GB2625546A - Optical characterisation of an object - Google Patents

Abstract

An optical characterisation method of an object such as a diamond or gemstone, the object being partially transparent at a first and second wavelength, in a first step, illuminating the object with a diffuse light of the first wavelength, and generating a digital image of transmitted light, in a second step, illuminating the object with light of a second wavelength approaching the object along an axis, detecting light propagating at a non-zero angle with respect to the axis and generating a second digital image, determining the presence of inclusions or cracks from the first and second images. Also disclosed is another system comprising a first optical detection apparatus, comprising a first sample chamber, diffuse light source and first detector, a second optical detection apparatus comprising a second sample chamber, second light source and second detector.

Description

OPTICAL CHARACTERISATION OF AN OBJECT

Field of invention

The invention relates to the optical characterisation of materials, in particular to the optical characterisation and subsequent classification of gemstones such as diamonds.

Background

Reference is made to manual inspection of objects, such as gemstones, for classification of the objects. The objects may be classified by quality, whereby a higher quality generally corresponds to a higher purity.

Statement of invention

In one aspect there is provided an optical characterisation method of an object, the object being at least partially transparent at a first wavelength of light and at a second wavelength of light, the method comprising: in a first step of the method, illuminating the object with a light source at the first wavelength, wherein the light source is a diffused light source; detecting transmission of light through the object at the first wavelength with a first optical detector, and generating a first digital image; in a second step of the method, illuminating the object with a light source at the second wavelength, wherein the light approaches the object along an optical axis; detecting light propagating at a nonzero angle with respect to the optical axis after interaction with the object, and generating a second digital image; determining the presence of inclusions or cracks in the object from the first digital image and the second digital image.

The method may comprise detecting the silhouette of the object in the first step or in the second step.

The first step may comprise illuminating the object with a plurality of light sources from a plurality of different directions.

The light used in the first step, and/or the second step, may comprise near infrared, and a plurality of visible spectrum wavelengths.

The first wavelength and the second wavelength may be the same or different, or may be part of a broader spectrum of wavelengths.

The first step may be carried out before or after the second step.

The second step may further comprise circularly polarising the light before the light illuminates the object, and blocking circularly polarised light before the detecting.

The non-zero angle may be a right angle.

The method may comprise rotating the object and collecting a plurality of measurements during the rotation.

The method may be a computer-implemented method.

The method may comprise classifying, based on the first and/or second digital image, the object in a quality class.

The classifying step may be carried out using a convolutional neural network.

The convolutional neural network may comprise a first branch configured to process the first digital image, a second branch configured to process the second digital image, and a third branch configured to process the outputs of the first branch and second branch.

The determining step may be a superpixel segmentation algorithm and said classifying step may comprises extracting statistical descriptors from the superpixel segmentation; and classifying the object, using the statistical descriptors, by using a support vector machine.

The object may be a gemstone, preferably a diamond, optionally a rough diamond.

In another aspect there is provided a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of the above aspect.

In a still further aspect there is provided an optical characterisation system for characterising an object, the system comprising: a first optical detection apparatus, comprising a first sample chamber, a diffused light source and a first optical detector; a second optical detection apparatus, comprising a second sample chamber, a second light source, and a second optical detector; wherein the second light source has a main optical axis of light propagation, and wherein the second optical detector has a main optical axis of light detection, and wherein the main optical axis of light propagation is at a non-zero angle to the main optical axis of light detection.

The system may comprise a computer system for processing output signals from the first optical detector and the second optical detector.

The diffused light source may comprise a plurality of light emitting diodes, LEDs, and a diffusion chamber.

The first optical detection apparatus may comprise circular polarising filters configured to filter out reflections from outer surfaces of the object.

The second optical detection apparatus may comprise a dark background, the dark background being arranged on a side of the object opposite to the side at which the second optical detector is arranged. The dark background may be absorbent in the detection wavelengths.

The first optical detection apparatus may further be configured to obtain silhouette images of the object.

The system may comprise a mount configured to support and rotate the object.

The system may comprise a classifier configured to classify, based on the output signals from the first and second optical detectors, the object in a quality class.

The object may be a gemstone, preferably a diamond.

Brief Description of Figures

Some embodiments of the invention will now be described by way of example only and with reference to the accompanying drawings, in which: Figure 1 illustrates a method of optically characterising an object; Figure 2 illustrates a system for carrying out optical characterisation of an object; Figure 3 illustrates an optical arrangement; Figure 4 illustrates an optical arrangement; Figure 5 illustrates an optical arrangement; Figures 6A, 63, 60, 6D and 6E illustrate an optical arrangement; Figures 7A, 73 and 7C illustrate aspects of optical arrangements; Figure 8 illustrates a computer system for carrying out optical characterisation of an object; Figure 9 illustrates classical machine learning; Figures 10A and 103 illustrate feature extraction using classical machine learning; and Figure 11 illustrates Deep Learning -Multi-view CNN architecture

Detailed Description

Described herein with reference to Figures 1 to 11 is an optical characterisation method of an object. The object is at least partially transparent at a first wavelength of light and at a second wavelength of light.

