Ground-To-Aerial Cross- Vi Ew Image Geolocalization and Semantic Transfer by Deep Learning and Transfer Learning

View Statistics Email Alert RSS Feed

Information

Details

Project title

Code/計畫編號

MOST106-2221-E019-050

Translated Name/計畫中文名

應用深度學習與遷移學習於地面至空拍跨視角影像地理定位與跨視角高階語意遷移技術

Project Coordinator/計畫主持人

Chin-Chun Chang

Funding Organization/主管機關

National Science and Technology Council

Department/Unit

Department of Computer Science and Engineering

Website

https://www.grb.gov.tw/search/planDetail?id=12488032

Year

2017

Start date/計畫起

01-08-2017

Expected Completion/計畫迄

31-07-2018

Bugetid/研究經費

617千元

ResearchField/研究領域

資訊科學--軟體

"本計畫擬發展地面至空拍(ground to aerial)跨視角地理定位(cross-view image geolocalization) 技術。此技術藉由事先學習之具GPS 標示的空拍影像與地面視角影像間的關係，來定位無GPS 標示的地面視角查詢影像其可能的地理位置。雖然這種跨視角影像地理定位技術不是著重於精確地理定位，但是對於應用地面視角影像來設定無人機搜索範圍，與分析地理高階語意標籤分布和變遷非常地重要。由於這種跨視角影像對映關係非常複雜，因此本計畫擬應用深度學習(deep learning)與遷移學習(transfer learning)於開發地面至空照跨視角影像地理定位技術。在此三年計畫，我們預計發展下面技術。  第一年研究主題:發展地面至空拍跨視角影像地理定位技術我們將從社群與地圖搜尋網站(如Google Street Views, Google Street Views, Flickr)裡自動蒐集具有 GPS 標示、豐富語意資訊的街景視角影像來建立具GPS標示的空拍影像與地面視角影像間的關係。並將應用深度學習和遷移學習來學習跨視角影像對映關係，以建立地面至空拍跨視角影像地理定位技術。  第二年研究主題:發展應用地面至空拍跨視角影像地理定位技術之跨視角地理高階語意遷移技術第二年，我們將應用空拍影像間非常強烈的地理空間關係於提升跨視角影像地理定位的準度。並且，我們也將從街景視角影像所得到的語意分析資訊，藉由其所在地的空拍影像，自動找到有類似空拍影像的區域，並推論出它們也是可能具有類似語意的區域，辦到地理高階語意遷移。  第三年研究主題:應用空拍跨視角影像地理定位技術之高階語意地理標籤分布與變遷分析技術第三年，我們將藉由地面至空拍跨視角影像地理定位技術，與應用蒐集到的空拍及地面視角影像，對實驗地理範圍內的區域，作高階語意地理標籤分布與變遷分析。""This project aims at developing the core technology of ground-to-aerial cross-view image geolocalization. By the relation between geo-tagged aerial images and ground-level images, this technology can be used to localize ground-level query images without GPS information. Although this technology cannot provide precise geolocalization, this technology can be used to specify patrol zones for auto-piloting drones by ground-level query images. Besides, this technology is useful for analyzing the geospatial distribution and distribution change of high-level semantic meanings extracted from aerial and ground-level images. Because the relationship between aerial images and ground-level images is complicated, deep learning and transfer learning will be used to learn this cross-view relationship. In this three-year project, we shall develop the following technology.  In the first year, a system will be developed to automatically collect geo-tagged aerial and ground-level images from photo-sharing websites and social networks, such as Google Street Views, Flickr, and Bing Map. We shall develop the core technology of ground-to-aerial cross-view image geolocalization by using deep learning and transfer learning to learn the ground-to-aerial cross-view relation.  In the second year, we shall use the spatial relation among aerial images to improve the accuracy of ground-to-aerial cross-view image geolocalization. Additionally, we shall transfer the semantic meanings of ground-level images to aerial images by the learned ground-to-aerial cross-view model.  In the third year, we shall apply the technology of cross-view image geolocalization and semantic transfer to analyze the geospatial distribution and distribution change of high-level semantic meanings extracted from aerial and ground-level images sampled around a city, such as Keelung City."

Keyword(s)

智慧物聯網
深度學習
遷移式學習
跨視角影像地理定位
smart internet of things
deep learning
transfer learning
cross-view image geolocalization

DSpace CRIS

Ground-To-Aerial Cross- Vi Ew Image Geolocalization and Semantic Transfer by Deep Learning and Transfer Learning

Details

Description