GB2140603A

GB2140603A - Adaptive pattern recognition

Info

Publication number: GB2140603A
Application number: GB08413384A
Authority: GB
Inventors: Peter John Taylor
Original assignee: PA Consulting Services Ltd
Current assignee: PA Consulting Services Ltd
Priority date: 1983-05-27
Filing date: 1984-05-24
Publication date: 1984-11-28
Also published as: EP0127445A2; EP0127445A3; GB8314778D0; JPS6036940A; US4637055A; GB8413384D0; GB2140603B

Description

1

SPECIFICATION

The adaptive pattern recognition 5This invention concerns automated inspection and particularly optical inspection systems adapted to recognise selected objectsfrom groups of objects and systemswhich can be adapted to learn what charac teristicsto lookforfrom a training set of such objects so as thereafter to be able to recognise such objects readily and quickly.

High speed image processing or pattern recognition of components and the like is and will become even more necessaryto cope with high throughput rates likely to be encountered on production lines and the like. To date this is achieved through the use of a conventional video camera or solid-state photodetec tor array coupled to a very powerful highspeed computer processor.

Often in microscopy it is necessary to search for objects of interest on slide preparations such as for example to countthe number of deformed cells amongstthe population of normal cells.

In roboticsthe object is usually more clearly defined and comprises the task of selecting quickly and accurately each ofa number of differently shaped component parts from a group of component parts which may be presented to the robot either in a static array or on a conveyor belt.

In quality control it is also necessaryto be able to watch for irregularities and imperfections in finished products and automated inspection of the output of assembly lines and the like will become more com monplace.

In all such applications it is necessaryforthe 100 inspection apparatusto be able to recognise both quicklyand reliablyfrom the shape of the objects presentedto it, those which are of interest, and to generate a signal indicating the presence of such an objectwithin itsfield of view. It is also desirable that such systems shall be capable of being taughtto determine what characteristics of shape should be lookedfor in orderto select each of a number of different articles from a set of articles. Thus in addition to providing a high speed recognition apparatus it is another object of the present invention to provide an apparatus which is capable of learning from a set of articles presented to it, what features are required to be looked for subsequently in order to be able to distinguish one object from another within the group.

It is an overriding object of the present invention to reduce the amount of information which has to be utilized by the system to determine the characteristics of objects presented thereto so as to speed up the processof identification.

Hitherto atypical pattern recognition system has comprised atelevision camera orsimilarscanning device linked to video signal processing circuitsand a computer adapted to produce from the scanning a digitised signal for processing bythe computer to permit a number of different measurementsto be made on the feature contentfound in the image of the field of view. Such processing and computation techniques arewell known and a number of image analysing computers are commercially available cap130 GB 2 140 603 A 1 able of making numerous different measurements of area, perimeter, diameteretc on objects in a field of view. In this way cells on microscope slides have been counted and classified according to shape.

Such systems however are inherently slow by comparison with the human eye basically in view of the large amount of information which hasto be processed before the measurements can be made on any given field. Even with the use of parallel proces- sors and full-field image stores, a considerable length of time may be needed between the presentation of an objecttothe system before a decision can be made by the system asto what is present in thefield of view.

Wherethe object of the inspection process is merely to lookforone particulartype of feature which perhaps may only be presentto the extent of 1 in 1000 or 10,000 objects presented to the inspection system orwhere the objectof the inspection is to identify selectively differentones of a group of known objects to determinethe position of each such object and therebyfor example assist a robotic deviceto locate and pick upthe objects,for exampleto facilitate in the assembly of a mass-produced product, the recognition process can be considerably speeded up if instead of using electronic processing, optical processing is employed instead. Thus it is possibleto, make an optical filterwhich can be compared with optical images of objects presentedto a viewing system and for example byforming atransparency of the desired objectand focussing light from each of the series of objects to be inspected by a system through the transparency, and by measuring the amount of light transmitted through the transparency, so an indication of correlation of the shape of the object in thefield of view can be made relative to the shape of the object in the transparency. If the object in the field of view is the same as that pictu red in the transparency then a given amount of light will be transmitted bythe transparency and for other shapes which do not completely fit the object shape in the transparency, different light levels will be transmitted. By looking for a given range of light level transmissions so an object can be identified in relation to its correlation with the transparency.

