卷积自编码器中粗粒度池化特征提取研究
发布时间：2018-09-05 10:33
【摘要】：卷积自编码器(Convolutional Auto Encoder,CAE)提取的粗粒度池化特征具有一定范围内旋转和平移的不变性,因而得到广泛使用.然而,目前CAE仍主要依靠经验调节内部参数以获取满足要求的粗粒度池化特征.本文将CAE看作一个整体,从概率上分析了影响其表现的具体原因,构建了一个通用框架用于调节其中的主要参数以获取更好的粗粒度特征.首先从概率上权衡了粗粒度特征在池化层上的判别性与不变性,并在CAE中选择合适的卷积范围和白化参数.然后通过分析池化域内特征的稀疏度选择相应的池化方法以获取具有更好可分离性的粗粒度池化特征.在两个公开数据库(STL-10和CIFAR-10)的实验结果表明本文提出的方法可以指导CAE提取到更好的粗粒度池化特征并在多类分类任务中表现得更好.
[Abstract]:Coarse-grained pool feature extracted by convolution self-encoder (Convolutional Auto Encoder,CAE) has the invariance of rotation and translation in a certain range, so it is widely used. However, at present, CAE still mainly relies on the experience to adjust the internal parameters to obtain the coarse-grained pool characteristics that meet the requirements. In this paper, CAE is regarded as a whole, and the specific reasons that affect its performance are analyzed in terms of probability, and a general framework is constructed to adjust the main parameters in order to obtain better coarse-grained features. In this paper, we first weigh the discrimination and invariance of coarse-grained features on the pool layer from the point of view of probability, and select the appropriate convolution range and whitening parameters in CAE. Then, by analyzing the sparsity of the feature in the pool domain, the corresponding pool method is selected to obtain the coarse-grained pool feature with better separability. Experimental results on two open databases (STL-10 and CIFAR-10) show that the proposed method can guide CAE to extract better coarse-grained pooled features and perform better in multi-class classification tasks.
【作者单位】： 空军工程大学防空反导学院;94691部队;
【基金】：国家自然科学基金(No.71501184)
【分类号】：TP181;TP391.41
本文编号：2224006
[Abstract]:Coarse-grained pool feature extracted by convolution self-encoder (Convolutional Auto Encoder,CAE) has the invariance of rotation and translation in a certain range, so it is widely used. However, at present, CAE still mainly relies on the experience to adjust the internal parameters to obtain the coarse-grained pool characteristics that meet the requirements. In this paper, CAE is regarded as a whole, and the specific reasons that affect its performance are analyzed in terms of probability, and a general framework is constructed to adjust the main parameters in order to obtain better coarse-grained features. In this paper, we first weigh the discrimination and invariance of coarse-grained features on the pool layer from the point of view of probability, and select the appropriate convolution range and whitening parameters in CAE. Then, by analyzing the sparsity of the feature in the pool domain, the corresponding pool method is selected to obtain the coarse-grained pool feature with better separability. Experimental results on two open databases (STL-10 and CIFAR-10) show that the proposed method can guide CAE to extract better coarse-grained pooled features and perform better in multi-class classification tasks.
【作者单位】： 空军工程大学防空反导学院;94691部队;
【基金】：国家自然科学基金(No.71501184)
【分类号】：TP181;TP391.41
【相似文献】
相关期刊论文 前10条
1 杨子煜;赵鹏;王大伟;李思昆;;关键循环到粗粒度可重构体系结构的存储感知映射[J];国防科技大学学报;2012年06期
2 黄雅菁;高建华;;基于粗粒度切片度量的重构定位方法[J];计算机工程;2011年11期
3 屈志毅;吴换霞;刘瑜;;粗粒度多尺度属性约简启发式算法[J];昆明理工大学学报(理工版);2007年04期
4 柴玉梅;刘东昊;王黎明;;对象级粗粒度切片方法[J];计算机工程与设计;2012年03期
5 宋立国,姜玉宪;CTaiJi——新的粗粒度可配置计算结构研究[J];微电子学与计算机;2005年09期
6 李常青;唐世渭;李红燕;;基于关联分析的粗粒度级个性化信息挖掘[J];计算机科学;2002年01期
7 孟伟,杨耀忠;油藏数值模拟软件的粗粒度层并行计算方法[J];计算机应用与软件;2004年08期
8 徐佳庆;邬贵明;窦勇;;二维DCT在粗粒度可重构处理器上的实现[J];计算机工程;2008年20期
9 肖建;刘波;梅晨;朱敏;杨军;刘雷波;魏少军;;粗粒度部分动态可重构的人脸检测[J];应用科学学报;2012年03期
10 王华;EJB环境下远程方法粗粒度调用的实现[J];计算机工程与科学;2004年07期
相关硕士学位论文 前1条
1 杨小溪;多层次程序周期行为分析[D];复旦大学;2011年
，本文编号：2224006
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2224006.html
