准确高效的麦粒计数对小麦育种和产量评估具有重要意义。传统人工计数方法费时费力且易出错。目前的自动计数方法主要基于二维图像处理技术,但在处理麦粒遮挡和获取立体形态特征方面存在局限。点云数据能够完整记录麦穗的三维几何结构,...准确高效的麦粒计数对小麦育种和产量评估具有重要意义。传统人工计数方法费时费力且易出错。目前的自动计数方法主要基于二维图像处理技术,但在处理麦粒遮挡和获取立体形态特征方面存在局限。点云数据能够完整记录麦穗的三维几何结构,为解决这些问题提供了新的思路。本文针对现有点云目标检测算法在处理密集分布麦粒时的不足,提出了一种改进的3DSSD网络用于麦穗点云中的麦粒检测与计数。该方法充分利用麦粒的形态学特征,设计了2个核心创新模块:一是提出局部形状感知采样策略(Local shape-aware sampling,LSAS),通过分析点云的局部几何结构来指导采样过程,有效缓解了传统最远点采样(Farthest point sampling,FPS)算法在密集目标场景下的特征退化问题;二是引入部件感知损失函数(Part-aware loss function,PALF),将麦粒建模为具有多个关键部位的目标,增强了网络对局部特征的感知能力。实验结果表明,改进后的方法在麦粒检测任务中AP@25达到72.68%,较基线3DSSD提升14.02%,计数任务MAE降至3.87,较3DSSD下降了85.54%,Recall提升至93.21%,从而在处理形态复杂、目标密集的麦穗点云时表现出显著优势。本研究为实现麦穗表型的快速、准确测量提供了新的技术方案,并成功地在马兰国家农业科技园区应用该方法。展开更多
随着深度学习的迅速发展,图像识别技术也随之日益提高,其中目标检测在辅助驾驶系统、医学领域和车流监测系统等占有重要地位。大多目标检测算法对大目标较为敏感,且并未考虑特征与特征之间的相互关系及重要程度,然而小目标在图像中覆盖...随着深度学习的迅速发展,图像识别技术也随之日益提高,其中目标检测在辅助驾驶系统、医学领域和车流监测系统等占有重要地位。大多目标检测算法对大目标较为敏感,且并未考虑特征与特征之间的相互关系及重要程度,然而小目标在图像中覆盖区域小,分辨率低,携带信息较少,导致小目标的误检或漏检率较高。针对以上问题,对小目标检测的难点进行研究,提出了一种基于改进的DSSD(deconvolutional single shot detector)的小目标检测算法。该算法引入混合注意力机制,在通道维度上增加权重分量进行加权求和表示信息相关度,并将图片中的空间域信息做对应空间变换,提取关键信息,突出局部重点区域,有利于前景小目标的特征学习。实验结果表明,该算法在VOC2007测试集上的精确度达到81.02%,比原DSSD算法高出1.3%,且均优于其他对比算法,证明了算法的有效性。展开更多
Double-crossed-step-stress(DCSS) accelerated life test(ALT) method is widely used for estimating the lifetime of products with high reliability and long lifetime. In order to further reduce the test time and test cost...Double-crossed-step-stress(DCSS) accelerated life test(ALT) method is widely used for estimating the lifetime of products with high reliability and long lifetime. In order to further reduce the test time and test cost, a double-synchronous-step-stress(DSSS) ALT method which combines a double-synchronous-step-downstress(DSSDS) ALT method and a double-synchronous-step-up-stress(DSSUS) ALT method is proposed. The accelerated stresses decrease and increase in a synchronous way with one step in the DSSDS-ALT and DSSUSALT methods, respectively. Monte Carlo method is adopted to simulate the two methods, and the validity and efficiency of them are demonstrated by the simulation results. In addition, a comparison analysis of efficiency between DSSDS-ALT method and DSSUS-ALT method is carried out. The result shows that the DSSDS-ALT method compared with the DSSUS-ALT method can significantly improve the test efficiency under the same test condition.展开更多
文摘准确高效的麦粒计数对小麦育种和产量评估具有重要意义。传统人工计数方法费时费力且易出错。目前的自动计数方法主要基于二维图像处理技术,但在处理麦粒遮挡和获取立体形态特征方面存在局限。点云数据能够完整记录麦穗的三维几何结构,为解决这些问题提供了新的思路。本文针对现有点云目标检测算法在处理密集分布麦粒时的不足,提出了一种改进的3DSSD网络用于麦穗点云中的麦粒检测与计数。该方法充分利用麦粒的形态学特征,设计了2个核心创新模块:一是提出局部形状感知采样策略(Local shape-aware sampling,LSAS),通过分析点云的局部几何结构来指导采样过程,有效缓解了传统最远点采样(Farthest point sampling,FPS)算法在密集目标场景下的特征退化问题;二是引入部件感知损失函数(Part-aware loss function,PALF),将麦粒建模为具有多个关键部位的目标,增强了网络对局部特征的感知能力。实验结果表明,改进后的方法在麦粒检测任务中AP@25达到72.68%,较基线3DSSD提升14.02%,计数任务MAE降至3.87,较3DSSD下降了85.54%,Recall提升至93.21%,从而在处理形态复杂、目标密集的麦穗点云时表现出显著优势。本研究为实现麦穗表型的快速、准确测量提供了新的技术方案,并成功地在马兰国家农业科技园区应用该方法。
文摘随着深度学习的迅速发展,图像识别技术也随之日益提高,其中目标检测在辅助驾驶系统、医学领域和车流监测系统等占有重要地位。大多目标检测算法对大目标较为敏感,且并未考虑特征与特征之间的相互关系及重要程度,然而小目标在图像中覆盖区域小,分辨率低,携带信息较少,导致小目标的误检或漏检率较高。针对以上问题,对小目标检测的难点进行研究,提出了一种基于改进的DSSD(deconvolutional single shot detector)的小目标检测算法。该算法引入混合注意力机制,在通道维度上增加权重分量进行加权求和表示信息相关度,并将图片中的空间域信息做对应空间变换,提取关键信息,突出局部重点区域,有利于前景小目标的特征学习。实验结果表明,该算法在VOC2007测试集上的精确度达到81.02%,比原DSSD算法高出1.3%,且均优于其他对比算法,证明了算法的有效性。
基金the National Natural Science Foundation of China(Nos.51265025 and 51665029)
文摘Double-crossed-step-stress(DCSS) accelerated life test(ALT) method is widely used for estimating the lifetime of products with high reliability and long lifetime. In order to further reduce the test time and test cost, a double-synchronous-step-stress(DSSS) ALT method which combines a double-synchronous-step-downstress(DSSDS) ALT method and a double-synchronous-step-up-stress(DSSUS) ALT method is proposed. The accelerated stresses decrease and increase in a synchronous way with one step in the DSSDS-ALT and DSSUSALT methods, respectively. Monte Carlo method is adopted to simulate the two methods, and the validity and efficiency of them are demonstrated by the simulation results. In addition, a comparison analysis of efficiency between DSSDS-ALT method and DSSUS-ALT method is carried out. The result shows that the DSSDS-ALT method compared with the DSSUS-ALT method can significantly improve the test efficiency under the same test condition.