Whilst optical matching methods such as described can indicate instantaneously whether or not there is identity between the image of an object and a transparency containing a picture of the same object there is a considerable amount of time needed to set up the system to ensure thatthe object is presented correctly orientated and in correct focus and scale before any comparison can be effected. Consequently such optical matching methods tend to be relatively slowoverall.

It is of course known that if a two-dimensional object is illuminated by coherent light a diffraction pattern can be obtained which is unique tothe object and which is in turn dictated bythe shape of the object. Although such diffraction patterns are orientation sensitivethey are relatively insensitive to the position of an objectwithin thefield of view and one form of pattern recognition system based entirely on optical techniques involves the production photographically of a transparency derived from the diffraction pattern produced by a given object and to position this

2 GB 2 140 603 A 2 transparency at a suitable point in an optical system which is setto view a population of objects using coherent light so as to form a diffraction pattern for each of the different ones of the objects presented thereto. By rotating the transparency relativeto an image of each diffraction pattern produced for each of the population of objects, so it is possible to checkfor identity between the transparency diffraction pattern and that arising from each of the series of objects.

Although the optical pattern matching is consider ablyfasterthan if ordinary optical pictorial trans parencies are compared with full images of the field of view containing thetest objects, the process has very limited applications since a transparency has to be prepared for each objectwhich is to be presented to the system and if each object of a series of objects has to be identified uniquely, then each diffraction pattern obtainedfrom each of the objects hasto be compared with each one of the transparencies contained within the system, in turn, and each transparency rotated in orderto determinethe identity of each objectso inspected.

It is an object of the present invention to combine the advantages of optical pattern recognition withthe high speed and power of digital computing pattern recognition systems so asto obtain a recognition systernwhich is capable of operating atvery high speed andwhich is also capable of beingadaptedto recognise different objectsfrom a population of objects by simply showing the population of the 95 objects to the system in turn during a learning mode.

According to one aspect of the invention an optical inspection system for analysing the object content of a field of view comprises i. an optical system adapted to produce either from the said object content or an image thereof, a diffraction pattern; ii. light level detection means on which the diffrac tion pattern isfocussed for prciducing electrical signals corresponding to light intensity at different points within the pattern; iii. electrical circuit means responsive to said electrical signals for comparing selected signals with a reference pattern.

In its simplestform therefore the invention com- bines the high speed of an optical regonition system using a diffraction pattern derivedfrom each of the objectsto be classified with the high speed of operation of an electronic detection deviceforcheck ing the content of the diffraction pattern and looking for correlation with electrical signals indicative of reference diffraction patterns previously stored in an electronic memory associated with the inspection system.

It should be noted thatthe image of the object 120 content which is to be transformed into a diffraction pattern may be formed entirely optically or electroni cally or by any combination of optical and electronic techniques and may be assembled from the output of a lens system as in a camera, or from the output signal 125 from a scanner such as a television camera or flying spot scanner or from an electron microscope or the like. It is merely necessary that the image so formed shall be capable of being illuminated by coherent light so as to form a diffraction pattern thereof and the 130 illumination maybe byway of transmitted light or reflected I ight.

Where two-dimensional objects are to be classified, it is simply necessary to ifluminate each object using a coherent light source so as to form a diffraction pattern corresponding to each object in turn and to electronically scan each diffraction pattern so produced and pattern match the electrical signals derived from the scanning with stored signals to classify each diffraction pattern.

Wherethe objects are not sufficiently two-dimensional for diffraction patterns relating solelyto shape or surface characteristics of the objectsto be produced,the latter may be imaged using conventional optics onto a device forforming a two-dimensional reproduction of the optical image of the object and a coherent light source is provided to illuminate the two-dimensional representation to produce a diffraction pattern fromthe content ofthe optical image.

Thefirst arrangementis thus eminently suited tothe direct study offorexample, microscope slideswhich can normally be considered to represent two-dimensional objects whilstthe latterapproach involving the use of additional optics is bettersuited to the inspection of three-dimensional objects such as components on a conveyor belt orthe like.

One device forforming a two-dimensional reproduction of an optical image comprises a Hughes liquid crystal lightvalve. Such devices rely on the local electricfield in a wafer of semi-conductorto control the local reflectivity of an adjoining liquid crystal with an optical barrier between thetwo to prevent light incident on the semi-conductor wafer from inpinging on the liquid crystal display and vice-versa. By focussing an image of the field ofview onto the semi-conductor wafer so a two-dimensional reproduction of the image is formed in the liquid crystal display associated therewith and by using the latter to reflect coherent light so a diffraction pattern can be obtained.

