Detector of dimmed objects of interest

Abstract

Method of inspecting cargo using penetrating (X-ray) radiation, comprising: obtaining an image of the cargo (Fig.4); detecting objects in the image 11,13; using a trained deep learning algorithm (1,Fig.1) to calculate a quality or dimming vector for each detected object, the dimming quality vector comprising coefficients indicating quality of the object image wherein each coefficient is calculated as a dimmed value obtained from the cargo image as compared to an ideal reference value learned by the deep learning algorithm; and determining whether the objects is an object of interest based on the quality dimming vector. Coefficients may include: statistical measures; descriptors of grey levels; entropy values; texture descriptors; geometrical characteristics; Hu moments values; and / or Fast Fourier Transforms (FFT) features. Objects of interest may be identified by comparing the quality dimming vector to a reference vector. A dimming (quality) score may be generated from the quality vector and displayed to a user. The object of interest may comprise weapons and / or contraband product, such as a drugs, cigarettes or alcohol. The deep learning algorithm may be trained on images with known quality coefficients, the difference being minimised by a mean square error loss function.

Description

Field of Invention The invention relates but is not limited to a device configured to detect an object of interest in an inspection image of cargo. The invention also relates but is not limited to a method for training a deep learning algorithm to detect objects of interest in cargo inspection images by calculating a dimming vector that captures distinguishing object features. The invention also relates but is not limited to producing a device configured to determine whether an object of interest is present in an inspection image generated using penetrating radiation. The invention also relates but is not limited to corresponding methods, devices and computer programs or computer program products. Background Inspection images of containers containing cargo may be generated using penetrating radiation, such as X-rays. These images allow detection of objects of interest, such as weapons or contraband, within the cargo. However, the appearance of such objects may be obscured or "dimmed" in the image due to superposition with the cargo container walls or other objects in the cargo. This dimming effect can make reliable detection challenging. In ambiguous cases, manual inspection of the cargo may be required, slowing the screening process. Worse, dimming may cause objects of interest to go undetected entirely, posing safety and security risks, particularly when screening for threats like weapons. Summary of Invention Aspects and embodiments of the invention are set out in the appended claims. These and other aspects of the invention, and aspects and embodiments which are useful in understanding the invention set out in the appended claims, are also described in the disclosure herein. Any feature in one aspect of the disclosure may be applied to other aspects of the disclosure, in any appropriate combination. In particular, method aspects may be applied to device and computer program aspects, and vice versa. Furthermore, features implemented in hardware may generally be implemented in software, and vice versa. Any reference to software and hardware features herein should be construed accordingly. Brief Description of Drawings Embodiments of the present disclosure will now be described, by way of example, with reference to the accompanying drawings, in which: Figure 1 schematically illustrates an example system and an example device; Figure 2 shows a flow chart illustrating an example method according to the disclosure; Figure 3 illustrates an example reference inspection image according to the disclosure; Figure 4 schematically illustrates an example of an inspection image with detection boxes around detected objects and masks over detected objects; Figure 5 shows a flow chart illustrating another example method according to the disclosure; Figure 6 shows a flow chart illustrating another example method according to the disclosure; and Figure 7 shows a flow chart illustrating another example method according to the disclosure. In the figures, similar elements bear identical numerical references. Description of Example Embodiments Applying the deep learning algorithm to perform dimming vector calculation and object detection As illustrated in Figure 1, the disclosure discloses an example device 15 configured to detect an object of interest in an inspection image of cargo 155. An example inspection image is schematically represented in Figure 4, with a plurality of objects 11-1 to 11-8 of interest. The inspection image is a 2D image generated using penetrating radiation 2 transmitted from an inspection system 154 through the cargo 155 under inspection and projected on detectors 3. In the inspection image, the representations of the objects of interest may be dimmed, e.g.: by a superposition with walls of a container of the cargo and / or other objects in the cargo, and / or by a deformation of shapes due to the nature of the projection of the radiation on the detectors; and / or by a low extent of a penetration of the cargo and / or object by the radiation. As a non-limiting example, the object of interest may comprise at least one of a threat object, such as a weapon, and / or a contraband product, such as a drug product and / or a cigarette product and / or an alcohol bottle. In the present disclosure, in order to assist the detection of the objects of interest in the inspection image of the cargo, a