660 PHOTOGRAMMETRIC ENGINEERING

the loss from the reduced scale.A fair comparison is afforded between the

results from the A-9 Autograph and those nor­mally expected from an A-7. Both are built bythe same company, both operate on the samemechanical principle, and the grid tests indi­cate that the precision of the A-9 tested isabout equivalent to that of the A-7 Auto­graph. Judging from results over several yearsfrom the A-7 at GIMRADA and othersowned by private companies under contract,and from the results with the A-9 in thesetests, the vertical accuracy of ultra-wideangle models is about t that of standardmapping photography.

Since the accuracy of a given system isconsidered to be essentially a straight-linefunction of altitude, it follows that accuracyequal to that of standard photography couldbe achieved by flying ultra-wide-angle photog­raphy at t the altitude. This, then, wouldred uce the coverage ad van tage so t hat thearea per photo would he only a little morethan double that of regular photography.

An Approach to AutomaticPhotographic Interpretation *

Still, cutting in half the number of picturesrequired for a given area seems to he a \\'orth­while achievement.

Considerable discussion has centeredaround the fact that in rough terrain theremay be hidden areas caused by the low slopeof rays from those areas. This is true, ofcourse, but the problem is not limited toultra-wide-angle photography. There will behidden areas wherever the slope of the groundis away from the camera station and issteeper than the rays to the camera. This con­dition is currently plaguing those trying tomap rough areas such as the Rocky Moun­tains, and would be worse if ultra-wide-anglephotography were being used. The conclu­sions are that there are vast areas in the worldthat are not too rough to be mapped by ultra­wide-angle photography, that it is not to beexpected that it will completely supplantregular mapping, but that there will be suffi­cient advantage in its coverage aspects toinsure its widespread acceptance as an impor­tant new tool of photogrammetry.

DR. AZRJEL ROSENFELD,

Budd Electronics,Long Island City 1, N. Y.

ABSTRACT: This p(tper describes an approach to the automatic identification ofthe images of targets on aerial photographs. The approach described involvesthe extraction from the photographic image of two basic types of information,one of them relating to the presence in the image of figures having given shapesand sizes, and the other to the "textnral" nature of the image.

TARGET RECOGNITIO~ CRITERIA

AMAJORaimof photographic interpretation isthe recognition of the presence or absence

of given types of features ("targets," in themilitary case) on given aerial photographs.To perform photographic interpretation auto­matically, the photographs must be acceptedas inputs by the data handling system whichwill make the recognition decisions. A majorobstacle to effective automatic photographicinterpretation is the wealth of information,measurable in the billions of bits, which a

high quality photograph contains. The humaninterpreter certainly does not consciouslyprocess all of this information in making flashrecognition decisions; for recognition pur­poses, most of the information content of thephotograph is redundant. It seems reasonableto conclude that a practical method of auto­matic target recogni tion will have tomarkedly reduce the amount of redundantinformation contained in the photograph.The result of such a reduction in informationmay be thought of as a simplified description

• This paper was presented at the 1962 National Convention on Military Electronics.

ACTOMATIC PHOTOGRAPHIC INTERPRETATION 661

of the photograph \\-hich tells just enoughabout it to make target recognition possible.

Of the many types of extractable descrip­tive information, the choice of those whichare actually extracted must be based on ap­parent or proved usefulness as aids to targetrecognition. An analysis of methods of strate­gic photographic interpretation suggests thatthere are at least t\\·o important kinds of de­scriptive information which can contribute totarget recognition, namely,

a) Shape and size information-that is, thepresence of images having characteristicranges of shapes and sizes. Examples oftargets for which shape/size informationcontributes to recognition are airfields.storage tanks, shi ps and planes, as wellas regions of terrain which contain cul­tural features (these last almost invari­ably contain straight lines and rightangles). Targets of these types areillustrated in Figures 1-2.

b) "Textural" information-that is, param­eters which describe local distributionsuf densi ties and densi ty con trasts inthe photographic image. The si m plestexamples of such parameters are themean density and the mean "level ofdetail" (=number of contrasts per unitarea). Targets for which textural in­formation contributes to recognition in­clude urban and industrial areas (highlevel of detail) and targets such asharbors, ships and bridges which areassociated with hydrographic features(negligible level of detail in the absenceof sun glitter). Such targets are illus­trated in Figures 3-4.

Most target types require for their identifi­cation the presence of particular combinationsof shape/size and textural information in par­ticular relatil'e locations on the photograph. Auseful automatic photographic interpretationscheme should thus extract from a givenphotograph at least the follo\\'ing types ofinformation:

(1) The locations of images having variousshapes (in particular: straight lines,circles, right angles, etc.) and sizes ap­propriate to the scale of the photo­graph.

(2) The locations of regions h(t1'in~ I'Minustextural natures correspondi ng to ter­rain and cultural feature types whichare rele\'an t to target iden ti fica tion.

