Anomaly capture system and process

Abstract

An anomaly capture system comprising: an illumination system; and image collection apparatus; wherein the illumination system comprises at least one light source configured for directionally illuminating a surface, the illumination system being configured to selectively illuminate the surface from a selected direction when seen in a plan view of the surface and / or at a selected illumination angle to the plane of the surface; and the image collection apparatus is configured to image the surface whilst the surface is being illuminated from the selected direction and / or at the selected illumination angle to the plane of the surface.

Description

[0001] ANOMALY CAPTURE SYSTEM AND PROCESS

[0002] FIELD

[0003] The present disclosure relates to anomaly detection, such as detection of cracks and other defects in structures, particularly but not exclusively on surfaces of structures, which could be concrete surfaces.

[0004] BACKGROUND

[0005] Detection and characterisation of anomalies, such as defects, cracks, deterioration, damage and the like, is of critical importance in many industries. For example, detection and characterisation of anomalies in civil and architectural structures, particularly concrete structures, is becoming increasingly important due to the increased environmental and economic pressure to extend the lifespan of aging infrastructure.

[0006] Manual visual inspection is a common approach when inspecting for anomalies. Manual inspection has a number of advantages, such as the flexibility and the adaptable, real-time decision-making ability of humans. However, there are also disadvantages, including cost, repeatability, safety, availability, time to train, the chance of human error, and others.

[0007] Automated inspection systems and data analysis techniques can be used to address many of these shortcomings in manual inspection, potentially providing a safe, remote, consistent and time efficient alternative to manual inspection. However, anomalies can be diverse and are found in a range of different conditions and situations. Some automated solutions may lack the adaptability required to reliably identify, assess and characterise anomalies across the entire diversity of anomaly types and environments.

[0008] At least one example described herein seeks to improve automated inspection systems and methods.

[0009] SUMMARY

[0010] The present inventors have identified that directional lighting can be of significant benefit in the automated identification of anomalies in surfaces, and that particular ways of applying the directional lighting can be used to further enhance those benefits. Directional lighting is generally incident on a surface in particular directions and / or angles to the surface when seen in a plan view of the surface. This is in contrast to diffuse lighting which has no preferred direction. Described herein are example in which light is projected on to a surface at a purposely varied angle and direction in order to improve contrast between the anomaly and background. This is in contrast to ambient or diffused illumination, where light is projected on to a surface from all directions and all angles at once. Furthermore, the present inventors have identified ways of processing images of surfaces to enhance detection and / or characterisation of anomalies, particularly but not exclusively in combination with directional lighting techniques.

[0011] At least one aspect of the present disclosure is defined in the independent claims. Preferred features are defined in the dependent claims.

[0012] According to a first example of the present disclosure is an anomaly capture system comprising: an illumination system; and image collection apparatus; wherein the illumination system comprises at least one light source configured for directionally illuminating a surface, the illumination system being configured to selectively illuminate the surface from one or more selected directions when seen in a plan view of the surface and / or at one or more selected illumination angles to the plane of the surface; and the image collection apparatus is configured to image the surface from one or more directions whilst the surface is being illuminated from the one or more selected directions and / or at the one or more selected illumination angles to the plane of the surface.

[0013] The illumination system may be configured to illuminate the surface from any of a plurality of directions and / or from any of a plurality of illumination angles to the plane of the surface. The illumination system may be a non-diffuse illumination system. However, in some examples, the illumination system may be configured to selectively illuminate the surface with diffuse light, e.g. in order to collect one or more images of the surface illuminated from the one or more selected directions and / or at the one or more selected illumination angles to the plane of the surface and one or more other images of the surface illuminated the surface with diffuse light. Each light source may be arranged to emit light in a limited range of angles. Each light source may be configured to emit a light profile centred on an optical axis of the light source. Each light source may be configured to emit light preferentially in a specific direction, which may be or may not be centred on the optical axis of the light source. Each light source may comprise at least one, and optionally a plurality of LEDs, such as white or RBG LEDs. However, in other examples, different types of light source could be used. The LEDs or other light emitters may be provided in an array, such as a linear array. The light emitters in the array may be individually controllable, e.g. to selectively illuminate individual light emitters in the array or subsets of the light emitters in the array

[0014] The illumination system may comprise a support. The image collection apparatus may be mounted on, or otherwise supported by, the support. The one or more light sources may be supported by the support, e.g. directly or indirectly mounted to the support.

[0015] At least one or each of the light sources may be mounted to the support via at least one manipulator for manoeuvring the light sources, e.g. to adjust the location and / or orientation of the respective light source. Respective light sources may be mounted to the support via at least one respective manipulator.

[0016] The at least one manipulator may comprise an arm, such as an articulated arm, which may be a robotic arm. The at least one arm may comprise at least one, and optionally a plurality of joints. The arm may be movable at each joint, e.g. under the action of a driver, such as a servo-motor or other motor, an actuator such as a pneumatic or hydraulic actuator, a piston, and / or the like. The arm may comprise at least one or a plurality of rigid sections. At least one joint may be configured to rotate at least one of the joints relative to at least one other joint and / or the support. At least one of the manipulators may comprise a rotator configured to rotate or otherwise reorient the respective light source. At least one of the manipulators may be configured to reposition the respective light source, e.g. to move the at least one light source to a different location.

[0017] The manipulators may be controllable by a controller, e.g. so that the at least one light source is selectively and controllably movable and / or re-orientable under the control of the controller. The manipulators may be controllable by the controller so as to control and / or vary the direction and / or illumination angle that the light from the respective light source is incident on the surface. Although manipulators can beneficially be used, other means for controlling and / or varying the direction and / or illumination angle that the light from the respective light source is incident on the surface could be used. For example, selective illumination of the surface by differently facing light sources could be used, or one or more of the light sources could movable and / or reorientable by robotic entities, crawlers, drones or other transportation and / or reorientation mechanisms.

[0018] The illumination system may comprise a plurality of light sources. At least one or each light source may be configured to illuminate the surface from a different direction to at least one of each other light source, e.g. without re-positioning or re- orienting the light sources. At least one or each of the light sources may be mounted in the illumination system, e.g. on the support, so that it faces in a different direction to at least one or each other light source. Each light source may be mounted on a different manipulator, e.g. on a different arm. Each light source may be independently or individually movable and / or re-orientable. Each light source may be independently or individually selectively activatable to provide illumination and deactivatable to turn off illumination from that light source. Each light source may be independently or individually selectively activatable to provide illumination and deactivatable to turn off illumination from that light source to selectively illuminate the surface from different directions, e.g. from different sides of the surface when seen in a plan view, and / or at different illumination angles to the plane of the surface, whilst different images or parts of the video are collected.

[0019] Although directional and / or angular illumination of the surface is described above, other differential lighting arrangements could be used in addition to the directional and / or angular illumination, e.g. illuminating the surface with different colours I spectral components (e.g. collecting different images of the surface whilst illuminated with different wavelengths of light), different types of lighting (e.g. illuminated by two or more of: LED, Fluorescent, halogen or laser light), illuminating with different projection methods or patterns (e.g. spot, dome, line and / or the like, which may be chosen to match the feature), different polarisations, different neutral densities, other different filter effects, and / or the like.

[0020] The illumination system may be configured to illuminate the surface with a selected colour and / or with a selected wavelength or wavelength band, which may be selected to according to a material of the surface or a material on the surface or a type of anomaly. For example the images may be selectively illuminated with blue light, which may be particularly suitable for imaging red corrosion. As another example, illumination with green lighting can be used to reduce or render negligible the effect of plant or other organic growth such as moss or algae on the surface. The anomaly capture system may comprise a shroud for at least partially blocking environmental or background light, such as sunlight, general room, background diffuse lighting and / or the like. The image collection apparatus may be mounted on a fixed support or the shroud. The illumination system may be mounted on the fixed support or the shroud. The shroud may comprise one or more facets, e.g. the shroud may be multifaceted. The shroud may comprise at least three, e.g. four or more sides. The shroud may define an inner surface, which may be concave, e.g. may define a recess. The inner surface may comprise a back surface and one or more side walls. The side walls may extend around, e.g. completely around, a perimeter of the back wall. The side walls may extend obliquely or perpendicularly to the back wall. The one or more facets may be provided in the side walls of the inner surface of the shroud. At least one image collection apparatus may be provided in or on the back wall. The at least one light source may be mounted or otherwise provided on an inner surface of the shroud. At least one light source may be provided on one or more or each facet. At least one light source, e.g. at least one LED or a strip or array of LEDs, may be provided on each of two or more or each facet. At least one or each facet may be angled differently to at least one or each other facet. A facet may extend around at least one, e.g. two or more or each side of the side wall. A light source may be provided on at least one or each facet on at least one or each side of the shroud. In this way, the surface may be selectively illuminated from different directions and / or different angles to the plane of the surface by selectively activating different light sources provided on different facets of the inner surface of the shroud. At least part or all of the inner surface of the shroud may be black and / or matt, e.g. matt black. This may reduce reflection, improve the directionality of the illumination and avoid significant diffuse illumination.

[0021] In an alternative arrangement, the light source, e.g. at least one light source per side could be provided on a motorized rail or other manoeuvring system in order to move the light source, e.g. to vary the direction and / or angle of illumination. In another variation, the light sources may be mounted on the at least one manipulator inside the shroud.

[0022] The image collection apparatus may be controllable, e.g. by the controller, to capture at least one or a plurality of images or video of the surface. The images may comprise greyscale and / or colour images. The illumination system may be configured to selectively illuminate the surface from different directions, e.g. from different sides of the surface when seen in a plan view, and / or at different illumination angles to the plane of the surface, whilst different images or parts of the video are collected. The illumination system may be configured to selectively illuminate the surface with different types of illumination. The different types of illumination may comprise diffuse lighting and / or hard (non-diffuse) lighting. Respective images or parts of the video may show the surface being illuminated from different directions, e.g. from different sides when seen in a plan view, and / or from different illumination angles with respect to the plane of the surface and / or with different types of illumination. Respective images of the at least one, e.g. a plurality of images or parts of the plurality of parts of the video may show the surface being illuminated from different directions that are opposite and / or perpendicular to each other in a plan view of the surface. Respective images or parts of the video may show the surface being illuminated from at least two, three four or even eight or more sides, e.g. all of: left, right, up and / or down when viewed in a plan view of the surface. At least one of the images or at least part of the video may be of the surface whilst being illumination by one direction when viewed in a plan view of the surface, and another image or another part of the video may be of the surface being illuminated from a different direction when viewed in a plan view of the surface and one or more further images or parts of the video may each respectively be of the surface being illuminated from one or more further different directions when viewed in a plan view of the surface. The image collection apparatus may be controllable, e.g. by the controller, to capture at least one, e.g. a plurality of images or video of the surface sequentially illuminated from different directions, e.g. sequentially over at least two and optionally at least three or at least four or at least eight or more different directions. Different images or different parts of the video may show the surface being illuminated from different directions from among the at least two and optionally at least three or at least four different directions.

[0023] The illumination system may be configured to dynamically change the direction in which the surface is illuminated, e.g. by moving and / or re-orienting one or more or each of the light sources or by providing different light sources facing in different directions and / or at different angles to the plane of the surface and selectively illuminating only one or a sub-set of the light sources to vary the direction and / or angle of illumination of the surface. The illumination system may be configured to dynamically change the direction in which the surface is illuminated based on a quality value of a resultant image, e.g. to maximise or optimise or improve the quality value of the images collected. The quality value may be a defined measure of image quality and / or any other metric indicative of the system’s ability to distinguish anomalies, such as uncertainty, training efficacy, and / or the like.

[0024] The illumination system may be configured to obliquely illuminate the surface. The illumination system may be configured to arrange the light sources so that they are angled obliquely to the plane of the surface, e.g. when illuminating the surface. The illumination system may be configured to arrange the light sources so that the optical axis of the light sources is angled obliquely to the plane of the surface, e.g. when illuminating the surface. The light sources may be arranged, or the illumination system may be configured to arrange the light sources, so as to illuminate the surface at an angle to the plane of the surface in a range from 0° to 90° but preferably from 5° to 50°, e.g. from 20° to 50°, such as from 30° to 40°, inclusive. For example, the illumination system may be configured to arrange the light sources so that the optical axis of the light sources is at an angle to the plane of the surface in a range from 5° to 50° when illuminating the surface. The illumination system may be configured to illuminate the surface from different illumination angles to the plane of the surface when different images or parts of the video being collected, or at the same illumination angle to the plane of the surface. The illumination system may be configured to reorient the light sources so as to illuminate the surface from different angles to the plane of the surface when different images or parts of the video being collected. Respective images or parts of the video may show the surface being illuminated from different angles to the plane of the surface.