dimming vector is calculated, as explained below. In Figure 1, the device 15 comprises a memory 151 storing a trained deep learning algorithm 1 configured to calculate a dimming vector indicative of features of objects in an inspection image of cargo. The device 15 also comprises a controller 152 coupled to the memory 151. The controller 152 is configured to, in operation, perform the method 100 as illustrated in Figure 2. The method 100 of Figure 2 comprises: obtaining, at S1, the inspection image of the cargo; detecting, at S2, one or more objects in the obtained inspection image (e.g., the objects 11-1 to 11-8 as illustrated in Figure 4); and using, at S3, the trained deep learning algorithm to calculate a dimming vector for each of the one or more detected objects. The dimming vector comprises a set of coefficients that capture visual features of each detected object. These features may include shape, texture, material properties, and other attributes that characterize an object of interest, even when visually obscured. While e.g., 36 coefficients may be used in an example embodiment, the number of coefficients may be varied as needed to capture relevant features for a given application. Each coefficient in the dimming vector is expressed relative to a corresponding reference value learned by the deep learning algorithm during training. This allows the system to recognize objects of interest even when their appearance is "dimmed" relative to an ideal unobstructed view. The reference value, with respect to which each coefficient of the dimming vector is calculated, corresponds to a value of a coefficient of a reference vector calculated for a reference object 11 in a reference inspection image 1000 illustrated in Figure 3, as explained in greater detail below. Figure 3 represents an example reference object 11 of interest (here a weapon of the “rifle” type). In Figure 3 the reference object 11 is represented as detected in the reference inspection image 1000 by an inspection system (for example the inspection system 154 of Figure 1), when the reference object 11 of interest is positioned along a reference axis A-A and in a reference orientation O, and when the reference object 11 of interest is exposed, without any obstruction, to penetrating radiation, perpendicularly to the reference axis A-A and the reference orientation O. An annotation to the reference inspection image 1000 of Figure 3 indicates the reference axis A-A and the reference orientation O of the object 11 of interest which is represented as projected in 2D in the reference inspection image 1000. In Figure 3 the reference axis is “Diagonal” (that is slanted and parallel to the plane of the inspection image 1000). Other non-limiting examples of reference axes are envisaged, such as “Orthogonal” (that is perpendicular to the plane of the reference inspection image 1000). In Figure 3 the reference orientation O is “Left-Right”, that is the orientation shows the sides of the object with respect to the direction of travel of the object 11 of interest under an inspection scan. Other non-limiting examples of reference orientations are possible, such as: “Front-Rear” where the object of reference is oriented to show the front and rear of the object with respect to the direction of travel of the object during the inspection scan, or “Up-Down” where the object of reference is oriented to show the upper and lower sides with respect to the direction of travel of the object during the inspection scan. The reference axis A-A and the reference orientation 0 are standardized, so as to obtain the reference values of the one or more coefficients of the reference vector. As illustrated in Figure 2, the method 100 also comprises determining, at S4, whether the one or more detected objects correspond to an object of interest, based on the calculated dimming vector. In order to determine at S4 whether the one or more detected objects correspond to an object of interest, the controller may compare the one or more calculated dimming vectors to one or more reference dimming vectors of objects of interest, as explained in greater detail below. The inspection image of the cargo comprises pixels of different intensity or grey levels, and the one or more objects detected in the inspection image also comprise pixels of different intensity or grey levels. As explained in greater detail below, the one or more coefficients of the dimming vector calculated for each of the one or more detected objects may comprise at least one of: one or more statistical measures; and / or one or more descriptors of grey levels; and / or one or more values indicative of entropy; and / or one or more descriptors of texture; and / or one or more geometrical characteristics; and / or one or more values indicative of Hu moments; and / or one or more features of Fast Fourier Transforms, FFT. The one or more statistical measures may comprise at least one of: an average of image intensity; and / or a median indicative a central intensity value; and / or one or more quartiles (for example Q25 and / or Q75), indicative of a distribution of intensity; and / or a standard variation, indicative of dispersion of intensity; and / or a skewness, indicative of asymmetrical distribution of intensity; and / or a kurtosis, indicative of flattening of distribution of intensity; and / or a range, indicative of internal contrast within intensity; and / or a variation coefficient, indicative of normalised value of dispersion. The one or more descriptors of grey levels are indicative of a co-occurrence matrix of the grey levels and comprise at least one of: a value indicative of energy, such as an angular second moment, ASM; and / or a