It is to be understood thatthe invention is not limited to the use of a Hughes liquid crystal light valve but any device capable of producing a two-dimensional reproduction of an optical which can then interact with a coherent light beam to produce a diffraction pattern, maybeemployed:

The reference pattern may comprise a set of electrical signals stored in a form which can be addressed for subsequent comparison with patterns of electrical signals derived by a scanning of the detection means onwhich the diffraction pattern is focussed. Typicallythe reference patterns are obtained by presenting specimen objects to the inspection system and storing a pattern of electrical signals obtained by scanning selected points within the detector. These signals maythemselves be processed by computing means to produce a mathematical algorithm which can then be usedto classify subsequent signal patterns corresponding to diffraction patterns of subsequently scanned fields ofview.

The diffraction pattern typically comprises a Fourier transform of the object content ofthefield ofview. Where this is the case, relative rotation either actual or effective must be introduced between each reference signal pattern and patterns of electrical signals

1 i 3 GB 2 140 603 A 3 obtainedfrom scanning the light level detection means. This arisesfrom thefactthata Fourier transform is sensitivetothe orientation of the object producing thetransform relativeto the optical axis of the viewing system.

Wherethe light level detection means comprises a photo-sensitive detector having different regions or elements which can be addressed separatelyto allow a scanning to be performed overthe area on which the diffraction a pattern is formed, relative rotation is simply effected by altering the control signals em ployed to effect the said scan ni ng.

Typically an inspection system as aforesaid will include memory meansfor storing a plurality of differentsignal patterns as electrical information signals and means will be provided for comparing the electrical signal patterns arising from a scanning of the detection meanswith the electrical signals obtained byaddressing the stored reference signal patterns to facilitate the identification of each signal pattern arising from a scanning of a detector when presented with an unknown diffraction pattern.

An optical system as aforesaid may be rendered adaptive therebyto enable the system to recognise different ones of a group of objects by providing computing means programmedto perform multivari ate anaylsisof signal patterns suppliedthereto and by providing means for supplying electrical signaisfrom the detectorto the inputof the computing means during a learning mode so that a discriminant function 95 can be obtained forthe set of objects (forming a training set) as the latter are presented in turn to the system during the said learning mode. Using the discriminantfunction during subsequent scans so the diffraction patterns arising from other objects in the 100 group can be identified and classified typically using a look-uptable.

Typicallythe computing means is programmed so asto perform a least mean square anayisis of the electrical signals derived from each of the diffraction patterns from the training set of objects.

Further means may be provided responsive to the discriminantfunction to identify selected points with in the area of the said detection means so as to effect a reduction in the number of points within the said area which must be inspected during subsequent scans for classifying diffraction patterns arising during said subsequent scanstherebyto speed up the said classification procedure.

5Q In addition to determining the selected points within 115 the area of the said detection means from which information isto be derived during subsequent scans, the cliscriminantfunction additionally provides an algorithm to dictatethe weighting to be attributed to the electrical signals obtained from each said selected point. Accordinglythe system as aforementioned preferably further comprises signal processing means responsiveto the said descriminant function to process the electrical signals produced during the scanning of the light level detector means to solve the 125 discriminantfunction algorithm for each diffraction pattern and means is provided for comparing the value of each such solution of the said algorithm with a lock-up table to determine the category or identity of each said diffraction pattern andthereby the category 130 or identity of the object producing same.

Thus in accordance with anotherfeatu re of the invention, an optical inspection system as aforesaid comprises a memory associated with computing means by which reference information obtained using the discriminantfunction is stored for different objects in the training set in the form of a look-up table so that subsequently computed values forthe discriminant function, using different objects, can be compared with the reference information stored in the memory and each subsequent object identified according to the identity stored in the look-up table.

According to another aspect of the invention a method of identifying the object content of a field of view comprises the steps of:

i. producing a diffraction pattern of the said object content, fl. imaging the diffraction pattern onto a detector adapted to generate a first electrical signal, one parameter of which varies in dependance on the light level at different points in the diffraction pattern image; and iii. comparing the said first electrical signal with one or more stored signels, the output of the comparison determining the identity of the said object content. In the method as aforesaid, a second electrical signal maybe derived from the first electrical signal and the identity comparison may be effected between the second electrical signal and the said stored signal or signals. Thus the electrical signals derived from a scanning of the diffraction pattern as seen by the detector may be processed according to some algorithm and a value computed for each diffraction pattern seen by the detector, for comparison with signals stored from previous computations (using the same algorithm) on signals arising from the scanning of diffraction patterns obtained using a training set of objects.

