当前位置: X-MOL 学术Future Gener. Comput. Syst. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
City space type recognition by leveraging deeply-learned topological features
Future Generation Computer Systems ( IF 6.2 ) Pub Date : 2020-06-13 , DOI: 10.1016/j.future.2020.06.004
Sufang Liu , Lei Shi , Simian Liu , Chuan Wang , Chulin Chen

Accurately recognizing the rich variety of geometric spaces with different functions is an indispensable but difficult task in modern artificial intelligence systems. In this work, we propose a novel city geometric space classification system, which is inspired by human visual perception toward different geometries within each city. First, we extract multiple subgraphs from different sub-regions inside each city’s aerial photo. We call these subgraphs graphlets and they can optimally capture the geometric feature inside each city. Subsequently, it is observable that there are an exponential number of graphlets within each region. And thereby we designed a ranking algorithm to select a few highly discriminative graphlets for city space recognition. Such selection process follows human visual attention, where only a small proportion of visually/semantically salient regions are selected for human visual cognition. Based on the selected graphlets from each city subspace, we integrate them into an image kernel for city subspace recognition. Comprehensive experimental results as well as several visualization results have demonstrated the efficiency and effectiveness of our method.



中文翻译:

利用深度学习的拓扑特征识别城市空间类型

准确地认识具有不同功能的丰富几何空间是现代人工智能系统中必不可少但艰巨的任务。在这项工作中,我们提出了一种新颖的城市几何空间分类系统,该系统受人类对每个城市内不同几何形状的视觉感知的启发。首先,我们从每个城市的航拍照片中的不同子区域提取多个子图。我们称这些子图为图小图,它们可以最佳地捕获每个城市内部的几何特征。随后,可以观察到在每个区域内都有指数数量的graphlet。因此,我们设计了一种排序算法,以选择一些具有高度区分性的小图进行城市空间识别。这样的选择过程遵循人类的视觉注意力,仅选择一小部分视觉/语义显着区域用于人类的视觉认知。基于从每个城市子空间中选择的图小图,我们将它们集成到用于城市子空间识别的图像内核中。全面的实验结果以及一些可视化结果证明了我们方法的有效性和有效性。

更新日期:2020-06-13
down
wechat
bug