[0025] The illumination system may be configured to dynamically change the illumination angle to the plane of the surface at which the surface is illuminated, e.g. by moving and / or re-orienting one or more or each of the light sources. The illumination system may be configured to dynamically change the illumination angle to the plane of the surface at which the surface is illuminated based on a quality value of a resultant image, e.g. to maximise the quality value of the resultant image.

[0026] The one or more image collection apparatus may be digital, e.g. a digital camera. The image collection apparatus may comprise an RGB camera, an infra-red camera, a hyperspectral camera, a multispectral camera, a UV camera and / or the like. The one or more image collection apparatus may comprise a CMOS, CCD or other suitable image collection sensor.

[0027] The one or more image collection apparatus may be configured to image the surface from at least one of: one or more directions; at one or more angles; and / or from one or more different locations, e.g. whilst the illumination system illuminates the surface. The one or more image collection apparatus may be reconfigurable or otherwise configured to image the surface from a different direction and / or at a different angle to the plane of the surface. The one or more image collection apparatus may be configured to collect at least one, e.g. a plurality of images of the surface whilst being illuminated by the illumination system, wherein at least one of the images may be collected by at least one of the one or more image collection apparatus from at least one of: a different direction, a different location, a different type of illumination and / or a different angle to the plane of the surface, to at least one other image. The at least one of the images may be collected by at least one of the image collection apparatus that faces a different direction, is located at a different location and / or is oriented at a different angle to at least one other of the image collection apparatus that collects another image. The at least one of the images may be collected by at least one of the image collection apparatus, wherein the at least one of the image collection apparatus is reconfigurable such that it faces a different direction, is located at a different location and / or is oriented at a different angle to the plane of the surface, and then operable to collect a different image whilst it faces the different direction, is located at the different location and / or is oriented at the different angle.

[0028] In examples, different images of the plurality of images may be of the surface being illuminated from opposing directions, and further different images may be of the surface in at least one pair of different opposing directions, which may be perpendicular to at least one other pair of opposing directions shown in the images. The images of the surface illuminated from different directions may be under non-diffuse (hard) lighting. At least one other of the images may be of the surface under diffuse lighting. In one example, the images may comprise images of the surface taken when illuminated using non-diffuse (hard) light from the up, down, left and right directions when viewed in a plan view and of the surface taken under diffuse light, with all images optionally being greyscale images.

[0029] The image collection apparatus may be mounted to the support via one or more manoeuvring apparatus, such as an actuator, robotic arm, one or more rotatable joints and / or the like. The manoeuvring apparatus may be configured to relocate and / or reorient at least one of the one or more the image collection apparatus, e.g. so as to vary the direction and / or angle at which the at least one image collection apparatus images the surface. The manoeuvring apparatus may comprise powered manoeuvring apparatus, e.g. comprising a motor such as a stepper motor or servo motor or the like, an actuator, a pneumatic or hydraulic actuator, a piston arrangement, an electroactive polymer actuator, or the like, which may be controllable by the controller. The manoeuvring apparatus may be configured to relocate and / or reorient at least one of the one or more the image collection apparatus, e.g. responsive to commands from the controller. The manoeuvring apparatus may be configured to adjust at least one of: pitch, roll and / or yaw of at least one of the imaging apparatus. At least one or each of the imaging apparatus may be individually and / or separately reconfigurable, e.g. under the action of an associated manoeuvring apparatus.

[0030] At least one or each of the imaging apparatus may be individually and / or separately reconfigurable The one or more image collection apparatus may be configured to collect depth, surface normal or other 3D data. The one or more image collection apparatus may be configured for photometric stereo, stereoscopic or other 3D imaging. The anomaly capture system may comprise a plurality of image collection apparatus, wherein each image collection apparatus may be configured to image the surface from at least one of different directions, from different locations and / or from different angles to the plane of the surface, to image the surface, e.g. to collect the depth, surface normal or other 3D data. For example, the other 3D imaging may comprise stereo photogrammetry, infrared imaging (such as that used in Kinect systems by Microsoft), LIDAR, RADAR, magnetic or electrical field sensors, laser scanners and / or the like.

[0031] The anomaly capture system may be configured to collect images for use with photometric stereo image processing, e.g. in order to determine normals to the surface at a plurality of locations. The anomaly capture system may be configured to collect at least one or a plurality of images of the surface from a fixed point of view, with each image of the plurality of images being captured under different lighting conditions. The number of images collected may be one or more, e.g. from 1 to 100 images, for example from 5 to 20 images, such as 10 images. This number of images has been found to give a good trade-off between accuracy and processing efficiency. The anomaly capture system may be configured to determine the surface normals of at least part of the surface from the at least one or plurality of images, e.g. by providing the at least one or plurality of images as inputs to a suitable process, which may be a white-box or black-box process. A white-box process may be a process comprising one or more defined equations or other mathematical processes that together derive the surface normals of at least part of the surface from the input images. The defined equations or other mathematical processes may be those known in the art of photometric stereo techniques. The black-box methods may comprise suitable machine learning or artificial intelligence models, such as, but not limited to, one or more suitably trained machine learning or artificial intelligence models, or a selflearning model, such as an unsupervised learning model, a semi-supervised learning model, self-supervised learning model, a reinforcement learning model, an adversarial model or the like.

[0032] The anomaly capture system may be configured to determine a 3D model or mesh of at least part of the surface, which may be by using the photometric stereo, stereoscopic or other 3D imaging, e.g. by using the determined normals. For example, 3D data of the surface, which may be generated from the determined normals, and / or from a plurality of pairs of stereoscopic images or from the other 3D imaging, may be stitched, merged or otherwise combined together to form the 3D model of mesh. In an example, the determined normals may be used as inputs to a further black-box or white-box process in order to derive the 3D model or mesh from the determined normals. The further black-box or white-box process may be analogous to those described above.

[0033] The plurality of pairs of stereoscopic images or the other 3D images may comprise images of at least part of the surface taken whilst the surface is being illuminated from the one or more selected directions and / or at the one or more selected illumination angles to the plane of the surface. That is, the 3D model or mesh of at least part of the surface may be formed from a plurality of pairs of stereoscopic images, wherein different pairs of stereoscopic images are taken whilst the surface is being illuminated from different directions and / or at different illumination angles to the plane of the surface. Optionally at least one of the pairs of stereoscopic images may comprise images of at least part of the surface under diffuse lighting.

[0034] The anomaly capture system may be configured to detect and / or characterise the anomalies from the 3D model or mesh. The characterization of the anomalies from the 3D model or mesh may comprise determining one or more geometric properties of the anomaly, such as depth, volume, length, width or other extent of the anomaly. For example, the characterization of the anomalies from the 3D model or mesh may comprise determining volume and / or height of an anomaly such as spalling or delamination.

[0035] The anomaly capture system may be configured to predict areas of the surface where anomalies are more likely to occur in future, e.g. by analysing the 3D model or mesh in order to identify features indicative of likely future anomalies, such as localised elevation of the surface or changes in surface height. The features indicative of likely future anomalies may be stored in a library defining features of likely future anomalies. The anomaly capture system may be configured to compare parts of the 3D model or mesh of the surface with the features of likely future anomalies in order to identify areas of the surface matching any of the features of likely future anomalies in order to identify areas of the surface indicative of likely future anomalies. However, the present disclosure is not limited to this and other suitable techniques for identifying areas of the surface indicative of likely future anomalies would be apparent based on the present disclosure. The anomaly capture system may be configured to produce a visualisation of the 3D model or mesh, which may have any anomalies highlighted. The anomaly capture system may be configured to process the at least one or plurality of images in order to locate the anomaly, e.g. using a black-box process. The anomaly capture system may be configured to translate the location of the anomaly in the at least one or plurality of images into a location in the 3D model or mesh. Alternatively, the 3D model or mesh may be processed, e.g. using a black-box process, to locate the anomaly in the 3D model or mesh.

[0036] At least one image collection apparatus of the plurality of image collection apparatus may be configured to image a different wavelength range or be of a different type to at least one other image collection apparatus. At least one or each of the image collection apparatus may comprise, or be configured to receive a filter, such as a polarisation filter or a spectral filter. Different image collection apparatus of the plurality of image collection apparatus may comprise or be provided with different filters, e.g. having different polarisation or to pass a different spectral range, to at least one or each other of the image collection apparatus.

[0037] The anomaly capture system may comprise, or be configured to communicate with, an analysis system. The analysis system may be, or may be comprised in, the controller. The analysis system may be or may comprise a separate analysis system to the controller, which may be a local or remote analysis system. The analysis system may be distributed between the controller and an external analysis system. The analysis system may be configured to analyse at least one or more of the images or video in order to analyse the respective different images or the different parts of the video of the surface when illuminated from respective different directions in order to analyse any anomalies in the surface. The analysing of any anomalies in the surface may comprise at least one or more of: identifying any anomalies and / or characterizing any anomalies. The identifying of any anomalies may comprise detecting the presence or otherwise of any anomalies. The characterizing of any anomalies may comprise at least one of: locating any anomalies such as a position of any anomalies on the surface, classifying any anomalies e.g. determining a type of the anomaly, determining an extent of any anomaly, e.g. determining one or more dimensions or extent of the anomaly such as length, width, depth, shape or area of the surface associated with any anomaly and / or the like.

[0038] The analysis system may be configured to determine a confidence value for at least one or each step of the analysis, e.g. one or more of: a confidence that a detected anomaly is there, a confidence that the anomaly is located at the identified location, a confidence that the anomaly is the identified type of anomaly, a confidence that the anomaly is of the identified extent, and / or the like. The analysis system may be configured to provide the determined confidence value for at least one of the steps above (e.g. to assist operators in making a decision) and / or to make any decisions based at least in part on the determined confidence value for at least one of the steps above.

[0039] The analysis system may be configured to flag that collection of further images is required based on the determined confidence value, e.g. if a determined confidence value is below a set or pre-set threshold, then the analysis system may be configured to flag that collection of further images is required. The flag that collection of further images is required may be provided to the controller, e.g. so that the controller can collect one or more further images of the surface responsive to the flag and / or to vary at least one of: the direction at which the surface is illuminated, the illumination angle of the surface with respect to the plane of the surface, the direction, location and / or angle at which the surface is imaged by the image collection apparatus. The controller may be configured to collect further images with the varied at least one of: the direction at which the surface is illuminated, the illumination angle of the surface with respect to the plane of the surface, the direction, location and / or angle at which the surface is imaged by the image collection apparatus. The controller may be configured to display or otherwise provide the flag to an operator.

[0040] The analysis system may be configured to aggregate or combine a plurality of images or parts of the video, wherein respective images of the plurality of images or parts of the video show the surface illuminated from different directions and / or at different illumination angles to the other images or parts of the video.

[0041] The analysis system may be configured to distinguish portions, e.g. areas, blocks or pixels, of the images or parts of the video representative of an anomaly from portions, e.g. areas, blocks or pixels, of the images or parts of the video representative of background, e.g. the majority, expected and / or regular surface, such as a concrete surface. The analysis system may be configured to segment the portions, e.g. pixels, of the images or parts of the video representative of the anomaly from the portions, e.g. pixels, of the images or parts of the video representative of background.

[0042] The analysis system may comprise and / or be configured to implement one or more models, which may be machine learning models. At least one of the models may be or comprise an identifier model. The identifier model may be configured to identify the presence of an anomaly. The identifier model may be configured to identify portions of the images or parts of the video representative of the anomaly, e.g. against the portions of the images or parts of the video representative of background. The identifier model may be configured to identify a region of the surface containing the identified anomaly, e.g. the identifier model may be configured to assign one or more bounding boxes around the identified portions of the images or parts of the video representative of the anomaly. The identifier model may be or comprise a neural network, such as a convolutional neural network and preferably (but not essentially) a region-based convolutional neural network (R-CNN) or the like. The identifier model may comprise a two-part or two-shot machine learning model. One part of the identifier model may be configured to detect features of interest against the background and / or may comprise a region proposal network. Another part of the identifier model may comprise a classifier model for classifying the features of interest as belonging to the portions of the images or parts of the video representative of the anomaly or not. The classifier model may comprise a convolutional neural network (CNN) or the like.

[0043] The identifier model may be trained on training data, such as but not limited to manually annotated training data or training data with manually assigned descriptive data, e.g. images of surfaces of similar material with one or more portions of the images or parts of the video representative of anomalies pre-identified, e.g. manually pre-identified and / or may be trained at least in part using artificially generated and / or automatically labelled training data. The identifier model may be configured to receive at least one of the images or at least part of the video as inputs, and may be trained to identify, and optionally apply the one or more bounding boxes around, the portions of the at least one of the images or combination of images or the at least part of the video representative of the anomaly. The at least one identifier model may be trained to at least roughly or approximately detect and / or locate the portions of the at least one of the images or the at least part of the video representative of the anomaly, e.g. within the limit of accuracy the one or more bounding box. The one or more bounding boxes may optionally be of a set or pre-set size and may be quadrilateral. The identifier model may be a fast and / or computationally efficient machine learning model, e.g. relative to at least one other of the machine learning models.