value indicative of contrast; and / or a value indicative of homogeneity; and / or a value indicative of correlation; and / or a value indicative of dissimilarity The one or more values which are configured to be indicative of entropy may be indicative of texture complexity. The one or more descriptors of texture may comprise at least one of: one or more values indicative of local binary patterns, LBP; and / or one or more values indicative of Gabor energy. The one or more geometrical characteristics may comprise at least one of: one or more values indicative of an area; and / or one or more values indicative of a perimeter; and / or one or more values indicative a mass centre; and / or one or more values indicative of an aspect ratio: and / or one or more values indicative of a solidity; and / or one or more values indicative of circularity. The values indicative of Hu moments comprise seven constant values indicative of the shape of the one or more objects. The one or more features of the FFT may comprise at least one of: a value indicative of an average of the FFT; and / or a value indicative a standard deviation of the FFT. After the controller has detected the one or more objects (e.g., the objects 11-1 to 11-8 of Figure 4) in the inspection image (for example by using one or more detection boxes 13 as shown in Figure 4), the controller may build a mask over the one or more detected objects. Figure 4 schematically illustrates a plurality of masks 12-1 to 12-8 built on a corresponding plurality of detected objects 11-1 to 11-8. As can be seen in Figure 4, the masks 12-1 to 12-8 may correspond to a polygon representing the detected object, one or more detected objects being located in the detection boxes 13. The detection boxes 13 may be obtained using segmentation of the inspection image. As shown in Figure 2, the controller may calculate, in an optional step S5, a dimming score of the one or more detected objects, based on the calculated dimming vector. The controller may further cause a display of a user interface of the inspection system to display the calculated dimming score for a user, for example next to the inspection image, in order to assist an operator of the inspection system to detect one or more objects of interest in the inspection image, based on the dimming score of the one or more detected objects. A non-limiting example of calculating and using the dimming score of the one or more detected objects, based on the calculated dimming vector is disclosed below. A function indicative of a dimming score may be defined. To calculate a dimming score, the dimming score function has two vectors as input: the calculated dimming vector (as predicted by the trained deep learning algorithm for a detected object), and a reference dimming vector (corresponding to a reference object of interest) of a plurality of reference dimming vectors (corresponding to a plurality of reference objects of interest). The controller computes two similarity measures using the dimming score function: a direction similarity; and a magnitude similarity. The direction similarity is calculated using a cosine similarity between the two input vectors. The direction similarity is indicative of an angle between the two input vectors. A cosine similarity of: 1 indicates that the vectors point in the same direction, -1 indicates that the vectors point in opposite directions, and 0 indicates that the vectors are orthogonal. The calculated cosine similarity is then scaled from [-1, 1] to, e.g., [0, 1], to obtain the direction similarity. The direction similarity may be scaled for example using the formula: direction similarity = (calculated cosine similarity + 1) divided by 2 The magnitude similarity is calculated as the ratio of the smaller vector magnitude to the larger vector magnitude. The magnitude similarity is indicative of how similar the respective lengths of the two vectors are, regardless of their direction. The controller uses the dimming score function to then compute an average dimming score of the calculated direction similarity and the calculated magnitude similarity. A direction weight parameter may determine a relative importance of direction similarity vs. magnitude similarity in the average dimming score. A direction weight parameter of 0.5 (e.g., a default value) is indicative of an equal importance of both the direction similarity and the magnitude similarity. For example a formula may be: dimming score = (direction similarity * direction weight parameter + magnitude similarity * (1 - direction weight parameter)) As an exampie, for each detected object, the controller: calculates the dimming vector of the detected object, using the trained deep learning algorithm; and compute the dimming score between the calculated dimming vector and each reference dimming vector in the database, as described above. If the highest of the plurality of calculated dimming scores across all of the reference vectors exceeds a predefined threshold (e.g., a predefined threshold of 0.8, but other thresholds may be envisaged), the controller classifies the detected object as an object of interest corresponding to the most similar reference dimming vector. For example, if the dimming vector of a detected object has a dimming score of 0.9 with the reference dimming vector for a rifle, then the detected object would be classified as a rifle (and e.g., flagged to an operator of the inspection system for further inspection). The use of both direction similarity and magnitude similarity in the dimming score calculation helps to robustly match detected objects, even if their orientation or size in the inspection image differs from the reference objects. The direction