基于分形和分理论的分形池化算法
Fractal pooling algorithm based on fractal sum theory
肖莎莎 1高哲 2贾凯 1焦芷媛 1柴浩宇1
作者信息
- 1. 辽宁大学数学与统计学院,辽宁沈阳 110036
- 2. 辽宁大学数学与统计学院,辽宁沈阳 110036;辽宁大学轻型产业学院,辽宁沈阳 110036
- 折叠
摘要
传统池化操作既不能客观地评价池化区域中数据之间的差异性,也不能有效地保留池化区域中鉴别性特征.为解决这类问题,提出了一种基于分形和分理论,且能够根据每个特征图各通道中数据间的差异性,自行地选择最优池化策略的分形池化算法.首先,引入分形和分的定义,构造分形池化算子和训练误差的反向传播算法.该算子不仅包括最大池化、平均池化,还能够降低训练误差.然后,在算法实现的过程中,根据每个特征图各通道中数据间的差异性自适应地整定阶次,以确定池化区域中每个数据的训练权重.最后,在不同数据集和不同架构上进行了大量分类性能实验,验证了所提出的方法比传统池化方法和混合池化都取得了更好的分类效果.
Abstract
Traditional pooling operations can neither objectively evaluate the differences among data in the pooled region nor effectively retain discriminative features in the pooled region.To solve these problems,a fractal pooling algorithm based on fractal sum theory is proposed,which can choose the optimal pooling strategy according to the variability among data in each channel of each feature map.Firstly,the fractal pooling operator and the back-propagation algorithm of training error are constructed by introducing the definition of fractal sum.The operator not only includes the max pooling and the average pooling,but also can reduce the training error.Then,during the implementation of the algorithm,the order is adaptively adjusted based on the differences between data in each channel of each feature map to determine the training weights for each data in the pooled region.Finally,a large number of classification performance experiments are carried out on different datasets and different architectures to verify that the proposed method achieves better classification results than traditional pooling methods and the mixed pooling.
关键词
分形和分/最大池化/平均池化/分形池化/分类Key words
fractal sum/max pooling/average pooling/fractal pooling/classification引用本文复制引用
基金项目
辽宁省教育厅科研基金(LJC2010)
辽宁省自然科学基金(20180520009)
沈阳市中青年科技创新人才支持计划(RC210082)
出版年
2024
NETL
NSTL
万方数据