[0044] The analysis system may be configured to extract or isolate the portions of at least one of the images or the at least part of the video within any of the bounding boxes. The analysis system may be configured to label, e.g. in metadata or otherwise identify, any portions of the at least one of the images or the at least part of the video that are outside the bounding boxes as background, e.g. as normal surface.

[0045] At least one other of the models may be or comprise a locator and / or classifier model. The locator and / or classifier model may be or comprise a machine learning model. The locator and / or classifier model may be configured to receive as inputs the portions of the at least one of the images or the at least part of the video that are within any of the bounding boxes, e.g. as identified by the identifier model. The locator and / or classifier model may be trained to locate and / or classify parts, e.g. blocks, of the at least one of the images or the at least part of the video within the bounding boxes that represent at least part of an anomaly.

[0046] The locator and / or classifier model may be more accurate and / or have a higher resolution than the at least one identifier machine learning model. The locator and / or classifier model may be more computationally intensive and / or slower than the identifier machine learning model. The locator and / or classifier model may be configured to identify one or more blocks that comprise the parts of the at least one of the images or the at least part of the video representative of the anomaly. Each block of the one of more blocks may be smaller, e.g. have fewer pixels, than the bounding box. Each block may represent a plurality of pixels. Each block may be of a set or preset size, and may contain a set or pre-set number of pixels. Each block may be the same size. In this way, the blocks output by the locator and / or classifier model may be subsets of the bounding boxes output by the identifier model. The locator and / or classifier model may be trained by transfer learning. The locator and / or classifier model may be trained using training data, such as but not limited to manually annotated training data or training data with manually assigned descriptive training data, e.g. images of surfaces of similar material with one or more parts of the images or parts of the video representative of anomalies pre-identified, e.g. manually preidentified. The locator and / or classifier model may comprise a convolutional neural network (CNN) such as a VGG-16 neural network but other models or variations of model may be used. For example, in another example, the YOLO algorithm may be used. Additionally or alternatively, different channels and / or different numbers of channels may be input into the locator and / or classifier model.

[0047] A plurality of channels may be provided as inputs into the locator and / or classifier model. At least some or each of the input channels may comprise images of the surface under different illumination conditions, such as but not limited to images of the surface being illuminated from the one or more different directions when seen in a plan view of the surface and / or at the one or more different illumination angles to the plane of the surface. The images may comprise greyscale and / or colour images. For example, there may be three channels input into the locator and / or classifier model, which may corresponding to red, green and blue of a colour image, but there could be more or less or different channels, e.g. channels comprising images corresponding to one or more of: different illumination directions, different illumination angles, different types of illumination, different directions, locations and / or angles of image collection apparatus and / or the like. The channels input into the locator and / or classifier model corresponding to different illumination directions may comprise images of the surface illuminated from one or more different sides with respect to the surface, e.g. one or more or each of: illumination from at least one, two or more pairs of opposing lateral sides with respect to a plan view of the surface. For example, at least one or some of the channels may comprise at least one image of the surface whilst illuminated from at least one or some or each of: left, right, top and / or bottom of the surface when viewed in a plan view of the surface. The different types of illumination may comprise diffuse lighting and / or hard (non-diffuse) lighting. As a specific example, the locator and / or classifier model may comprise a plurality of input channels, in which each of the input channels correspond to images collected whilst the surface is illuminated from different directions, e.g. from opposing lateral sides with respect to a plan view of the surface, with one of diffuse or hard lighting and at least one other respective input channel collected with illumination of the surface under the other of hard or diffuse lighting. An example may comprise five input channels into the locator and / or classifier model, which may comprise channels corresponding to illumination of the surface at least from two pairs of opposing directions, such as up, down, left and right with respect to the surface when viewed in a plan view under hard lighting and an input channel corresponding to the surface when illuminated by diffuse lighting.

[0048] The analysis system may be configured to perform pixel level segmentation, e.g. in order to segment out the one or more anomalies from the rest of the surface. The at least one or plurality of images may be used as inputs to the pixel level segmentation. The pixel level segmentation may comprise the application of a suitable white-box model to the at least one or plurality of images in order to segment out the one or more anomalies from the rest of the surface.

[0049] The analysis system may be configured to apply a two-tier machine learning model approach. The analysis system may be configured to roughly locate the portions of the at least one of the images or the at least part of the video representative of the anomaly, and provide one or more bounding boxes around the located portions of the at least one of the images or the at least part of the video representative of the anomaly. The rough locating of the portions of the at least one of the images or the at least part of the video representative of the anomaly may be carried out using the at least one locator machine learning model. The analysis system may be configured to more accurately identify the parts, e.g. blocks, of the at least one of the images or the at least part of the video in the one or more bounding boxes that are representative of the anomaly, e.g. using the at least one locator and / or classifier machine learning model. Any the blocks in the one or more bounding boxes of the at least one of the images or the at least part of the video other than those identified to be representative of at least part of the anomaly may be designated as background, i.e. normal or expected parts of the surface.

[0050] The analysis system may be configured to implement one or more white-box or classical image processing methods. The one or more white-box or classical image processing images may be non-machine learning and non-artificial intelligence techniques, mathematical filters, mathematical transforms and / or the like. The whitebox or classical techniques may comprise a segmentation algorithm for segmenting pixels representing anomalies from pixels representing background. The white-box or classical techniques may comprise edge detection, e.g. for detecting edges of any anomalies. The white-box or classical techniques may be configured to process outputs from one of the machine learning models, e.g. from the locator and / or classifier machine learning model. The white-box or classical techniques may be applied to the blocks output by the locator and / or classifier machine learning model and may be configured to further refine the blocks identified as comprising portions of the at least one of the images or the at least part of the video representative of the anomaly.

[0051] The analysis system may be a hybrid analysis system that comprises both one or more machine learning or black-box models and white-box or classical techniques. The black-box models may comprise trained or statistical models, e.g. rather than manually defined absolute mathematical relations. The white-box or classical techniques may be configured to refine or accept as inputs the outputs of the one or more machine learning models.

[0052] Although the above advantageous arrangement in which two stages of machine learning model are applied and then white-box or classical techniques, any combination of order of any of these techniques may be applied, e.g. only one machine learning model may be applied or only white-box or classical techniques, or a combination of one machine learning model and the white-box techniques, or the white-box techniques may be applied before one or both of the machine learning models, and / or the like. As such, any of the black-box or machine learning techniques and / or any of the white-box or classical techniques or others may be applied in any order, or the functionality described in relation to any techniques could be combined into a common model or step, or certain models or steps could be omitted, repeated or otherwise altered.

[0053] The analysis system may be configured to receive as input a plurality of images or parts of the video. Respective images of the plurality of images or parts of the plurality of parts of the video may show the surface being illuminated from different directions when seen in a plan view of the surface and / or at different illumination angles with respect to the plane of the surface. Respective images of the plurality of images or parts of the plurality of parts of the video may show the surface being illuminated from different directions, which may be opposite and / or perpendicular to each other, but are not limited to this. Respective images or parts of the video may show the surface being illuminated from at least two, three or all of: left, right, up and / or down directions when seen in a plan view of the surface, and / or may comprise additional or other directions. A plurality of the images or parts of the video in which different images or different parts of the video show the surface being illuminated from different directions when seen in a plan view of the surface and / or at different illumination angles with respect to the plane of the surface may be analysed by the analysis. The analysis system may be configured to combine the results of the analysis of the plurality of images or parts of the video, e.g. the output of at least one of: the identifier model and / or the at least one locator and / or classifier model and / or the white-box or classical techniques. For example, the combining may comprise applying an ‘OR’ function, e.g. a bit-wise OR function to the results of the analysis of the plurality of images, but other forms of combining could be used. For example an alternative form of combining may comprise summing the intensity of the locator and / or classifier model’s outputs from multiple directionally-lit images to arrive at a probability heat map indicating where anomalies are.

[0054] The analysis system may be configured to dynamically determine an optimal or preferred configuration, the optimal or preferred configuration comprising at least one of: an optimal or preferred direction and / or angle of illumination of the surface, an optimal direction, location and / or angle of the image collection apparatus, an optimal or preferred colour or wavelength of illumination and / or the like. The analysis system may be configured to dynamically determine the optimal or preferred configuration by collecting a plurality of images of the surface illuminated from different directions and / or illumination angles and / or different direction, location and / or angle of the image collection apparatus and determining a quality value of each image. The illumination system may be controllable, e.g. responsive to the controller, to dynamically illuminate the surface from the direction and / or angle of illumination of the surface and / or different direction, location and / or angle of the image collection apparatus determined to be optimal or preferred. The illumination system may be controllable to dynamically reposition and / or re-orient the one or more light sources in order to dynamically illuminate the surface from the direction and / or angle of illumination of the surface determined to be optimal or preferred, and / or to dynamically reposition, re-orient or use the one or more image collection apparatus to the determined optimal or preferred arrangement. The analysis system may be configured to dynamically determine an optimal or preferred direction and / or angle of illumination for each surface in a plurality of surfaces.

[0055] The anomaly capture system utilises directional lighting to illuminate the surface. In this way, the anomaly capture system may be configured to utilise the illumination system to illuminate the surface from different directions and / or at different illumination angles with respect to the plane of the surface when different images or parts of the video are being collected by the image collection apparatus. The analysis system may analyse each of the plurality images or parts of the video, where respective different images or parts of the video show the surface being illuminated from a different direction to other images or parts of the video. The analysis system may analyse each image or part of the video, e.g. using the at least one machine learning model and / or the white-box or classical techniques, in order to identify and / or characterise any anomalies in the surface. In this way, the identification and / or characterisation of anomalies in the surface can be enhanced.

[0056] The analysis system may be configured to process a plurality of images or parts of video of the surface, wherein different images or parts of video are collected by image collection apparatus that one or more of: face different directions, are located at different locations and / or are oriented at different angles. Different images of the plurality of images may be collected by the same or different image collection apparatus. The analysis system may be configured to process a plurality of images or parts of video of the surface, wherein different images or parts of video are of the surface illuminated from different directions and / or at different illumination angles to the plane of the surface. The analysis system may be configured to process the plurality of images or parts of video to determine surface normals for the surface from the plurality of images or parts of video, e.g. from the plurality of images or parts of videos in which different images or parts of the video show the surface being illuminated from different directions and / or at different illumination angles to the plane of the surface. The analysis system may be configured to determine the surface normals for the surface using photometric stereo techniques (e.g. https: / / en.wikipedia.org / wiki / Photometric stereo). The surface normals for the surface may be used to form a surface map and / or a 3D model or mesh of at least part of the surface, which may be the 3D model or mesh referred to above. The surface normals may be derived from at least one or the plurality of images using photometric stereo techniques, e.g. as described above.

[0057] According to a second example of the present disclosure is a method of identifying and / or characterising anomalies in a surface, the method comprising: providing an illumination system; and at least one image collection apparatus; wherein the illumination system comprises at least one light source configured for directionally illuminating a surface, the method further comprising selectively illuminating the surface from at least one selected direction when seen in a plan view of the surface and / or at one or more selected illumination angles to the plane of the surface using the illumination system; and imaging the surface from at least one direction using the at least one image collection apparatus whilst the surface is being illuminated from the at least one selected direction and / or at the one or more selected illumination angles to the plane of the surface.

[0058] The method may comprise providing an anomaly capture system according to the first example, wherein the illumination system and the image collection apparatus are comprised in the anomaly capture system.

[0059] The method may comprise illuminating the surface from any of a plurality of directions and / or from any of a plurality of illumination angles to the plane of the surface. The method may comprise illuminating the surface from a different direction and / or from a different angle to the plane of the surface by manoeuvring the at least one light source using the at least one manipulator described in relation to the first example, and / or by selectively activating and deactivating selected light sources that face different directions and / or are oriented at different angles. The method may comprise using other differential lighting arrangements in addition to the directional and / or angular illumination, e.g. illuminating the surface with different colours I spectral components (e.g. collecting different images of the surface whilst illuminated with different wavelengths of light), different types of lighting (e.g. illuminated by two or more of: LED, Fluorescent, halogen or laser light), illuminating with different projection methods or patterns (e.g. spot, dome, line and / or the like, which may be chosen to match the feature), different polarisations, different neutral densities, other different filter effects, diffuse and non-diffuse light, and / or the like.

[0060] The method may comprise illuminating the surface with a selected colour and / or with a selected wavelength or wavelength band, which may be selected to according to a material of the surface or material on the surface or a type of anomaly. The colour or wavelength or wavelength range of the light used for illumination may be selected to accentuate or highlight or alternatively to reduce or diminish the anomaly or a material on the surface. For example the images may be selectively illuminated with blue light, which may be particularly suitable for imaging red corrosion, as the red corrosion may be accentuated in the images. As another example, illumination with green lighting can be used to reduce, diminish or render negligible the effect of plant or other organic growth such as moss or algae on the surface.

