In recent years,optoelectronic synapses have garnered significant attention in the field of neuromorphic computing due to their integration of optical sensing and synaptic functions.In this work,we propose an optoelec...In recent years,optoelectronic synapses have garnered significant attention in the field of neuromorphic computing due to their integration of optical sensing and synaptic functions.In this work,we propose an optoelectronic synapse based on IGZO/Bi_(3.25)La_(0.75)Ti_3O_(12)heterojunction.Under UV light stimulation,this device can simulate a range of synaptic behaviors,including paired-pulse facilitation,spike-intensity-dependent plasticity,spike-number-dependent plasticity,spike-width-dependent plasticity,and the transition from short-term memory to long-term memory.The majority of perceptible information for humans is acquired through the visual system.The 3×3 retinal morphology synapse arrays constructed based on plasticity behaviors not only integrates light perception and storage functions but also exhibits adaptive adjustment capabilities to address image blurring caused by object movement.At the same time,in CNN recognition training,the device successfully simulates the learning-relearning mechanism of the human brain.These findings highlight the device's immense potential for applications in artificial vision systems.展开更多
航空光电装备平面光学窗口在确定构型尺寸和组合形式后,窗口厚度对装备成像性能影响至为关键。从设计变量、组合形式、作用载荷3个方面对窗口设计问题进行分析,确定了基于结构刚性和RMS(root mean square)波前误差的平面窗口厚度设计目...航空光电装备平面光学窗口在确定构型尺寸和组合形式后,窗口厚度对装备成像性能影响至为关键。从设计变量、组合形式、作用载荷3个方面对窗口设计问题进行分析,确定了基于结构刚性和RMS(root mean square)波前误差的平面窗口厚度设计目标;从光学窗口一阶结构谐振频率及其在温度和压力载荷作用下的RMS波前误差量化求解出窗口厚度尺寸范围11.9 mm~18.4 mm;提出针对光学窗口组合工况性能计算的集成分析流程,计算得到光学窗口一阶谐振频率为151 Hz,在飞行工作时各向组合工况-流场/重力/振动作用下光学窗口RMS波前误差不大于1/10参考波长。实验测试结果表明:光学窗口RMS波前误差小于0.07参考波长,实物样机振动和飞行条件下成像清晰、稳定,证明窗口厚度尺寸设计分析方法正确有效。展开更多
目标跟踪作为图像处理领域的重要组成部分,广泛应用于智能视频监控、军事侦察等领域。但在面对物体形变以及遮挡等复杂应用场景时,相关滤波算法由于缺乏目标和背景判别区分以及遮挡状态判断等策略,存在跟错目标、缓慢漂移到背景等现象,...目标跟踪作为图像处理领域的重要组成部分,广泛应用于智能视频监控、军事侦察等领域。但在面对物体形变以及遮挡等复杂应用场景时,相关滤波算法由于缺乏目标和背景判别区分以及遮挡状态判断等策略,存在跟错目标、缓慢漂移到背景等现象,在遮挡后目标重新出现时,缺乏重检测机制,这些问题导致了跟踪性能在实际工程中大幅下降。针对以上问题进行改进设计,首先在跟踪过程中,使用网络优化器更新多层深度特征提取网络,优化损失函数提高目标与背景的判别能力;其次,采用多重检测抗遮挡优化机制,确定跟踪器状态更新机制;最后,基于深度学习进行检测跟踪识别一体化设计,实现跟踪前典型目标的自动捕获,目标受遮挡后重新出现时实现对典型目标的重新捕获定位。在实验分析中,分别从跟踪精度、可视化定量损失以及算法速度等方面进行了性能验证。实测数据显示,本文采用的方法在以上方面性能表现良好,优于改进前的ECO(efficientconvolution operators for tracking)算法。展开更多
文摘现有的基于深度学习的医学图像分割方法,大多是利用大量的训练数据拟合检测网络,以获得优异的检测性能。这些方法往往需要较大的模型参数,导致检测实时性较差。为此,提出了基于局部上下文引导特征深度融合轻量级医学分割网络(local context guided feature deep fusion lightweight medical segmentation network,LCGML-net)。LCGML-net通过精确的特征选择与特征融合来减少模型拟合所需的参数数量,从而在保证检测精度的同时实现更小的模型参数。在特征提取阶段和映射阶段,分别通过提取和融合目标的多层次多尺度局部上下文特征来丰富特征表达和精准分割。在STARE、CHASEDB1和KITS19等多个基准数据集上开展的实验证明,与其他方法相比,所提出的LCGML-net具有最佳的检测性能和最小的模型参数。
基金financially supported by the National Natural Science Foundation of China(Grant Nos.11574057 and12172093)the Guangdong Basic and Applied Basic Research Foundation(Grant No.2021A1515012607)。
文摘In recent years,optoelectronic synapses have garnered significant attention in the field of neuromorphic computing due to their integration of optical sensing and synaptic functions.In this work,we propose an optoelectronic synapse based on IGZO/Bi_(3.25)La_(0.75)Ti_3O_(12)heterojunction.Under UV light stimulation,this device can simulate a range of synaptic behaviors,including paired-pulse facilitation,spike-intensity-dependent plasticity,spike-number-dependent plasticity,spike-width-dependent plasticity,and the transition from short-term memory to long-term memory.The majority of perceptible information for humans is acquired through the visual system.The 3×3 retinal morphology synapse arrays constructed based on plasticity behaviors not only integrates light perception and storage functions but also exhibits adaptive adjustment capabilities to address image blurring caused by object movement.At the same time,in CNN recognition training,the device successfully simulates the learning-relearning mechanism of the human brain.These findings highlight the device's immense potential for applications in artificial vision systems.
文摘目标跟踪作为图像处理领域的重要组成部分,广泛应用于智能视频监控、军事侦察等领域。但在面对物体形变以及遮挡等复杂应用场景时,相关滤波算法由于缺乏目标和背景判别区分以及遮挡状态判断等策略,存在跟错目标、缓慢漂移到背景等现象,在遮挡后目标重新出现时,缺乏重检测机制,这些问题导致了跟踪性能在实际工程中大幅下降。针对以上问题进行改进设计,首先在跟踪过程中,使用网络优化器更新多层深度特征提取网络,优化损失函数提高目标与背景的判别能力;其次,采用多重检测抗遮挡优化机制,确定跟踪器状态更新机制;最后,基于深度学习进行检测跟踪识别一体化设计,实现跟踪前典型目标的自动捕获,目标受遮挡后重新出现时实现对典型目标的重新捕获定位。在实验分析中,分别从跟踪精度、可视化定量损失以及算法速度等方面进行了性能验证。实测数据显示,本文采用的方法在以上方面性能表现良好,优于改进前的ECO(efficientconvolution operators for tracking)算法。