As illustrated in the example of Figure 1, the method comprises a first step of illuminating the object with a light source at the first wavelength, wherein the light source is a diffused light source; detecting transmission of light through the object at the first wavelength with a first optical detector, and generating a first digital image. The method comprises a second step of illuminating the object with a light source at the second wavelength, wherein the light approaches the object along an optical axis; detecting light propagating at a non-zero angle with respect to the optical axis after interaction with the object, and generating a second digital image. The presence of inclusions and/or cracks in the object is determined from the first digital image and the second digital image. An apparatus for carrying out the method is also disclosed.

The first step may be preceded by a step of taking a silhouette image with the light source and detector at the first wavelength, thereby determining the presence of the object, the outline of the object, and the orientation of the object. The first step may further be repeated, during relative rotation of the stone such that multiple images are created from different angles. Each time an image is taken of light transmitted through the object, an accompanying silhouette image may be collected of the outline of the object such that the transmission data can be related to the silhouette and orientation of the object. The second step may also be combined with a relative rotation of the object during repeated image collection, whether by rotating the object while the measurement apparatus is at rest, or by rotating the measurement apparatus around the object.

The first and second steps may be carried out within one measurement chamber, and the object would then remain at the same location when first and second steps are carried out. The different measurements required different light sources, which may be provided within the same chamber and which may be selectively activated depending on the measurement that takes place. Alternatively, the first and second steps may be carried out in separate locations or measurement chambers, whereby the object is moved from a first measurement location to a second measurement location. If the object is moved, the second step may be followed by a step of returning the object to the initial location of the first step such that another silhouette image can be taken to determine whether the object has moved compared to the initial silhouette image taken at step 1. The object is returned to the same relative location by returning an object holder or measurement apparatus back to the same location. If it is confirmed that the object is returned to the same location, within acceptable boundaries of accuracy, then the measurements are considered valid and progressed to the next step of image processing described below. If the object cannot be returned to the same location by returning the object holder, or measurement apparatus, back to the original orientation, then it can be concluded that the object has moved with respect to the holder or measurement apparatus and the measurement process needs to be repeated.

In one example, the non-zero angle is around ninety degrees; said another way, an optical axis of the optical detector is generally perpendicular to the optical axis of the light source.

Figure 2 illustrates one example of a system suitable for carrying out at least aspects of the above-described optical characterisation method. The system generally comprises optical arrangements for "bright-field" (BF) and "dark-field" (OF) imaging of an object. In this specific example, the system also comprises an optical arrangement for silhouette (S) imaging of an object; in another example (not shown here), the optical arrangement for silhouette imaging is omitted. The "bright-field" imaging is used to detect piques in the object and the "dark-field" imaging is used for crack detection. The structural features of these optical arrangements will be described in more detail below.

The inventors have developed a method for detecting a combination of impurities in objects, such as gemstones, and the detection of cracks is part of the overall method, or may be carried out independently. A crack is formed when the regular lattice structure of a crystal is interrupted within the body or at a surface of the crystal. Cracks may also form in an amorphous structure such as glass, which lacks the long range structure of crystals, and some of the same considerations as for crystal structures apply when detecting cracks in amorphous structures.

The crack may be sufficiently large for there to be a cavity between two surfaces facing each other on either side of the crack. The cavity could be small, maybe only a few atomic crystal layers wide. The cavity may be filled with air, or with another solid or liquid material. The change in material will cause a change of refractive index, and therefore cause reflection at one or both of the surfaces. In an amorphous structure such as glass, the presence of a different material also causes reflection despite the lack of crystal layers, and indeed a crack in a glass window tends to be visible because of the presence of air in the crack. The parallel surfaces on either side of the crack may also cause Bragg reflection of incident light, when the spacing between the parallel surfaces is a multiple of the wavelength of the incident light.

A discontinuity in the lattice structure of a crystal also modulates the incident light, even if there is no gap between the sides of the crack. The modulation may be scattering or diffraction of the light. An example of studies of scattering and diffraction from a crack of a square lattice are: Scattering on a square lattice from a crack with a damage zone; Basant Lal Sharma and Gennady Mishuris; Proc Math Phys Eng Sci. 2020 Mar; 476(2235): 20190686; and Diffraction of Waves on Square Lattice by Semi-Infinite Crack; Basant Lal Sharma, SIAM J. Appl. Math., 75(3), 1171-1192. It is noted that diamonds and most other gemstones do not have a square lattice structure, but cubic, but these references are provided as background only for a better understanding of possible mechanisms for light modulation.

