摘要
为利用高阶条件随机场有效标注室内场景,文中提出一种稀疏字典驱动高阶依赖的RGB-D颜色-深度图像语义分割法。首先,利用融合深度的多尺度组合成组的全局概率边缘超度量图分层法过分割彩色-深度图像。然后,提取场景中各个超像素区域的视觉特征,构建超像素标签池并用于训练支持向量机分类器。接着,计算超像素一元势能和相邻超像素成对项势能;同时,以每一类超像素区域内关键点特征的稀疏编码子之和的直方图统计作为高阶势能。最后,利用融合自顶向下的判别性类别成本的条件随机场模型推理实现语义标注。实验表明,与其他方法相比,该方法能得到视觉表现力更强、准确率更高的语义标签图。
A sparse dictionary-driven high-order dependent RGB-D（ color-depth） image semantic segmentation method is proposed to annotate the given indoor scene using high-order conditional random field. Firstly,the global probability of boundary-ultrametric contour map is exploited using depth-fused multi-scale combinatorial grouping for hierarchical over-segmentation of the given RGB-D scene. Secondly,the regional visual feature of each super pixel is extracted to build super pixel label pool to train support vector machine classifier. Then,the unary potential energy of each super pixel and the pairwise potential energy between the adjacent super pixels are calculated,while accumulating the statistical histograms of the sparse code of keypoint features in each super pixel for each class as high-order potential energy.Finally,the semantic segmentation is implemented by exploiting the conditional random field model inference with the top-down discriminative category cost. Compared with other state-of-the-art methods,the presented method can obtain semantic label map with stronger visual expression and higher accuracy.
作者
刘天亮
徐高帮
戴修斌
曹旦旦
罗杰波
LIU Tianliang XU Gaobang DAI Xiubin CAO Dandan LUO Jiebo(Jiangsu Province Key Lab on Image Processing ＆ Image Communication, Nanjing University of Posts and Telecommunications, Nanjing 210003, China Department of Computer Science, University of Rochester, Rochester NY 14627, USA)
出处
《南京邮电大学学报（自然科学版）》
北大核心
2017年第5期13-18,共6页
Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition
基金
国家自然科学基金(61001152
31200747
61071091
61071166
61172118)
江苏省自然科学基金(BK2012437)
国家留学基金
南京邮电大学校级科研基金(NY214037)资助项目
关键词
语义分割
条件随机场模型
稀疏字典学习
结构化支持向量机
semantic segmentation
conditional random field models
sparse dictionary learning
structural support vector machine
