Effect of the Three Gorges Reservoir on vertical distribution of suspended sediment concentration and estimation accuracy of depth-average concentration in downstream reaches

The sediment regime in the middle Yangtze River has been significantly changed from quasi-equilibrium to nonequilibrium since the impoundment of the Three Georges Reservoir (TGR). To understand the effects of the TGR on vertical distribution of suspended sediment concentration (SSC) and estimation accuracy of mean concentration, vertical sediment concentration and flow velocity data at the Shashi and Jianli hydrological stations in the reach before and after the impoundment were collected. Comparisons and analysis of vertical profiles of SSC before and after the TGR impoundment show that after the impoundment of the TGR, due to the coarsening of sediment particle size, the reduction in sediment load, and the significant decrease in sediment saturation, the vertical distribution of SSC in the downstream reaches became more uneven under medium and low water flows, which was reflected in the vertical gradient and the fluctuation degree of SSC significantly increased. In addition, the depth-average sediment concentrations were calculated by the selected-point method and the mean values calculated by the “multi-point method” were regarded as the “true mean” to evaluate the accuracy of the mean value calculated by the “few-point method.” It was found that the relative errors for the selected-point method were mainly positive before impoundment but mainly negative after impoundment. Additionally, the correction factors of one-point, two-point, and three-point methods and the position of the near-bed substituted point for the five-point method were given to reduce the error when the point measurements were used to calculate the depth-average sediment concentration in the downstream reaches.     

基于注意力和卷积特征重排的图像修复

近年来,基于UNet与GAN（生成对抗网络）的深度学习网络模型在图像修复领域展现出了独特的优势,但是修复结果中仍然存在伪影、模糊、纹理细节退化、对于大面积破损难以修复、修复后的孔洞与背景图像不相容等现象。为了解决现有模型对大面积破损图像修复不友好以及修复后的图像存在退化现象等问题,通过对现有方法进行研究,改进了Shift-UNet（移位网络）模型：在UNet与GAN的基础上,在每一层编码器和解码器之间增加了改进之后的注意力机制Attention-UNet并融入shift-UNet,形成Attention-Shift-UNet;并且通过研究,将原来下采样部分的激活函数由Leaky_ReLU改为了SiLU函数。改进模型不仅在64×64的中心遮罩上取得了较好的效果,还实现了随机遮罩,遮罩面积由20%增加到80%。通过实验结果证明,该模型的修复效果更优,特别是针对大面积破损图像的修复。在CelebA、ParisArchitecture以及Paris Streetview数据集上经过测试,各评价指标都有明显的提高,其中结构相似性（SSIM）由原来的0.944 5提高到0.947 1,峰值信噪比（PSNR）由原来的27.992 7提高到28.553 6,L2损失由原来的0.001 7降低到0.001 5。     

基于混合采样的图对比学习推荐算法

在推荐系统领域中,图卷积网络具有对于图结构数据更强的信息抽取能力。然而,现有的图卷积网络推荐算法主要关注改进模型结构,忽视了提高原始样本采样质量与挖掘用户—项目间隐式关系的重要性。针对上述问题,提出一种基于混合采样的图对比学习推荐算法。首先使用混合采样方法,提取出正样本中部分信息并将其注入负样本,从而生成全新的富含信息的难负样本;其次,通过轻量图卷积网络对难负样本进行特征提取,得到用户和项目的节点表征,采用邻域对比学习方法挖掘样本隐式关系;最后,利用多任务策略对推荐监督任务和对比学习任务进行联合优化。在真实数据集Yelp2018和Amazon-book上进行实验,采用recall和NDCG指标进行评估,实验结果表明,提出的模型相较其他基准模型取得了更好的效果。     

LFSCA-UNet：基于空间与通道注意力机制的肝纤维化区域分割网络

目的 肝纤维化是众多慢性肝脏疾病的常见表现,如不及时治疗可发展为肝硬化甚至引发肝癌。肝纤维化的准确评估对临床治疗和预后评估等至关重要。目前,肝纤维化的诊断通过肝穿活检判断,有创且有并发症危险。为此,基于影像学的无创诊断方法越来越受到关注。本文提出一种基于通道注意力与空间注意力机制改进的用于肝纤维化区域的自动化分割U-Net （liver fibrosis region segmentation network based on spatial and channel attention mechanisms,LFSCA-UNet）。方法 依据Attention U-Net的改进方式,围绕U-Net的跳跃连接结构进行基于注意力的改进,在AG （attention gate）的基础上,加入以ECA （efficient channel attention）模块为实现方式的通道注意力机制,依据加入ECA的位置,LFSCA-UNet分为A、B、C共3个子型。结果 在肝数据集上与其他实验网络进行评估对比,本文提出的LFSCA-UNet网络结构平均Dice系数达到了93.33%,相比原始U-Net的Dice系数提高了0.539 6%。结论 本文方法将空间注意力机制与通道注意力机制进行结合,有效提高了肝纤维化区域的分割精度,对空间注意力模块使用通道注意力模块优化输入和输出,增加了网络的稳定性,提升了网络的整体效果。     