The exact underlying mechanism of light modulation does not need to be determined for the purpose of quantitative and qualitative determination of cracks. Cracks are typically also irregular structures, and a combination of reflection and diffraction mechanisms may occur as a result. The inventors have realised that what the different effects of cracks on the propagation of light have in common is an off-axis propagation of light after interaction with a crack in an object, such as a gemstone.

In one example, the gemstone is therefore illuminated by a light beam, preferably a collimated light beam, along an axis. A collimated light beam comprises generally parallel rays which diverge minimally over a short distance. The illumination takes place at a wavelength, or a spectrum of wavelengths, within a transmission spectrum of the gemstone. For example, a diamond is typically transparent in the visible spectrum, so one or more illumination wavelengths in the visible spectrum can be selected. A white-light light emitting diode (LED), emitting a spectrum with main peaks around 450nm and 600nm, together forming white light, and/or light in the near-infrared spectrum for example, between around 750 nm and 2500 nm, is one example of a suitable light source, used in combination with lenses or curved mirrors to collimate the beam. A corresponding optical detector, such as a camera, sensitive in the same spectral range is used for detecting off-axis propagation of reflected or diffracted light.

The inventors have realised that the signal to noise ratio for crack detection in objects such as gemstones, for example diamonds, can be improved further by using a set of circular polarisers around the diamond. A pure diamond is isotropic and the polarisation of light is not changed while it propagates through the diamond. A crack usually causes stress in the surrounding lattice structure, and the stress causes birefringence. The birefringence changes the polarisation of light when the light is reflected or diffracted at a crack, unlike the reflection from an outer surface, whether total internal reflection or reflection from the outer surface. When using a set of circular polarising filters, one before the diamond and one after the diamond, the reflection from outer surfaces is substantially filtered out while the reflection or diffraction from a crack is transmitted to the optical detector. The circular polariser can be provided by a combination of linear and quarter wave plates.

Figure 3 shows a particular setup, optical arrangement or system for carrying out the measurements of cracks described above. Such an arrangement is sometimes referred to as "dark-field" imaging. In this illustrated example, the object 10 to be measured is a rough diamond, but the invention is not limited thereto. As described previously, the inventors have realised that cracks cause incident light to propagate away from the central axis of illumination.

The system illustrated in Figure 3 comprises an optical detector 20 (a camera and a lens), two light sources 30, 40, and a set of polarisers 50. In this example, both light sources are telecentric light sources provided substantially collimated light. An optical arrangement illustrating the use of polarisers 50, as discussed above, is further described with reference to Figure 4.

In this particular example, the main optical axis A of the optical detector 20 is generally perpendicular to an optical axis B of the illumination light from the two light sources, 30, 40, as shown in Figure 3. As discussed above, such an arrangement is sometimes

referred to as "dark-field" imaging.

The setup of Figure 3 will now be described in more detail. Two telecentric light sources 30, 40 are arranged with their main optical axis B substantially perpendicular to the optical axis A of the optical detector 20. Each light source 30, 40 provides collimated light in a same part of the spectrum. The wavelength selection is made depending on the type of stone to generally coincide with the parts of the spectrum where stones are most transparent. In this example, one of the light sources is arranged to illuminate the stone at a wavelength in the visible part of the optical spectrum, for example at around 520 nm, and preferably between 470 nm and 550 nm, while the other one of the light sources is arranged to illuminate the stone at a wavelength in the near infrared (NIR) part of the spectrum, for example, at wavelengths of around 850 nm and preferably between 800 nm and 900 nm, optionally 750 nm to 2500 nm. This wavelength choice is particularly useful for analysing diamonds, because many diamonds are transparent in the visible part of the spectrum, while some diamonds are more transparent in the near infrared part of the spectrum than the visible part.

The setup of Figure 3 can be provided within a chamber (not shown here) to reduce background noise. In this case, the walls of the chamber are provided with a material that absorbs the part or parts of the spectrum used for illumination. This feature further improves the signal to noise ratio of the optical measurement. In the visible part of the spectrum, the corresponding absorption spectrum of the walls is also referred to as a 'black' background, but for the near infrared there is not a particular term. An absorbing coating or material may be used, and the skilled person will be able to select the coating or material to match the particular measurement spectrum. In addition, filters may be used in front of the optical detector to block parts of the spectrum which are not of interest, while letting the relevant parts through to the optical detector.

Referring back to Figure 3, the stone 10 is held by partial vacuum on a nozzle 60 described elsewhere, whereby the nozzle 60 does not obstruct the view of any part of the stone from the optical detector 20 or from the light sources 30, 40. Alternative mounting/holding arrangements may be provided. The stone 10 is rotated by the nozzle 60 around an axis generally perpendicular to a plane in which the optical axes A, B of the optical detector 20 and light sources 30, 40 are arranged. The axis is preferably the main central axis of the nozzle and the stone. Said another way, an axis of rotation of the nozzle 60 is perpendicular to the optical detector 20 and to the two light sources 30, 40. A plurality of two-dimensional images, for example five images, or twenty-five images, is collected during rotation of the nozzle 60 and stone 10. In practice, 5 images is the minimum for image reconstruction with an acceptable confidence level. The combined set of images obtained while the stone 10 is under visible and/or NIR illumination increases the chance of detecting a crack, and also increases the amount of information available for one particular crack.