The amount of information to be processed by the algorithm and the amount of information to be stored for subsequent comparison can be significantly reduced by carefully selecting a relatively small number of points within the area on which the diffraction patterns are focussed and only processing signals from the selected regions within the detector so as to generate a value to be attributed to the diffraction pattern concerned for identifying same. If random access scanning is employed, the time required to determine the signal content to be fed to the algorithm can be reduced dramatically and itwill be seen that if a sing le numerical value expressed perhaps in binary digital form is to be computed and stored for each training article and a similar binary digital value obtained for comparison therewith from subsequent scans, the amount of storage needed and the time for comparison of one value with another can be very significantly reduced by comparison with other comparison techniques which involve point by point comparison of analogue or digitised video signals.

According to a preferred feature of the invention a method is provided for selecting those regions of a detector in an optical inspection system from which electrical signals areto be derived to facilitate the classification of objects subsequently imaged bythe system, the system being one in which a diffraction 4 pattern of the object content of each field of viewto be inspected and classified is focussed onto the detector and in which an electrical signal is generated having a parameterwhich varies in dependence on the light level at different points in the diffraction pattern, the method comprising the steps of:

L presenting to the inspection system one of a plurality of objects from a group ii. producing a diffraction pattern of the object content in the field of view as each object is presented 75 tothesystern iii. imaging the diffraction pattern onto a detector having a plurality of different light sensitive regions and generating from a scanning ofthe said regions a first electrical signal for each diffraction pattern from each said object.

iv applying the said first electrical signals to computing means programmed to produce from such electrical signals a discriminantfunction, the discrimi nantfunction determining the positions within the image area of the said detector of selected ones of a plurality of regions therein, from which the said first signal is to be generated during subsequent inspec tion scans, and v. limiting the electrical signals from the said 90 detector means during inspection scans to electrical signals arising from the said selected regions of the detector to thereby reduce the amount of information which must be processed during the inspection scans forthe subsequent identification of objects of the said group when presented to the inspection systems.

The said limitation may be effected by replacing the detector with one which is only sensitive atthe positions of the said selected regionsthereof.

Alternatively the limitation may be effected by masking the said detectorwith a mask having windowstherein which correspond in position to the selected regions of the detector.

Where electrical control signals are generated to effect a scanning of the different regions of the detector,the limitation may be effected by adjusting the control signals so asto cause onlythe selected regionswithin the detectorto be addressed during the scanning.

According to a further aspect of the invention in an optical inspection system comprising optical means forforming a diffraction pattern of the object content of thefield of view of the inspection system, light level detection means on which the diffraction pattern is focussed for producing electrical signals correspond ing to light intensity at different points within the diffraction pattern, electrical circuit means responsive to the said electrical signalsfor comparing selected signalswith a reference pattern, the outcome of each such comparison determining the identity of the object content of the field of view, there is provided an improved method of operation for reducing the time to identify objects of a given class, comprising the steps of:

i. presenting to the system examples from the said 125 group of objects so that each said object occupies the said field of viaw in turn, ii. forming a diffraction pattern of the object content of each said field of view and at least temporarily storing the first electrical signat obtained 130 GB 2 140 603 A 4 from the detectorforeach saidfield of view, iii. computing from the stored signals a discriminantfunction forthe group and determining therefrom the positions of selected windows in the diffraction pattern imagefrom which information must be derivedto permit identification of the different objects within the group.

iv. storing the discriminantfunction or signals which will allow for identification orcreation of said selected windows in the diffraction pattern image, and v. limiting the information presented to the said electrical circuit means during subsequent inspections to the electrical signals derived from the said selected windows in the diffraction pattern.

The invention thus provides an optical inspection system which reduces the optical content ofthe image on which pattern recognition is to be performed by utilising diffraction patterns of the objectswhich are to be identified and further reducesthe information to be handled bythis system when classifying different objects within a group of known objects, by determining onlythose points within each diffraction pattern which must be checked to permit each such pattern to be identified uniquelyfrom the other patterns in the group to thereby reducethe amount of signal processing and computation which must be performed during each such identification and classification step.