weight parameter allows the system to be tuned to prioritize shape (direction) or material properties (magnitude) as appropriate for the application. Therefore, the deep learning algorithm of the disclosure, after it has been trained, enables more accurate detection of objects of interest such as weapons or contraband products, thus improving safety. After the device 15 has been configured with the deep learning algorithm 1, the device 15 can thus use the deep learning algorithm 1 based on locally acquired inspection images to determine whether an object of interest is present in an inspection image generated using penetrating radiation, based on the dimming vector (and e.g., the dimming score) of the one or more detected objects. In the inspection images, the deep learning algorithm is configured to calculate the dimming vector (and e.g., the dimming score) of the one or more detected objects, the inspection images comprising one or more features at least similar to the training images used to generate the deep learning algorithm by the machine learning algorithm, as described in greater detail below. In general, the deep learning algorithm 1 is configured to calculate the coefficients of the dimming vector (and e g., the dimming score) of the one or more objects detected in the inspection images, in a way similar to the calculation of the coefficients of the dimming vector (and e.g., the dimming score) performed during the training described in greater detail below. In some examples, the disclosure may be applied for inspection of a real container containing the cargo 155, and at least some of the methods of the disclosure may comprise obtaining inspection images by irradiating, using penetrating radiation, one or more real containers configured to contain the cargo 155, and detecting radiation from the irradiated one or more real containers. Generation of the trained deep learning algorithm The disclosure also discloses an example method for generating a trained deep learning algorithm 1 configured to calculate a dimming vector indicative of features of an object detected in an inspection image of cargo. The trained deep learning algorithm 1 is obtained after training. The machine learning algorithm is trained using: a plurality of training images which are similar to inspection images and comprising one or more objects of interest, and one or more annotated reference inspection images 1000. Each annotated reference inspection image 1000 represents a reference object 11 of interest (as shown in Figure 3) among a plurality of objects of interest. As already stated, the annotations indicate a reference axis A-A and a reference orientation of the object of interest which is represented as projected in 2D in the reference inspection image 1000. After the deep learning algorithm has been trained, the controller is configured to detect one or more objects of interest, regardless of whether each object of interest is represented dimmed in the inspection image, e.g., because of its superposition with the walls of container and / or other objects in the cargo, the deformation of shapes due to the nature of the projection of the radiation on the detectors, or the low extent of the penetration of the cargo by the radiation. Figure 5 shows a flow chart illustrating an example method 200 according to the disclosure. The method 200 is for generating a trained deep learning algorithm configured to calculate a dimming vector indicative of features of an object detected in an inspection image of cargo, the inspection image being generated using penetrating radiation. In Figure 5, the method 200 comprises: obtaining, at S21, a plurality of training images of cargo comprising one or more objects of interest; and training, at S22, the deep learning algorithm using the obtained training inspection images. The trained deep learning algorithm 1 is generated based on the training images obtained at S21. The learning process is typically computationally intensive and may involve large volumes of training images (such as thousands or tens of thousands of images). As explained in more detail below, the training step S22 mainly involves inferring one or more coefficients of a dimming vector, based on the training images, and encoding the inferred features in the form of the trained deep learning algorithm 1. As shown in Figure 6, the training at S22 comprises: detecting, at S221, one or more objects in the obtained training inspection images; for each detected object, predicting, at S222, one or more coefficients of a dimming vector; determining, at S223, a difference between each of the predicted coefficients and a corresponding training dimmed coefficient, and minimising, at S224, the determined difference by minimising a loss function. The dimmed training coefficients of the dimming vector are calculated as explained below. First one or more observed coefficients are calculated. The one or more observed coefficients correspond to the one or more coefficient which are desired in the dimming vector to be calculated by the trained deep learning algorithm. The one or more coefficients are observed for each detected object on the obtained training inspection images. After the one or more observed coefficients are calculated, the calculated one or more observed coefficients are divided by corresponding one or more reference values. As already stated, each reference value corresponds to a value of a coefficient of a reference vector calculated for a reference object of interest as detected in a reference inspection image (e.g. the reference inspection image 1000 of Figure 3) by an inspection system when the reference object of interest is positioned along