图像深度层次特征提取算法  

Deep Hierarchical Feature Extraction Algorithm

作　　者：李钦[1] 游雄[1] 李科[1] 汤奋[1] 

机构地区：[1]郑州测绘学院地理信息工程系,郑州450000

出　　处：《模式识别与人工智能》2017年第2期127-136,共10页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.41201390);河南省科技创新(中原学者)项目(No.142101510005)资助~~

摘　　要：完成众多视觉任务的关键是提取具有较强表达能力的图像特征,传统的图像特征仅描述图像某一方面的信息,表达能力受到很大限制.文中基于卷积神经网络提出图像深度层次特征(DHF)提取算法,通过对图像的层层抽象表达,可以有效挖掘隐藏在图像内部的本质信息.首先基于卷积神经网络产生图像特征图,选取卷积输出层的特征图构建图像阶层结构.然后基于匹配实验选择最佳的层级组合,采用信息熵描述低层级特征图,采用区域平均的方法描述高层级特征图,最终构建具有较强表达能力的DHF特征.实验表明,相比已有特征,DHF特征优势明显,可以高效准确地完成图像匹配任务.Extracting the image features with strong representation is critical to complete different vision tasks. The traditional features only describe one aspect of the image information, and therefore their representation capability is limited. In this paper, deep hierarchical feature （DHF） extraction algorithm based on the convolutional neural networks （CNN） is proposed. The essential information hidden inside the image is effectively mined by abstractly expressing the image in different layers. Firstly, the feature maps of the image are created based on CNN, and those in the convolutional layers are selected to construct the hierarchical structure of the image. Then, the best layer combination is determined according to the matching experiment. The feature maps in the low layers are described by the information entropy, and the ones in high layers are described by averaging the pixels in specified region, the DHF with strong representative capability is ultimately constructed. The experiment demonstrates that the proposed DHF has evident advantages compared with the existing features, and it can complete the matching task with high efficiency.

关 键 词：深度层次特征(DHF) 卷积神经网络(CNN) 特征图 特征表达能力 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]