Additionallythe invention provides an adaptive pattern recognition system for usewith a group of objects by providing at least during the learning process a computing means adapted to perform multivariate analysis of signal patterns supplied thereto to derive forthe group of objects a discriminantfunction forthe group which not only determines the algorithm which must besolved to produce different unique numerical values for each of the different objects within a group of objects but also will dictatewhich of the points within the scanned area must be investigated during each exploratory scan from which electrical signals must be derived for processing within the algorithm to produce a numerical value attributable to each diffraction pattern so analysed.

The invention will now be described by way of example with referenceto the accompanying drawings in which Figure 1 is a blockschernatic diagram illustrating one embodiment of the invention, Figure 2 is a similar blockschernatic diagram of another embodiment of the invention, Figure 3 is a further blockschernatic diagram of a third embodiment of the invention, Figure 4 is a block schematic diagram illustrating a stillfurther embodiment of the invention, Figure 5 is a diagram illustrating theapplication of the invention to object recognition in microscopy and Figure 6 illustrates howthe invention may be applied to thefield of robotics.

In Figure 1 a standard mu Iti-element two-di mensional photodetector 10typicallya General Electric CID (charge injection device) scanner, is located in the Fouriertransform plane 12 of an optical inspection system. The lattercomprises a lens 14forforming a real and in focus image of an object 16 supported by a neutral background 18 on a Hughes liquid crystal light

1 1 GB 2 140 603 A 5 valve 20. Parallel coherent fight 22 also fails on the lightvalve 20 and the two optical systems of the light from the valve 20 and the parallel coherent light are separated by a half reflecting mirror 24.

The two-dimensional reproduction of the image of the object 16 is produced bythe lightvalve 20which is itself illuminated bythe parallel light 22 and produces a diffraction pattern constituting a Fou rier transform which isfocussed by a Fou rier transform lens 26 onto detector1O.

The latter is addressed by a scanner 28 and this photodetector samples the spatial frequency spectrum of the object 16 at a large number of points, thus translating the optical spectrum into an electrical signal.

Where appropriate analogueto digital conversion is provided for byway of a converter 30 to provide digital signals relating to the frequency spectrum seen bythe detectorfor supply to a digital processor32. The latter is typically a digital computer having stored therein a reference spectrum withwhich the incoming digital signals may be correlated rapidly.

In Figure 2 the lightvalve 20 is replaced by a transparent object 34which is illuminated bythe parallel coherent light22 so as to form as before a Fou rier transform diffraction pattern which is brought tofocus by lens 26 on the detector 10. the latter may comprise a Vidicon television camera ora General Electric CID device as referredto with regard to Figure 1. Scanning of the different photosensitive regions of the device may be effected on a random access basis or on a raster scanning basis or circularly or in a spiral manner.

Scanning is controlled by a scan control circuit 36 and as shown diagrammatically by switch 38, the output of the detector during any single frame scan maybe supplied either to a reference memory 40 of a bufferstore42.

During a preliminary learning scan, a known object is located at34 and switch 38 is setto supplythe scan outputsignal to reference store 40. The object 34 is then replaced by other objects which are to be checked againstthefirst andthe switch 38 is changed to supply the outputfrom the scanning of each of thetest objects in turn to the buffer store 42. Atthe end of each scanning,the contents of the buffer 42 and the reference store 40 are compared by a comparator 44 the output of which indicates a measure of the correlation of thetwo signals stored in 40 and 42.

Depending on the degree of correlation so the 115 subsequent objects can be identified as the same as or differentfrorn the original specimen object.

In Figure 3 a Fourier transform diffraction pattern is formed by either of the arrangements shown in Figures 1 and 2 on a detector 10 having ascanner associated therewith for addressing the different points within the scanner.

A gate 46 in the output of the scanner is opened and closed by means of a control circuit48 attached to a memory 50 containing information relating to selected positions in the scanned area of the detector. The control circuit48 usesthis information to control the opening and closing of the gate 46 so asto limitthe information passing to a computer 52, to that arising from the selected regions of the scanned area of the detector 10 using the information contained in the memory 50.

Comparator 52 compares the information arising du ring a scanning of the area with the information stored in the reference memory 54 to produce an output signal the value of which indicates whether or notthere is correlation between the information arising during the scan and the reference memory content.

