Describing the Where – Improving Image Annotation and

Search through Geography

Ross S. Purves, Alistair Edwardes and Mark Sanderson

Department of Geography, University of Zurich, Suiss

Department of Information Science, University of Sheffield, UK

Abstract. Image retrieval, using either content or text-based techniques, does

not match up to the current quality of standard text retrieval. One possible

reason for this mismatch is the semantic gap – the terms by which images are

indexed do not accord with those imagined by users querying image databases.

In this paper we set out to describe how geography might help to index the

where facet of the Pansofsky-Shatford matrix, which has previously been

shown to accord well with the types of queries users make. We illustrate these

ideas with existing (e.g. identifying place names associated with a set of

coordinates) and novel (e.g. describing images using land cover data)

techniques to describe images and contend that such methods will become

central as increasing numbers of images become georeferenced.

1 Introduction and Motivation

1.1 Image Retrieval and the Semantic Gap

The performance of image retrieval systems is currently recognized as lagging behind

that of text retrieval in terms of both the quality of the results returned for individual

queries and the overall ubiquitousness of image retrieval techniques as a first port of

call for search for a particular illustration [11]. The two main approaches to image

retrieval – content-based image retrieval (CBIR) and retrieval of images based on text

surrogates [13] have significant limitations. In general, CBIR techniques work best

within domains where the expected structures and forms of the images are limited, for

example in the retrieval of medical images [7]. However, at the time of writing, CBIR

is not in a state where it can be applied to large collections of what might be termed

“general photographs”, such as those typically held in large stock image collections.

Text-based image retrieval (TBIR) is predicated on the use of appropriate terms to

describe an image. These terms can be drawn from a range of possible sources,

including terms freely chosen by indexers [1]; terms selected by indexers from some

controlled keyword list and text thought to be related to the content of an image (e.g.

figure captions or filenames associated with images embedded in documents).

Although in principle images annotated in such a way are amenable to identical

approaches to those applied in full text document search, the quality of TBIR is

handicapped by the quality of the annotation. This appears to be primarily due to the

relative sparseness of the description of an image, which will tend to be based on the

S. Purves R., Edwardes A. and Sanderson M. (2008).

Describing the Where – Improving Image Annotation and Search through Geography.

In Metadata Mining for Image Understanding, pages 105-114

DOI: 10.5220/0002338401050114

 SciTePress

purpose for which the image itself was being indexed and the cultural background of

the indexer. This in turn means that image search requires that users are, for example,

familiar with the controlled keyword lists used to describe images or that they have

similar backgrounds and expectations to those describing images with free text. A

similar mismatch between user expectations and indexing methods has been identified

in the context of CBIR, where it is termed the semantic gap. Smeulders et al. [24]

defined this gap as follows:

“…the lack of coincidence between the information that one can extract from the

visual data and the interpretation that the same data have for a user in a given

situation.” (p.1353)

In effect, the key problem with image search is centered on this mismatch between

the data describing content and the expectations of those searching for content. In

contrast to document search, both CBIR and TBIR index images using proxies for

content, which it is hoped will describe the images indexed as fully as possible. Thus,

a key challenge in image retrieval must be to develop methods which will describe

images in as universal and rich a way as possible, in order to bridge this semantic gap.

1.2 Describing Images

Clearly, if we wish to describe images universally, we must first formalize the ways

in which images may be described. Shatford [21] set out to do exactly this, refining

the work of Pansofsky to develop the Pansofsky-Shatford facet matrix (Table 1). This

matrix contains three levels, termed the specific of, the generic of and about. Each of

these levels has four associated facets: who, what, where and when. We will term

individual entries in the matrix (e.g. “who/specific of”) as elements.

Table 1. The Pansofsky-Shatford facet matrix (Shatford [21], p. 49).

Facets

Specific Of Generic Of About

Who?

Individually named

persons, animals, things

Kinds of persons,

animals, things

Mythical beings, abstraction

manifested or symbolised by

objects or beings

What? Individually named events Actions, conditions

Emotions, abstractions

manifested by actions

Where?

Individually named

geographic locations

Kind of place

geographic or

architectural

Places symbolised, abstractions

manifest by locale

When?

