Improved object detection reduction

Abstract

The disclosed method obtains a first set of representational data of an object class and applies a generative adversarial network, trained on the object class, to generate a second set of representati

Description

A Convolutional Neural Network (CNN) is a type of artificial neural network typically used to evaluate visual information of an image for object detection and / or classification. The CNN is typically trained using a supervised learning approach on an object class such that the CNN can identify and locate an object of the object class within an image. CNNs typically comprise a number of layers that interact to detect and / or classify an object in an image. In particular, the layers of a CNN are often referred to as convolutional layers, pooling layers, and fully-connected layers. The convolution layers typically extract various features from the input images by applying convolutional mathematical operations, the pooling layer typically decreases the spatial size of the features extracted by a convolution layer, for example, by applying a max pooling operation, and the fully-connected layer typically includes a number of weights / biases and a fully-connected set of neurons to provide an output of the detection / classification of an object in the input image. CNNs are well known in the art and as such a detailed description of a CNN is not provided herein. The Applicants have recognised that in order to prevent, or substantially reduce, the detection of an object of an object class by a detector utilising a CNN, it is preferable to target the ability of the CNN to detect and / or classify the object, rather than, for example, target the sensors that feed into the CNN. In other words, if the CNN can be targeted such that it is unable to detect and / or classify an object in an image, then the CNN can be prevented, or substantially impaired, meaning that the object would not be detectable, or substantially less likely to be detectable by the CNN of the detector system. Thus, one or more embodiments are directed to, inter alia, determining one or more combinations of observable features of the object class that would not be detectable, or difficult to detect, by a CNN of a detector system as an object of the object class in question. An observable feature is a feature of the object that is observable, e.g. “seen”, by one or more sensors used as input to a CNN of a detection system. Furthermore, the observable feature should preferably also be one that is an interpretable feature, e.g. a feature that is modifiable and is also understandable, or interpretable, to a human. For example, if the sensor is a camera then an observable (and interpretable) feature would be a physical, external, attribute of the object that is visible in an image, or a video, recorded by the camera and provided as an input to the CNN of the detection system. For example, if the object is a drone the observable (and interpretable) features may include the number of rotors, the shape of the canopy, the shape of the landing gear, the position of the camera, the colour, and so on. If the sensor is a radar or microwave sensor, then an observable (and interpretable) feature may be a unique, external attribute identifiable within the radar or microwave measurements, such as a radar cross-section (RCS) of the object, which quantifies the extent to which an object can be detected by radar or microwave systems. The RCS may be influenced by various parameters, i.e. observable features, including the objects dimensions, structural design, and the materials used in its construction. If the sensor is a Light Detection and Ranging (LiDAR) system, the observable (and interpretable) features may include the three-dimensional geometric configuration (including shape information) of the object. For example, if the object is a drone then the three-dimensional geometric configuration (i.e. observable and interpretable features) may encompass the main body, propellers, and any ancillary equipment, all of which would be discernible within the LiDAR's point cloud data. If the sensor is an infrared (IR) sensor, the observable (and interpretable) features may include the thermal footprint emanating from the object. For example, if the object is a drone, the thermal footprint may emanate from a propulsion system and any electronic components, wherein such a thermal footprint (or signature) may provide a basic outline of the object’s (e.g. drone) proportions and configuration, contingent upon the sensor's resolution. The observable features of an object of the object class that are interpretable may be collated in an interpretable feature map for the given object, wherein the interpretable feature map may be used as a reference to interpret and understand extracted latent feature map(s), which will be discussed in more detail hereinbelow. Accordingly, there may be an interpretable feature map generated for each object in the object class. For example, if the object class is a drone class, then an interpretable feature map may be generated for each drone model within the drone class. Preferably, each interpretable feature map is tagged with a reference to the associated object of the object class. In the embodiments, the Applicants have recognised that a Generative Adversarial Network (GAN) in combination with an explainable artificial intelligence (AI) such as a dimension reduction algorithm (DRA), may be utilised to determine one or more combinations of observable features that would not be detectable, or difficult to detect, by a CNN of a detector system as an object of the object class in question. Figure 1 shows a schematic of a system 101, for example, a computer system, that includes a processor 102, a memory 103, a GAN 104, a DRA 105, an analysis module 106, and a manufacturing process 107. The GAN 104 is a machine learning model that utilises two neural networks, typically referred to