Although not shown the gate 46 may include signal processing means (notshown) for modifying the amplitude or phase orfrequency or some other parameter of the electrical signals released bythe gate 46 according to some pre-arranged algorithm.

Itwill be seen that in each of these embodiments there is a reduction in the amount of information needed to effect a comparison between one object and anotherthereby allowing a considerable increase in the speed of classification.

The self learning or adaptive capability of the invention is demonstrated in its simplestform in Figure 2 and is further demonstrated in the embodiment illustrated in Figure 4.

In this arrangementthe inspection system is arranged to ascertain from each of a group of objects, the salientshape characteristics from the Fourier transform diffraction pattern obtained for each of the objects, by which each objectcan be distinguished from other members of the group and an algorithm of descrim inant function as it is more commonly called is computed from the preliminary measurements made bythe detector on the different Fouriertransforms presentedto it bythe training set of objects.

The outputfrorn the detector 10 can be supplied to a computer 56 which may for example be programmed to perform multivariate analysis of signal patterns supplied thereto, typically by means of at least mean square anaylsis, so as to determine a diacriminant function (typically a linear discriminant function) for the group of objects which function is stored in a memory 58 associated with the computer.

The objects mayfor example be a series of photographic transparencies of microscope slides shown at60 and the Fouriertransform diffraction pattern is obtained by illuminating each transparency in turn by parallel coherent light.

After computing and storing the discriminantfunction in memory 58 the computer no longer serves any useful purpose otherthan as perhaps a centralised control for the system, if required, and during subsequent scans, during which known transparencies are converted into diff raction patterns on the detector 10, the scanning of the latter is controlled by means of a scan control system 62 itself programmed in accordance with the discriminant function stored in the memory 58 and the outputfrom the detector 10 is supplied via the alternative position of switch 64to a signal processing circuit 66for producing valuefrom information arising during the scanning of each fresh diffraction pattern, with the weighting of each individual component arising during the scanning being determined bythe discriminant function which has thetypicalform 1.30 A = B.m + C.n + D.p + E.q + F.r...

6 GB 2 140 603 A 6 whereA isthevalueto be computed and B, C, D, E and Farewhole numbers orfractions determining the weighting of the individual compo nents and m. n. p, q, r etc constitute the electrical 70 signals atselected points arising during the scanning of the detector 10.

The output from the processing stage 66 is supplied to a comparator 68 associated with a second memory 70 arranged as a look-up table and the output signals at 72 comprise the values in the look-up table corresponding to the computed values of the discrini nant function as produced by processing stage 66.

Providing that the computer 56 provides the values forthe look-up table memory 70 via signal path 74 during the initial learning phase, it will be seen thatthe system shown in Figure 4 is completely self adaptive to recognise each of the members of any group of objects and all members of the group can be supplied tothe system as atraining set.

Often in microscopy it is necessaryto search for objects of interest on a slide preparationJor example to couritthe number of deformed cells amongst a population of normal cells.

Figure 5 Illustrates howthis recognition process can 90 be automated. Abeam of coherent fight76 is projected through an area of a slide78 and focused by lens 80 onto a photodetector array 82.

If, for example, normal cells are circular and deformed cells are elliptical, a detector array consist ing of a plurality of concentric photosensitive rings 84 can be used as shown in Figure 5a.

The detector 82 is situated in the Fourier plane of the lens 80 so thatthe outputfrom each ring 84 corresponds to the content of the field of view over a range of spatial frequencies defined by the inner and outer radii of each ring. Varying the orientation of an object within thefield of view does not affectthe signals from the detector rings arising from scanning thedetector82.

In the case of a deformed cell in which one diameter reduces to become the minor axis of an ellipse, the reduced dimensions parallel to that axis will result in a higher spatial frequency content and will produce a relatively larger signal in the outer rings of the detector, whateverthe orientation of the ellipse, than isthe case for a circular cell.

For normal cells the signals from the various ring elements will fall within certain limits defined by the variations in acceptable cells. For deformed cells the signals will fall within other limits outside the first set of limits and can therefore be distinguished using a comparator and reference memory (not shown).

When the microscopic slide is scanned pastthe field of view, celiswithin the field of view can be classified 120 bycomparing the signals they producewithin the defined limits.

When deformed cells are detected they can be counted, or held forfurther investigation, for example by eye.