Linear time; dates or

periods

Cyclical time;

seasons, time of day

Emotions or abstraction

symbolised by or manifest by

The Pansofsky-Shatford matrix has been extensively used in information science,

particularly in the classification of image queries. For example, Armitage and Enser

[3] examined queries posed to a number of image libraries, and allocated each query

to one or more elements of the matrix (e.g. the query “Churchill’s funeral” is

allocated to the “specific of/ what” element). Armitage and Enser demonstrated that

both the “specific of” and “generic of” levels were commonly used in queries

submitted to image archives, whilst those represented by the more abstract “about”

level were rarely identified.

106106

The matrix suggests a variety of ways in which images might be queried or

indexed. For example, CBIR techniques capable of face recognition [27] might allow

us to annotate the “specific of/ who” element (a picture of Jim), whilst CBIR

techniques capable of face spotting [30] would allow annotation of the “generic of/

who” element (a picture of some people). Moreover, the matrix also suggests how

proxy data might be profitably be used to help us describe images – for instance use

of a time-stamp and location associated with an image and a local almanac would

allow generation of annotation related to the “generic of/ when” (a picture at night).

Such tools, taking account of ancillary information to annotate images have been

developed as part of, for example, the MediAssist and NameSet projects [21, 20, 19].

The where facet is, we would argue, particularly interesting and relevant to image

annotation for a number of reasons. Firstly, previous studies have shown that location

is an important element in both indexing and searching for images (e.g. [3]).

Secondly, the volume and content of spatial data describing the semantics of locations

has grown exponentially in recent years, providing a wide variety of potential sources

of ancillary data to describe the where facet. Thirdly, and crucially, images

increasingly carry information related to where within metadata created at the time of

capture, storing at a minimum a set of image coordinates, but potentially also

information about the camera’s orientation (azimuth, pitch and roll).

In this paper we aim to briefly introduce ideas from existing work on the notion of

place from a geographic perspective, which provides a potential theoretical

framework in considering how location can be used to annotate images. We then

explore existing work from a variety of research areas which has attempted to

described locations and introduce ongoing research within the European Project

Tripod which aims specifically to describe images through their location. All of this

research is positioned within the framework of the where facet of the Pansofsky-

Shatford matrix which we contend, together with an understanding of place, will

allow us to address challenge of developing methods to richly and universally

describe images.

2 Geographic Perspectives on Place

The explosion of devices which are location aware and of resources which contain

references to location has led to a wide variety of research in, for example, Location-

Based Services and Geographic Information Retrieval. In general, much of this

research has been dominated by researchers from computing, information and

geographic information science and has paid relatively little attention to more

theoretical work deriving mainly from human geography. This has in turn, led to a

conflation of terminology with location, place and space being used interchangeably

by many researchers and generally being considered to be represented by the

assignment of a set of geometric coordinates or a toponym

to a resource.

However, in geography place is considered to lie at one end of a continuum of

viewpoints with the other extreme being space. Place relates geography to human

existence, experiences and interaction and therefore cannot be considered as purely an

Toponyms are names allocated to some location on the Earth’s surface

107107

abstract property of a set of geometric coordinates [9]. Space on the other hand

encompasses a more abstract and objective geometric view of the world, such as is

typically encoded in spatial data stored in computers. Thus a key challenge in

describing images is to include not only the objective and geometric notion of space

but also the more subjective and potentially everyday idea of place.

3 Methods to Describe Where

We set out here to consider a range of methods to achieve our aim of describing both

space and place, and consider how these methods can be positioned within the

Pansofsky-Shatford matrix. These methods can be considered to be a means of

addressing many of the issues set out by Egenhofer in his discussion of a semantic

geospatial web [7].

3.1 Methods to Describe “Where/ Specific of”

“Where/ specific of” is characterized in the Pansofsky-Shatford matrix as the use of

terms describing “individually named geographic locations”. Thus, a caption which

associates a toponym with an image (e.g. “A church in Bristol” can be considered to

be describing this matrix element. An initial challenge in deriving image metadata

from geographic information is therefore simply to find the most appropriate toponym

to describe an image. This task at first seems trivial, requiring a database lookup to

identify the nearest place name from a gazetteer

and applying the selected toponym

to the image. However, typical gazetteer data will include objects of widely varying

granularities (varying from individual houses to the centroids of large administrative

regions) and representing very different feature types (from mountain summits to the

names of individual pubs). Research on salience in navigation, that is to say

perceptually or cognitively prominent objects, has an important role to play in

deciding which toponym is most appropriate in the labeling of an image. However, to

date most research in this area has focused on objects of similar, relatively fine

granularities which are appropriate to navigational systems.