[0061] The method may comprise selectively illuminating the surface from different directions, e.g. from different sides of the surface when seen in a plan view, and / or at different illumination angles to the plane of the surface, whilst different images or parts of the video are collected.

[0062] The method may comprise dynamically changing the direction in which the surface is illuminated based on a quality value of a resultant image, e.g. to maximise or optimise or improve the quality value of the resultant image.

[0063] The method may comprise arranging the light sources, so as to illuminate the surface at an angle to the plane of the surface in a range from 0° to 90° but preferably from 5° to 50°, e.g. from 20° to 50°, such as from 30° to 40°, inclusive.

[0064] The method may comprise collecting a plurality of images of the surface, wherein at least one of the images may be collected by at least one of the one or more image collection apparatus from at least one of: a different direction, a different location and / or a different angle to the plane of the surface, to at least one other image.

[0065] According to a third example of the present disclosure is a computer program product configured such that, when implemented on a controller of an anomaly capture system, such as the anomaly capture system of the first aspect, causes the controller to control the illumination system to selectively illuminate the surface from a selected direction when seen in a plan view of the surface and / or at a selected illumination angle to the plane of the surface; and the image collection apparatus is configured to image the surface whilst the surface is being illuminated from the selected direction and / or at the selected illumination angle to the plane of the surface. The computer program product may be configured to cause the controller to implement the method of the second example.

[0066] According to a fourth example of the present disclosure is a method of analysing images of a surface to determine anomalies in the surface, the method comprising applying a hybrid analysis that comprises applying one or more machine learning or black-box models to an image of the surface to determine blocks of the image containing at last a part of the anomaly and to apply one or more classical or white-box techniques to the blocks of the image determined to contain at least part of the anomaly to identify portions of the blocks that represent the anomaly from portions of the blocks that don’t represent the anomaly.

[0067] The method may comprise analysing at least one or more of the plurality of images or video in order to analyse the respective different images or the different parts of the video of the surface when illuminated from respective different directions in order to analyse any anomalies in the surface. The analysing of any anomalies in the surface may comprise at least one or more of: identifying any anomalies and / or characterizing any anomalies. The identifying of any anomalies may comprise detecting the presence or otherwise of any anomalies. The characterizing of any anomalies may comprise at least one of: locating any anomalies such as a position of any anomalies on the surface, classifying any anomalies e.g. determining a type of the anomaly, determining an extent of any anomaly, e.g. determining one or more dimensions or extent of the anomaly such as length, width, depth, shape or area of the surface associated with any anomaly and / or the like.

[0068] The method may comprise determining a confidence value for at least one or each step of the analysis, e.g. one or more of: a confidence that a detected anomaly is there, a confidence that the anomaly is located at the identified location, a confidence that the anomaly is the identified type of anomaly, a confidence that the anomaly is of the identified extent, and / or the like. The method may comprise providing the determined confidence value for at least one of the steps above (e.g. to assist operators in making a decision) and / or to make any decisions based at least in part on the determined confidence value for at least one of the steps above. The method may comprise flagging that collection of further images is required based on the determined confidence value, e.g. if a determined confidence value is below a set or pre-set threshold, then the method may comprise flagging that collection of further images is required. The flag that collection of further images is required may be provided to the controller, e.g. so that the controller can collect one or more further images of the surface responsive to the flag and / or to vary at least one of: the direction at which the surface is illuminated, the illumination angle of the surface with respect to the plane of the surface, the direction, location and / or angle at which the surface is imaged by the image collection apparatus. The method may comprise collecting further images with the varied at least one of: the direction at which the surface is illuminated, the illumination angle of the surface with respect to the plane of the surface, the direction, location and / or angle at which the surface is imaged by the image collection apparatus. The method may comprise displaying or otherwise providing the flag to an operator.

[0069] The method may comprise aggregating or combining a plurality of images or parts of the video, wherein respective images of the plurality of images or parts of the video show the surface illuminated from different directions and / or at different illumination angles to the other images or parts of the video.

[0070] The method may comprise distinguishing portions, e.g. areas, blocks or pixels, of the images or parts of the video representative of an anomaly from portions, e.g. areas, blocks or pixels, of the images or parts of the video representative of background, e.g. the majority, expected and / or regular surface, such as a concrete surface. The method may comprise segmenting the portions, e.g. pixels, of the images or parts of the video representative of the anomaly from the portions, e.g. pixels, of the images or parts of the video representative of background.

[0071] The method may comprise implementing one or more models, which may be machine learning models.

[0072] At least one of the models may be or comprise an identifier model. The identifier model may be configured to identify a region of the surface containing the identified anomaly, e.g. the identifier model may be configured to assign one or more bounding boxes around the identified portions of the images or parts of the video representative of the anomaly. The identifier model may be or may comprise at least one or any or all features of the identifier model described above in relation to the first example. The method may comprise extracting or isolating the portions of at least one of the images or the at least part of the video within any of the bounding boxes determined by the identifier model. The method may comprise labelling, e.g. in metadata or otherwise identify, any portions of the at least one of the images or the at least part of the video that are outside the bounding boxes as background, e.g. as normal surface.

[0073] At least one other of the models may be or comprise a locator and / or classifier model. The locator and / or classifier model may be or may comprise at least one or any or all features of the locator and / or classifier model described above in relation to the first example. The locator and / or classifier model may be configured to receive as inputs the portions of the at least one of the images or the at least part of the video that are within any of the bounding boxes, e.g. as identified by the identifier model. The locator and / or classifier model may be trained to locate and / or classify parts, e.g. blocks, of the at least one of the images or the at least part of the video within the bounding boxes that represent at least part of an anomaly.

[0074] The locator and / or classifier model may be or comprise a machine learning model. The locator and / or classifier model may be configured to receive as inputs the portions of the at least one of the images or the at least part of the video that are within any of the bounding boxes, e.g. as identified by the identifier model. The locator and / or classifier model may be trained to locate and / or classify parts, e.g. blocks, of the at least one of the images or the at least part of the video within the bounding boxes that represent at least part of an anomaly.

[0075] The method may comprise applying a two-tier machine learning model approach. The method may comprise roughly locating the portions of the at least one of the images or the at least part of the video representative of the anomaly, and provide one or more bounding boxes around the located portions of the at least one of the images or the at least part of the video representative of the anomaly. The rough locating of the portions of the at least one of the images or the at least part of the video representative of the anomaly may be carried out using the at least one locator machine learning model. The method may comprise more accurately identifying the parts, e.g. blocks, of the at least one of the images or the at least part of the video in the one or more bounding boxes that are representative of the anomaly, e.g. using the at least one locator and / or classifier machine learning model. Any the blocks in the one or more bounding boxes of the at least one of the images or the at least part of the video other than those identified to be representative of at least part of the anomaly may be designated as background, i.e. normal or expected parts of the surface.

[0076] The method may comprise implementing one or more white-box or classical image processing methods. The one or more white-box or classical image processing images may be or may comprise at least one or any or all features of the white-box or classical image processing methods described above in relation to the first example. The white-box or classical methods may comprise edge detection, e.g. for detecting edges of any anomalies. The white-box or classical methods may be configured to process outputs from one of the machine learning models, e.g. from the locator and / or classifier machine learning model. The white-box or classical methods may be applied to the blocks output by the locator and / or classifier machine learning model and may be configured to further refine the blocks identified as comprising portions of the at least one of the images or the at least part of the video representative of the anomaly.

[0077] The method may comprise using a hybrid analysis system that comprises both one or more machine learning or black-box models and white-box or classical techniques. The black-box models may comprise trained or statistical models, e.g. rather than manually defined absolute mathematical relations. The white-box or classical techniques may be configured to refine or accept as inputs the outputs of the one or more machine learning models.

[0078] Although the above advantageous arrangement in which two stages of machine learning model are applied and then white-box or classical techniques, any combination of order of any of these techniques may be applied, e.g. only one machine learning model may be applied or only white-box or classical techniques, or a combination of one machine learning model and the white-box techniques, or the white-box techniques may be applied before one or both of the machine learning models, and / or the like. As such, any of the black-box or machine learning techniques and / or any of the white-box or classical techniques or others may be applied in any order, or the functionality described in relation to any techniques could be combined into a common model or step, or certain models or steps could be omitted, repeated or otherwise altered.

[0079] The method may comprise receiving as input a plurality of images or parts of the video. Respective images of the plurality of images or parts of the plurality of parts of the video may show the surface being illuminated from different directions when seen in a plan view of the surface and / or at different illumination angles with respect to the plane of the surface. Respective images of the plurality of images or parts of the plurality of parts of the video may show the surface being illuminated from different directions, which may be opposite and / or perpendicular to each other, but are not limited to this. Respective images or parts of the video may show the surface being illuminated from at least two, three or all of: left, right, up and / or down directions when seen in a plan view of the surface, and / or may comprise additional or other directions. A plurality of the images or parts of the video in which different images or different parts of the video show the surface being illuminated from different directions when seen in a plan view of the surface and / or at different illumination angles with respect to the plane of the surface may be analysed by the analysis. The method may comprise combining the results of the analysis of the plurality of images or parts of the video, e.g. the output of at least one of: the identifier model and / or the at least one locator and / or classifier model and / or the white-box or classical techniques. For example, the combining may comprise applying an ‘OR’ function, e.g. a bit-wise OR function to the results of the analysis of the plurality of images, but other forms of combining could be used. For example an alternative form of combining may comprise summing the intensity of the locator and / or classifier model’s outputs from multiple directionally-lit images to arrive at a probability heat map indicating where anomalies are.

[0080] The method may comprise dynamically determining an optimal or preferred configuration. The optimal or preferred configuration may comprise at least one of: an optimal or preferred direction and / or angle of illumination of the surface and / or an optimal direction, location and / or angle of the image collection apparatus and / or an optimal colour or wavelength of illumination, or the like. The determining of the optimal or preferred configuration may comprise collecting a plurality of images of the surface illuminated from different directions and / or illumination angles and / or different direction, location and / or angle of the image collection apparatus and / or at a different colour or wavelength, and determining a quality value of each image. The method may comprise controlling an illumination system to dynamically illuminate the surface from the direction and / or angle of illumination of the surface and / or different direction, location and / or angle of the image collection apparatus determined to be optimal or preferred. The method may comprise dynamically controlling the illumination system to dynamically reposition and / or re-orient the one or more light sources in order to dynamically illuminate the surface from the direction and / or angle of illumination of the surface determined to be optimal or preferred, and / or to dynamically reposition, reorient or use the one or more image collection apparatus to the determined optimal or preferred arrangement. The method may comprise dynamically determining an optimal or preferred direction and / or angle of illumination for each surface in a plurality of surfaces.

[0081] The method may comprise processing a plurality of images or parts of video of the surface, wherein different images or parts of video are of the surface illuminated from different directions and / or at different illumination angles to the plane of the surface. The method may comprise processing the plurality of images or parts of video to determine surface normals for the surface from the plurality of images or parts of video, e.g. from the plurality of images or parts of videos in which different images or parts of the video show the surface being illuminated from different directions and / or at different illumination angles to the plane of the surface. The method may comprise determining the surface normals for the surface using photometric stereo techniques (e.g. https: / / en.wikipedia.org / wiki / Photometric stereo). The photometric stereo techniques may be or comprise black-box or white-box processes. The method may comprise forming a map of the surface normals representing at least part of the surface and / or forming a 3D model or mesh of at least part of the surface using the determined surface normals. The method may comprise detecting and / or characterising the anomalies from the map of the surface normals and / or the 3D model or mesh. The method may comprise producing a visualisation of the 3D model or mesh, which may have any anomalies highlighted. The method may comprise identifying the one or more anomalies and / or areas with a high likelihood of future anomaly formation from the 3D model or mesh and / or from the map of surface normal. The method may comprise raising an alarm, alert or flag responsive to identification of an anomaly.

[0082] According to a fifth example of the present disclosure is an analysis system configured to implement the method of the fourth aspect.

[0083] According to a sixth example of the present disclosure is a computer program product configured such that, when implemented on a computer implemented analysis system, causes the analysis system to implement the method of the fourth aspect.

[0084] The individual features and / or combinations of features defined above in accordance with any example of the present disclosure or below in relation to any specific embodiment of the invention may be utilised, either separately and individually, alone or in combination with any other defined feature, in any other aspect or embodiment of the invention.