In Figure 6 an object 86 is imaged by lens 88 via beamsplitter 90 onto a Hughes light valve 92 which produces a two-dimensional reproduction in the form of a transparency of the object 86.

Parallel coherent light 94 is projected via beamsplit- 130 ter 91 onto the light valve 92 and a diffraction pattern of the two- dimensional reproduction in the light valve is focused by lens 96 onto a detector 98 and detector 100 via beamsplitter 102.

Item 98 is a large area detector, shown in plan in Figure 6a, which integrates all the lightfrom thefield of viewcontaining the object86. Item 104 amplifies the signal and gives an odtput at 106 corresponding to the area of the objectin thefield of view.

Detector 100 is obscured centrallyby an opaque stop 108 as shown in Figure 6b. As detector 100 is in Fou rier transform plane of the lens 96, the effectof the stop isto removethe light produced by low spatial frequencies in the field of view. The lightfalling onto the detector correspondsto the high spatial frequency content of the scene atthe edges of the object 86. Item 110 is an amplifierwhich gives an output at 112 correspondingto the length of each of the edges In the object, ie its perimeter.

Different objects can be characterised bythe values of their area and perimeter and the ratio between the square of the perimeter andthe area. Foreach class of objectsone can define upperand lower limitsfor each feature parametersuch as area; perimeter; and perimeter squared over area. These limits would be derivedfrom measurements on a training set of known objects.

When a new object enters thefield of view, its features are calculated. If all three fall within the limits

Claims

defined for one object type, the new object is recognised as being of thattype. CLAIMS

1. An optical inspection system for analysing the object content of a field of view comprising:

i. an optical system adapted to produce either from the said object content oran image thereof, a diffraction pattern; ii. light level detection means on which the diffraction pattern is focussed for producing electrical signals corresponding to light intensity at different pointswithinthe pattern, and; iii. electrical circuit means responsive to said electrical signals for comparing selected signals with a reference pattern.

2. An optical inspection system as claimed in Claim 1 for use with 2dimensional objects comprising a coherent light source for illuminating the objects directly and forming thereby the said diffraction patterns thereof.

3. An optical inspection system as claimed in Claim 1 wherein the optical system comprises:

i. a first focussing system for forming an opticat image of an objectto be inspected, ii. means for forming a 2-dimensional reproduction of the said optical image; iii. means for illuminating the said 2-dimensional reproduction with coherent light, the said 2-dimensional reproduction constituting a spatial light modulatorto produce a diffraction patternfrom the content of the optical image.

4. An optical inspection system as claimed in Claim 3 wherein the said spatial light modulator comprises a liquid crystal device.

5. An optical inspection system as claimed in any of the preceding claims, whereinthe said reference k -h v 7 GB 2 140 603 A 7 pattern comprises a set of electrical signals stored in a form which can be addressed for subsequent com parison and which have been obtained by presenting a specimen objectto the system previously.

6. An optical inspection system as claimed in any of the preceding claims, wherein the diffraction pattern comprises a Fou rier transform of said object content.

7. An optical inspection system as claimed in Claim 5 or 6 further comprising means for introducing 75 effective relative rotation between the reference pattern of signals and the electrical signals obtained from the light level detection means.

8. An optical inspection system as claimed in claim 7 in which the light level detection means 80 comprises a photosensitive detector having different regions or elementswhich can be addressed separate lyto allow a scanning to be performed overthe area on which the diffraction is formed, and relative rotation is effected by altering the control signals employedto 85 effect the said scanning.

9. An optical inspection system as claimed in any of the preceding claims further comprising means for storing a plurality of reference signal patterns and means forcomparing the said electrical signals with 90 each of the stored reference signal patternsto facilitate the identification of signal patterns arising from the detector.

10. An optical inspection system as claimed in any of the preceding claims further comprising computing 95 means programmed to perform multivariate analysis of signal patterns supplied thereto, and wherein means is provided for supplying the said electrical signals to the input of the computing means in a learning mode, whereby a discriminant function can be obtained fora set of objects (forming a training set) by presenting each object in the training set in turn to the inspection system in the learning mode.

11. An optical inspection system as claimed in claim 10 wherein further means is provided respon sive to the said discriminant function to identify selected points within the area of the said detection means, so asto effect a reduction in the number of points within the said area which must be inspected during subsequent scans to enable the classification of diffraction patterns to be made during the said subsequent scans, therebyto speed up the said classification procedure.

