Academic journal article Cartography and Geographic Information Science

An Automated System for Image-to-Vector Georeferencing

Academic journal article Cartography and Geographic Information Science

An Automated System for Image-to-Vector Georeferencing

Article excerpt

Introduction

For modern imagery, true world location (such as latitude and longitude coordinates) is usually captured at the source and integrated with the image. Nevertheless, much potentially useful imagery lacks location information such as historic imagery, scanned aerial photographs and maps, and construction blueprints (often termed "as builts"). Additionally, locational information for modern imagery may be lost, as often occurs when data are downloaded from websites that fail to transmit locational information, or when images are obtained by capturing screenshots.

Manual georeferencing, a standard feature of most current GIS commercial software packages such as ArcGIS (ESRI 2011), can provide this missing location information. The procedure involves the creation by the user of a set of control point pairs that link pixel locations on the raw image with corresponding geospatial locations on the reference target, usually but not necessarily a vector map. Unfortunately, this remains largely a manual process that is time-consuming, tedious and subject to user error (Jensen 2005: 236). Most critically, locations may be impossible to find without prior information on the image's approximate location. Consider the example used in this paper of locating images covering small urban neighborhoods of less than one square mile within an urban county (i.e., Dallas County, Texas) covering 871 square miles. This task would be impossible manually unless handled by a specialist who is familiar with both the scene in the image and the whole Dallas County area, and is also lucky or skilled enough to recognize the scene depicted in the imagery even though it may be distorted.

Automated solutions for acquiring missing or imprecise locational information for images have been proposed (Brown 1992; Cox and De Jager 1993; Wood 1996; Zitova and Flusser 2003) and are available in some current software packages such as IMAGINE AutoSynch (ERDAS 2011). Most of these existing algorithms involve image-to-image registration whereby an image lacking location information is matched with a reference image of approximately the '"same" scene (Brown 1992; Araiza et al. 2002; Rao et al. 2004; Navy et al. 2006; Yu et al. 2008). Their time complexity is usually high: O([N.sup.2]) (Navy et al. 2006), O([N.sup.3]) (Ton and Jain 1989) or higher (N is the number of extracted points), and therefore not scalable when the size of a point dataset is increased significantly. Assuming datasets are of the same geographical area effectively limits the size of the searching space and hence reduces problem complexity. However it significantly hinders the general applicability of the automated solution since the same scene (that is, approximate geographic location) is often not known for image-to-vector georeferencing. The absence of this information necessitates a far smarter matching algorithm, able to efficiently search an arbitrarily larger vector space to identify the location for the image. Approaches such as Stockman (1982), Ranade and Rosenfeld (1980), Ton and Jain (1989), Seedahmed and Martucci (2004), and Cordon and Damas (2006) fail this efficient search requirement. Other image-to-image registration methods use cross-correlation, SIFT, or Fourier methods to exploit image intensities/gradients for matching (Araiza et al. 2002; Xie et al. 2003; Kim and Fessler 2004; Li and Leung 2004; Zografos and Buxton 2005; Yu et al. 2008; Kim et al. 2010). These cannot be applied to the image-to-vector georeferencing problem because there is no image intensity/gradient information in the reference vector dataset. Consequently, georeferencing of images to a vector base still remains largely a manual process, relying on humans to achieve a match by identifying common locations in the image and the reference map.

As a substitute for this manual process and its inherent potential errors, this paper proposes an efficient, fully-automated solution for massively asymmetric image-to-vector georeferencing whereby an image of a relatively small geographic area is automatically located relative to a substantially larger vector map base. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.