异构网络光传感通信跨域口令认证方式优化

为了解决异构网络通信跨域口令认证效率低、安全性差的问题,提出了改进传统C2C-PAKE认证协议安全性能,改善切换切入阶段时延的跨域口令认证方式优化方法.改近传统C2C-PAKE认证协议,建立分立式子层,形成多层扩展子层认证系统,引入私钥机制及随机序列数解密用户信息,生成跨域通信用户特有的认证口令,用自身私钥加密语义信息和认证协议,重复用户验证步骤,排除其它恶意攻击者,提高认证安全指数;在预切换和接人认证传输安全口令阶段,应用增加算法得到口令认证域间信任列表,提高认证效率,建立异构网络跨域通信实验环境,仿真对比结果表明,在安全性能和认证效率两方面,所提方法遭遇的认证威胁与认证开销均优于其它方法,具有广阔的应用空间.     

基于深度学习的集群式供应链应急物资需求预测研究

在突发事件和大数据情景下,建立基于数据流模糊C均值聚类算法的集群式供应链应急物资需求重要度决策算法,有助于辨识集群式供应链子系统应急物资需求的重要程度。针对集群式供应链中各子供应链之间的耦合特性和预测指标的快速变化数据流特征,提出基于长短期记忆网络的集群式供应链应急物资需求动态预测算法,提取集群式供应链多个子系统应急物资需求参数的时序特征,动态地、分布地对互联大系统的应急物资需求不确定性进行系统辨识估计。仿真实验结果表明了基于长短期记忆网络的集群式供应链互联大系统应急物资需求动态预测算法的可行性和精确性。     

基于显著性的遥感图像舰船目标检测

舰船作为海洋信息感知中的重要目标,其检测在军舰探测、精确制导等军用领域以及海面搜救、渔船监测等民用领域具有极其重要的战略意义.海洋遥感图像受云雾、风浪、海杂波和光照等干扰使得舰船检测具有挑战性.根据可见光遥感图像舰船目标检测特点提出粗检测和细鉴别相结合的技术路线.先基于视觉显著性的谱残差法对图像进行增强以提取目标候选区域,后根据舰船与干扰因素差异采用舰船方向梯度直方图特征对目标候选区域进行鉴别,提取真正的舰船目标.实验结果表明,上述算法舰船检测率高,对光照、海杂波干扰具有一定程度的鲁棒性,且能有效剔除碎云岛屿等干扰物,显著降低虚警率.     

Finite Time Synchronization of Delayed Quaternion Valued Neural Networks with Fractional Order

Neural Processing Letters - In this article, the finite time (FT) synchronization problem of fractional order quaternion valued neural networks with time delay is investigated. Without separating...     

An improved K-means algorithm for underwater image background segmentation

Conventional algorithms fail to obtain satisfactory background segmentation results for underwater images. In this study, an improved K-means algorithm was developed for underwater image background segmentation to address the issue of improper K value determination and minimize the impact of initial centroid position of grayscale image during the gray level quantization of the conventional K-means algorithm. A total of 100 underwater images taken by an underwater robot were sampled to test the aforementioned algorithm in respect of background segmentation validity and time cost. The K value and initial centroid position of grayscale image were optimized. The results were compared to the other three existing algorithms, including the conventional K-means algorithm, the improved Otsu algorithm, and the Canny operator edge extraction method. The experimental results showed that the improved K-means underwater background segmentation algorithm could effectively segment the background of underwater images with a low color cast, low contrast, and blurred edges. Although its cost in time was higher than that of the other three algorithms, it none the less proved more efficient than the time-consuming manual segmentation method. The algorithm proposed in this paper could potentially be used in underwater environments for underwater background segmentation.

     

基于多时序特征和卷积神经网络的农作物分类

近年来,以卷积神经网络为主的深度学习模型在各种遥感应用中都显示出巨大的潜力。以加州帝国郡为研究区,以Landsat 8 OLI年内时序遥感影像计算时序植被指数NDVI、EVI、RVI以及TVI,组合后输入到构建的一维卷积神经网络 模型,以实现作物的高精度精细分类。为了验证卷积模型的优越性,另搭建了基于递归神经网络及其变体的深度学习模型。结果表明：①引入其他时序特征后,能够有效地提高卷积神经网络的分类精度。NDVI+EVI+TVI+RVI组合特征总体精度和Kappa系数最高,分别是89.667 4%和0.856 0,对比NDVI时序特征总体精度和Kappa系数提高了近4%和0.6。②在与其他深度学习模型的对比中,一维卷积神经网络分类精度最高,能够从时序数据中较为准确捕捉作物时序特征信息,尽管递归神经网络被广泛应用于序列数据的研究,但分类结果要略差于卷积神经网络。实验表明在NDVI的基础上引入其他植被指数辅助,能够有效地提高分类精度。基于一维卷积神经网络的深度学习框架为长时间序列分类任务提供了一种有效且高效的方法。