基于改进噪声检测的迭代中值滤波算法

提出了一种基于改进噪声检测的迭代中值滤波算法。该算法首先检测出噪声点,然后利用信号点对噪声点滤波。其中噪声检测方法改进了IMFLED方法,进一步降低了误检率和漏检率。经过仿真实验并与其他滤波算法进行比较表明,该算法可以有效地去除图像中的脉冲噪声并且保留原图像的细节。     

差分离散方法分布式并行计算的重叠边界优化

针对分布式存储环境下显式差分方法的并行计算问题 ,依据分布式存储多处理机的体系结构特点 ,提出了一个重叠边界优化模型 ;该模型目前已成功应用于中国科学院第三代海洋环流模式的分布式并行计算优化。     

协调式六自由度运动平台的运动学分析

从平台的基本变换着手,用运动平台和固定平台间的矢量关系分析了协调式六自由度运动平台的运动学性能.给出了平台进行各种运动(平动、转动)时作动筒的位移、速度和加速度变化规律,并以一个具体实例对其进行了验证.     

未来中远距协同空战多目标攻击决策研究

根据先进战斗机的技术特点和未来中远距协同空战发展方向,提出一种基于先敌发现、先敌发射、先敌摧毁能力的空战态势分析模型。同时,采用一种将并行遗传算法与分布估计算法相结合的并行分布遗传算法并将其用于多目标攻击决策。最后利用具体算例进行仿真。结果表明空战态势分析模型能够较为准确地描述未来中远距多机协同空战,采用的算法具有较好的可行性和实时性,为中远距多机协同空战的多目标攻击决策提供了新的思路和方法。     

神经网络架构搜索研究进展与展望

神经网络架构搜索旨在针对不同任务,自动化地搜索得到性能最优的神经网络结构,是深度学习、计算机视觉技术结合当前现实需求应运而生的一大重要科学问题。对近年来神经网络架构搜索研究进行梳理、归类和评述;阐述神经网络架构搜索的定义和意义,全方位剖析当前研究所面临的难点与挑战;以此为基础,对主流的搜索策略进行阐述和归纳;探讨研究潜在的问题及未来颇具潜力的研究方向,以期推动该领域的进一步发展。     

低空复杂场景红外弱小目标快速精准检测

针对低空复杂场景下红外弱小动目标检测难度大、虚警率高等问题,面向探测系统中高帧频图像实时处理应用需求,提出基于全卷积网络的弱小目标精准检测方法和基于现场可编程逻辑门阵列(field programmable gate array, FPGA)的低时延并行处理方法。采用轻量化全卷积网络对红外图像中弱小目标进行空域检测,对相邻图像帧疑似目标进行时域轨迹关联以进一步降低虚警率。实验结果表明:上述方法相比于五种传统方法在检测率和虚警率性能方面均有显著提升,并在单片FPGA上完成100 Hz图像实时处理,处理时延低于1.8 ms,实现低空复杂场景弱小目标高精度高鲁棒快速实时检测。     

并行生成卷积网络图像修复算法

为解决修复纹理精细、背景复杂图像中大面积不连续语义缺失时存在的边缘伪影和语义不连续的缺陷,提出一种并行生成卷积的残差连接图像修复算法。将残缺图像输入一个两列平行卷积的结构修复网络得到两个具有不同感受野大小的图像分量,通过共享解码合并两个图像分量并计算输出的L2损失优化网络。将结构修复网络的输出送入包含残差连接与注意力机制的细节修复网络,融合上下文信息,改善修复细节能力。使用全局与局部鉴别器和预训练视觉几何组网络计算损失,对修复网络进行整体判别优化,增强修复结果的整体与局部一致性。在国际公认数据库上验证提出算法的性能,实验结果表明：提出算法可以有效修复复杂背景且包含精细纹理的大面积不规则缺失区域,提升图像细节、语义和结构的真实性与完整性,其峰值信噪比和结构相似度优于经典的对比算法。     

高速并行Gardner算法设计与实现

随着空间探测任务逐步增加、空间信道频谱资源日趋紧张,传统Gardner定时同步算法已经无法满足高速数传系统高通量、高可靠性的需求。为了提高Gardner定时同步算法的吞吐率并增大可纠正误差范围,提出一种高速并行Gardner算法。为了保证插值精度同时减少乘法器消耗,设计了一种并行分段抛物线插值滤波器;为了便于并行流水线设计和最佳采样点选取,构建了计数模块和定时缓存调整模块;为了提高等价吞吐率,重构了流水线并行环路滤波器结构和并行数控振荡器结构。结果表明,该算法等价吞吐率可达1 739.13 Msps,数字信号处理器资源消耗可减少44%,可纠正2×10^-3的定时误差。     

Three-dimensional coordinates test method with uncertain projectile proximity explosion position based on dynamic seven photoelectric detection screen

To objectively obtain the three-dimensional coordinates of the projectile fuze proximity explosion when projectile intersects the head of missile target, we propose a dynamic seven photoelectric detection screen test method, which is made up of six plane detection screens and a flash photoelectric dynamic detection screen. The three-dimensional coordinates calculation model of the projectile proximity explosion position based on seven plane detection screens with dynamic characteristics is established. According to the relation of the dynamic seven photoelectric detection screen planes and the time values, the analytical function of the projectile proximity explosion position parameters under non-linear motion is derived. The projectile signal filtering method based on discrete wavelet transform is explored in this work. Additionally, the projectile signal recognition algorithm using an improved particle swarm is proposed. Based on the characteristics of the time duration and the signal peak error for the projectile passing through the detection screen, the signals attribution of the same projectile passing through six detection screens are analyzed for obtaining precise time values of the same projectile passing through the detection screens. On the basis of the projectile fuze proximity explosion test, the linear motion model and the proposed non-linear motion model are used to calculate and compare the same group of projectiles proximity explosion position parameters. The comparison of test results verifies that the proposed test method and calculation model in this work accurately obtain the actual projectile proximity explosion position parameters.     

Due‐window assignment with unit processing‐time jobs

In due‐window assignment problems, jobs completed within a designated time interval are regarded as being on time, whereas early and tardy jobs are penalized. The objective is to determine the location and size of the due‐window, as well as the job schedule. We address a common due‐window assignment problem on parallel identical machines with unit processing time jobs. We show that the number of candidate values for the optimal due‐window starting time and for the optimal due‐window completion time are bounded by 2. We also prove that the starting time of the first job on each of the machines is either 0 or 1, thus introducing a fairly simple, constant‐time solution for the problem. © 2004 Wiley Periodicals, Inc. Naval Research Logistics, 2004