融合隐向量对齐和Swin Transformer的OCTA血管分割

目的 光学相干断层扫描血管造影（optical coherence tomography angiography,OCTA）是一种非侵入式的新兴技术,越来越多地应用于视网膜血管成像。与传统眼底彩照相比,OCTA技术能够显示黄斑周围的微血管信息,在视网膜血管成像邻域具有显著优势。临床实践中,医生可以通过OCTA图像观察不同层的血管结构,并通过分析血管结构的变化来判断是否存在相关疾病。大量研究表明,血管结构的任何异常变化通常都意味着存在某种眼科疾病。因此,对OCTA图像中的视网膜血管结构进行自动分割提取,对众多眼部相关疾病量化分析和临床决策具有重大意义。然而,OCTA图像存在视网膜血管结构复杂、图像整体对比度低等问题,给自动分割带来极大挑战。为此,提出了一种新颖的融合隐向量对齐和Swin Transformer的视网膜血管结构的分割方法,能够实现血管结构的精准分割。方法 以ResU-Net为主干网络,通过Swin Transformer编码器获取丰富的血管特征信息。此外,设计了一种基于隐向量的特征对齐损失函数,能够在隐空间层次对网络进行优化,提升分割性能。结果 在3个OCTA图像数据集上的...     

Dual-terminal event-triggered dynamic output feedback H∞ control for aero engine networked control systems with quantization

This study deals with the problem of dual-terminal event-triggered dynamic output feedback (DOF) control for aero engine networked control systems (NCSs) subject to network-induced delay, external disturbance, and quantization effects. First, we established a generalized mathematical model of the aero engine. Second, a dual-terminal dynamic event-triggered mechanism (DETM) was designed to reduce the utilization of the network bandwidth. In addition, a DOF controller with a compensation function was proposed to stabilize the system. By utilizing the Lyapunov-Krasovskii (LK) method, the stability criteria were determined. Accordingly, the design conditions of the DOF controller and DETMs were presented. Furthermore, based on a genetic algorithm (GA), a parameter tuning method was proposed to obtain the allowable delay upper bound with less conservatism. Finally, some examples were presented to show the effectiveness and superiority of the presented scheme.     

PDE-based event-triggered containment control of multi-agent systems with input delay under spatial boundary communication

This article studies the containment control problem for multi-agent systems with input delay under spatial boundary communication by employing an event-based approach. Firstly, the collective dynamics of multi-agent systems are described as parabolic partial differential equations (PDEs). Applying the integral transformation method developed in PDEs, the delayed parabolic PDEs are transformed into a series of new coupled PDE-PDE systems. Then, by taking the spatial boundary communication scheme into account and using the local boundary information, two boundary containment control protocols together with a dynamic event-triggered scheme (DETS) are designed, such that the states of all followers converge to a convex hull formed by multiple leader agents with and without input delay. The optimal protocols are given by minimizing the 2-norm of the designed control gain matrix, and the exclusion of the Zeno behavior is analyzed. Finally, a numerical simulation example is provided to support the main results.     

Semi-supervised transformable architecture search for feature distillation

Pattern Analysis and Applications - The designed method aims to perform image classification tasks efficiently and accurately. Different from the traditional CNN-based image classification methods,...     

Graph-based fine-grained model selection for multi-source domain

Pattern Analysis and Applications - The prosperity of datasets and model architectures has led to the development of pretrained source models, which simplified the learning process in multi-domain...     

边界信息保持的全染色肾脏切片多粒度分割

目的 肾小球图像的准确分割对肾脏病理学的疾病诊断和定量分析起到关键作用,然而全染色肾脏切片图像存在由肾小球个体差异大导致的空间尺度和上下文形状变化大,以及图像分辨率过高的问题,给高精度、高性能分割任务带来挑战。为此,提出一种边界信息保持的全染色肾脏切片多粒度分割方法。方法 使用一种多粒度上下文的空间注意力机制生成多粒度和多形状变化的空间注意力图,以限制上下文特征,减弱背景对目标的影响,强化网络对目标的感知能力,使网络更多地关注小目标特征;将原图像切分为若干小图来解决全染色图像分辨率高的问题,使用增广路径边界补零策略处理卷积核存在的贡献偏移效应,解决了肾小球目标处于图像边界所导致的分割困难问题,保证图像块的信息无损失地向高层传递,提高处于图像块边界的肾小球目标的分割精度;进一步地,针对图像块拼接带来的边缘肾小球容易漏检、计算开销大的问题,采用特征复用的概率累积滑窗策略,同时提高了分割精度和效率。结果 在小鼠肾脏细胞切片和HuBMAP（human biomolecular atlas program）人体肾脏数据上,本文方法提高了分割精度,并使预测速度提高50%左右。结论 对于全染色肾脏切片的肾小球分割问题,多粒度上下文特征和增广路径边界补零策略解决了边界区域肾小球目标分割困难、分割精度低的问题,并通过概率累积滑窗策略提高分割速度,相较传统的分割方法有更优秀的性能。     

