重 识别算法(second-order mixed attention module,SMAN)。二阶混合注意力模块(second-order mixed attention module,SOMA)由二阶通道注意力(second-order channel attention,SOCA)和二阶空间注意力模 块(second-order information,SOSA)组成,该方法将全局协方差池函数嵌入到SOCA和SOSA模块中,学习特征中二阶信息 。SOCA模块学习特征图之间相关性,SOSA模块则重新为特征图分配权重,关注特征图空间域 的细粒度信息。SMAN算法在Market-1501和 DukeMTMC-ReID数据集上的首位准确率分别 为 94.3%和87.1%,mAP分别达到85.7%和74.5%,同时使用类激活图验证SOMA模块的影响 ,实验表 明SMAN算法充分利用特征图的通道域和空间域中二阶信息。算法的性能优于现有的一些基于 注意力机制行人重识别方法,甚至接近某些优秀的方法。     

融合可变形卷积网络的鱼眼图像目标检测

环视鱼眼图像具有目标形变大和图像失真的缺点,导致传统网络结构在对鱼眼图像进行目标检测时效果不佳.为解决环视鱼眼图像中由于目标几何畸变而导致的目标检测难度大的问题,提出一种基于可变形卷积网络的鱼眼图像目标检测方法.将Cascade_RCNN中固定的卷积层和池化层分别替换为可变形卷积层和可变形池化层,使用Resnet50网...     

渤海海域浅层油气运移新认识与勘探新进展

浅层新近系是渤海海域油田储量和产量的主要层系。随着勘探程度的提高，大—中型油气田发现的难度也越来越大。基于渤海海域的勘探实践，结合浅层大—中型油田源外成藏的特性，通过分析油气深层运移与浅层富集的耦合关系，明确了古近系烃源层油气初次、二次运移所形成的"汇聚脊"对其上浅层油气富集、聚集的控制作用，并根据输导体系的组合划分出油气汇聚样式。在"汇聚脊"控藏模式的指导下，渤海海域多个构造带的油气勘探都取得突破，获得一系列亿吨级新发现，由此保证了渤海海域浅层规模优质油气储量持续发现，指导后续勘探进展。     

四川盆地震旦纪—早寒武世克拉通内裂陷地质特征

随着四川盆地震旦系—寒武系天然气勘探取得重大突破,地质认识也有了新的变化,其中之一就是在川中高石梯一磨溪地区西侧发现了一个近南北向展布的负向构造带,但对其构造性质、边界特征与展布范围、沉积充填特征、形成时期与演化等方面的认识还存在着较多分歧。由于该负向构造带与安岳震旦系寒武系特大型气田的形成关系十分密切,因此有必要对其进行深入研究。通过分析该区震旦系—寒武系钻井资料和野外剖面观察,利用27 000 km二维地震资料和最新钻井资料(高石17井),结合该盆地区域构造背景,研究了绵竹—长宁克拉通内裂陷边界与展布、地层沉积特征和演化过程。结果表明:①该裂陷总体呈南北向展布,裂陷东边界陡,相对稳定发育,裂陷西边界缓且不同时期发育位置不同;②震旦纪灯二期末该裂陷表现为近似对称的坳陷形态特征,分布面积为3×10~4 km~2;③灯四期末其又表现为受裂陷东边界断层控制的东陡西缓的断陷形态特征,分布范围达8×10~4 km~2;④该裂陷演化分为震旦纪灯一期一灯二期坳陷雏形形成、灯三期—灯四期裂陷形成发育、早寒武世麦地坪期—筇竹寺期裂陷充填与沉降、早寒武世沧浪铺期—龙王庙期裂陷萎缩与消亡等4个阶段。结论认为:该裂陷两侧震旦系—寒武系天然气成藏条件优越,是该区今后的主要勘探方向。     

Using functional programming to recognize named structure in an optimization problem: Application to pooling

Branch‐and‐cut optimization solvers typically apply generic algorithms, e.g., cutting planes or primal heuristics, to expedite performance for many mathematical optimization problems. But solver software receives an input optimization problem as vectors of equations and constraints containing no structural information. This article proposes automatically detecting named special structure using the pattern matching features of functional programming. Specifically, we deduce the industrially‐relevant nonconvex nonlinear Pooling Problem within a mixed‐integer nonlinear optimization problem and show that we can uncover pooling structure in optimization problems which are not pooling problems. Previous work has shown that preprocessing heuristics can find network structures; we show that we can additionally detect nonlinear pooling patterns. Finding named structures allows us to apply, to generic optimization problems, cutting planes or primal heuristics developed for the named structure. To demonstrate the recognition algorithm, we use the recognized structure to apply primal heuristics to a test set of standard pooling problems. © 2016 The Authors AIChE Journal published by Wiley Periodicals, Inc. on behalf of American Institute of Chemical Engineers AIChE J, 62: 3085–3095, 2016