There is a growing requirement for automated image analysis and in many applications this involves checking the size and shape of an object constituting an image. Checks of this sort are suitable for computer-control and, to present the image in a form suitable for analysis by computer, the image must be digitized. The digitization may be effected by taking the image acquired by a television camera or similar device and dividing it into picture elements (hereafter called pixels). The pixels are arranged ideally in some regular pattern, of which rectangular or hexagonal arrays are the most common, though others are possible. In an embodiment to be described, a rectangular array of pixels is employed, but the arrangement can readily be modified to apply to other geometries to define a measuring grid or lattice.

The resolution available in a digitized image depends to some extent on the density of the pixel array. The most common systems use square arrays of 256 x 256 or

512 x 512 pixels, but some have rectangular windows or square arrays of up to 4096 x 4096 pixels; however, with normally practicable lenses and cameras it is doubtful whether the true optical resolution is better than can be achieved with a 512 x 512 pixel raster. Since, in the ideal case of perfect contrast between an object and background on the field of view, any particular pixel would be either "on" or "off", i.e. register as being either inside or outside the object, it may be considered that the resolution of a digitized picture is directly proportional to the inverse of the pixel density and that, for example, the length of an object detected by a 512 x 512 pixel array could be determined to only about 1 part in 500. This conclusion results from the practice of determining lengths by a point-to-point calculation with the positions of the two points in question each carrying, in this instance, an error of up to unit inter-pixel distance.

It will be seen from the particular embodiments of the present invention to be described that it is possible to determine the position of a line or edge in a digitized picture to a much greater degree of accuracy than plus or minus one unit of distance (the inter-pixel distance) by the use of at least one of three procedures:

(a) by employing many points to make a statistical analysis of the position of the line,

(b) in the case of a straight line (or edge), by positioning the line so that it makes a very small angle with one of the co-ordinate axes and

(c) by repeating the measurement an appropriate number of times (e.g. 10 times) with small relative displacements of the camera,, the object or the measuring grid or lattice

(equivalent to fractions of the inter-pixel distance) between successive measurements.

The distance between two straight lines or edges, which might be the length or breadth of a rectangular object in a digitized picture, can then be obtained as the difference between their respective positions within an error which is a compound of the two separate small errors. When the positions of lines or edges are determined in one of these ways, secondary features such as angles or positions of corners can be determined to a high degree of accuracy. Furthermore, the calculations of line position can be simplified by making use of the distinction in an image between the lattice, on the points of which the pixels are centred, and a grid, which exists between the pixels. This simplification applies particularly to positioning curved lines or edges. In the case of a straight line or edge it is straightforward to calculate the difference in position between a line calculated from lattice points and that calculated from grid points, if the angle between the line and a co-ordinate axis is known. In either case, the setting of this angle permits the measurement to be made in a manner similar to the employment of a vernier and, by reducing the angle, the -measurement can be of an indefinitely high accuracy, although in a practical situation they might be expected to be to only about 20 times the

pixel density, i.e. to about 1 part in 10000 on a 512 x 512 pixel array.

A method of making a measurement in accordance with the present invention will be described below, by way of example, with reference to the accompanying drawings in which:-

Fig 1 is a representation of a picture obtained by digitizing an image of an object viewed by a television camera, and Fig 2 is a representation in more detail of a part of the picture shown in Fig 1.

Apparatus used in carrying-out the invention will be described below, by way of example, with reference to Fig 3 of the accompanying drawings, which shows a block schematic circuit arranged with a diagrammatic side view of an object. Referring to Fig 1, it will be understood that the requirement is to monitor the length and breadth of a panel represented by an image having sides 1, 2, 3 and 4 and the parallelism of its sides. Referring to Fig 2 which shows a part of the side 1, the filled circles 5 mark the centres of pixels forming an interior "border" of the digitized image of the panel, the* open circles 6 mark an exterior "border" and the line 1 shows where the true edge of the image of the panel f lls. These two sets of circles may be considered to mark lattice points. Either set may be used to calculate the position of a line, but either of the lines derived from such calculations may be distant from the line 1 marking the

true edge of the panel by an amount up to *«^ 2 times the unit length of the lattice-grid network. However, it is readily seen that the small crosses 7 in Fig. 2, marking the cross-over points of a grid 8,9 straddle the line of the true edge and a calculation of the best line through the grid points is very close to the line 1 of the true edge of the image. Whether the true line 1 is estimated from lattice points 5 and 6 or grid points 7, the error depends on the number of points used and the angle between the true edge 1 and the lattice-grid network 8,9. Suppose that a sufficient number of lattice 5,6 or grid 7 points is available to minimise any variation due to the number of points, then, for a calculation based on lattice points, the error in positioning the line 1 falls from 0.71 units when the line is inclined at 45° to 0.50 units when it is at 0° to a line of lattice points 5,6. However, for a calculation based on grid points the positional error falls from 0.71 units at 45° inclination, linearly to zero as it approa¬ ches 0° inclination, to a grid line 8,9; although at exactly 0° the error is 0.50 units, as for the lattice point calcula¬ tion. The method of measurement, in the particular embodiment, is to observe the panel at an inclination not more than a few degrees away from parallelism with the lattice grid network. The positions of the sides can be measured to within a small fraction of a unit and the dimensions of the panel obtained to high accuracy. At the same time the inclinations of the sides to respective lattice or grid lines can be measured and the parallelism or

otherwise of the sides ascertained.