Alternatively or additionally, an assembly of one or more optical detectors is rotated around the stone to obtain a plurality of images of the stone from different angular positions while the stone remains at rest. Alternatively or additionally, a plurality of optical detectors are equiangularly arranged around the stone, such that each detector obtains at least one image of the stone from a different angular position, and both the optical detectors and the stone remain at rest during the imaging process.

With reference to Figure 4, as discussed above, a set of circular polarising filters, one before the stone and one after the stone (here referred to as the "reflecting surface"), can be used to filter out the reflection from outer surfaces of the stone (or object), while the reflection or diffraction from a crack is transmitted to the optical detector. The circular polariser can be provided by a combination of linear and quarter wave plates.

A three-dimensional (3D) model of the stone (or other object) can be generated based on the plurality of two-dimensional images captured under visible and/or NIR illumination. An example of a method to reconstruct a three dimensional model from a plurality of two-dimensional projections is the inverse Radon-transform, but there are also other methods suitable for this purpose. The processing of the digital images is described in more detail below in an embodiment.

The set of images obtained by the optical arrangement of Figure 3 enables the determination of the presence of a crack, whether there are multiple cracks, and the shape and size of the cracks, as well as the location of the cracks on or within the stone. This determination can assist in deciding how an object such as a rough gemstone should be classified and/or cut.

Besides determining the presence of cracks, a plurality of 2D images of the silhouette of the stone (or other object) can also be detected with the setup and method described above. In this case, in addition to the light sources shown in Figure 3, a third light source 70 is supplied, as shown in Figure 5. Of course, a separate optical arrangement may be provided in which only silhouette images are captured, said arrangement comprising the optical detector 20, light source 70 and nozzle 60, with the other light sources 30, 40 omitted. The optical detector may the same as used for "dark-field" imaging, or may be different, e.g. the system illustrated with reference to Figure 2 may comprise an optical detector for "dark-field" imaging and a separate, second optical detector for silhouette imaging. It will be appreciated that each of these optical detectors may be configured according to their specific imaging function.

In one example, the silhouette images are obtained while the stone is under illumination by substantially collimated, visible light (for example at around 520 nm, and preferably between 470 nm and 550 nm) emitted by the third light source 70, which is a telecentric light source. As shown in Figure 5, the third light source 70 is positioned to be co-axial to the optical detector 20 main axis A. Although the use of polarisers 50 (not shown in Figure 5) will make the silhouette less prominent, the silhouette can still be detected well on the basis of the dark background described above. Optionally, the polarisers 50 can be removed when making the silhouette measurements. When measuring the silhouette, the perimeter (outline) of the stone 10 (or other object) is imaged.

As the digital images generated by the imaging system(s) of Figures 3, 4 and 5 are two-dimensional, a plurality of images from different angles can be used to reconstruct a three-dimensional model of the stone. For example, the silhouette images can be used to create a shape model of the stone.

In an example, twenty-five silhouette digital images of the stone are captured using the optical detector. Alternatively or additionally, one or more optical detectors form an assembly which is rotated around the stone, for example by 360 degrees, so as to obtain a plurality of silhouette images of the stone from multiple angular positions, while the stone remains at rest. Alternatively or additionally, the optical detectors are equiangularly positioned around the stone so that each optical detector captures at least one silhouette image from a different angular position. Alternatively or additionally, the stone is rotated by at least 360 degrees while the one or more optical detectors obtain a plurality of images of the stone at multiple angular positions.

It will be appreciated that the number of digital images obtained off-axis under visible and/or NIR illumination, and the silhouette images, optionally obtained co-axially under visible light illumination, can vary according to the application. Additionally, the number of images obtained under NIR illumination can be greater than the number of silhouette images, and vice versa. The sets of images obtained off-axis under visible and/or NIR illumination can be combined with the silhouette images obtained co-axially under visible illumination can be combined to produce one or more 3D models, by combining 2D images and/or by combining existing 3D models.

While obtaining silhouette images of an object has been described above in combination with obtaining "dark-field" images, alternatively or additionally, the silhouette images can be obtained in combination with obtaining "bright-field" images, described in more detail below.

The inventors have developed a method for detection of the inclusion of other materials in an object, such as a gemstone, for example a rough diamond. An inclusion may be accompanied by a crack as described above, because the inclusion of another material within a crystal lattice structure will cause strain in the lattice and in some instances the strain is sufficient to cause a crack. An inclusion is sometimes also referred to as a pique in the context of diamonds, or other gemstones, and sometimes a pique is understood as an inclusion visible to the naked eye. The present techniques are not limited to the detection of piques, even if reference is made to piques, and are instead applicable to inclusions more generally.