a reference axis and in a reference orientation and is exposed, without any obstruction, to penetrating radiation, perpendicularly to the reference axis and the reference orientation. After the division described above, the one or more dimmed coefficients are obtained. The obtained one or more dimmed coefficients are compared at S223 to the one or more coefficients predicted by the deep learning algorithm at S222. The training inspection images and the reference inspection image 1000 comprise pixels of different intensity or grey levels, and the one or more objects detected in the training inspection images and the reference object in the reference inspection image 100 also comprise pixels of different intensity or grey levels. The one or more observed coefficients, the one or more coefficients of the reference vector and the one or more predicted coefficients comprise at least one of: one or more statistical measures; and / or one or more descriptors of grey levels; and / or one or more values indicative of entropy; and / or one or more descriptors of texture; and / or one or more geometrical characteristics; and / or one or more values indicative of Hu moments; and / or one or more features of Fast Fourier Transforms, FFT. The one or more statistical measures may comprise at least one of: an average of image intensity; and / or a median indicative a central intensity value; and / or one or more quartiles, indicative of a distribution of intensity; and / or a standard variation, indicative of dispersion of intensity; and / or a skewness, indicative of asymmetrical distribution of intensity; and / or a kurtosis, indicative of flattening of distribution of intensity; and / or a range, indicative of internal contrast within intensity; and / or a variation coefficient, indicative of normalised value of dispersion. The one or more descriptors of grey levels are indicative of a co-occurrence matrix of the grey levels and comprise at least one of: a value indicative of energy, such as an angular second moment, ASM; and / or a value indicative of contrast; and / or a value indicative of homogeneity; and / or a value indicative of correlation; and / or a value indicative of dissimilarity. The one or more values which are configured to be indicative of entropy may be indicative of texture complexity. The one or more descriptors of texture may comprise at least one of: one or more values indicative of local binary patterns, LBP; and / or one or more values indicative of Gabor energy. The one or more geometrical characteristics may comprise at least one of: one or more values indicative of an area; and / or one or more values indicative of a perimeter; and / or one or more values indicative a mass centre; and / or one or more values indicative of an aspect ratio; and / or one or more values indicative of a solidity; and / or one or more values indicative of circularity. The values indicative of Hu moments comprise seven constant values indicative of the shape of the one or more objects. The one or more features of the FFT may comprise at least one of: a value indicative of an average of the FFT; and / or a value indicative a standard deviation of the FFT. The training images are annotated. In other words, in the training images, the values of the one or more observed coefficients are known. The values of the one or more coefficients of the reference vector are also known. In some examples, a domain specialist (such as a human operator) may manually annotate the training images with ground truth annotation. Inputting ground truth and annotations by an operator may use a man / machine interface, such as comprising a display, and input means such as a keyboard and / or a mouse and / or a tactile function of the display. As already stated in S224 training the deep learning algorithm comprises training the algorithm to minimise a loss function. As a non-limiting example, the loss function may comprise a mean square error. Other loss functions indicative of a distance between a predicted value and a training ground truth value (such as the value of a training dimmed coefficient) are envisaged. The deep learning algorithm can also be trained to calculate the dimming score as described above. As already stated, during training, the type of each reference object of interest is known. During training, the deep learning algorithm may thus also be taught to predict a dimming score of the one or more detected objects, based on the predicted dimming vector. During training, the deep learning algorithm learns to calculate a "dimming score" that quantifies the similarity between two vectors: the dimming vector predicted by the algorithm for a detected object in a training image; a reference vector corresponding to a known object of interest, selected from a database of reference vectors for a range of objects of interest. This score is calculated using a dimming score function that takes these two vectors as inputs. By learning to maximize the dimming score for matching objects and minimize it for non-matches, the algorithm learns to accurately recognize objects of interest. The deep learning algorithm learns to compute two similarity measures using the dimming score function: a direction similarity; and a magnitude similarity. The direction similarity is calculated using a cosine similarity between the two input vectors as already described, to calculate the direction similarity. As also already stated, the magnitude similarity is calculated as the ratio of the smaller vector magnitude to the larger vector magnitude. The deep learning algorithm learns to use the dimming score function to then compute an average dimming score of the calculated direction