as a generator and a discriminator, to become more accurate in their predictions such that it can artificially produce outputs that can be recognisable as real data. For example, a GAN can be trained, via unsupervised learning, to generate realistic images of an object that are recognisable as an object of the object class, e.g. human faces, animals, etc. The generator of the GAN is often a convolutional neural network, and the discriminator of the GAN is often a deconvolutional neural network. The generator generates new examples, or versions, of the object of an object class and the discriminator classifies the generated versions of the object based on extracted features of the generated versions of the object as real, i.e. recognisable as an object of the object class, or fake, i.e. unrecognisable as an object of the object class. As discussed above, traditionally the GAN is trained in an unsupervised manner until the versions of the object generated by the generator effectively fool the discriminator, meaning that the artificially generated versions of the object are classified, or detected, as an object of the object class. However, the Applicants have recognised that a GAN can be utilised in effectively the opposite manner to its normal function, that is to identify observable features, and, in particular, combinations of observable features of an object that are unrecognisable, or hard-to-learn for a CNN detector. In particular, a GAN has the advantage that it has a similar generalisation ability to convolutional layers in a CNN-based object detector / classifier and therefore, if the GAN is unable to generate artificial, or synthetic, versions of the object based on certain observable features, or combinations of observable features, of the object then those observable features, or combinations of observable features, of the object would be hard-to-leam by, or be substantially unrecognisable to, a CNN based detector system. In other words, if the GAN is unable to generate artificial, or synthetic, versions of the object based on one or more observable features, or combinations of observable features, of the object, then it is likely that a CNN based detector system would not be able to detect, or classify, an object as being part of the object class based on those observable features or combinations of observable features. Accordingly, those identified, or determined, observable features, or combinations of observable features, that would be unrecognisable, or hard-to-leam, as an object of the object class, utilising the GAN 104, the DRA 105, and the analysis module 106, can be used to manufacture, via a manufacturing process 107, an object, or part thereof to refit an object, such that a CNN based object detection system would not be able to detect and / or classify the object as being part of the object class. With reference to Figure 2, in embodiments, the GAN 202 obtains, or receives, a first set of representational data 201, where the first set of representational data 201 refers to data that is representational of the object and relates to real world examples, or versions, of one or more objects of the object class. For example, the first set of representational data 201 may be of visual images (still or video) relating to the object class, the representational data may be of radar output relating to the object class, the representational data may be of microwave output relating to the object class, the representational data may be of a Light Detection and Ranging (LiDAR) output relating to the object class, the representational data may be of acoustic output relating to the object class, the representational data may be of InfraRed output relating to the object class, and so on. Thus, representational data includes any data that represents the object class in a manner on which the GAN is being trained on the object class. The GAN 202 may be trained 209 on the object class using the first set of representational data 201, or may be trained on the object class using a further different set of representational data, or a combination of the first set and the further set of representational data of the object class. The trained GAN typically includes a plurality of layers. The GAN 202, via the generator 203, generates the second set of representational data 205.For example, using the knowledge in the generator 203 gained by training the GAN on the object class, the generator 203 can generate the second set of representational data by transforming a given noise, by each layer of the GAN generator 203, into one or more observable features of the object class. In essence the generator 203 is generating further representational data of the object class by transforming, or modifying, observable features of the object class represented by the first set of representational data 201 based on a given input noise. Thus, the second set of representational data 205 relates to artificial, or synthetic, examples, or versions, of a plurality of objects of the object class. For example, the generator 203 may generate synthetic versions of the object given a random noise (preferably sampled from a latent space and following a certain data distribution). During the generation, the input noise traverses through a number of fractionally-strided convolutional layers comprising linear and nonlinear operations, gradually forms the observable data of the target object with visual feature distributions learned during GAN training. The GAN 202, via the discriminator 204, then extracts 206 a first latent feature map for each of the first set of representational data of the object class to generate a first set of latent feature maps, wherein each first latent feature map corresponds to combinations of observable features of the object class in the first set of representational data, and extracts 206 a second latent feature map for each of the second set of representational data of the object class to generate a