The inventors have realised that an inclusion generally causes absorption of light which otherwise would be transmitted through the stone. In some cases, the inclusion causes diffraction, reflection, or other propagation of light which is not absorption, and in those cases the inclusion can be detected with the method described previously in relation to cracks. In most cases, however, absorption is the primary effect of an inclusion.

Absorption prevents the propagation of light along the direction of travel of the light, and absorption can therefore be measured indirectly by detecting the absence of propagated light. The stone, or other object, is illuminated by a light source and an optical detector measures the transmitted light. In particular, an optical detector such as a camera can be used to image the stone, and the inclusions will appear as dark areas within the stone.

As discussed above, the stone is held on a mount, such as a nozzle or dop, whereby the mount does not obstruct the view of any part of the stone from the optical detector or from the light source.

The illumination light source may in general be a collimated or un-collimated light source such as a point source. However, as described above in relation to the detection of cracks, the incident light will be reflected or diffracted at the multiple outside surfaces of the stone as well as at any internal cracks or at surface cracks. Those reflections are not relevant for the detection of inclusions, and may even make the detection of inclusions more challenging. For example, a strong reflection at an internal crack will also appear as a darker region when facing the light source because the straight propagation along the optical axis is reduced or prevented. When using automated detection, an algorithm analysing the image may not be able to tell the difference between an inclusion and a crack.

The inventors have realised that the perturbation from cracks or surfaces can be reduced by using a diffused light source. While optical effects such as reflection or diffraction strongly depend on the direction of light, the optical effect of absorption only weakly depends on the direction of incident light. By using a diffused light source, for example, a white light or an NIR diffused light source, the light approaches the stone from many different directions and the effects from reflection or diffraction, or the like, are 'averaged out', but the effect from absorption is not diminished and will be more prominent relative to the other effects. The above-described illumination is sometimes referred to as bright-field imaging.

A single diffused light source can be used, for example a single LED with a diffuser placed in front. The detection of inclusions can be improved further by using multiple diffused light sources illuminating the stone from multiple different directions. The multiple sources can be distributed evenly around the stone, and also above and below the stone. The effect of averaging out described above is maximised by distributing the directions of incoming light across the entire sphere surrounding the stone, as much as practical while leaving space for other parts of the apparatus such as the holder and the optical detector. The skilled person will be able to implement this concept with different light sources. A plurality of LEDs with diffusers is one example. Another example is using a leaky optical fibre, which emits light coupled into the fibre at one end along the length of the fibre. A leaky fibre can be created by bending a fibre, for example by winding an optical fibre in one or more loops around the measurement location. One or more reflectors can also be used to distribute light over multiple directions.

The wavelength of the light is selected in a part of the spectrum where the main body of the stone is transparent, but the inclusions are absorbent. In one example, the same examples are provided as described for the crack detection above, such as a 'white light' source (e.g. a visible light source emitting light at around between 380 to 700 nm) in combination with a near infrared (NIR) light source, emitting light in the range 870 nm +150 nm, optionally 750 nm to 2500 nm.

Figures 6A and 6B illustrate some specific aspects of an apparatus, or optical arrangement, for carrying out the "bright-field" measurement of piques and/or inclusions. Figure 6A illustrates a side view of a "bright-field" light cell 400 and an optical detector 200, which in this example comprises a camera and lens.

Figure 6B provides an alternative view of the arrangement of Figure 6A, by illustrating an interior of the light cell 400 in a vertical cross section.

In this illustrated example, the light cell 400 comprises a set of LED panels 41 arranged to illuminate a chamber with stone 10. The LED panels comprise a set of LEDs on a board such as a PCB. Other light sources than LEDs may also be used, such as lasers or incandescent light bulbs, but LEDs are particularly suitable due to their advantages of size, low heat emission and durability. The LED panels are held in place by a frame 43. Figure 6C is a perspective view of the light cell, while Figure 6D is a perspective view of the cell cut open along a vertical plane. Figure 6E is a perspective view of a diffusing filter 42. The diffuser has a cylindrical body, with flat top and bottom portions closing the top and bottom of the cylinder. The diffuser is arranged between the LED panels and the stone, such that the light emitted by the LED panels is diffused before illuminating the stone. The emitted light propagates through the diffuser and leaves the diffuser distributed over a large solid angle. The cylinder of the diffuser has an opening 47 for the camera to be able to view the stone. Figure 6D shows parts of the LED panels, each comprising 20 to 30 LEDs, but a larger or smaller number of LEDs can be used for each panel. Figure 60 shows that the panels define a hexagonal chamber arranged around the diffuser 42, whereby 5 of the walls of the chamber comprise LED panels, and the sixth side defines an opening 48 that is aligned with opening 47 of the diffuser. A larger or smaller number of panels than 5 may be used. A hinged top 51 is provided, connected to the rest of the housing by way of hinges 49, so that the chamber can be opened, for example to retrieve a stone that has fallen off a holder. A hinged electrical connection 50 is included for providing an electrical connection to the LEDs in the hinged top 47.