[0085] Furthermore, the present invention is intended to cover apparatus configured to perform any feature described herein in relation to a method and / or a method of using or producing, using or manufacturing any apparatus feature described herein. Any method described herein above or below can be performed by one or more programmable processors executing a computer program to perform functions of the invention by operating on input data and generating output. Method steps can also be performed by special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit) or other customised circuitry. Processors suitable for the execution of a computer program include CPUs and microprocessors, and any one or more processors. Generally, a processor will receive instructions and data from a read-only memory or a random access memory or both. The essential elements of a computer are a processor for executing instructions and one or more memory devices for storing instructions and data. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magnetooptical disks, or optical disks. Information carriers suitable for embodying computer program instructions and data include all forms of non-volatile memory, including by way of example semiconductor memory devices, e.g. EPROM, EEPROM, and flash memory devices; magnetic disks, e.g., internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks. The processor and the memory can be supplemented by, or incorporated in special purpose logic circuitry.

[0086] To provide for interaction with a user, the method can be implemented on a device having a screen, e.g., a CRT (cathode ray tube), plasma, LED (light emitting diode) or LCD (liquid crystal display) monitor, for displaying information to the user and an input device, e.g., a keyboard, touch screen, a mouse, a trackball, and the like by which the user can provide input to the computer. Other kinds of devices can be used, for example, feedback provided to the user can be any form of sensory feedback, e.g., visual feedback, auditory feedback, or tactile feedback; and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0087] Any method described herein can be implemented on a device, such as a mobile or network enabled device, comprising or configured to implement the controller or processing system of the first aspect. The device may be or comprise or be comprised in a mobile phone, smartphone, PDA, tablet computer, laptop computer, and / or the like. The controller or processing system may be implemented by a suitable program or application (app) running on the device. The device may comprise at least one processor, such as a central processing unit (CPU), maths co-processor (MCP), graphics processing unit (GPU), tensor processing unit (TPU) and / or the like. The at least one processor may be a single core or multicore processor. The device may comprise memory and / or other data storage, which may be implemented on DRAM (dynamic random access memory), SSD (solid state drive), HDD (hard disk drive) or other suitable magnetic, optical and / or electronic memory device. The at least one processor and / or the memory and / or data storage may be arranged locally, e.g. provided in a single device or in multiple devices in in communication at a single location or may be distributed over several local and / or remote devices. The device may comprise a communications module, e.g. a wireless and / or wired communications module. The communications module may be configured to communicate over a cellular communications network, Wi-Fi, Bluetooth, ZigBee, near field communications (NFC), IR, satellite communications, other internet enabling networks and / or the like. The communications module may be configured to communicate via Ethernet or other wired network or connections, via a telecommunications network such as a POTS, PSTN, DSL, ADSL, optical carrier line, and / or ISDN link or network and / or the like, via the cloud and / or via the internet, or other suitable data carrying network. The communications module may be configured to communicate via optical communications such as optical wireless communications (OWC), optical free space communications or Li-Fi or via optical fibres and / or the like. The device and / or the controller or the at least one processor or processing unit may be configured to communicate with the remote server or data store via the communications module. The controller or processing unit may comprise or be implemented using the at least one processor, the memory and / or other data storage and / or the communications module of the device.

[0088] BRIEF DESCRIPTION OF THE DRAWINGS

[0089] For a better understanding of the present disclosure and to show how embodiments may be put into effect, reference is made to the accompanying drawings in which:

[0090] Figure 1 shows a schematic view of an anomaly capture system;

[0091] Figure 2 shows a plan view of the anomaly capture system of Figure 1;

[0092] Figure 3 shows a perspective view of the anomaly capture system of Figure 1;

[0093] Figure 4 shows a schematic representation of a part of the anomaly capture system of Figure 1 ; Figure 5 shows a schematic representation of an image collection apparatus of the anomaly capture system of Figure 1 ;

[0094] Figure 6 is a schematic showing the field of view of the image collection apparatus of Figure 5;

[0095] Figure 7 shows examples of a surface that comprises an anomaly being illuminated from different directions and different angles.

[0096] Figure 8 shows a contour chart showing an optimum lighting direction and an optimum angle for an exemplary surface.

[0097] Figure 9 shows a histogram of optimised quality of image scores for diffuse and optimized directional lighting conditions;

[0098] Figure 10 shows a flowchart of an analysis method for analysing images collected using the system of Figure 1;

[0099] Figure 11 shows the effect of applying a white-box analysis to images collected using the system of Figure 1;

[0100] Figure 12 shows examples of outputs of stages of the method of Figure 10;

[0101] Figure 13 is a chart of accuracy of crack detection for different lighting directions;

[0102] Figure 14 is a chart of accuracy of crack detection for different lighting angles t to the plane of the surface;

[0103] Figure 15 is a chart of overall accuracy of crack detections;

[0104] Figure 16 shows confusion values for different methods of combining results obtained from different images of the surface illuminated from different directions and angles; Figures 17 to 19 show an alternate anomaly capture system, Figures 17 and 18 showing an inner surface view and Figure 19 showing a reverse I back side surface view.

[0105] DETAILED DESCRIPTION

[0106] In the following detailed description, reference is made to the accompanying drawings that form a part hereof, and in which is shown by way of illustration specific embodiments in which the inventive subject matter may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice them, and it is to be understood that other embodiments may be utilized, and that structural, logical, and electrical changes may be made without departing from the scope of the inventive subject matter. Such embodiments of the inventive subject matter may be referred to, individually and / or collectively, herein by the term “invention” merely for convenience and without intending to voluntarily limit the scope of this application to any single invention or inventive concept if more than one is in fact disclosed.

[0107] The following description is, therefore, not to be taken in a limited sense, and the scope of the inventive subject matter is defined by the appended claims and their equivalents.

[0108] In the following embodiments, like components are labelled with like reference numerals.

[0109] In the following embodiments, the term data store or memory is intended to encompass any computer readable storage medium and / or device (or collection of data storage mediums and / or devices). Examples of data stores include, but are not limited to, optical disks (e.g., CD-ROM, DVD-ROM, etc.), magnetic disks (e.g., hard disks, floppy disks, etc.), memory circuits (e.g., EEPROM, solid state drives, random-access memory (RAM), etc.), and / or the like.

[0110] As used herein, except wherein the context requires otherwise, the terms “comprises”, “includes”, “has” and grammatical variants of these terms, are not intended to be exhaustive. They are intended to allow for the possibility of further additives, components, integers or steps.

[0111] The functions or algorithms described herein are implemented in hardware, software or a combination of software and hardware in one or more embodiments. The software comprises computer executable instructions stored on computer readable carrier media such as memory or other type of storage devices. Further, described functions may correspond to modules, which may be software, hardware, firmware, or any combination thereof. Multiple functions are performed in one or more modules as desired, and the embodiments described are merely examples. The software may be executed on a digital signal processor, ASIC, FPGA, microprocessor, microcontroller, on a dedicated electrical circuit, or other type of processing device or combination thereof.

[0112] Furthermore, references are made herein to a processing module. The processing module may comprise software and / or be at least partly implemented using software. The processing module may comprise, and / or be at least partially implemented using, hardware for processing, which may comprise one or more devices, a system of devices, which may comprise a plurality of processing devices that may be distributed or localized, may comprise different processing devices in different apparatus or may all be contained in a single apparatus. The processing devices may comprise one or more processors, which may be multi-core or single core processors, ASICs, one or more hardware programmable devices (e.g. FPGAs), electrical, electronic or other logic circuits, and / or the like, or any combination thereof.

[0113] Specific embodiments will now be described with reference to the drawings.

[0114] Figures 1 to 4 show an anomaly capture system 5 that comprises an illumination system 10 and an image collection apparatus 15 in the form of a digital camera. The illumination system 10 and the image collection apparatus are both mounted to a rigid support 20. The illumination system 10 is configured to illuminate a surface 25, such as but not limited to a concrete surface, from a plurality of directions and / or angles to the plane of the surface 25. The image collection apparatus 15 is configured to capture images (which may be comprised in different parts of a video) of the surface 25. Different images of the plurality of images show the surface being illuminated from the different directions and / or different angles to the plane of the surface 25.

[0115] The illumination system 10 comprises a plurality of light sources 30a-30d, each light source being mounted on the end of a corresponding robotic arm 35a-35d that extends from the support 20.

[0116] The support 20 is a rigid support, in this example in the form of a right angled steel plate, which can be attached to a manoeuvring system for manoeuvring the support 20. In this example, the manoeuvring system comprises a robotic arm, such as a six-axis robotic arm (not shown). However, other forms of rigid support and / or manoeuvring systems could be used (or no manoeuvring system could be used).

[0117] In this example, each light source 30a-30d is advantageously in the form of an elongate strip of LEDs, but the disclosure is not limited to this.

[0118] Each robotic arm 35a-35d to which a respective light source 30a-30d is mounted is a three-jointed servo-motorised arm, but again the disclosure is not limited to this and other means for manoeuvring the light sources into the desired directional and angular relationship with the surface 25 could be used. The location, direction and angle of each light source 30a-30d is individually and / or independently adjustable. That is, in the example shown, each robotic arm 35a-35d is individually and independently adjustable so as to manoeuvre the associated light source 30a-30d into the required location and angle with respect to the surface 25.

[0119] Specifically, each of the robotic arms 35a-35d comprises a plurality of elongate rigid sections 40a, 40b. A first rigid section 40a is connected at one end to the support 20 via a first articulated joint 45a. The other end of the first rigid section 40a is connected to an end of the second rigid section 40b via a second articulated joint 45b. The respective light source 30a-30d is mounted to the other end of the second rigid section 40b via a third articulated joint 45c. The first, second and / or third articulated joints 45a, 45b, 45c are rotatable in one or two or three axes under the action of a suitable controllable forcing mechanism, such as a motor (e.g. servo motor), pneumatic actuator, hydraulic actuator, and / or the like. For example, the third articulated joint 45c can be configured to rotate the respective light source 30a-30d so as to vary the angle that light from the light source 30a-30d is incident on the surface 25 (e.g. the angle 0 of the optical axis of the light source 30a-30d to the plane of the surface 35 shown in Figure 4). The direction from which the surface 25 is illuminated as seen in a plan view of the surface can be changed by arranging the light sources 30a-30d so that they face in different directions and selectively activating different light sources 30a-30d. For example, the light sources can be arranged so that each light source 30a-30d is perpendicular to at least one other light source 30a-30d, and then activated sequentially, one after the other, and images collected when each light source 30a-30d is illuminating the surface 25.

[0120] The motion provided by each robotic arm 25a-25d is shown specifically in Figure 4. The three joints 45a-45c include a shoulder 45a, an elbow 45b, and a wrist 45c, which can be adjusted using servo motors or the like to re-orient the light source 30a-30d to change the light’s incident angle, OL and proximity, P with respect to the surface 25. By considering the arms 25a-25d as vectors, the horizontal components are:

[0121] A+LICOSQI=L2COSQ2+PCOSQL, (1) and the vertical components are:

[0122] D=LiSin G>i+L2sin O2+Psin 03(2) where Li and L2 are the lengths of the arms, A is the distance from the centre of the camera lens of the image collection apparatus 15 to the shoulder joint 45a, and D is the working distance of the camera of the image collection apparatus 15 or distance from the shoulder joint 45a to the surface 25. 0? and 02 and are the angles of the rigid sections 40a, 40b to the horizontal and 03is the angle form the optical axis of the light source 30a-30d to the horizontal. OL is the angle of incidence of the light from the light source 30a-30d on the surface 25 (e.g. the angle of the optical axis of the light source 25 to the plane of the surface 25). In the particular example, the lengths of the rigid sections 40a, 40b are in the range of 100mm to 500mm and the shoulder joint 45a angle and the elbow joint 45b angle are constrained to operate in a range that falls in a range from -90°< 0?< 90° and 0°< 02<18O° respectively. In this way, for an input set of desired illumination angles OL, proximity P and working distance D it is possible to determine solutions for the joint angles 07, 02 and 03, e.g. from equations (1) and (2) above, for example by using suitable solution methods such as least-squares or other minimization techniques.

[0123] The support 20 can be moved, e.g. translated, by the manoeuvring system in order to manoeuvre the illumination system 10 into the required relationship with the surface 25. The robotic arms 35a-35d can be operated to fine adjust the light sources 30a-30d into locations in which they can illuminate the surface 25 from the required different directions and / or angles to the plane of the surface 25. The light sources are then activated and deactivated sequentially (or optionally in combination if diffuse light is required) such that the image collection apparatus 15 can collect a plurality of images, each image showing the surface 25 illuminated from a different direction (e.g. left, right, top, bottom when viewed in a plan view of the surface 25) and / or a different illumination angle OL to the plane of the surface (e.g. at angles in the range from 0 = 10° to 50°). The manoeuvring system, the servo motors or other suitable controllable forcing mechanism operating each of the joints 45a-45c of the robotic arms 35a-35d and the light sources 35a-d can be controlled by a controller 50. The controller 50 comprises a processor 55, data storage 60 and communications module 65 for communicating control commands to the manoeuvring system, the servo motors and light sources 35a- 35d so as to controllably operate the illumination system 15. The controller 50 may operate the illumination system 15 responsive to manual input via a suitable user interface 70, or according to automated control which may comprise a plurality of rules, instructions or other operating parameters, a machine learning model or the like, or a mixture of both manual and automated control.