Fully distributed optimization of second-order systems with disturbances based on event-triggered control

This paper studies the distributed optimization problem of second-order multiagent systems containing external disturbances. To reject the external disturbances and lead agents' states to converge to the optimal consensus point, an adaptive event-triggered controller is proposed based on the internal model principle. With the adaptive mechanism, both the controller and the event-triggering condition do not contain the parameters related to global information, such as the maximum Lipschitz constant and the minimum strongly convex constant of local cost functions, and hence the event-triggered controller is fully distributed. By utilizing the event-triggered scheme, the consumption of communication among neighbors and the computing resources are saved. Furthermore, with the Lyapunov analysis framework, the optimal consensus can be proved to achieve and Zeno behavior is excluded from the event-triggering condition. Finally, the effectiveness of the proposed protocol is verified by numerical simulations.     

基于IGA-BP神经网络的锂电池健康状态估算

针对锂离子电池健康状态(SOH)估算精度低、传统遗传算法(GA)易陷入局部最优、收敛速度慢的问题,为提高锂电池健康状态的估算精度,提出了交叉概率和变异概率自适应的调整策略对传统GA进行改进,在改进遗传算法(IGA)的作用下,使优良个体仍保持较好的进化能力,算法初期搜索范围、后期局部搜索能力以及收敛速度也得到加强。提取间接健康因子,再用改进的遗传算法对BP神经网络的初始参数寻优得到IGA-BP神经网络模型,基于NASA锂电池数据集分别用GA-BP与IGA-BP神经网络算法对SOH进行估算。结果表明:IGA-BP神经网络算法估算精度更高,且具备快速收敛的优势,平均绝对百分比误差和均方根误差分别下降了0.422%和0.412,拟合程度提高了8.1%。     

多智能体强化学习的机械臂运动控制决策研究

针对传统运动控算法存在环境适应性较差,效率低的问题。可以利用强化学习在环境中不断去探索试错,并通过奖励函数对神经网络参数进行调节的方法对机械臂的运动进行控制。但是在现实中无法提供机械臂试错的环境,采用Unity引擎平台来构建机械臂的数字孪生仿真环境,设置观察状态变量和设置奖励函数机制,并提出在该模型环境中对PPO(proximal policy optimization)与多智能体（agents）结合的M-PPO算法来加快训练速度,实现通过强化学习算法对机械臂进行智能运动控制,完成机械臂执行末端有效避障快速到达目标物体位置,并通过该算法与M-SAC（多智能体与Soft Actor-Critic结合）和PPO算法的实验结果进行分析,验证M-PPO算法在不同环境下机械臂运动控制决策调试上的有效性与先进性。实现孪生体自主规划决策,反向控制物理体同步运动的目的。     

基于线-超图神经网络的阿尔兹海默症分类

在阿尔兹海默症分类问题中,超图神经网络可以从被试间的超图关系中提取特征,在表示学习复杂图结构方面具有很好的优势,但大多数模型都直接或间接地将超图所表示的被试间的高阶复杂关系分解,转化为简单的二元关系进行特征学习,没有有效利用超边的高阶信息,因此提出了基于线-超图神经网络(line-hypergraph neural network, L-HGNN)的阿尔兹海默症分类模型,该模型利用稀疏线性回归表征被试间多元相关性,借助超图和线图的转换在神经网络模型中实现节点的高阶邻域信息传递和超边整体结构特征学习,同时,结合注意力机制生成更具区分性的节点嵌入,进而用于阿尔兹海默症的辅助诊断.在ADNI数据上与常用的两种方法比较,实验结果表明,该方法能有效提高分类准确率,在阿尔兹海默症早期诊断上具有重要的应用价值.