Figure 70 illustrates the top 51 being opened like a lid to access the interior of the chamber.

Figure 7A illustrates an LED 71, with a diffuser 72 placed in front of the emitting surface.

Figure 7B illustrates a different embodiment than the one illustrated in Figures 6A-E. The main difference with the Figure 6 embodiment is the number of LED panels, which is larger in this example. In the illustrated example, the frame 43 comprises a generally flat and circular top 44 connected to a generally flat and circular bottom section 45 by spacers 46.

In one example, the stone 10 (or other object) is held by a nozzle 600 (see Figure 6B, for example) under partial vacuum in the middle of the chamber and is rotated by 360 degrees to collect a plurality of digital images, captured by one or more optical detectors 200, such as cameras. Alternatively or additionally, one or more optical detectors, such as cameras, are rotated around the stone to obtain multiple images from different angular positions. Alternatively or additionally, the optical detectors are equiangularly positioned around the stone such that each optical detector obtains one or more images of the stone from a different angular position.

In one example, the nozzle 600 may be actuated to insert the stone 10 into the chamber 43 and to remove the stone 10 therefrom once imaging is complete. Once the stone 10 has been inserted into the chamber interior, the nozzle 600 and chamber 43 may be configured such that only light emitted by the one or more LEDs 41 enters the chamber interior.

The plurality of 2D digital images obtained of the stone under visible and/or NIR bright-field illumination can be used to create a 3D model of the stone, as discussed above.

As discussed above, the optical assembly which obtains "bright-field" images of an object, such as a gemstone, can in one example be modified to additionally obtain silhouette images of the object, for example, using an additional light source and/or additional optical detector.

It will be appreciated that the order of the steps of the method described above, in which digital images of an object, such as a gemstone, are obtained under different lighting conditions, can be varied according to the specific application.

The obtained images may be analysed by a user, or preferably by a computer. Figure 8 illustrates a computer 80, with a processor and memory 81, a data receiving port 82, a data emitting port 83, a user interface including a screen 84 and user input means such as a keyboard and mouse 85.

In one example illustrated in Figure 9 of a computer implemented method for optically characterising an object, such as a gemstone, the plurality of 2D dark-field digital images is passed into a crack feature extraction algorithm 96. Similarly, the plurality of 2D bright-field images is passed into a pique feature extraction algorithm 94. In one example, the plurality of 2D digital silhouette images is passed into the crack feature extraction algorithm 96 and/or the pique feature extraction algorithm 94. Alternatively, the plurality of 2D silhouette images is passed into a shape feature extraction algorithm 92 or is incorporated or otherwise combined with an output of one or both of said pique feature extraction algorithm 94 and/or crack feature extraction algorithm 96. Where used, silhouette images are used to determine the shape and/or boundary of the gemstone or other object.

Both feature extraction algorithms may be a superpixel segmentation algorithm. In one example said superpixel segmentation algorithm may be a Normalized Cut algorithm. Other superpixel segmentation algorithms are possible, such as Watershed, Constant Intensity Superpixels, Eikonal Region Growing Clustering, or Topology Preserving Superpixels. The skilled person will appreciate that any superpixel segmentation algorithm can be similarly applied. It will also be appreciated that the pique feature extraction algorithm and crack feature extraction algorithm may utilise different superpixel segmentation algorithms.

The superpixel segmentation algorithm may output a set of superpixels from which various statistics of each superpixel can be computed. The statistics may include pique and/or crack area, pique and/or crack distance, pique and/or crack distance to area ratio, clean area, clean distance, and clean distance to area ratio. Furthermore, statistical descriptors for each of these statistics may be computed such as minimum, maximum, mean, standard deviation, skewness, and kurtosis.

The statistical descriptors may be used to classify the shape and quality of the gemstone. Preferably, the gemstone may be classified using a Support Vector Machine (SVM). The gemstone may alternatively be classified using a Random Forest algorithm or k-Nearest Neighbour. The skilled person will appreciate that any machine learning classification algorithm may be appropriate.

The superpixel segmentation may be useful for deciding how/where to cut a gemstone, for example, to maximise the quality of the resulting cut stone and/or to maximise the use of material.

Figure 10A provides an illustration of a bright field image of a diamond, and superpixel segmentation of the image in Figure 10B.

It will be appreciated that in one example the algorithms comprise software configured to be run by a processor of a computer and/or computer system. The computer system can be an integral component of the apparatus carrying out the characterisation, or can be provided separately. In one example, the apparatus and the computer system are connected via a wired or wireless connection over a network. In one example, the computer system is a cloud computer system. In one example, the processor is configured to perform actuation of one or more of: the light sources, the optical detectors, the rotation of the object.

