目标检测作为计算机视觉的重要研究方向,在智慧城市、无人驾驶等领域的作用越来越重要。传统目标检测算法中,根据交并比(Intersection over Union,IOU)的大小判断正负样本,但较低的IOU会引入噪声,降低检测器的精度;较高的IOU会保留少数...目标检测作为计算机视觉的重要研究方向,在智慧城市、无人驾驶等领域的作用越来越重要。传统目标检测算法中,根据交并比(Intersection over Union,IOU)的大小判断正负样本,但较低的IOU会引入噪声,降低检测器的精度;较高的IOU会保留少数高质量样本,造成过拟合;并且推荐区域和检测器的IOU阈值相差过大会引起质量不匹配问题。针对上述问题,提出了一种基于级联网络的平行级联检测网络,它由一系列检测器串并联而成,每个检测器设置递增的IOU阈值,从而在每个阶段都会得到一个更高质量的样本分布来训练下一级检测器,并逐步重采样减少过拟合。实验结果表明提出的平行级联检测网络的检测精度优于传统目标检测算法,在目标检测数据集Microsoft COCO上平均准确度(AP)提升了1.5个百分点左右。展开更多
Round method is the common method for discrete variable optimization in optimal design of complex mechanical structures;however,it has some disadvantages such as poor precision,simple model and lacking of working cond...Round method is the common method for discrete variable optimization in optimal design of complex mechanical structures;however,it has some disadvantages such as poor precision,simple model and lacking of working conditions＇ description,etc.To solve these problems,a new model is constructed by defining parameterized fuzzy entropy,and the rationality of parameterized fuzzy entropy is verified.And a new multidirectional searching algorithm is further put forward,which takes information of actual working conditions into consideration and has a powerful local searching capability.Then this new algorithm is combined with the GA by the fuzzy clustering algorithm（FCA）.With the application of FCA,the optimal solution can be effectively filtered so as to retain the diversity and the elite of the optimal solution,and avoid the structural re-analysis phenomenon between the two algorithms.The structure design of a high pressure bypass-valve body is used as an example to make a structural optimization by the proposed HGA and finite element method（FEM）,respectively.The comparison result shows that the improved HGA fully considers the characteristic of discrete variable and information of working conditions,and is more suitable to the optimal problems with complex working conditions.Meanwhile,the research provides a new approach for discrete variable structure optimization problems.展开更多
文摘目标检测作为计算机视觉的重要研究方向,在智慧城市、无人驾驶等领域的作用越来越重要。传统目标检测算法中,根据交并比(Intersection over Union,IOU)的大小判断正负样本,但较低的IOU会引入噪声,降低检测器的精度;较高的IOU会保留少数高质量样本,造成过拟合;并且推荐区域和检测器的IOU阈值相差过大会引起质量不匹配问题。针对上述问题,提出了一种基于级联网络的平行级联检测网络,它由一系列检测器串并联而成,每个检测器设置递增的IOU阈值,从而在每个阶段都会得到一个更高质量的样本分布来训练下一级检测器,并逐步重采样减少过拟合。实验结果表明提出的平行级联检测网络的检测精度优于传统目标检测算法,在目标检测数据集Microsoft COCO上平均准确度(AP)提升了1.5个百分点左右。
基金supported by Key Program for International S&T Cooperation Projects of China (Grant No. 2009DFA71860)Program for New Century Excellent Talents in Heilongjiang Provincial University of China(Grant No. 1153-NCET-005)
文摘Round method is the common method for discrete variable optimization in optimal design of complex mechanical structures;however,it has some disadvantages such as poor precision,simple model and lacking of working conditions＇ description,etc.To solve these problems,a new model is constructed by defining parameterized fuzzy entropy,and the rationality of parameterized fuzzy entropy is verified.And a new multidirectional searching algorithm is further put forward,which takes information of actual working conditions into consideration and has a powerful local searching capability.Then this new algorithm is combined with the GA by the fuzzy clustering algorithm（FCA）.With the application of FCA,the optimal solution can be effectively filtered so as to retain the diversity and the elite of the optimal solution,and avoid the structural re-analysis phenomenon between the two algorithms.The structure design of a high pressure bypass-valve body is used as an example to make a structural optimization by the proposed HGA and finite element method（FEM）,respectively.The comparison result shows that the improved HGA fully considers the characteristic of discrete variable and information of working conditions,and is more suitable to the optimal problems with complex working conditions.Meanwhile,the research provides a new approach for discrete variable structure optimization problems.