[0124] Although a specific apparatus for arranging the light sources 30a-30d with respect to the surface 25 is illustrated and described above with respect to Figures 1 to 4, it will be appreciated that the possibilities are not limited to this and that other arrangements are possible. Other apparatus can be used for arranging the light sources 30a-30d to selectively illuminate the surface 25, e.g. a concrete surface, from different directions and / or illumination angles OL to the plane of the surface 25. For example, although four light sources 30a-30d are shown, it is possible to use other numbers of light sources 30, e.g. one, two, three or more than four light sources 30. Although robotic arms 35a-35d for manoeuvring the light sources 30a-30d into the desired position and / or angle with respect to the surface 25 are provided, different arrangements for locating and orienting the light sources 30a-30d are possible. For example, no robotic arms 35a-35d could be provided and one or more light sources could be mounted to the support 20, e.g. via a joint such as a one, two or three axis rotating joint. In addition, although robotic arms 34a-35d with three joints 45a, 45b, 45c and two rigid sections 40a, 40b are shown and described, other types of robotic arms could be used, e.g. having more or less rigid sections 40a, 40b and / or more or less joints 45a-45c. In another example, a segmented arm could be used. As such, whilst a particularly beneficial and flexible arrangement of illumination system 10 is shown in Figures 1 to 4, it will be appreciated that other arrangements of illumination system could be used.

[0125] Furthermore, although collection of images with the surface 25 illuminated at different angles and / or directions is described and shown, other differential lighting arrangements could be used in addition to the directional and / or angular illumination, e.g. illuminating the surface with different colours I spectral components (e.g. collecting different images of the surface whilst illuminated with different wavelengths of light), different types of lighting (e.g. illuminated by two or more of: LED, Fluorescent, halogen or laser light), illuminating with different projection methods or patterns (e.g. spot, dome, line and / or the like, which may be chosen to match the feature), different polarisation, different neutral density, other different filter effects, with diffuse and nondiffuse (hard) lighting, and / or the like. However, at least for the particular case of concrete and some other forms of surface 25, it has been found that imaging the surface 25 with different directions and / or angles of illumination is particularly effective for enhancing the identification and characterisation of anomalies such as cracks and other defects.

[0126] In examples that have been found to be particularly effective, different images of the plurality of images may be of the surface being illuminated from at least two pairs of different opposing directions (e.g. up, down, left and right in a plan view), which as will be appreciated are under non-diffuse light, and at least one image of the surface taken under diffuse light, with some or all of the images optionally being greyscale images.

[0127] The image collection apparatus 15 can be operated to collect images (or video that can be considered as a series of images). The image collection apparatus 15 can comprise one or more digital cameras. Beneficially, the image collection apparatus may be configured to have a minimum spatial resolution of 0.3mm or less, e.g. 0.1mm or less. In this way, cracks and other defects that are of a size most likely to require further action can be suitably identified. The image collection apparatus 15 can comprise any suitable camera, such as a forward looking IR (FLIR) camera. An example of a suitable arrangement of image collection apparatus 15 is shown in Figures 5 and 6. The image collection apparatus 15 comprises an imaging sensor 75, such as a CMOS or CCD pixel array, for converting received light into electrical output, an aperture 80 and a focussing lens 85. In a specific example, the imaging sensor 75 is a 5472x3648 pixel array FLIR sensor and the lens is an 8mm focal length lens, which together provide a feature resolution of 0.1mm or less at a working distance D of 350mm and field of view FoV of 574mm x 383mm. However, other arrangements of components and / or components with different properties may be used to achieve the required feature resolution.

[0128] The depth of field DoF defines the distance at which objects remain in focus, in this case: where fn— fl is the f-number which is the ratio of the focal length, fi to aperture ad diameter ad, and C is the circle of confusion: a limit that defines an acceptable level of loss of focus. The circle of confusion C can be calculated by dividing the diagonal size of the camera sensor by 1500.

[0129] Equation (3) shows that high f-numbers result in a high DoF, but this comes with trade-offs, such as a decrease in image capture speed (as image sensor exposure is reduced). Particularly high f-numbers can cause diffraction effects in images, while low f-numbers can cause blurring at image edges.

[0130] The f-number and the distance Fd from the centre of the lens to the DoF can be selected to result in no discernible diffraction effects with acceptable levels of edge blurring. For example, in the specific example described above, f = 8 and Fd = 250 mm. give acceptably little diffraction effect and edge blurring and allows objects to be imaged clearly between lens-object distances of 200 mm to 350 mm.

[0131] The settings of the camera I imaging sensor 75 can be set for acceptable imaging. To compensate for lens image distortion, calibration images such as images of checkerboard patterns of known sizes at various distances and angles can be captured. Distortion correction coefficients can be calculated from these images using a publically available techniques such as a script based on Python’s OpenCV camera calibration module.

[0132] White colour balance and exposure settings can be automatically calculated and adjusted by a camera’s on-board algorithms. However, this is a slow process relative to the rapid changes in lighting conditions created by the illumination system 10. Instead, having preset values for exposure and white-balance for each lighting condition can ensure exposure changes are fast and consistent.

[0133] The present inventors have identified that exposure settings are particularly dependent on lighting angle when the surface 25 is illuminated from a single direction. The exposure setting EQLrequired for a lighting angle 0 is: where E50 is the exposure setting required during diffused lighting at a 50° illumination angle (calibrated once at the beginning of a scan).

[0134] The results of experiments to show the improvement of efficacy achievable by the above use of directional lighting are provided in Figures 7 to 9. To produce these results, loads were applied to reinforced concrete slabs producing various cracks on the surface with different lengths, widths and orientations. Cracks ranged from 0.1 mm to 1 mm wide and 10 mm to 500 mm long. Images of the sample surfaces were captured at a working distance of D = 250 mm, resulting in a FoV ~ 500 x 375 mm.

[0135] In total, 12 distinct 5472 x 3648 pixel images of surfaces 25 that were concrete surfaces of the slabs were acquired. Each area was captured with single direction lighting from orientations O = L, R, U, D, A, where U, D, L, R, A are lighting from up, down, left, right when viewed in a plan view of the surface 25, and diffused (or All directions) respectively. In other words, the L, II, R and D directions are four direction in which each direction is mutually perpendicular to each of the preceding and following directions. For each lighting orientation, incident light angles E)L ranged from 10° to 50° in steps of 10°. An example of an area of the slab during varying lighting conditions is shown in Figure 7, with the insets comparing a single 224 x 224 block within the larger image. Qualitatively, we can see that this block shows enhanced shadowing during the lower angle of 10° when compared to 50°.

[0136] The image quality achieved under various lighting conditions was assessed using the BRISQUE method (Mittal, Moorthy, & Bovik (2012), No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 27(12), 4695-4708), the contents of which are hereby incorporated as if set out in full herein, a non-reference metric that has previously been used to assess the quality of images of cracked concrete (Kim, Choi, Hu, Lee, & Serfa Juan (2021), Multivariate analysis of concrete image using thermography and edge detection. Sensors, 27(21), the contents of which are also hereby incorporated as if set out in full herein.

[0137] BRISQUE scores vary substantially with position in a directionally lit image. AS such, each image was split into sub-images (blocks) of size 224 x 224 pixels, and individual BRISQUE scores calculated for each block. The incident lighting angle is initially fixed at E)L = 50° and the BRISQUE scores are calculated for each block for each lighting direction B^ = BU,D,L,R,A, where U, D, L, R, A are respectively lighting from the up, down, left, and right directions when viewed in a plan view of the surface, and diffuse (all directions) lighting. This approach allows a quality value Qdir for each direction of directional lighting to be defined as follows: The highest quality value Qdir defines the optimum direction O of lighting for each block, where O = [U, D, L, R, A],

[0138] Figure 8 shows a typical colour contour map of the optimum lighting direction, O, overlaid on a concrete crack image. The map shows that while diffused lighting performs well in most regions, some regions distinctly benefit from being lit by directional lighting, particularly from the left or right.

[0139] Once the optimum lighting direction, O, has been found, the optimum incident lighting angle, QLcan be found in a similar fashion to the approach described above for direction, by defining a quality value Q©L as: where B0 &Lis the BRISQUE score for a block lit from its optimum direction, O, and at an incident lighting angle of 0 / .. In Figure 8, the optimum OL values have been added as numbers to each corresponding block. Figure 8 shows that directional lighting (that is, illuminating the surface 25 from a particular direction when seen in a plan view of the surface 25 and at an oblique illumination angle to plane of the surface 25) is almost never worse than diffused lighting, and in many cases may be significantly better. It also shows that there are some regions where a lower oblique illumination angle OL may be better for image quality and other areas where a higher oblique illumination angle OL may be better.

[0140] Figure 9 shows a histogram of the BRISQUE scores for all blocks during diffused, Bditf, and optimized, B0 QL, illumination and associated kernel density estimate (KDE). This shows that the optimised lighting conditions provide an overall improvement in the quality of the image.

[0141] Having identified that optimized directional lighting can significantly enhance the images of the surface 25, thereby making them more suited to identifying anomalies such as cracks and other defects, the present inventors also seek to further improve anomaly detection by employing enhanced image analysis methods to analyse the images of the surface to identify and characterize anomalies. The process for analyzing images of surfaces 25 is illustrated in Figure 10.

[0142] The process of Figure 10 is a hybrid image analysis process using both machine learning or Al techniques (so called “black-box” techniques) and classical techniques (so called “white-box” techniques). This hybrid process that combines both black-box and white-box techniques has been found to be particularly effective at processing images of the surface 25 to identify anomalies such as cracks and other defects. The process could be carried out by the controller 50 shown in Figures 1 and 2, or by an external processing resource, which could be local or remote to the controller 50, e.g. in a server, cloud computing resource, central monitoring facility or other suitable system, or the method could be performed by a combination of any two or more of the controller 50, the local processing resource and / or the remote processing resource, e.g. in a distributed computing arrangement.

[0143] In step 1005 one or more images 1010 of the surface 25 are received. Beneficially, the images 1010 received may comprise at least one image 1010 from a plurality of images, wherein different images of the plurality of images show the surface whilst being illuminated from different directions and / or at different angles to the plane of the surface 25 obtained from the anomaly capture system 5 shown, and described above, in relation to Figures 1 to 7. In one possible example, the method processes at least the image 1010 showing the surface 25 illumined from the direction and / or angle to the surface giving rise to the best quality value QQL, e.g. using the process described above. In another example, a plurality images for which different images show the surface 25 illuminated from different directions and / or angles are processed and the results from the different images are combined, e.g. using ‘OR’ or other suitable combining logic.

[0144] In step 1015, the image 1010 of the surface 25 is input into an identifier model configured to detect and roughly locate features (such as cracks or other anomalies) against a background (i.e. plain concrete in this example). An example of a particularly effective machine learning model that can be used for the identifier model is a region based convolutional neural network (R-CNN), which is a relatively fast neural network. The identifier model is configured to receive the full image 1010 of the surface 25 and process it to identify and roughly locate the portions of the image representing anomalies, such as cracks, and place a bounding box 1020 around the areas identified as representing each anomaly 1025 identified. As such, the output from the identifier model is an image 1030 that comprises the input image 1010 with the bounding box 1020 placed around the anomaly 1025.

[0145] In specific examples, the identifier model is a two-shot region-based CNN (R- CNN) comprising of a region proposal network (RPN), which detects features of interest against a background, followed by a CNN classifier network that classifies the features of interest as belonging to the anomaly or to the background. A specific example of a suitable R-CNN is Tensorflow’s Faster-R-CNN Inception v2 COCO Abadi et al. (2015), the contents of which are incorporated by reference in full, trained via transfer learning on manually-annotated concrete crack images, which may be specifically collected and manually annotated images subject to directional lighting and / or diffuse lighting.

[0146] The identifier model is a relatively fast operating model, e.g. it is computationally efficient and provides fast output, at least relative to the locator and / or classifier model described later.

[0147] According to the method, in step 1035, the output 1030 of the identifier model is sectioned into blocks of set or pre-set sizes. In this example, the blocks are 224x224 pixel blocks but could be of any other suitable size. The result of this is an image 1040 formed from the image 1030 output by the identifier model but sectioned into the blocks.

[0148] In step 1045, the image 1040 is effectively cropped by sorting the boxes in the image 1040 into those boxes that are within the bounding box 1020 and those boxes that are outwith the bounding box 1020. Specifically, any blocks lying within the bounding box 1020 with an overlap of 1 are labelled “in" and passed on to the next step for further processing and any blocks outside the bounding box have all their pixels labelled as "background", step 1047 (and optionally set aside). The labelled image resulting from this process is shown as 1050.