In another example of a computer implemented method for optically characterising an object, such as a gemstone, a multi-view Convolutional Neural Network (CNN) algorithm 110, such as illustrated in Figure 11, is used to classify the quality of the gemstone. The multi-view CNN 110 may comprise a dark-field branch 112 and a bright-field branch 114. In the dark-field branch, each of the plurality of the 2D dark-field images 116 are processed individually through separate CNNs 118. Each CNN takes as input a specific dark-field image corresponding to a specific view of the gemstone. Each CNN applies a sequence of convolution operations interspersed by pooling operations (such as Max Pooling, Max Unpooling, and Transpose Convolution). The number of different bright field images is 25, and the number of dark field images is also 25, or another number of images can be used.

The resulting output of each separate CNN is collectively passed through a view-pooling layer 120. The view-pooling layer 120 takes as input the output of each separate CNN of the dark-field branch 112 and output a single 2D image of the same height and width as each 2D dark-field view image. The view-pooling layer 120 creates a new image by comparing pixels with corresponding coordinates from each input image and write the value resulting from said comparison into the corresponding coordinate of the new image. Comparing the pixels may include selecting the maximum value, minimum value, or average value of the set of pixels with corresponding coordinates.

The output of the view-pooling layer may be input into a final dark-field branch CNN 122. The output of the final dark-field branch CNN 122 is the output of the dark-field branch 112 The bright-field branch 114 operates in the same manner as the dark-field branch 112, but using the bright-field views as input instead. As such, the bright-field branch 114 will not be discussed in in detail as well.

The output of the dark-field branch 112 and bright-field branch 114 are combined via a view-pooling layer 124. Said view-pooling layer operates in substantially the same way as the view-pooling layer in the dark-field branch and the bright-field branch.

The combination of the two branches provides a synergistic effect. The characterisation of a pique or crack can be carried out more accurately when both branches are combined, than when only one branch is present.

The output of said view-pooling layer may be input into a final CNN 126. Preferably, the final CNN 126 outputs a vector with the same number of entries as quality classes. The skilled person would understand that it is possible to use various flattening layers and dense layers to reach the appropriate output dimension required.

Each coordinate of the output vector is associated with a quality class. For example, the first coordinate is assigned to the highest quality class, and the last coordinate is assigned to the lowest quality class. The skilled person appreciates that the assignation is not important as long as it is consistent with each use.

The output vector is passed through a soft-max layer 128. The soft-max layer 128 converts the values of each coordinate of the output vector into a probability. The probability corresponds to the probability that the gemstone falls within the quality class corresponding to that coordinate.

The quality class with the highest associated probability may be chosen to be the classified quality 130 of the gemstone.

In order for the multi-view CNN 110 to accurately classify the gemstone, the multi-view CNN 110 may be trained in a supervised fashion via gradient descent. A dataset of bright-field images and a dataset of dark-field images are collected and classified by hand, the datasets are then split into three subsets labelled training dataset, validation dataset, and testing dataset. Each image in the dataset has a corresponding quality class to which the gemstone belongs to.

Training occurs through at least one epoch. In each epoch, the training dataset is split into batches of equal size. Each batch is passed through the multi-view CNN 110. Once the soft-max output is computed, the vector of probabilities is compared to the actual quality class using a loss function. The quality class is encoded as a vector with a value of zero in each coordinate and a value of one in the coordinate corresponding to the actual quality class.

Examples of loss functions for classification may include multi-class cross-entropy loss, square loss, or Kullback-Leibler Divergence loss. Using the computed loss, the parameters of the multi-view CNN 110 can be further optimised using backpropagation. Examples of optimisation algorithms are Adam, AdaGrad, and RMSProp.

After all the training dataset has been passed through the multi-view CNN 110 and the loss has been backpropagated, the validation step can occur. During this step, the validation dataset is passed through the multi-view CNN 110, the output is then compared to the corresponding actual quality class and the loss can be computed as well as any other desired validation metric such as accuracy.

The training step and validation step form a single epoch. Once a certain stopping criterion is met (for example a maximum number of epochs or the validation metric has not improved over some previous number of epochs), the training of the multi-view CNN 110 is terminated and the testing step begins. The resulting multi-view CNN 110 is fully trained and can be subsequently used in the manner described above to classify gemstone quality.

The testing step consists of passing the test dataset through the multi-view CNN 110 and comparing the output to the corresponding actual quality class. From this the test loss and test metrics are computed. These then form the final metrics for the trained model.

The method of analysing the digital images is implemented on a computer, a distributed computer system, a cloud computing system, or other standard systems known to the skilled person.