second set of latent feature maps, wherein each second latent feature map corresponds to combinations of observable features of the object class in the second set of representational data. For example, all but the last fully connected layer of the discriminator 204 may be utilised as a feature extractor, wherein the observable features traverses through the feature extractor to provide the latent feature map(s) in the form of matrices. Each floating-point in the matrix may represent the degree to which the current observable feature contains a certain latent feature. Each latent feature map relates to the observable features of a given representational data (e.g. real or synthetic representational data). In order to identify the hard to learn combinations of observable features for a CNN based detector system, a dimension reduction algorithm 207 is then applied to the first set of latent feature maps and to the second set of latent feature maps. The dimension reduction algorithm 207 generates a set of nodes, or discrete entities, wherein each node, or entity, comprises one or more latent feature maps from the first set of latent feature maps and / or the second set of latent feature maps relating to an associated combination of observable features. Each node includes latent feature maps (relating to the first set of representational data and the second set of representational data) with the same, or similar, observable features, wherein the observable features are determined, or predetermined, by the dimensional reduction algorithm. In embodiments, the dimensional reduction algorithm is a Topological Data Analysis (TDA) and therefore the observable features focus on topological features such that each node includes latent feature maps that have the same, or similar, topological features. However, as will be appreciated, different dimensional reduction algorithms may focus on different observable features in generating the set of nodes. The dimensions reduction algorithm 207 compares the data distribution of the first set of latent feature maps (relating to the first set of representational data of real-world versions of one or more objects in the object class) with the data distribution of the second set of latent feature maps (relating to the second set of representational data generated by the GAN of a plurality of synthetic versions of objects in the object class). The comparisons enable the dimension reduction algorithm to cluster data, in the nodes, that have similar, or associated, combinations of observable features and each node may include the representational data associated with the latent feature map for the combination of observable features. Figure 3 shows a schematic diagram of the clustered data in nodes 301. As will be appreciated, Figure 3 shows a small subset of example nodes where, in reality, the clustered data may include tens, hundreds, or thousands of nodes. Each node 302, 303, 304, may include a number of latent feature maps for representational data from the first set of representational data and a number of latent feature maps for representational data from the second set of representational data with the same, or similar, observable features. Examples of the number of latent feature maps for representational data from the first set of representational data and the number of latent feature maps for representational data from the second set of representational data is shown for an example subset of nodes 305a-d. A ratio of the number of first latent feature maps and the second latent feature maps for each node of the generated set of nodes is then determined. Each node may then be classified based on the ratio of the number of extracted latent feature maps of the first set of representational data and number of extracted latent feature maps of the second set of representational data in each node, and a comparison with at least one predetermined threshold. A node may be classified as a weak node 303, a balanced node 302, or a failed node 304, wherein a weak node 303 includes a ratio of the number of latent feature maps from the second set of representational data to the number of latent feature maps of the first set of representational data that is below the predetermined threshold, a balanced node 302 includes a includes a ratio of the number of latent feature maps from the second set of representational data to the number of latent feature maps of the first set of representational data that is above the predetermined threshold, and a failed node 304 only includes latent feature maps from the second set of representational data. The Applicants have recognised that the nodes of interest for determining, or identifying, combinations of observable features for the object that are hard-to-learn, are the weak nodes 303 as a weak node 303 means, or represents, that the generalisation ability of the GAN on the combinations of observable features in the node is weak. Accordingly, one or more weak nodes 303 from the set of nodes is then identified based on a comparison of the ratio for each node to a predetermined threshold, wherein the identified one or more weak nodes 303 include a ratio of the number of second latent feature maps to the number of first latent feature maps that is below the predetermined threshold for the combination of observable features The connections, or links, 306 between the nodes indicates the distance from another node, wherein the distance may indicate a degree of difference between the observable features, or combination of observable features, of the two nodes. Thus, a weak node 303 may be considered to be a node that includes none, or a low number, defined by the predetermined threshold, of the second latent feature maps, which means that it is difficult for the GAN to generate artificial, or synthetic, representational data with the combination of observable features represented by the node. In other words, the one or more weak nodes are indicative