In the following sections of this paper, auto­matic shape recognition techniques as theyapply to photographic interpretation are

briefly reviewed; approaches to statisticalimage analysis are then discussed In somedetail.

SHAPE RECOGKITlO:-l

Two basic methods of detecting the pres­ence of an image of a gi\'en shape on a photo­graph may be referred to as direct matchingand property matching. In direct matching,the photographic image is correlated \\'ith asuitable physical or mathematical templatein such a way as to produce signal maxima atpoints corresponding to the locations ofimages of the given shape. An important ad­vantage of the direct matching approach isthat it selects and recognizes the given shapedimages in a single operation; it does not re­quire that portions of the photograph first beselected ("figure extraction") and then testedto determi ne whether or not they have thegiven shape ("figure recognition"). This ap­proach has the disadvantage of being sensi­tive to the relative orientation of the photo­graph and template; however, it can be per­formed for all possible orientations at fairl\'high speeds.

There exists a wide variety of propertymatching techniques for shape recognition.These techniques generally presuppose thatfigure extraction has been performed. Theythen proceed to measure certain descriptiveproperties of the extracted figures and to com­pare these measurements with those of theshape which is sought. Many of the mostuseful of such measurements relate to theboundary of the figure in question. For ex­ample, an arbitrary shape is completely de­termined by specifying the curvature of itsboundary (as a function of arc length meas­ured along the boundary). If the shape \\'hichis to be recognized is of a relatively simpletype, special-purpose recognition techniquescan be devised for it; for example, straightlines can be recognized by applying "track­ing" techniques. The examples just givenindicate that the property matching approachto shape recognition can be made independentof the position and orientation of the shapebeing sought. However, this approach stillrequires that the shapes in question be some­how picked out from the photographic imageas a whole; this will in general be a highlynontrivial task.

ANALYSIS OF IMAGE TEXTURE

A variety of parameters can be used tomeasure the "texture" of a given portion of aphotograph. Among the operations which canhe performed on the image density function toyield useful texture information are harmonic

662 PHOTOG1L\MMETlUC ENGINEERING

FIG. 1

FIG. 2

AUTOMATIC PHOTOGRAPHIC INTERPRETATION

FIG. 3

FIG. 4

663

664 PHOTOGRAMMETRIC ENGINEERING

analysis; autocorrelation; di fferentiation andlevel slicing; and measurement of statisticalmoments (mean, variance, etc.). The relativemeri t of these parameters as texture descri p­tors has yet to be completely determined.HO"'ever, the limited studies which have beenperformed* indicate that such parameters canprovide useful texture type classificationswhich are suitable for automatic photographicin terpreta tion purposes.

A shortcomi ng of most of the past work onimage texture analysis is that it has related tothe measuremen t of texture parameters for apreselected portion of the image. To performtesture analysis for the image as a whole, twoapproaches are possible:

a) The photograph can first be divided intofixed portions (small squares, for ex­ample) and the texture parameters inquestion measured for each of these por­tions. As a means for extracting texturalinformation for purposes of photo­graphic interpretation, this approachhas the disadvantage that the portionsare not preselected with any regard tothe content of the-photograph. If one ofthe portions o\'erlaps two differentlytextured types of regions on the photo­graph, its parameters will correspond toneither of the types, and may mislead­ingly correspond to a third type whichis not actually present. As a result,important information relating to thebou ndaries between di fferen t types ofregions is lost.

I,) Alternati\'ely, the texture parameterscan be measured for an "aperture"which is systematically scanned throughevery possible posi tion over the photo­graph. This approach gives completetextural information about the photo­graphic image. However, this informa­tion is actually greater in quantity thanthe information contained in the imagedensity function, since it includes sev­eral texture parameter measurementsmade over an aperture centered at everypoint of the photograph. This approachthus increases, rather than decreases theinformation content of the image, and assuch is hardly practical for use in anautomatic photographic interpretationsystem.

The disadvantages of the t,,·o approaches

• See, for example, the author's AutomaticRecognition of Basic Terrain Types from. AerialPhotographs, PHOTOGHAMMETHIC ENGINEERI1\G,

\'01. XXVIII, \10. 1, March, 1962.

just described can he avoided by taking aglobal, rather than a local, approach to tex­tural image analysis. In this approach, thephotograph is automatically divided intoregions of appreciable size, in each of whichthe texture parameters of interest are suh­stantially constant, but which are such thatthese parameters differ significantly fromregion to region. If this can be done, the re­sulting regions will be essentially the partsinto which a human observer might dividethe photograph if asked to report on its ap­pearance after very briefly viewing it; theyshould in general correspond to differenttypes of terrain. The texture parameters ofinterest can then be measured for these re­gions. This will result in a radical quantitativereduction in the information content of thephotograph; at the same time, very littleuseful information will be lost, since theparameter measuremen ts taken over suchuniform regions will not differ widely fromthe local measured values of the parameterswithin these regions. The important informa­tion relating to the boundaries between differ­ent types of regions will also be preserved.

