基于多特征融合的图像区域几何标记

Geometric Labeling of Image Regions Based on Combination of Multiple Features

提出一种基于多特征融合的图像区域几何标记方法.首先,提出了一种新型卷积网络结构——多尺度核卷积网络用于提取像素点的多尺度特征信息,推断像素点的几何类别,并结合图像超像素分割获得图像超像素区域的几何标记;其次,将提取的多尺度特征与超像素区域传统特征相结合,建立超像素区域的特征表达.最后,建立超像素图像的条件随机场(conditional random field,CRF)模型,对超像素区域的几何类别进行推断.在公开数据集Geometric Context(GC)上的实验结果表明,同已有算法相比,所提方法提高了图像区域几何标记的准确率.

A geometric labeling method of image regions was proposed based on combination of multiple features. First of all, according to the requirement of multi-scale feature information extraction,a novel network structure-multi-scale kernel convolutional network （MSKCN） was proposed. The multi-scale feature information was used for inferring geometric label of pixel. The geometric labeling of super-pixel regions with the image super-pixel segmentation was achieved. Then a feature representation of super-pixel regions was established by combining multi-scale features proposed and traditional features of super-pixel regions. Finally, a CRF （ conditional random field） model was constructed for the super-pixel image to infer geometric label of super-pixel regions with the image super-pixel segmentation. The experiments on public database Geometric Context （GC） indicated that the accuracy of geometric labeling was improved by using the proposed method compared with the existing state- of- art.