[0149] The boxes of the image 1050 that are within the bounding box 1020 are input to a locator and / or classifier model in step 1055. The locator and / or classifier model acts on each block within the bounding box 1020 in order to more accurately locate and classify the blocks within the bounding box 1020 that represent at least part of an anomaly. Specifically, the locator and / or classifier model is configured to identify those blocks within the bounding box 1020 that overlap with an anomaly.

[0150] In an example, a particularly beneficial locator and / or classifier model comprises a convolutional neural network such as VGG16-CNN. The locator and / or classifier model is run on each block within the bounding box 1020 to identify if the block contains part of an anomaly. If the block does contain part of an anomaly, then the block is passed for pixel level analysis, e.g. by a classical or “white-box” analysis. If the block is determined not to contain part of the anomaly (e.g. a negative detection result), then all the pixels in those boxes are labelled as “background”, step 1047, and set aside.

[0151] In more detail, the objective of the locator and / or classifier model is to filter the blocks (e.g. the 224 x 224 blocks) within a bounding box (determined by the identifier model) into “positive” and “negative” categories. In this case, “positive” disseminates a crack detected within the block, a “negative” block is a block within the bounding box with no crack in it.

[0152] The identifier model is highly computationally efficient and more computationally efficient than the locator and / or classifier model. As such, by using the identifier model prior to the locator and / or classifier model, the computational burden resulting from the locator and / or classifier model can be lowered and processing time improved by reducing the number of predictions required.

[0153] The VGG-16 model used in some examples herein is described in Simonyan & Zisserman (2014) Very deep convolutional networks for large-scale image recognition. arXiv, and the contents of which are incorporated by reference as if set out in full herein. It uses 13 convolutional and 5 pooling layers which feed into 3 fully connected layers. In the present example, transfer learning is used to train the VGG-16 model on an extensive data set gained using directional lighting apparatus and / or training data showing diffuse lighting. Although the VGG-16 model has been found to be particularly effective for this process, the present disclosure is not limited by this and other suitable machine learning or other models can be used.

[0154] The output 1060 of the locator and / or classifier model is the boxes from the image 1050 that represent part of an anomaly. This is provided as input to a pixel level analysis model in step 1065. In this example, the pixel level analysis model uses classical or “white-box” techniques to refine the boxes that represent part of an anomaly to the pixels that represent the anomaly, e.g. the crack. The classical or “white-box” techniques can include the likes of thresholding, edge detection and / or the like.

[0155] In more detail, the input to the pixel level model comprises the blocks (e.g. the 224 x 224 blocks) previously labelled “positive” for cracks by the locator and / or classifier model (e.g. the VGG-16 algorithm). This hybrid approach of running a classical or white-box analysis on a reduced number of blocks filtered by machine learning or Al models such as the R-CNN and VGG-16 models described above, improves the accuracy of the final binary image output as it can completely remove any background noise that would have been apparent in “negative” blocks. It may also remove false positive pixels due to features that merely resemble cracks but can be filtered out by the machine learning models (e.g. the R-CNN or VGG-16 models), for example, due to markings on the surface. An example of a suitable classical or “white-box” technique that can be used for pixel level analysis is edge detector spatial domain methodology used for crack detection, such as that described in Dorafshan, Thomas, & Maguire (2019) in Construction and Building Materials, 186, 1031-1045, which is incorporated herein by reference as if set out in full herein. Other white-box techniques can be applied to other types of anomaly. Beneficially the approach proposed in Dorafshan can be enhanced by (1) applying pre-processing contrast enhancements to each image; and (2) using a 3 x 3 Laplacian kernel is used as the edge detector (however, different kernel sizes and edge detector types could be used instead of the above). The preprocessing enhancements could comprise, for example, one or more of: image multiplication of diffused and directional images, anisotropic diffusion to add a slight blur to areas in the image deemed "not an edge", and / or exposure correction (not doing this currently but may be good to include).

[0156] Figure 11 provides an illustration of the classical or “white-box” technique without previous filtering by machine learning models (such as the R-CNN or VGG-16 models described above). Although the images shown in Figure 11 are full surface images, it will be appreciated in the present example, that the classical or white-box techniques are applied only to the blocks determined to include anomaly 1060 output from the locator and / or classifier model applied in step 1055. The classical or whitebox crack detection image processing techniques are separately applied to each directional lighting image (right, left, up and down), 1070a, 1070b, 1070c, 1070d respectively, to produce a segmented image 1075a, 1075b, 1075c, 1075d, for each lighting direction, with pixels labelled as “anomaly” or “background”. The segmented images for all separate directionally lit directions are then combined, e.g. by using a bit wise OR operation or other suitable technique such as combined by summing the images and normalising. For example, if a pixel appears as part of a crack in all images it will have a confidence value of 1. If a pixel appears as part of a crack in half of the images it will have a confidence value of 0.5, and so on.

[0157] The resulting combined image 1080 highlights the extremities of the crack in all lighting directions. Noise removal can be used to remove false positives. For example, for the bitwise OR combination above, this could involve removing groups of pixels that are not part of the anomaly. For image summing (confidence method) this could involve removing pixels of low confidence that are not part of a pixel group that contains higher confidence pixels, but other techniques could be used. As noted above, for input images filtered by the identifier model and / or the locator and / or classifier model, this process is done only on selected blocks, rather than the entire image. This produces a much more accurate combined output image 1080 with the majority of noise removed.

[0158] Figures 12 to 16 show experimental data illustrating the effectiveness of the directional lighting methods and systems described herein. Figure 12(a) shows a sample area of a concrete surface 25 used as the test sample. The surface 25 contains two anomalies 1205, and 1210 in the form of cracks and a superficial surface line 1215 that looks like a crack. Figure 12(a) also shows bounding boxes 1020a, 1020b and 1020c determined for each crack 1205, 1210 and the surface line 1215 by the identifier model (step 1015 in Figure 10). In this case, the identifier model is a R- CNN having a confidence value threshold set at 40%. Notably, the surface 25 comprises superficial surface markings 1220 in the top left hand corner. The identifier model has successfully identified that these markings are not cracks and has filtered these out. However, had classical or white-box methods been applied directly to this image without the prior filtering using the identifier model, then these markings would have given rise to false positives and been incorrectly identified as anomalies. Similarly, the casting markings 1225 in the bottom left hand corner would also likely have given rise to false positive anomaly detections had the image been fed directly as input to classical I white-box methods without prior filtering using the identifier model. As such, not only may the hybrid approach discussed herein reduce computational burden and increase detection speed, it may also reduce false positive rates and improves accuracy.

[0159] Images for each lighting direction (i.e. up, left, down and right in a plan view of the surface 25) are fed into the identifier model and corresponding images 1030a, 1030b, 1030c with associated bounding boxes 1020a, 1020b and 1020c are output.

[0160] The identifier model (e.g. R-CNN) output 1030 for each lighting direction is taken separately. Following this, all blocks within the bounding boxes 1020a, 1020b and 1020c (and any overlapping) are classified as “in” and are sent to the locator and / or classifier model (e.g. the VGG-16 model) for classification. In this example, 32.5% of blocks were determined as “in” by the locator and / or classifier model (e.g. the VGG-16 model), saving a considerable amount of processing time. In this example, a false positive bounding box 1020c was produced over the surface line 1215 created due to casting that resembles a crack. This is eliminated when the required confidence of the identifier model is set >80%. However, as the identifier model is the first step, it is preferable to have a higher recall than precision to ensure nothing is missed. For this reason, a lower confidence bar is set.

[0161] The output of the identifier model is input into the locator and / or classifier model, which filters out blocks within the bounding box that do not contain crack features. The remaining blocks that are determined by the locator and / or classifier model to contain a part of the anomaly are output for classical I white-box processing. In this case, the false positive bounding box 1020c contains no crack feature blocks according to the locator and / or classifier model predictions. Figure 12 (b) shows the filtered blocks output by the locator and / or classifier model and Figure 12 (c) shows the final binary output from the application of classical I white-box techniques.

[0162] Area under curve (AUC) of a precision recall (P-R) plot can be used to assess the accuracy of the identifier model (in this case a faster R-CNN). A higher AUC value corresponds to a more accurate model. The results for each of three test samples are shown for directional lighting in each direction in Figure 13. For each sample, it can be seen that one of the images showing a directionally lit surface 25 is the most accurate. However, the direction giving rise to the greatest accuracy is sample dependent, e.g. it may depend on the anomaly in question and the way the light shines on the anomaly. If the image with the most suitable or best lighting direction is determined (e.g. using the Qdir value discussed above) and used, then significantly improved results over the diffuse situation are potentially achievable.

[0163] A corresponding analysis was carried out to show the effect of angle of illumination with respect to the plane of the surface 25 and the results shown in Figure 14. It can be seen from this that angles from 20° to 50° and more particularly from 30° to 40° gave the best results. For the next steps, only images with the angle of illumination being 40° were used in the analysis of the effectiveness of the locator and / or classifier model and the classical I white-box techniques.

[0164] Figure 15 shows the variation of accuracy of the locator and / or classifier model (in this case a VGG-16 neural network). Accuracy is assessed against a manually labelled ground truth. This shows that appropriate use of directional lighting (particularly an aggregation of lighting from different directions or by selecting a “best” direction) can potentially achieve significant benefits over diffuse lighting (the ‘A’ sample in Figure 15).

[0165] Figure 16 shows a confusion matrix for predicted label against true label for (a) a prior art method by Dorfshan et. al. (2018) Comparison of deep convolutional neural networks and edge detectors for image-based crack detection in concrete. Construction and Building Materials, 186, 1031-1045, and (b) directional images combined using the bitwise OR process described above. These results show that the predictions of background pixels (majority of pixels) is extremely accurate regardless of the method used. However, the conventional Dorafshan method achieved a 54% accuracy in predicting crack pixels, whereas the bitewise OR combination of directionally lit images described above achieved a 69% accuracy.

[0166] From the above, it can be seen that the use of directional lighting (direction and / or illumination angle when viewed in a plan view of the surface 25) can yield significant improvements over diffuse imaging. However, the optimum direction can be different depending on the sample. As such, determining a best direction to illuminate form by using a quality value such as Qdir in order to dynamically select the lighting direction for a given sample may be beneficial. Similarly, although angle appears to be less dependent on sample, there may also be scope for dynamically selecting an optimum angle to use, e.g. based on a quality value Q© / ..

[0167] The use of the robotic arm based system 5 shown in Figures 1 to 4 allows easy manoeuvring of the light sources 30a-30d into the required position and orientation to provide a desired illumination direction and angle to a selected portion of a surface 25. However, variations of the system 5 of Figures 1 to 4 are possible. For example, different numbers of robotic arms 35a-d and light sources 30a-d may be used, such as one, two, three, five or more robotic arms 35a-d and light sources 30a-d. Furthermore, the robotic arms 25a-d may have more or less joints 45a-c and / or rigid sections 40a-b than those shown in Figures 1 to 4. Indeed the light sources 30a-d need not be mounted on robotic arms and other mechanisms to manoeuvre or otherwise provide the light sources 30a-d into or in the position and orientation required to provide the desired direction and angle of illumination of a surface 25 could be used.

[0168] For example, an alternative arrangement for imaging a surface to be inspected whilst selectively illuminating the surface from different directions and / or orientations to the plane of the surface is shown in Figures 17 to 19.

[0169] In this example, no robotic arms are used (although the entire device could be mounted on a robotic arm or other suitable manipulator if required). The system 5’ comprises an illumination system 10’ and an image collection apparatus 15’ in the form of a digital camera. The illumination system 10’ and the image collection apparatus 15’ are both mounted to a rigid support 20’, in this example in the form of a concave shroud but in other examples could comprise different shapes or arrangements. The illumination system 10’ is configured to illuminate a surface 25’, such as but not limited to a concrete surface, from a plurality of directions and / or angles to the plane of the surface 25. The image collection apparatus 15’ is configured to capture images (which may be comprised in different parts of a video) of the surface 25’. Different images of the plurality of images show the surface being illuminated from the different directions and / or different angles to the plane of the surface 25’.

[0170] In this example, the support 20’ in the form of a concave shroud is multifacetted, i.e. at least an inner surface of the support 20’ is provided with a plurality of facets 3T, 32’, 33’, 34’ in which different facets 3T, 32’, 33’, 34’ are oriented differently from at least one or each other facet 3T, 32’, 33’, 34’.