12. An optical inspection system as claimed in claim 9 or 10 wherein the said computing means is programmed so asto perform a least mean square analysis of the electrical signals derived from each of the diffraction patterns from a training set of objects.

13. An optical inspection system as claimed in any of claims 10 to 12 further comprising a memory associated with the said computing means, in which reference information obtained using the discriminant function for different objects in the training set is stored in the form of a look-up table, so that subsequently computed values forthe discriminant 125 function (using different subsequent objects) can be compared with the reference information stored in the memory, and each subsequent object identified according to the identity stored in the look-u p table.

14. A method of identifying the object content of a 130 field of view comprising the steps of-- L producing a diffraction pattern of the said object; ii. imaging the diffraction pattern onto a detector adapted to generate a first electrical signal one parameter of which varies independence on the light level at different points in the diffraction pattern image, and iii. comparing the said first electrical signal with one or more stored signals, the output of the comparison determining the identity of the said object content.

15. A method as claimed in Claim 14 wherein a second electrical signal is derived from the said first electrical signal and identity comparison is effected between the said second electrical signal and the said stored signal orsignals.

16. A method as claimed in either claim 14 or claim 15 comprising means for selecting some of the regions of the detector on which the image of the diffraction pattern is focussed, and generating the said first electrical signal from the light level in the diffraction pattern atthe selected regions.

17. A method of selecting regions of a detector in an optical inspection system inwhich a diffraction pattern of object content to be inspected is focussed onto the detector and in which an electrical signal is generated having a parameter which varies in depenclenceon the light level at selected points in the said diffraction pattern, comprising the steps of:L presenting to the inspection system one of a plurality of objects from a group; ii. producing a diffraction pattern of the object content in the field of view as each object is presented to the system; iii. imaging the diffraction pattern onto a detector having a plurality of different light sensitive regions and generating from a scanning of the said regions a first electrical signal for each different diffraction pattern; iv. supplying the said first electrical signals to computing means programmed to produce from such electrical signals a discriminant function, the discriminantfunction determining the positions within the image area of the said detector, of selected ones of a plurality of regions therein, from which the said first signal is to be generated during subsequent inspection scans, and v. limiting the electrical signals from the said detector means during inspection scans to electrical signals arising from the said selected regions in the cletectorto thereby reduce the amount of information which must be processed cluringthe inspection scans to permit identification of the objects of thesaid group when presented to the inspection system.

18. A method as claimed in Claim 17 wherein the said limitation is effected by replacing the detector with one which is only sensitive atthe positions of the said selected regionsthereof.

19. A method as claimed in Claim 17 wherein the said limitation is effected by masking the said detector with a mask having windows therein which correspond in position to the selected regions.

20. A method as claimed in Claim 17 wherein electrical control signals are generated to effect a 8 GB 2 140 603 A 8 scanning of the regions of the detector, and the limitation is effected by adjusting the control signals so asto cause onlythe selected regions within the detectorto be addressed.

21. Inan optical inspection system comprising optical means for forming a diffraction pattern of the object content of the field of view of the system, light level detection means on which the diffraction pattern is focussed for producing electrical signals corres- ponding to light intensity at different points within the diffraction pattern, electrical circuit means responsive to the said electrical signals for comparing selected signalswith a reference pattern,the outcome of each such comparison determining the identity of the object content of the field of view, an improved method of operation for reducing the time to identify objects of a given class, comprising the steps of:- i. presenting to the system examples from the said group of objects so that each said object occupies the said field of view in turn, ii. forming a diffraction pattern of the object content of each said field of view and at least temporarily storing the first electrical signal obtained from the detectorfor each said field of view, iii. computing from the stored signals a discriminantfunction forthe group and determining therefrom the positions of selected windows in the diffraction pattern image from which information must be derived to permit identification of the different objects within the group, iv. storing the discriminant function or signals which will allow for identification or creation of said selected windows, and v. utilising the stored function or signals to limit the information presented to the said electrical circuit means during subsequent inspections, to the electrical signals derived from the said selected windows in the diffraction pattern.

Printed in the United Kingdom for Her Majesty's Stationery Office, 8818935, 11/84, 18996. Published at the Patent Office, 25 Southampton Buildings, London VVC2A 1AY, from which copies may be obtained.

7 ^9 I l