基于物体间支撑语义关系的室内场景彩色深度图像分割

Indoor red green blue-depth segmentation based on object-object supportive semantic relationships

中文关键词  自主机器人  场景理解  图像分割  先验知识  支撑语义关系

英文关键词  autonomous robots  scene understanding  image segmentation  prior knowledge  supportive semantic relationship

中文摘要

      场景理解是智能自主机器人领域的一个重要研究方向, 而图像分割是场景理解的基础. 但是, 不完备的训练数据集, 以及真实环境中的罕见情形, 会导致在图像分割时存在先验知识不完备的情况, 进而影响图像分割的效果. 因此, 提出在RGB-D图像上使用抽象的支撑语义关系来解决多样的物体形态所面对的先验知识不完备问题. 在先验知识不完备情况下, 针对自底向上的图像分割过程中被过度分割出的物体块, 首先对物体块间的支撑语义关系进行建模并计算其支撑概率, 然后构造能够度量场景总体稳定性的能量函数, 最后通过SWC(Swendsen-Wang cuts)随机图分割算法最小化该能量函数的值, 将物体块间的支撑概率转化为强支撑语义关系并完成物体块合并, 实现先验知识不完备情况下的图像分割. 实验结果证明, 结合支撑语义关系的图像分割能够在先验知识不完备的情况下, 将同一物体被过度分割的部分重新合并起来, 从而提升了图像分割的准确性.

英文摘要

      One of the most important research fields for intelligent autonomous robots is scene understanding which requires a foundation of image segmentation. However, it usually faces a condition of incomplete prior knowledge due to incomplete training dataset and uncommon situations in the real world, which affects the segmentation quality. Therefore, a method that solves the incomplete prior knowledge problem caused by the diversity of objects is presented, which is based on extracting abstract supportive semantic relationships in RGB-D images. This method aims at the over-segmented object parts during the bottom-up image segmentation process with incomplete prior knowledge. First, the supportive semantic relationships between object parts are modelled by calculating supportive probabilities. Second, an energy function is built to measure the total stability of the entire scene. Last, SWC(Swendsen-Wang cuts) algorithm is implemented to minimize the energy function so that the supportive probabilities between object parts are converted into strong supportive semantic relationships. According to the relationships, the object parts are merged to accomplish image segmentation with incomplete prior knowledge. The experiments show that the over-segmented parts are merged into one single object with incomplete prior knowledge by simply introducing supportive semantic relationships to image segmentation. With our method more precise image segmentation results are achieved.