Naaman et al.[20] took the problem of identifying appropriate toponyms to

describe images one step further, and asked the question, “given a set of diverse

geographic coordinates, find a textual name that describes them best”. Their system,

NameSet, identified appropriate toponyms from a polygon-based dataset by testing

for containment within regions such as cities and parks and nearby cities. They

included a proxy for salience by using what they termed the “Google count” (number

of documents in Google that match a query word) and the population for individual

city names to weight distances, allowing locations with larger populations and higher

Google counts to have a larger zone of influence.

Typically when we describe locations we do so with qualifiers which represent the

spatial relationship between the object of interest and the referencing toponym. These

spatial relationships may be metric (10 km from Edinburgh), directional (north of

A gazetteer is a dictionary of toponyms, usually with associated coordinates and a hierarchy of related

toponyms

108108

Berlin), topological (in Belgium) or vague (near Bern). In practice, combinations of

spatial relationships in natural language are used to reduce ambiguity and refine

information (e.g. 10 km east of Edinburgh on the A1 road). Representing metric

spatial relationships is straightforward and was implemented in the NameSet

prototype [20]. However, the representation of vague spatial relationships is less

trivial and requires development of both computational techniques to represent and

process vagueness and empirical research to identify how people use spatial

relationships [29].

Typically, in everyday language we commonly use vernacular names which are

not found in gazetteers and whose spatial extent is ill-defined [18]. This problem has

been recognized by those working on administrative gazetteers as a pressing issue

[14]. Recent research has used datamining techniques to identify toponyms with

entries in gazetteers which co-occur with known vernacular names and to define

potential spatial extents related to vernacular names [16]. However, most work has so

far addressed relatively large regions (such as Mid-Wales or the South of France),

though work is currently ongoing on the automated identification and definition of

vernacular names with finer granularities [22]. The techniques so far developed have

not reached a level of accuracy such that they can be used to automatically generate

appropriate vernacular names for any given set of coordinates, and also do not address

the issue of identifying vernacular names.

3.2 Methods to Describe “Where/ Generic of”

The “where/generic of” element of the Pansofsky-Shatford matrix is characterized as

representing “kinds of geographic place or architecture”. The first task in defining

this element is therefore to understand what “kinds of geographic place or

architecture” are, before we consider how we can develop techniques to annotate

images.

A logical first step is to identify basic levels of geographic kinds – that is to say

informative exemplars which particularly characterize a geographic scene in terms of,

for example, typical attributes, types of related activities and component parts [28].

Within geography, previous research has examined the terms most commonly used as

basic levels in empirical experiments by asking subjects to give exemplars of natural

earth formations, with a number of researchers finding that “mountain” was a

particularly popular term [4, 26]. The advent of large volunteered datasets as part of

Web 2.0 gives rise to a new sources of data for investigating such questions. We have

been experimenting with data obtained from Geograph (www.geograph.org.uk), a

project with the aim to collect “geographically representative photographs and

information for every square kilometer of the UK and the Republic of Ireland.” The

project allows contributors to submit photographs representing individual 1km grid

squares, and after moderation these images are uploaded together with descriptions to

a publicly available web site. Using these data, we have identified the most commonly

given terms from a set of basic levels derived from earlier empirical research [8].

Table 2 illustrates the top 20 nouns identified in the Geograph data, together with

their frequencies in the collection. Here, we assume that a reference to a road, whilst

possibly naming a specific location (e.g. “London Road”) is also, in most cases, likely

109109

to illustrate a generic example of the matrix element. Further work will be necessary

to test this assumption.

Table 2. Most common terms occurring in Geograph and their frequencies.

45768 road

21119 farm

17242 lane

16232 hill

16157 church

15815 bridge

14737 river

14150 square

13690 house

12707 village

9892 railway

9829 building

9327 centre

9240 park

9234 footpath

9060 line

8563 valley

8532 station

8416 way

8331 track

Within the Tripod project, we are currently developing an ontology of scene types

together with their relationships, qualities, elements and related activities through a

three-pronged approach utilizing analysis of existing datasets such as Geograph,

empirical experiments where subjects are asked to describe images and a literature

study of previous work from a diverse range of fields ranging from landscape

architecture through psychology to remote sensing in order to explore how landscapes

are classified and described. This concept ontology can be seen as a description of the

“where/ generic of” and is illustrated in Fig. 1 for land cover and landforms.

Conce

Land cover Landforms

Agricultural land

crops

farmland

agriculture

Forest

plantation

wood

woods

…

Arable land

field

fields

wheat

…

hic eminences

Mountains

beinn

mountain

sgurr

....

