在反卷积网络中引入数值解可视化卷积神经网络
发布时间：2018-11-25 11:29
【摘要】：经典的反卷积可视化模型通过反池化、反激活、反卷积将特征图像还原至原图像空间,可视化网络节点从输入图像学习到的特征,有助于探究卷积神经网络运行良好的机制,但是由于采用近似处理,还原特征不明显。本研究引入数值求解方法来代替原模型中直接用卷积核的反转近似反卷积核的方法。先构造数据集:随机生成大小、形状、位置不一的结构简单、角点特征明显的三角形和矩形,用于组成层次结构逐渐复杂的数据集,并利用该数据集测试模型的可视化效果。实验表明,改进后的可视化模型提取的特征更明显,引入的噪音更少,可以更为精确地将激活网络节点从原图像学习的特征可视化。在更大的数据库上进行实验来验证结果,并利用这种结果进一步探究准确率与网络结构之间存在何种关系。
[Abstract]:The classical deconvolution visualization model restores the feature image to the original image space through deconvolution, deactivation and deconvolution, and visualizes the features that the network node learns from the input image, which is helpful to explore the mechanism of the convolution neural network running well. However, due to the approximate treatment, the reduction characteristics are not obvious. In this paper, numerical method is introduced to replace the inverse approximate deconvolution kernel of convolution kernel in the original model. First, construct the data set: the simple structure with different size, shape and position, and the triangle and rectangle with obvious corner feature, are used to form the gradually complicated data set with hierarchical structure, and the visualization effect of the model is tested by using the data set. The experimental results show that the improved visual model can extract more obvious features and introduce less noise, which can more accurately visualize the features of active network nodes from the original image. Experiments are carried out on a larger database to verify the results, and the results are used to further explore the relationship between accuracy and network structure.
【作者单位】： 中山大学数学学院;
【基金】：国家自然科学基金(61272338)资助
【分类号】：TP183
本文编号：2355947
[Abstract]:The classical deconvolution visualization model restores the feature image to the original image space through deconvolution, deactivation and deconvolution, and visualizes the features that the network node learns from the input image, which is helpful to explore the mechanism of the convolution neural network running well. However, due to the approximate treatment, the reduction characteristics are not obvious. In this paper, numerical method is introduced to replace the inverse approximate deconvolution kernel of convolution kernel in the original model. First, construct the data set: the simple structure with different size, shape and position, and the triangle and rectangle with obvious corner feature, are used to form the gradually complicated data set with hierarchical structure, and the visualization effect of the model is tested by using the data set. The experimental results show that the improved visual model can extract more obvious features and introduce less noise, which can more accurately visualize the features of active network nodes from the original image. Experiments are carried out on a larger database to verify the results, and the results are used to further explore the relationship between accuracy and network structure.
【作者单位】： 中山大学数学学院;
【基金】：国家自然科学基金(61272338)资助
【分类号】：TP183
【相似文献】
相关期刊论文 前8条
1 张新家,戴冠中,卢满宏;基于反卷积的系统跳变故障的检测方法[J];信息与控制;1991年02期
2 沈晓安;;基于支持向量回归机的反卷积技术[J];机床与液压;2009年03期
3 王世刚;孙小君;邓自立;;多模型多传感器信息融合稳态白噪声反卷积估值器[J];科学技术与工程;2008年02期
4 张三宝,刘文江,聂泽昆;固定区间最小方差反卷积及其Systolic阵列实现[J];西安交通大学学报;1999年11期
5 尤富强,田作华,施颂椒;线性时滞系统的H_∞反卷积设计[J];系统仿真学报;2005年07期
6 李孝文,唐志刚;利用反卷积改善非线性传感器的性能[J];传感技术学报;1995年04期
7 王欣;李云;邓自立;;基于Kalman滤波方法的多传感器信息融合最优白噪声反卷积滤波器[J];黑龙江大学自然科学学报;2006年02期
8 ;[J];;年期
相关博士学位论文 前1条
1 孙小君;最优和自校正多传感器信息融合白噪声反卷积估值器[D];黑龙江大学;2010年
相关硕士学位论文 前2条
1 王佳伟;统一和通用的信息融合白噪声反卷积估值器[D];黑龙江大学;2007年
2 孙赫;基于现代时间序列分析方法的信息融合反卷积滤波器[D];黑龙江大学;2013年
，本文编号：2355947
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2355947.html