of one or more combinations of observable features of the object that the trained GAN has a low object classification success rate and / or a low generalisation ability on. The generalisation ability of the GAN on a given object of an object class is reflected by the GANs ability to generate (diverse) synthetic versions of the object in the object class. For example, the generalisation ability of the convolutional layer(s) of the GAN to the one or more combinations of observable features, e.g. physical or geometry attributes, can be considered to be whether the GAN is able to generate a significant number (e.g. above the predetermined threshold) of representative synthetic versions of the object including the one or more combinations of observable features. The predetermined threshold may be any suitable predetermined threshold, for example, it may be 5%, 8%, 10%, 15%, 20%, and so on. A lower predetermined threshold for a node to be considered a weak node, say 5%, that is a node includes a total of 5% or less second latent feature maps in comparison to the number of first latent feature maps, indicates that the combination of observable features associated with the node may be hard-to-leam by a CNN detector, as shown in weak node 303 and associated data 305c. The combinations of observable features represented by the second latent feature maps in the weak nodes that are below the predetermined threshold may be analysed, manually or computationally (e.g. via simulation) for incorporating, singularly or in any combination thereof, into the object, or for retrofitting an object, to make the object less detectable by a CNN based detector system. A higher predetermined threshold, say 15%, that is a node includes a total of 15% or less second latent feature maps in comparison to the number of first latent feature maps indicating that the combination of observable features associated with the node are hard-to-learn, means more nodes than the lower 5% predetermined threshold mentioned above will be considered to be weak nodes and as such a greater number of combinations of observable features may be considered to be hard-to-leam by a CNN detector, as shown in weak nodes 303 and associated data 305c and 305d. In the above examples, the predetermined threshold is described as a percentage, however, as will be appreciated, any suitable predetermined threshold may be used, for example, an integer value. In embodiments, once a weak node has been identified the latent feature maps relating to the first set of representational data included in the weak node are obtained along with the associated first representational data for each latent feature map in the weak node, for example, from memory 103. The first representational data of each latent feature map identifies which object of the object class the first representational data represents, or is associated with. Referring back to Figure 2, once each of the objects of the object class that the first representational data of each latent feature map represents is identified, then the interpretable feature map for each of the objects are obtained and analysed computationally, for example, by the analysis module 208, to determine one or more of the combinations of observable (and interpretable) features that are common between each of the identified object(s). As will be appreciated, the analysis may alternatively, or additionally, be performed manually. The determined one or more of the combinations of observable (and interpretable) features that are common between each of the identified object(s) are likely, with a high probability, to be a combination of observable features that are hard-to-learn by a CNN detector. One or more potential design options, e.g. modifications to the object, based on the one or more combinations of observable features determined from the weak nodes for the object may be generated by the analysis module 208 (and / or determined manually) and tested / validated (e.g. by a validation emulator, a simulator, or any other suitable test means) to determine which potential design option(s), e.g. modifications to the object, are functional as well as being hard to detect. One or more of the potential design options, e.g. modifications, for the object, or for a part of the object to retrofit the object, may be selected, for example, based on a selection criteria, and manufactured, for example, by transmitting the selected design option (e.g. modification) to a 3D printer or to any other suitable manufacturing system or process 107. The object incorporating one or more of the hard to learn observable features would, in use, prevent, or substantially reduce, the ability of the CNN based detector system from detecting, or classifying, the object as part of the object class. In the following examples, the object class is an unmanned ariel vehicle (UAV), such as a drone, and the representational data used relates to visual images. However, as will be appreciated, the object class could be any object class for which hard-to-learn features in relation to a CNN detector is important, whether that is for evasion or for an early diagnosis tool. For example, this could include medical applications, e.g. analysing hard-to-learn features relating to early cancerous cell detection. A first set of visual images are obtained wherein each of the visual images in the first set represent a set of 14 different real drone models that are currently available. However, as will be appreciated, any number of different real drone models may be analysed as part of the drone object class dataset from which the first set of visual images (i.e. first set of representational data) 201 are obtained. The visual images in this example are images that are captured based on computer aided design (CAD) drawings of each of the 14 drone models to enable a significant number of images of all observable features of the drone models to be captured. However, as will be appreciated, the