Using the methods and systems described herein, an object such as a gemstone (e.g. a rough diamond) may be optically characterised, optionally providing a classification of the gemstone and/or a determination of how/where the stone should be cut. In some examples, a 3D digital model of the object may be produced.

Although the invention has been described in terms of preferred embodiments as set forth above, it should be understood that these embodiments are illustrative only and that the claims are not limited to those embodiments. Those skilled in the art will be able to make modifications and alternatives in view of the disclosure which are contemplated as falling within the scope of the appended claims. Each feature disclosed or illustrated in the present specification may be incorporated in the invention, whether alone or in any appropriate combination with any other feature disclosed or illustrated herein.

Although some embodiments are described with reference to gemstones, such as diamonds, the skilled person will appreciate that the apparatus and methods described herein can be similarly applied to other objects.

Claims (25)

1. CLAIMS: 1. An optical characterisation method of an object, the object being at least partially transparent at a first wavelength of light and at a second wavelength of light, the method comprising: in a first step of the method, illuminating the object with a light source at the first wavelength, wherein the light source is a diffused light source; detecting transmission of light through the object at the first wavelength with a first optical detector, and generating a first digital image; in a second step of the method, illuminating the object with a light source at the second wavelength, wherein the light approaches the object along an optical axis; detecting light propagating at a non-zero angle with respect to the optical axis after interaction with the object, and generating a second digital image; determining the presence of inclusions or cracks in the object from the first digital image and the second digital image.
2. 2. The method of claim 1, further comprising detecting the silhouette of the object in the first step or in the second step.
3. 3. The method of claim 1 or 2, wherein the first step comprises illuminating the object with a plurality of light sources from a plurality of different directions.
4. 4. The method of any one of claims 1 to 3, wherein the light used in the first step, and/or the second step comprises near infrared, and a plurality of visible spectrum wavelengths.
5. 5. The method of any one of claims 1 to 3, wherein the first wavelength and the second wavelength are the same or different, or part of a broader spectrum of wavelengths.
6. 6. The method of any one of the preceding claims, wherein the first step is carried out before or after the second step.
7. 7. The method of any one of the preceding claims, wherein the second step further comprises circularly polarising the light before the light illuminates the object, and blocking circularly polarised light before the detecting.
8. 8. The method of any one of the preceding claims, wherein the non-zero angle is a right angle.
9. 9. The method of any one of the preceding claims, further comprising rotating the object and collecting a plurality of measurements during the rotation.
10. 10. The method of any one of the preceding claims, wherein the method is a computer-implemented method.
11. 11. The method of any one of the preceding claims, further comprising classifying, based on the first and/or second digital image, the object in a quality class.
12. 12. The method of claim 11, wherein said classifying step is carried out using a convolutional neural network.
13. 13. The method of claim 12, wherein the convolutional neural network comprises a first branch configured to process the first digital image, a second branch configured to process the second digital image, and a third branch configured to process the outputs of the first branch and second branch.
14. 14. The method of claim 11, wherein said determining step is carried out with a superpixel segmentation algorithm and said classifying step comprises: extracting statistical descriptors from the superpixel segmentation; and classifying the object, using the statistical descriptors, by using a support vector machine.
15. 15. The method of any one of the preceding claims, wherein the object is a gemstone, preferably a diamond.
16. 16. A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of claims 10 to 15.
17. 17. An optical characterisation system for characterising an object, the system comprising: a first optical detection apparatus, comprising a first sample chamber, a diffused light source and a first optical detector; a second optical detection apparatus, comprising a second sample chamber, a second light source, and a second optical detector; wherein the second light source has a main optical axis of light propagation, and wherein the second optical detector has a main optical axis of light detection, and wherein the main optical axis of light propagation is at a non-zero angle to the main optical axis of light detection.
18. 18. The system of claim 17, further comprising a computer system for processing output signals from the first optical detector and the second optical detector.
19. 19. The system of claim 17 or 18, wherein the diffused light source comprises a plurality of light emitting diodes, LEDs, and a diffusion chamber.
20. 20. The system of any one of claims 17 to 19, wherein the second optical detection apparatus comprises circular polarising filters configured to filter out reflections from outer surfaces of the object.
21. 21. The system of any one of claims 17 to 20, wherein the second optical detection apparatus comprises a dark background, the dark background being arranged on a side of the object opposite to the side at which the second optical detector is arranged, and wherein the dark background is absorbent in the detection wavelengths.
22. 22. The system of any one of claims 17 to 21, wherein the first optical detection apparatus is further configured to obtain silhouette images of the object.
23. 23. The system of any one of claims 17 to 22, further comprising a mount configured to support and rotate the object.
24. 24. The system of any one of claims 18 to 23, further comprising a classifier configured to classify, based on the output signals from the first and second optical detectors, the object in a quality class.
25. 25. The system of any one of claims 17 to 24, wherein the object is a gemstone, preferably a diamond, optionally a rough diamond.