【作者】 胡泊； 王勇； 邹逸群；

【Author】 Po Hu;Yong Wang;Yiqun Zou;School of Automation, Central South University;

【机构】 中南大学自动化学院；

【摘要】 透明物体图像分割在机器人导航、智慧家居等领域有着广泛的应用。由于透明物体的外观容易受到背景、光照等因素的影响,现有的深度学习算法在分割过程存在语义信息难提取、边缘分割不准确等问题。针对上述问题,本文提出了一种基于差分边界注意力和区域注意力的透明物体分割算法。首先,提出了一个差分边界注意力模块,它可以提取并融合多尺度的边缘特征,以得到精准的边缘分割图。其次,提出一个区域注意力模块,对高分辨率特征和低分辨率特征进行类别层面的上下文关系建模,显式地增强了透明物体区域的语义信息。我们收集并整理了一个高质量的透明物体分割数据集,并在该数据集上对比了本文提出的算法和目前主流的分割算法。实验结果表明,本文提出的算法具有比现有分割算法更好的性能。更多还原

【Abstract】 Transparent object image segmentation is widely used in robot navigation, smart home, and other fields. As the appearance of transparent objects is easily affected by background, illumination and other factors,existing deep learning algorithms have problems such as difficult semantic information extraction and inaccurate edge segmentation results. To solve the above problems, this paper proposes a transparent object segmentation algorithm based on differential boundary attention and regional attention. First, a differential boundary attention module is proposed, which can extract and fuse multi-scale edge features to obtain accurate edge segmentation images. Then, a regional attention module is proposed to model the context relationship between high-resolution features and low-resolution features at the category level, which explicitly enhances the semantic information of transparent object regions. We have collected and organized a high-quality transparent object segmentation dataset, and compared the proposed algorithm with current mainstream segmentation algorithms on this dataset.Experimental results demonstrate that the proposed algorithm performs better than existing segmentation algorithms.更多还原