Should the object to be examined be bounded by sides that are not .straight, its edges can be monitored within tolerance limits by the same method, provided that the lines of the edges bounding the object can be expressed by a mathematical equation. Circles, for example, with radii of 100 units can be measured to 1 part in 10000 in respect of radius, or to within 0.01 units in respect of the location of the centre. The method can be applied to circular arcs (so obtaining curvature) but with an accuracy depending on the length of arc available.

Segments, or lengths of edge, of flat objects for which there may be a need to check dimensions normally lie either between distinctive marks which may be recognized by (or be made recognizable by) a TV camera system, or else between corners or- indentations. The use of a grid boundary of an object (rather than either an interior or exterior lattice boundary) makes the identification of corners or nodes very easy. There are many ways of arriving at an identification, but passing a sequence of grid boundary elements through an appropriate digital filter enables the result to be achieved simply and economically. Once the points have been identified, by whatever method, the position of the edge between two adjacent points, the distance along it, the shape of the edge, or the area of a segment, can be computed, as mentioned above, using the grid points either for a straight edge (as in Fig. 2) or for a curved one.

Referring now to Fig 3, there is shown an object 12 arranged on a surface 13. Above the object, there is a camera 14 having an array of light sensitive elements upon which an image of the object 12 appears. The lighting of the object and the colour of the sampling surface are such that the image formed in the camera is sharp and well contrasted. The light sensitive elements in the camera are arranged in an arra of required pixel density that forms the basis of a digitizable co-ordinate system. The output of the camera is applied to the input of an image analyser and computer

15. A suitable image analyser and computer is sold by

Buehler Ltd of Illinois, U.S.A. under the trade name Omnimet.

In a preferred embodiment, the picture signals are stored in the analyser and computer 15 while the analysis takes place.

The information resulting from the analysis, which relates the representation of the object 12 to the co-ordinate system, is obtained from an output 16 and may be fed, for example, to a printer where it is recorded. In a particular application where a tolerance is being checked, an output signal from 16 is arranged to operate an alarm device which indicates that a tolerance has been exceeded.

It will be understood that the digitizable co-ordinate system may be separate from the means, such as camera 14, for changing the visual image into digital data. The co¬ ordinate system may be an -array associated with the object 12 and the representation of the object which is converted into digital data may be a shadow of the object whose position

bears a known relationship to the co-ordinate system array.

In arrangements measuring the length or position of a straight line, or the straight edge of object, although it is possible to operate the method with the line arranged at any angle less than 45° with respect to an axis of the co-ordinate system, improved results are obtained when the angle is less than 10° and preferably less than 5" .

It will be appreciated that by relating digital data derived from one part of a representation of an object to digital data derived from a co-ordinate system of high resolution, such as a 512 x 512 pixel raster, it is .possible to produce information about the object to a high degree of accuracy. By relating digital data derived from a second part of the representation of the object to the data derived from the co-ordinate system, it is possible to produce second information about the object to a similar high degree of accuracy and to relate the two pieces of information to one another in order to determine the distance between the two parts.

It is also possible to repeat these operations a number of times either with or without relative movement between the imaging device, the object or the measuring grid or lattice, and thereby to make a statistical analysis of the information and improve its accuracy.

Furthermore, by arranging that there is a comparatively small, i.e., less than 45° and preferably less than 5°

angular relationship between a straight edge and an axis of co-ordinate system (lattice or grid) it is possible to achieve extremely accurate results, as the following analysis explains . Considering a digitized picture of a flat object with a pair of parallel or near-parallel edges, the method outlined above can determine the position of each edge accurately. There are many ways of writing the equation of the edge, but for simplicity of calculation we can take the form x sinθ. - y cosθ.. = p .

In this form θ_ is in the angle between the edge and the x-axis and p ₁ is the length of the perpendicular from the origin of co-ordinates onto the line. The calculated position of the edge, obtained from the digitized picture will be x sinθ' - y cosθ' = p' where p' ^. , = P-, ⁺ ι ^anc ^ ®'ι ⁼ ®ι ⁺ ^i ^and oi ^an( ^ ι ^are observational errors. Similarly, the second edge will be calculated to be x sin θ * ~ - y cosθ' = p ' . In monitoring a rectangular panel, the first check is of the parallelism of the sides. A modest number of grid points (less than 100) will give angular errors a

^-1 and £_ less than 0.02°, which would be accurate enough for most purposes, allowing divergence or convergence of the edges at a rate of less than 5 parts per 10000 (i.e. 5mm in 10m). Assuming that the edges were within the tolerance of parallelism, their distance apart would be given by p'. - p' and the error in this estimate would be

? _τ = °- ⁰⁷ - ^τhe S -values are absolute, so it is advisable to make p'_ - p ¹ , as large as conveniently possible. Thus with p' - P' _^ -.. 200, the error is about 1 part in 3000 and it would be correspondingly less if p'., - p' were greater.

In the case of measuring both length and breadth for a rectangular panel it might not be possible to have p' ₂ - P'ι for the width, say, large enough if the correspon¬ ding length measurement, p" - p"-,, say, were to be measured from the same digitized picture. In such cases, it might be necessary to make separate measurements, at different magnifications, of length and width.

It will be understood that, although the invention has been described with reference to particular arrangements, by way of example, variations and modifications may be made within the scope of the accompanying claims; for example, where there is relative movement between an imaging device, the object or the co-ordinate system, the information may be derived during or after the movement. It will be appreciated that views from other directions than above may be taken and analysed, for example from the side.