基于结构森林的RGB-D图像轮廓提取-计算机应用研究.PDF

优先出版 计 算 机 应 用 研 究 第34 卷 基于结构森林的RGB-D 图像轮廓提取* 张 恒，徐 刚，刘艳丽，温珍强 (华东交通大学 信息工程学院，南昌 330013) 摘 要：为了获取更加清晰、更多细节的轮廓特征，充分利用Kinect 传感器获取的RGB-D 图像信息，将结构化的随机 森林作为分类器，提出一种更加精确的轮廓提取器。首先，将RGB-D 图像的多种信息利用数学式表示出来；然后，利 用BSD500 数据集以及NYU 深度数据集训练结构化的随机森林算法，核心是将给定节点的结构化标签映射到一组离散 标签；最后，利用该随机森林算法对RGB-D 图像信息进行分类，得到图像轮廓。针对细节不同的四种场景图像进行对 比实验，结果表明，经本文改进后的算法得到的轮廓效果更加清晰、准确。 关键词：轮廓提取；深度信息；结构森林；RGB-D 图像 中图分类号：TP391.4 Contour extraction from RGB-D image based on structured forests Zhang Heng, Xu Gang, Liu Yanli, Wen Zhenqiang (School of Information Engineering, East China Jiaotong University, Nanchang 330013, China) Abstract: In order to obtain more clear and more detailed contour features, and fully use the RGB-D image information from Kinect, this paper proposed a more precise contour extractor using structured random for ests as a classifier. First, the various information from RGB-D image was expressed with mathematical formulas . Then, the BSD500 Segmentation dataset and NYU Depth dataset were used to train the structured random forests . The core idea is to map all the structured labels at a given node into a discrete set of labels. Finally, used the random forests to classify the RGB-D image information and get the image contours. In the experiments, four images with different details were taken. The results of the experiments demonstrate significant improvements of our improved algorithm over the state-of-the-art. Key Words: contour extraction; depth information; structured forests; RGB-D image [7]