Hills

hill

down

cnoc

....

Fig. 1. Excerpt from a concept ontology – dashed lines indicate that other concepts exist at this

level, and indented terms were identified in Geograph.

Our working hypothesis is that using such a concept ontology, it will be possible to

develop methods which exploit spatial data to describe “where/ generic of”. This

hypothesis can be illustrated through two examples, one exploring the identification

of land cover and land forms.

Land cover is typically described in spatial data which ensure that every location is

allocated a single land cover value. Within Europe, the CORINE project has produced

a dataset describing land cover for 12 countries at a nominal scale of 1:100000, with a

resolution of 100m. CORINE has 3 levels of description, a top level with 5 classes, an

intermediate level with 15 classes and a detailed level with 44 classes. Given either a

point location, or a bounding box it is possible to retrieve the associated land cover

classes with this location. Fig. 2 shows a set of georeferenced images taken in

Peloponnes, Greece and the land cover classes associated with the point locations of

these images. In these 3 cases, it can be seen that the land cover classes describe, to

different degrees, the content of the image. The third image which lies in coniferous

forest clearly illustrates some of the challenges of this approach. Firstly, the position

110110

of the photographer is different from that which was photographed and the

photographer’s location may not reflect the image contents. Secondly, this image

contains two dominant land covers – natural grassland and coniferous forest, and a

method purely based on associating a point with a land cover cannot represent

multiple land covers. Thirdly, this simple approach does not consider errors either in

GPS position or classification and the likely cumulative error in the associated land

cover. The first picture illustrates a further problem – CORINE has a resolution of

100m and an associated scale of 1:100000 – therefore close-up images of objects such

as this fallen tree are not represented.

Fig. 2. Comparing Corine landcover classes with images from georeferenced images

Peloponnes, Greece.

To assess more quantitatively whether this approach can represent the “where/

generic of” we carried out a preliminary study to rate the accordance of land cover

classes with images for 225 georeferenced images from Greece, Italy and the

Netherlands. For the top level descriptions (5 classes), we found that 73% of land

cover classes had a good or fair accordance with the images whilst for the detailed

descriptions (44 classes) 47% of land cover classes had a good or fair accordance with

the images.

Although data sets describing land cover at a European scale exist, this is not the

case for land forms (e.g. mountains, valleys, plains, etc). Research to, for example,

answer the question “Which locations within this region can we delineate as

mountains?” has investigated “mountainousness” through empirical studies [25],

methods which seek to delineate features using digital elevation models (DEM)

[15]

and methods which recognize that the scale at which one observes the surface of the

earth influences the nature of the features that can be identified [12]. At the simplest

level it is possible to assign an image whose coordinates lie at the peak of a mountain

A Digital Elevation Model (DEM) is a regular, usually gridded, tessellation of space where each grid cell

represents a single height value. Attributes of topography such as gradient and aspect can easily be

derived from DEMs)

111111

(and thus are near to a toponym representing an object belonging to the feature class

mountain) to the class mountain. However, at what point do we move from being on a

mountain to being in a valley? Research to investigate this issue must consider

perceived properties of mountains from a particular viewpoint and aim to provide a

probability function describing the extent to which a location belongs to the class

mountain for a set of users with a particular background.

3.3 Methods to Describe “Where/ About”

The “where/about” element of the Pansofsky-Shatford matrix is described as

representing “places symbolized, abstractions manifest by locale”. In their analysis of

queries submitted to image libraries Armitage and Enser [3] found that the abstract

facet in general was a rarely used query form. However, they emphasized that this

result is probably related to the nature of the image archives studied, and suggested

that for stock-photo libraries providing images to, for example, advertising agencies,

abstract concepts, such as “peaceful scenes”, are important. Such qualities are good

examples of abstract properties of place and also relate closely to the geographic

notion of place as being related to experience and interaction.

Although at first glance it may appear unlikely to be possible to describe such

qualities using spatial data, there are in fact a number of examples of research to

define such qualities. For instance, a recent study in the UK has, through participative

research, firstly explored what tranquility is, and secondly attempted to map variation

in tranquility within the UK [17]. Other, similar studies have explored how qualities

such as wilderness can be modeled in space [5]. A common factor of such research is

that locations are placed on a continuous scale describing some relative quality, but

that the identification of a location as being, for instance, tranquil is dependent on the

perceptions and experiences of those who visit a location.