visual images may be obtained, or captured, directly from the real drone models by, for example, a camera. For each drone model, the drone image represented by an associated CAD file is manipulated, for example, by applying random 3-Dimensional rotations, and images are captured for the first set of visual images. In this example, 300 visual images are captured for each of the 14 drone models giving a total of 4200 images in the first set of visual images. The images captured as the first set of images may be of any suitable resolution, for example, 1800 x 1500 pixel resolution. An equal number of visual images are captured of each drone in this example in order to avoid bias in the GAN generalisation ability by an unbalanced dataset of the first set of visual images. Each captured visual image forming the first set of visual images may be labelled to identify the particular drone model and the captured image number of that drone model for the purpose of backtracking, if required. Further labels, or tags, may be added to the first set of visual images to indicate each of the observable features, e.g. physical attributes, of the drone, for example, wingspan, rotor shape / height, body shape / dimensions, and so on. Alternatively, or additionally, an interpretable feature map may be determined or generated for each drone model of the drone class, wherein the interpretable feature map may be used as a reference to interpret and understand extracted latent feature map(s) and the associated observable features. The GAN 202 may be any suitable GAN for the purpose of identifying hard-to-leam features for an object of an object class, and, in embodiments, the GAN is a Deep Convolutional GAN (DCGAN). The DCGAN 202 may be trained 209 based on the first set of images 201, thereby utilising the first set of images as a training data set for the DCGAN. The DCGAN may be trained on the images as captured, or the images may be resized appropriately for the DCGAN, for example, to a resolution of 64 x 64 pixels. The training data set of the first set of visual images may also be randomly divided into two or more batches. The DCGAN 202 typically includes a generator 203 and a discriminator 204 and the learning rates for each may be set at any suitable value for the DCGAN being used, and / or for the object class being analysed, for example, the learning rate for the discriminator in this drone example may be set at 4e-3 and the learning rate for the generator in this drone example may be set at 2e-3. Each of the discriminator and the generator may be trained simultaneously or consecutively, and may be trained the same, or different number of times. For example, the discriminator may be trained once and the generator may be trained twice in order to balance the ability of the generator and discriminator in generating synthetic images and discriminating images, and to improve the convergence of the DCGAN. Once trained, the DCGAN effectively learns the data distribution of the raw dataset, i.e. the first set of visual images, and represents the learned data distribution by a DCGAN generated second set of synthetic visual images, i.e. the second set of representational data 205. In this example, the number of generated synthetic visual images in the second set of representational data is the same as the number of visual images in the first set of representational data, namely 4200 synthetic visual images such that the number of synthetic visual images in the second set of representational data is sufficient to express the learned data distribution. Latent feature maps may then be extracted 206 for both the first set of visual images and the second set of synthetic visual images. The latent feature maps may be extracted utilising the convolutional layers of the DCGAN discriminator 204. An advantage of using the convolutional layers of the DCGAN discriminator is that the later, or higher, convolutional layers extract higher dimensional knowledge for an overall understanding of the visual image. The final layer of the convolutional layers may not be used in the latent feature map extraction as this final layer is typically used for result inference. A dimension reduction algorithm 207, such as Topological Data Analysis (TDA), is then applied to the extracted latent feature maps to effectively visualise the data distribution represented by the latent feature maps. The extracted latent feature maps from both the first set of visual images and the second set of synthetic images may be input to a mapper, such as a Kepler Mapper, in order to visualise and compare their data distribution in a uniform dimension reduction (e.g. TDA) space. The dimension reduction algorithm may analyse the combinations of observable features, e.g. the physical attributes and / or structural characteristics of the drones (i.e. the object), in the visual images and the synthetic images by topological information across the combinations of observable features from their latent feature maps, and may cluster data into nodes with similar topological combinations of observable features. Each node may include a number of latent feature maps for visual images from the first set of visual images and an indeterminate number of latent feature maps for synthetic visual images from the second set of synthetic visual images with the same, or similar, observable features. Each node may then be classified based on the ratio of the number of extracted latent feature maps of the first set of visual images and number of extracted latent feature maps of the second set of synthetic visual images in each node, and a comparison with at least one predetermined threshold. A node may be classified as a weak node 303, a balanced node 302, or a failed node 304, wherein a weak node 303 includes a ratio of the number of latent feature maps from the second set of synthetic visual images to the number of latent feature maps of the first set of visual images that is below