Methods of automatically dividing a photo­graph into texture (or terrain) types will no,,'be discussed.

TEXTURE TYPE SEPARATION

The subdivision of a photograph into terraintypes is not always easy even for a humaninterpreter. In some cases the boundarieshetween terrain types will be sharply definedand the terrain on opposite sides of theseboundaries will be significantly different. Tnother cases the transitions between adjoiningterrain types may be quite gradual. For illus­trations of both si tuations, see Figures 3-4.

Tn automating the process of terrain typesubdivision, it is reasonable to begin by treat­ing situations in which the desired subdivisionis relatively obvious. Further, it is convenientto consider a one- rather than a two-dimen­sional case. Suppose, specifically, that an"image density" function is given whichvaries along a Ii ne rather than over a plane.It is desired to subdivide the line into seg­ments "'ithin each of which certain "textural"parameters of the given density function re­main approximately constant while differingsignificantly from segment to segment. Thiscorresponds to a subdivision into clearlydefined, significantly differing terrain types.Such a subdivision can be achieved if it ispossible to determine those points along theline (if any) on opposite sides of which one ormore of the parameters in question are sig-

AUTOMATIC PHOTOGIL\I'HIC INTERPRET.\TlON 665

nificantly different.Mathematical models for the detection of

subdivision points can be formulated usingany of the textural parameters defined above.r\ simple model which used mean density andmean contrast frequency as parameters hasbeen constructed and successfully tested asdescribed elsewhere.* This model may be re­garded as a first approximation in the sensethat it employed only first statistical mo­ments (means).

In devisi ng more sophisticated su bdi visionmodels, it is convenient to make the furthersimplifying assumption that the input photo­graph can be adequately approximated by aquantization into elements which are eitherblack or white. This assumption, thoughdubious at first glance, is actually not toounreasonable; very good black-or-white ap­proximations to aerial photographs can bemade using high-resolution "screening" tech­niques which are standard practice in thegraphic arts. A one-dimensional black-or­white density function may be thought of astaking on the values zero and one only. Thegraph of such a function consists of a succes­sion of plateaus (ones) and valleys (zeros).

Higher order approximations to the defini­tion of texture can be very conveniently for­mulated in terms of black-or-white densityfunctions. Specifically, one can use as textureparameters the mean, variance and highermoments of the plateau \\'idths and of the\'alley widths in the function. It appears quiteplausible that these parameters do indeedrelate closely to the textural appearance ofthe region \\'hich has the given (cross-sec­tional) density function.

It may be noted that certain other usefultexture parameters are closely related to the"width moment" parameters just defined.Thus the mean density and mean contrastfrequency, are simple functions of the meanplateau and valley widths. The autocorrela­tion spectrum is not completely derivablefrom the first few II'idth moments; howe\'er,periodicities in the density function, which

* See the author's Automatic Recognition Tech­niques Applicable to High-ll1formation Pictorial In­puts, to appear in the Proreedings of the 1962 IREInternational Convention.

are perhaps the most useful properties whichcan be detected by autocorrelation, can alsobe detected by measuring the variances of theplateau and valley widths. The relationship ofharmonic components to statistical momentsis even less immediate. It seems likely. hOII'­ever, that statistical moments are morcclosely related to human perception of anddifferentiation among textures than are har­monic coefficients.

The one-dimensional case discussed abovecan be generalized to two dimensions bymeasuring one-dimensional texture param­eters in every direction through each point ofthe photograph. This straightforward gen­eralization is probably not completely ade­quate in the sense that it does not givespecial weight to inherently two-dimensionalaspects of the subdivision problem, such asthe shapes of elements and of regions, whichare "disproportionately" important to hu­man observers. A more complete generaliza­tion requires an analysis of informationalproperties of perceived shapes and is beyondthe scope of the presen t paper.

Texture-type separation schemes such asdescribed above can be simulated quite in­expensively. The given photograph is firstconverted to the black-or-white format byhigh-resolution screening. Plateau and valleywidths, and densities if desired, can be meas­ured on this quantized photograph using amedium-power microscope. The resultingnumerical data can be processed either manu­ally or on a small general-purpose digitalcomputer such as the IBM 1620. Real-timeprocessing is also possible using an ultrahigh-resolu tion Aying spot scanner or a motor­dri ven mechanical scan of the microscopestage; the resulting density signals can bequantized and processed using relativelysimple special purpose circuitry.

The techniques discussed in this paperconstitute an approach to automatic photo­graphic interpretation along several parallelfronts. \Vhile much work has yet to be donebefore fully automatic photographic interpre­tation even approaches realization, it is fel tthat present lines of effort II·ill play importantroles in bringing the ultimate goals closer topracticali ty.