We have attempted to explore abstract qualities of locations which might describe

the notion of place by investigating the co-occurrence of adjectives commonly used to

describe landscapes [6] with typical classes which we have identified in our concept

ontology (Fig 1.). For example, the following 10 adjectives were mostly commonly

used in the Geograph dataset in conjunction with the land cover class beach: sandy;

deserted; eroded; soft; rocky; warm; glacial; low; beautiful and lovely [8]. Thus, a

protypical beach picture might represent abstract qualities such as being deserted.

This property could be modeled using a similar approach to that adopted by Carver et

al. [5], using, for instance, accessibility models. Images identified as belonging to the

“where/generic of” element associated with beaches, through use of, for example,

land cover data as discussed in §3.2, might then be rated in terms of their accessibility

and thus assigned a probability of representing the abstract quality of desertion.

4 Conclusions

We have set out in this paper to consider how Text-based Image Retrieval (TBIR)

might be improved through the use of index terms describing the where facet of the

Pansofsky-Shatford matrix. We contend that such methods will become not only

112112

possible, but indispensable, as increasing numbers of images are georeferenced

Indeed, we would contend that the WorldExplorer system [1], implicitly describes the

where facet of the Pansofsky-Shatford matrix by aggregating Flickr tags from

multiple users to generate useful labels for groups of pictures. We have illustrated

how a wide range of existing methods might be used to describe images not only in

terms of their locations (as represented through a set of coordinates), but also in terms

of the notion of place. Broadly speaking, methods which go beyond analyzing

notionally objective datasets, such as administrative gazetteers or land cover data, can

be considered to address place-based geography. Thus, for example, methods to

identify vernacular names or describe prototypical scene types and their qualities rely

on the development of methods which can, for instance, exploit volunteered datasets

representing experiential data. However, an important note of caution must also be

sounded here – descriptions of place derived from such datasets are inevitably

situated according to the perspective of their contributors. This in turn requires that if

we wish to develop methods describing both space and place that we do so critically.

Acknowledgements

This research reported in this paper is part of the project TRIPOD supported by the

European Commission under contract 045335. Simone Bircher is thanked for her hard

work on image collection and the Corine experiments. We would also like to

gratefully acknowledge contributors to Geograph British Isles, see http://

www.geograph.org.uk/credits/2007-02-24, whose work is made available under the

following Creative Commons Attribution-ShareAlike 2.5 Licence (http://

creativecommons.org/licenses/by-sa/2.5/).

References

1. Ahern, S., Naaman, M., Nair, R., Yang, J.: World Explorer: Visualizing Aggregate Data

from Unstructured Text in Geo-Referenced Collections. In Proceedings, Seventh

ACM/IEEE-CS Joint Conference on Digital Libraries, (JCDL 07), June 2007, Vancouver,

British Columbia, Canada. (2007)

2. Ahn L. von, Dabbish, L.: Labeling Images with a Computer Game. In ACM Conference on

Human Factors in Computing Systems. ACM, New York (2004) 319-326

3. Armitage, L.H., Enser, P.G.B.: Analysis of user need in image archives. J. Info. Sci. 23

(1997) 287–299

4. Battig, W. F., Montague, W. E.: Category norms for verbal items in 56 categories: a

replication and extension of the Connecticut Norms. J. Expt. Psych. 80 (1969) 1–46.

5. Carver, S., Evans, A.J., Fritz, S.: Wilderness attribute mapping in the United Kingdom.

International Journal of Wilderness. 8 (2002) 24-29

6. Craik, K.H.: Appraising the Objectivity of Landscape Dimensions. In: Krutilla, J.V. (ed):

Natural Environments: Studies in Theoretical and Applied Analysis. John-Hopkins Uni.

Press, Baltimore (1971) 292–346

7. Deselaers, T., Müller, H., Clough, P., Ney, H., Lehmann. T.M.: The CLEF 2005 Automatic

Medical Image Annotation Task. International Journal of Computer Vision. 74 (2007) 51-58

8. Edwardes, A.J and Purves, R.S. A theoretical grounding for semantic descriptions of place.

To appear in Proceedings of W2GIS.

