查询结果:   张永库,杜帅川,孙劲光,周积林,金雅茹.基于多尺度特征学习的阴影检测[J].计算机应用与软件,2016,33(5):185 - 188.203.
中文标题
基于多尺度特征学习的阴影检测
发表栏目
图像处理与应用
摘要点击数
752
英文标题
SHADOW DETECTION BASED ON MULTI-SCALE FEATURE LEARNING
作 者
张永库 杜帅川 孙劲光 周积林 金雅茹 Zhang Yongku Du Shuaichuan Sun Jinguang Zhou Jilin Jin Yaru
作者单位
辽宁工程技术大学电子与信息工程学院 辽宁 辽宁工程技术大学研究生学院 辽宁 葫芦岛 125105 山东中医药大学护理学院 山东 济南 250000   
英文单位
School of Electronics and Information Engineering,Liaoning Technical University,Huludao 125105,Liaoning,China Institute of Graduate,Liaoning Technical University,Huludao 125105,Liaoning,China School of Nursing,Shandong University of Traditional Chinese Medicine,Jinan 250000,Shandong,China   
关键词
阴影检测 卷积神经网络 特征学习 条件随机场
Keywords
Shadow detection Convolutional neural network Feature learning Conditional random field
基金项目
国家科技支撑计划项目(2013bah12f01)
作者资料
张永库,副教授,主研领域:图形图像处理和多媒体,数据处理和数据挖掘。杜帅川,硕士生。孙劲光,教授。周积林,硕士生。金雅茹,本科。 。
文章摘要
针对传统阴影检测方法存在精心设计特征、训练时间长与阴影检出率低等问题,提出一种有监督学习的阴影检测方法。首先输入的图像经过拉普拉斯金字塔变换,确定聚类中心,分别以聚类中心为中心进行窗口提取;然后合成训练样本,训练样本在卷积神经网络中进行训练得到后验分布;最后将得到的后验分布反馈给条件随机场生成有标签的图像。实验结果表明,该方法有较好的场景适应性、训练时间短并且有较高的阴影检出率。
Abstract
Traditional shadow detection methods need careful hand-crafted features design and long training time. Specially, these methods have lower detection rate as well. In order to solve these problems, in this paper we propose a supervised learning method for shadow detection. Firstly, the inputted images are transformed through Laplacian pyramid to determine the clustering centres, and these clustering centres are then taken as the centres for extracting the windows separately. Secondly, the method synthesises the training samples, and trains these samples in convolutional neural network to generate the posterior distribution. Finally, it feeds the derived posterior distribution back to the conditional random field to generate the labelled image. Experimental results show that this method works well in different scenes, the training time is short and the shadow detection rate is high.
下载PDF全文