the predetermined threshold, a balanced node 302 includes a includes a ratio of the number of latent feature maps from the second set of synthetic visual images to the number of latent feature maps of the first set of visual images that is above the predetermined threshold, and a failed node 304 only includes latent feature maps from the second set of synthetic visual images. The Applicants have recognised that the nodes of interest for determining, or identifying, combinations of observable features for the object that are hard-to-learn, are the weak nodes 303 as a weak node means, or represents, that the generalisation ability of the DCGAN on the combinations of observable features in the node is weak. The generalisation ability of the DCGAN on a given drone model, or models, is reflected by the DCGANs ability to generate (diverse) synthetic versions of the drone model(s). For example, the generalisation ability of the convolutional layer(s) of the DCGAN to the one or more combinations of observable features, e.g. physical or geometry attributes of the drone model(s), can be considered to be whether the DCGAN is able to generate a significant number (e.g. above the predetermined threshold) of representative synthetic versions of the drone model(s) including the one or more combinations of observable features. The determined, or identified, weak nodes 303 may then be analysed, for example, by the analysis module 208, individually, or in combination, to determine, or identify, the combinations of observable features included in the respective weak node. For example, one or more combinations of observable features can be identified based on the associated interpretable feature map(s) and / or on labels that were previously added to the first set of visual images and added, or copied, either manually or computationally, to the corresponding second set of synthetic visual images. The labels applied to the visual images and the synthetic visual images, e.g. the original drone model along with the associated observable features (e.g. wingspan, curvature, rotor shape / height, and so on) enable the combinations of observable features of particular drone models within each weak node to be identified based on an analysis of the labels of the visual images and synthetic visual images associated with the extracted latent features maps included in the weak nodes. If the one or more combination of observable features are associated with two or more original drone models, then the process can perform further analysis, for example, apply a statistical regression, to determine and interpret which feature, e.g. physical attribute(s), is of the greatest relevance to the identified one or more combinations of observable features. As an example, if the weak node 303 relates to a particular drone model in which the latent feature maps correspond to the combination of observable features of a high curvature, white in colour, four rotors, a fixed landing gear, an X-shaped body and a bottom-mounted camera, then a drone having the above combination of observable features would be hard-to-learn by a CNN detector system. A drone (e.g. an object) can then be designed and manufactured with one or more of the determined, or identified, combinations of observable features included in one or more of the weak nodes. The drone can be either designed and manufactured with the one or more combinations of observable features, or one or more parts of the drone, e.g. a canopy, can be designed and manufactured to be retrofitted on the drone. In this example of a drone, a modification to the drone design, or a modification to the design of one or more parts of the drone, e.g. canopy, to be retrofitted to the drone can be performed manually or computationally. In the case of computationally generating the relevant design one or more design options, e.g. one or more modifications to the object, may be generated based on the one or more combinations of observable features (i.e. to incorporate one or more of the combinations of observable features in any combination thereof), and validating each design option, e.g. modification to the drone, using a validation emulator, or simulator, to test the functional performance of each modification in order to ensure that the one or more combinations of observable features provide a workable and functional drone (e.g. object) with one or more hard-to learn features (i.e. combinations of observable features) incorporated into the drone. The validation emulator, or simulator, can be any suitable known software. One or more of the design options, e.g. modifications, for the drone, or the one or more parts of the drone (e.g. a retrofit canopy), that are functional may then be selected. The selection of the one or more modification to the drone may be performed by any suitable means, for example, based on a selection criteria, where the selection criteria may include a degree of functionality of the modified drone, the ability to evade detection / classification, or any other suitable selection criteria. The selected modifications to the drone may then be transmitted to a manufacturing system, e.g. a 3D printer, or any other suitable manufacturing system or process, to be manufactured. Thus, the resulting drone modified by and incorporating at least one of the one or more combinations of observable features will have a higher ability and robustness in evading detection and / or classification as a drone by a CNN based detector system. In the foregoing embodiments and examples, features described in 5 relation to one embodiment and / or example may be combined, in any manner, with features of a different embodiment and / or example in order to provide a more efficient and effective object detection reduction. Note that, the above description is for illustration only and other embodiments and variations may be envisaged without departing from the scope of the invention as defined by the appended claims.