For example, by 3 Nov., ‘07, over 30 million Flickr images were associated with coordinates

113113

9. Edwardes, A.J.: Re-placing Location: Geographic Perspectives in Location Based Services,

Ph.D Thesis, University of Zurich (2007)

10. Egenhofer, M.J.: Toward the semantic geospatial web. In: Voisard, A. and Chen, S.C.

(eds.): Proceedings of the 10

ACM International Symposium In Geographic Information

Systems. ACM Press, New York (2002) 1–4

11. Enser, P.: Visual image retrieval: seeking the alliance of concept-based ad content-based

paradigms. Journal of Information Science. 26 (2000) 199-210

12. Fisher, P., Wood, J., Cheng, T.: Where is Helvellyn? Fuzziness of Multiscale Landscape

Morphometry. Transactions of the Institute of British Geographers. 29 (2004) 106-128

13. Goodrum, A. A.: Image Information Retrieval: An Overview of Current Research.

Informing Science. 3 (2000), 64–67

14. Hill, L.L., Frew, J., Zheng, Q.: Geographic names. The implementation of a gazetteer in a

georeferenced digital library. Dig. Lib. 5 (1999)

15. Iwahashi, J., Pike, R.J.: Automated classifications of topography from DEMs by an

unsupervised nested-means algorithm and a three-part geometric signature.

Geomorphology. 86 (2007) 409-440

16. Jones, C.B., Purves, R.S., Clough, P.D AND Joho, H.: Modelling Vague Places with

Knowledge from the Web. Int. J. of Geog. Info. Sci. (In press)

17. MacFarlane, R., Haggett, C., Fuller, D., Dunsford, H. and Carlisle, B. (2004). Tranquillity

Mapping: developing a robust methodology for planning support, Report to the Campaign

18. Montello, D., Goodchild, M., Gottsegen, J., Fohl, P.: Where's Downtown?: Behavioral

Methods for Determining Referents of Vague Spatial Queries. Spatial Cognition and

Computation 3 (2003)185–204.

19. Naaman, M., Harada, S., Wang, Q., Garcia-Molina, H., and Paepcke, A.: Context data in

geo-referenced digital photo collections. In: Proceedings of the 12th Annual ACM

international Conference on Multimedia (MULTIMEDIA '04) ACM Press, NY (2004) 196-203.

20. Naaman, M., Song, Y.J., Paepcke, A., Garcia-Molina, H: Assigning textual names to sets of

geographic coordinates, Comp. Env. and Urban Sys. 30 (2006) 418-435

21. O'Hare, N., Lee, H., Cooray, S., Gurrin, C., Jones, G.J.F., Malobabic, J., O'Connor, N.E.

Smeaton, A.F., Uscilowski, B.: MediAssist: Using Content-Based Analysis and Context to

Manage Personal Photo Collections. In: Proceedings of the 5th International Conference on

Image and Video Retrieval (CIVR 2006). Tempe, AZ, U.S.A., (2006) 529-532

22. Pasley, R.C., Clough, P. and Sanderson, M.: Geo-Tagging for Imprecise Regions of

Different Sizes. In: Proceedings of GIR07. ACM, New York (2007) 77-82.

23. Shatford, S: Analyzing the subject of a picture: a theoretical approach. Catalog. and Class.

Quart. 6 (1986) 39–62

24. Smeulders A.W.M., Worring, M., Santini, S., Gupta, A. Jain, R.: Content-Based Image

Retrieval at the End of the Early Years, IEEE Trans. on PAMI 22 (2000) 1349–1380

25. Smith B., Mark D. M.: Do mountains exist? Towards an ontology of landforms.

Environment and Planning B: Planning and Design. 30 (2003) 411-427

26. Smith, B. and Mark, D.M.: Geographical categories: an ontological investigation. Int. J. of

Geog. Info. Sci. to Protect Rural England, Countryside Agency, North East Assembly,

Northumberland Strategic Partnership, Northumberland National Park Authority and

Durham County Council, Centre for Environmental & Spatial Analysis, Northumbria

University. 15(2001) 59–612

27. Turk, M.A., Pentland, A.P.: Face Recognition Using Eigenfaces. In: Proceedings of the

Conference on Computer Vision and Pattern Recognition (CVPR). (1991) 586-591.

28. Tversky, B. and Hemenway, K., Categories of Environmental Scenes. Cogn. Psych. 15 (1983)

121–149.

29. Worboys, M. F., Nearness relations in environmental space. Int. J. of Geog. Info. Sci. 15 (2001)

633–651

30. Yang, M-H., Kriegman, D.J.; Ahuja, N.: Detecting faces in images: a survey. IEEE

Transactions on Pattern Analysis and Machine Intelligence. 24 (2002) 34-58

114114