[0171] The illumination system 10’ comprises a plurality of light sources 30’, wherein different light sources 30’ are mounted on differently angled facets 3T, 32’, 33’, 34’ of an inner surface of the concave support 20’. Furthermore, the support 20’ has a multisided shape, such as two, three, four or more sides (in this example generally square in a plan view). Different light sources 30’ are also provided on different sides on the inner surface of the support 20’ (and optionally on different facets on different sides). The inner surface of the concave support 20’ is open at one end. In this way, the system 5’ can be placed with the open end against the surface 25’ to be tested and different light sources 30’ on different sides of the inner surface of the support and / or different facets can be selectively activated and deactivated to provide illumination at different times. The image capture apparatus 15’ can capture images or parts of video of the surface 25’, wherein different images or parts of the video show the surface 25’ being illuminated by different light sources 30’ on different sides and / or different facets 3T, 32’, 33’, 34’ thereby illuminating the surface 25’ from different directions and / or angles to the plane of the surface 25’ respectively. In this way, the light surface 25’ can be illuminated from different directions in a plan view and / or different angles to the plane of the surface by simply selectively activating and deactivating selected light sources 30’ on different sides and / or facets 3T, 32’, 33’, 34’, and without necessarily requiring a robotic arm (although one could be used to manoeuvre the entire system 5’ if required). The inner surface of the shroud can be beneficially matt black or some other non-reflective finish, in order to ensure directional illumination and avoid diffuse illumination. The concave shroud form of the support 20’ allows diffuse light from the outer environment, e.g. room lighting, sunlight and / or the like to be better excluded from the images.

[0172] It will be appreciated variations on the above specific examples could be made. For example, whilst the examples described above could be used to produce a 2D combined output image, the techniques described above could be used with image collection apparatus configured to collect images suitable for photometric stereo analysis and / or other 3D imaging techniques, from which surface normals and / or 3D data representing the surface can be obtained.

[0173] In examples, any of the the anomaly capture systems described above can be configured to collect images for use with photometric stereo image processing, e.g. in order to determine normals to the surface at a plurality of locations. In this case, a plurality of images of the surface can be collected from a fixed point of view, with each image of the plurality of images being captured under different lighting conditions. The number of images could be at least one image, e.g. from 1 to 100 images, for example from 5 to 20 images, such as 10 images. This number of images has been found to give a good trade-off between accuracy and processing efficiency.

[0174] Any of the anomaly capture systems described above may be configured to determine a 3D model or mesh of at least part of the surface, which may be by using the photometric stereo, stereoscopic or other 3D imaging, e.g. by using the determined normals. For example, 3D data of the surface generated from the determined normals, can be stitched, merged or otherwise combined together to form the 3D model of mesh.

[0175] The 3D data of the surface could be stitched, merged or otherwise combined together to form a 3D model or mesh of part of the surface, which could be output as a visualisation.

[0176] The anomaly could be highlighted in the 3D model or mesh and the 3D model or mesh manipulated to further explore and / or characterise any anomalies or predict future anomalies (e.g. based on local elevations in surface height or localised changes in surface height over time). The images could be colour or greyscale images. In another example a map of surface normals can be created. This could optionally be used to visualise the surface and / or characterise any anomalies. The 3D model or mesh may be created in part, or supplemented, by using other 3D imaging techniques such as stereo photogrammetry, infra-red imaging (such as that used in Kinect systems by Microsoft), LIDAR, RADAR, magnetic or electrical field sensors, laser scanners and / or the like.

[0177] While various arrangements of different illumination directions and / or different illumination angles and optionally also other differential lighting arrangements are described above, in examples that have been found to be particularly effective, different images of the plurality of images may be of the surface being illuminated from at least two pairs of different opposing directions (e.g. up, down, left and right in a plan view), which as will be appreciated are under non-diffuse light, and at least one image of the surface taken under diffuse light, with some or all of the images optionally being greyscale images. The different channels, e.g. wherein each channel comprises images of the surface with different illumination directions or angles and / or under different lighting types (e.g. diffuse light or non-diffuse light), or images or data derived therefrom, could be used as inputs to a pixel segmentation process, e.g. to segment pixels representing anomalies from the those representing the rest of the surface.

[0178] In examples, the surface can be illuminated with a selected colour and / or with a selected wavelength or wavelength band, which can be selected to according to a material of the surface or material on the surface or a type of anomaly. For example, the colour or wavelength or wavelength range of the light used for illumination can be selected to accentuate or highlight or alternatively to reduce or diminish the anomaly or a material on the surface. As an example the images may be selectively illuminated with blue light, which may be particularly suitable for imaging red corrosion, as the red corrosion may be accentuated in the images. As another example, illumination with green lighting can be used to reduce, diminish or render negligible the effect of plant or other organic growth such as moss or algae on the surface.

[0179] As such, the drawings provided herein are for illustration only and the scope of protection is defined by the claims.

[0180] Method steps of the invention can be performed by one or more programmable processors executing a computer program to perform functions of the invention by operating on input data and generating output. Method steps can also be performed by special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit) or other customised circuitry. Processors suitable for the execution of a computer program include CPUs and microprocessors, and any one or more processors. Generally, a processor will receive instructions and data from a read-only memory or a random access memory or both. The essential elements of a computer are a processor for executing instructions and one or more memory devices for storing instructions and data. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto-optical disks, or optical disks. Information carriers suitable for embodying computer program instructions and data include all forms of non-volatile memory, including by way of example semiconductor memory devices, e.g. EPROM, EEPROM, and flash memory devices; magnetic disks, e.g., internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks. The processor and the memory can be supplemented by, or incorporated in special purpose logic circuitry.

[0181] To provide for interaction with a user, the invention can be implemented on a device having a screen, e.g., a CRT (cathode ray tube), plasma, LED (light emitting diode) or LCD (liquid crystal display) monitor, for displaying information to the user and an input device, e.g., a keyboard, touch screen, a mouse, a trackball, and the like by which the user can provide input to the computer. Other kinds of devices can be used, for example, feedback provided to the user can be any form of sensory feedback, e.g., visual feedback, auditory feedback, or tactile feedback; and input from the user can be received in any form, including acoustic, speech, or tactile input.

Claims

CLAIMS:

1. An anomaly capture system comprising: an illumination system; and at least one image collection apparatus; wherein the illumination system comprises at least one light source configured for directionally illuminating a surface, the illumination system being configured to selectively illuminate the surface from one or more selected directions when seen in a plan view of the surface and / or at one or more selected illumination angles to the plane of the surface; and the at least one image collection apparatus is configured to image the surface whilst the surface is being illuminated from the one or more selected directions and / or at the one or more selected illumination angles to the plane of the surface.

2. The anomaly capture system of claim 1, further comprising a support; and at least one or each of the light sources is mounted to the support via at least one manipulator for manoeuvring the light sources to adjust the location and / or orientation of the respective light source.

3. The anomaly capture system of claim 2, wherein the at least one manipulator comprises a robotic arm.

4. The anomaly capture system of claim 3, wherein the at least one arm comprises at least one joint, wherein the arm is movable at each joint under the action of a driver.

5. The anomaly capture system of any of claims 2 to 4, wherein the at least one manipulator is controllable by a controller so as to selectively vary the direction and / or illumination angle that the light from the respective light source is incident on the surface.

6. The anomaly capture system of any preceding claim, wherein the illumination system comprises a plurality of light sources configurable to face in different directions, wherein each light source is independently or individually selectivelyactivatable to provide illumination and deactivatable to turn off illumination from that light source.

7. The anomaly capture system of any preceding claim, wherein: the at least one image collection apparatus is controllable to capture a plurality of images or video of the surface; and the illumination system is configured to selectively illuminate the surface from different sides of the surface when seen in a plan view and / or at different illumination angles to the plane of the surface whilst different images or parts of the video are collected.

8. The anomaly capture system of any preceding claim, wherein: the one or more image collection apparatus are configured to image the surface from at least one of: different directions, different locations and / or different angles to the plane of the surface; and the one or more image collection apparatus are configured to collect a plurality of images or video of the surface whilst the surface is illuminated by the illumination system, wherein at least one of the images may be collected by the one or more image collection apparatus from at least one of: a different direction, a different location and / or a different angle to the plane of the surface, to that used to collect at least one other image.

9. The anomaly capture system of any preceding claim, wherein: the illumination system is configured to selectively illuminate the surface with one or more of: different colours; different spectral components or wavelengths of light; different types of lighting; different projection methods or patterns ; different polarisations; different neutral densities; and / or different filter effects.

10. The anomaly capture system of any preceding claim, wherein the illumination system is configured to dynamically change the direction in which the surface isilluminated based on a quality value of a resultant image to optimise the quality value of the images collected.

11. The anomaly capture system of any preceding claim, wherein the light sources are arranged, or the illumination system is configured to arrange the light sources, so as to illuminate the surface at an angle to the plane of the surface in a range from 5° to 50°'12. The anomaly capture system of any preceding claim, wherein the anomaly capture system comprises, or is configured to communicate with, an analysis system for identifying and / or characterizing anomalies in the surface from at least one image of the surface collected by the image collection apparatus.

13. The anomaly capture system of claim 12, wherein the analysis system is configured to distinguish portions of the images or parts of the video representative of an anomaly from portions of the images or parts of the video representative of background.

14. The anomaly capture system of claim 13, wherein the analysis system comprises and / or is configured to implement one or more models, wherein the one or more models comprise one or both of: an identifier model configured to identify portions of the images or parts of the video representative of the anomaly and assign one or more bounding boxes around the identified portions of the images or parts of the video representative of the anomaly; and / or a locator and / or classifier model configured to locate and / or classify those blocks of the at least one of the images or the at least part of the video within the bounding boxes that represent at least part of an anomaly.

15. The anomaly capture system of claim 14, wherein the identifier model comprises a region-based convolutional neural network, R-CNN, and the locator and / or classifier model comprises a VGG-16 neural network.

16. The anomaly capture system of any of claims 14 or 15, wherein the analysis system is configured to implement one or more white-box or classical imageprocessing methods configured to further refine the blocks identified as comprising portions of the at least one of the images or the at least part of the video representative of the anomaly.

17. The anomaly capture system of any of claims 12 to 16, wherein the analysis system is configured to combine output of at least one of: the identifier model and / or the at least one locator and / or classifier model and / or the white-box or classical techniques obtained from a plurality of images or parts of the video, wherein different images or parts of the video from the plurality of the images or parts of the video show the surface being illuminated from different directions when seen in a plan view of the surface and / or at different illumination angles with respect to the plane of the surface.

18. The anomaly capture system of any of claims 12 to 17, wherein the analysis system is configured to dynamically determine an optimal or preferred configuration, the optimal or preferred configuration comprising one or more of: an optional or preferred direction and / or angle of illumination of the surface; an optimal direction, location and / or angle of the image collection apparatus; and / or an optimal or preferred colour of wavelength of illumination, the analysis system being configured to determine the optimal configuration by collecting a plurality of images of the surface that are one or more of: illuminated from different directions and / or illumination angles; collected from different locations and / or angles of the image collection apparatus; and / or with different colour of wavelength of illumination and determining a quality value of each image and control the illumination system into the configuration determined to be optimal or preferred.

19. The anomaly capture system of any preceding claim, comprising a concave shroud and wherein the at least one image collection apparatus and one or more light sources of the illumination system are mounted on the inside of the shroud.

20. The anomaly capture system according to any preceding claim configured to determine a 3D model or mesh of the surface at least in part from imagescollected by the image collection system and / or surface normals derived therefrom.

21. A method of identifying and / or characterising anomalies in a surface, the method comprising: providing an illumination system; and at least one image collection apparatus; wherein the illumination system comprises at least one light source configured for directionally illuminating a surface, the method further comprising selectively illuminating the surface from at least one selected direction when seen in a plan view of the surface and / or at one or more selected illumination angles to the plane of the surface using the illumination system; and imaging the surface from at least one direction using the at least one image collection apparatus whilst the surface is being illuminated from the at least one selected direction and / or at the one or more selected illumination angles to the plane of the surface.

22. A computer program product configured such that, when implemented on a controller of an anomaly capture system of any of claims 1 to 20, causes the controller to control the illumination system to selectively illuminate the surface from at least one selected direction when seen in a plan view of the surface and / or at one or more selected illumination angles to the plane of the surface; and the at least one image collection apparatus is configured to image the surface from at least one direction whilst the surface is being illuminated from the at least one selected direction and / or at the one or more selected illumination angles to the plane of the surface.

23. A method of analysing images of a surface to determine anomalies in the surface, the method comprising applying a hybrid analysis that comprises applying one or more machine learning or black-box models to an image of the surface to determine blocks of the image containing at last a part of the anomaly and to apply one or more classical or white-box techniques to the blocks of the image determined to contain at least part of the anomaly to identify portions of the blocks that represent the anomaly from portions of the blocks that don’t represent the anomaly.

24. A computer implemented analysis system configured to implement the method of claim 23.

25. A computer program product configured such that, when implemented on a computer implemented analysis system, causes the analysis system to implement the method of claim 23.

Images