Claims

1. A method, comprising:obtaining a first set of representational data of an object class;applying a generative adversarial network trained on the object class, wherein the trained generative adversarial network is operable to:generate a second set of representational data of the object class, wherein one or more observable features of the object class are transformed by the generative adversarial network in the second set of representational data of the object class;extract a first latent feature map for each of the first set of representational data of the object class to generate a first set of latent feature maps, wherein each first latent feature map corresponds to one or more combinations of observable features of the object class; andextract a second latent feature map for each of the generated second set of representational data of the object class to generate a second set of latent feature maps, wherein each second latent feature map corresponds to one or more combinations of observable features of the object class;apply a dimension reduction algorithm to the first set of latent feature maps and the second set of latent feature maps to generate a set of nodes, wherein each node comprises one or more latent feature maps from the first set of latent feature maps and / or the second set of latent feature maps relating to an associated combination of observable features;determine, for each node, a ratio of a number of first latent feature maps and the second latent feature maps in each node;compare the determined ratio for each node to a predetermined threshold; andidentify one or more weak nodes for which the determined ratio is below the predetermined threshold, wherein each weak node is associated with one or more combinations of observable features.

2. The method of claim 1, in which the one or more combinations of observable features associated with the weak nodes relate to one or more combinations of observable features of the object class for which the generative adversarial network has a low generalisation ability.JO O J3. The method of claim 2, in which the generalisation ability of the generative adversarial network is the ability to generate representational data of the second set of representational data that includes the one or more combinations of observable features associated with the weak nodes.

4. The method according to any one of the preceding claims in which the first set of representational data corresponds to real world objects of the object class, and the second set of representational data corresponds to synthetic versions of objects of the object class.

5. The method according to any one of the preceding claims, further comprising: applying one or more labels relating to the object and to one or more observablefeatures of the object to the first set of representational data and to the second set of representational data; and / or generating for each object an interpretable feature map, wherein the interpretable feature map includes one or more observable features that are modifiable.

6. The method according to claim 5, in which the one or more combinations of observable features associated with a respective weak node are identified based on the one or more labels applied to the first set of representational data and to the second set of representational data; and / or based on the interpretable feature map associated with the one or more objects included in the respective weak node.

7. The method according to any one of the preceding claims, in which the generative adversarial network is trained using the first set of representational data.

8. The method according to any one of the preceding claims, in which the generative adversarial network includes a generator comprising a convolutional neural network formed of a plurality of convolutional layers; and wherein the second set of representational data is generated by the generator by transforming the one or more observable features of the object class by each convolutional layer of the generator.

9. The method according to any one of the preceding claims, in which the generative adversarial network includes a discriminator comprising a deconvolutional neural network;and wherein the discriminator is operable to extract the first latent feature map and extract the second latent feature map.

10. The method according to any one of the preceding claims, in which the set of nodes are generated based on a comparison of a data distribution of the first set of latent feature maps with a data distribution of the second set of latent feature maps to cluster data in the nodes that have associated combinations of observable features.

11. The method according to any one of the preceding claims, further comprising: storing in each node the representational data associated with the one or more latent feature maps included in the node.

12. The method according to any one of the preceding claims, further comprising: analysing the one or more combinations of observable features of each identified weak node to generate one or more modifications to the object.

13. The method of claim 12, in which a weak node is associated with two or more objects of the object class, and the step of analysing the one or more combinations of observable features comprises:identifying common combinations of observable features between the two or more objects associated with the weak node.

14. The method of claim 12 or 13, further comprising:validating each of the one or more modifications to the object; andselecting at least one of the one or more modifications to the object based, at least in part, on the validation of each of the one or more modifications to the object.

15. The method of claim 14, further comprising:transmitting the selected at least one of the one or more modifications to the object to a manufacturing system.

16. The method of any one of the preceding claims, in which the one or more observable features of the object class are transformed by the generative adversarial network based on an input noise.

17. A computer program product comprising computer readable executable code for implementing a method according to any one of claims 1 to 16.

18. A system comprising:a processor;a memory, wherein the memory is operatively connected to the processor, the memory comprising one or more instructions for implementing a method according to any one of claims 1 to 16.

Images