基于图像特征分析的物体轮廓提取

Object contour extraction based on image feature analysis

对物体的轮廓进行分析提取,是计算机视觉方向的基础问题之一,对其进行研究对于复杂场景的分析理解至关重要。本文对室内场景图像进行研究,基于图像特征进行图像分割,提取物体轮廓。在彩色场景图像全局轮廓后验边界概率(g Pb)提取算法的基础上,加入深度图像信息,对室内场景的彩色、深度(RGB-D)图像中的物体轮廓进行分析。通过多尺度信息融合,计算得到多尺度轮廓后验概率(m Pb)和谱后验概率(s Pb),两后验概率加权综合得到g Pb。而后结合超度量轮廓图与分水岭算法,对基于方向特征变化的g Pb图像融合处理,最终得到清晰的物体轮廓。本文所提方法在通用的RGB-D数据库基础上进行实验。实验结果表明,本文所提出的方法能提取出清晰的室内物体轮廓图。

Contour analysis and extraction is the fundamental problem in computer vision, and the research about it plays an important part in complex scene analysis and comprehension. In this paper, an algorithm for analyzing indoor scene images is studied. Based on the image features extracted from the images, the objects in the indoor scenes are segmented, and further the contours of the objects are extracted. Based on the globalized posterior probability of a boundary （gPb） method for the contour extraction on the RGB image, we introduce the depth information to enhance the performance of contour extraction on RGB-D data of indoor scenes. By combining muhi-seale cues, the multi-scale posterior probability （mPb） and spectral posterior probability （sPb） are obtained. The mPb and sPb results are summed and weighted to get the gPb information. Then, the gPb information is processed by ultrametrie contour and watershed algorithm, and the contours of the indoor scene objects are gained. The experiments presented in this paper are run on the general RGB-D dataset. The experimental results show that our method can extract the distinct contours of indoor objects.