similarity and the calculated magnitude similarity, using a direction weight parameter. As, during training, the type of each detected object and the type of each reference object of interest are known, the deep learning algorithm learn to compute the dimming score between the predicted dimming vector and each reference dimming vector in the database. The deep learning algorithm learns how to calculate a high dimming score for a detected object which corresponds to a type of reference object of interest (e.g., a dimming score greater than a predefined threshold of 0.8, but other thresholds may be envisaged). The training of the deep learning algorithm for the calculation of the dimming score may also use a loss function, the deep learning algorithm learning to minimise a difference between a predicted dimming score with a ground truth dimming vector, by minimising the loss function (such as a mean square error but other loss functions are envisaged). Any suitable deep learning algorithm may be used. For example, approaches based on a convolutional neural network may be used. As already stated, after it has been trained, the deep learning algorithm is used for determining whether an object of interest is present in an inspection image generated using penetrating radiation. The learned detection function may be used to detect objects of interest (e.g., weapons) that human operators (e.g. operators in customs organisations) may find difficult to detect in an inspection image, in particular when the objects of interest are at least partly dimmed (e.g., by a superposition with walls of a container of the cargo and / or other objects in the cargo, and / or by a deformation of shapes due to the nature of the projection of the radiation on the detectors; and / or by a low extent of a penetration of the cargo and / or object by the radiation). Computer system and detection device Figure 1 shows a device 15 configurable by the method 200 to generate the deep learning algorithm configured to calculate a dimming vector indicative of features of an object detected in an inspection image of cargo. Figure 1 also schematically illustrates an example computer system 110. Both the computer system 110 and the device 15 are configured to implement, at least partly, the example method 200 of Figure 5. In particular, in a preferred embodiment, the computer system 110 executes the machine learning to generate the deep learning algorithm 1 to be stored on the device 15. Although a single device 15 is shown for clarity, the computer system 110 may communicate and interact with multiple such devices. The training images may themselves be obtained using images acquired using the device 15 and / or using other, similar devices and / or using other sensors and data sources. In some examples, the training images may have been obtained in a different environment, e.g. using a similar device (or equivalent set of sensors) installed in a different (but preferably similar) environment, or in a controlled test configuration in a laboratory environment. In some examples, one or more training inspection images may be obtained by altering the reference inspection image 1000 of the reference object of interest or by altering one or more inspection images of objects of interest. Data augmentation techniques may be applied to generate additional training images that capture a range of expected visual variations. For example, Gaussian noise can be added to simulate sensor noise. Alternatively or additionally, grey or black image regions can be added to simulate the dimming effect of cargo container walls or contents obscuring an object of interest. By training on images with simulated dimming, the system learns to be robust to these expected obstructions. The computer system 110 of Figure 1 conventionally comprises a memory 121, a processor 122 and a communications interface 123. The system 110 may be configured to communicate with one or more devices 15, via the interface 123 and a link 30 (e.g. WiFi connectivity, but other types of connectivity may be envisaged). The memory 121 is configured to store, at least partly, data, for example for use by the processor 122. In some examples the data stored on the memory 121 may comprise data such as the training images (and the data used to generate the training images) and / or the machine learning algorithm. The detection device 15 of Figure 1 also comprises a communications interface 153 (e.g. Wi-Fi connectivity, but other types of connectivity may be envisaged) allowing connection to the interface 123 via the link 30. It should be understood that in order to determine whether an object of interest is present in the inspection image, the device 15 may be connected, at least temporarily, to the system 110, and the device 15 may access the memory 121 of the system 110. In a non-limiting example, the device 15 may also comprise an apparatus acting as the inspection system 154, as described in greater detail later. The inspection system 154 may be integrated into the device 15 or connected to other parts of the device 15 by wired or wireless connection. In other words, the inspection system 154 may be used to acquire one or more of the plurality of training images. In some examples, the processor 122 of the system 110 may comprise greater computational power and memory resources than the processor 152 of the device 15. The deep learning algorithm 1 generation is therefore performed, at least partly, remotely from the device 15, at the computer system 110. However, if sufficient processing power is available locally then the deep learning algorithm 1 learning could be performed (at least partly) by the processor 152 of the device 15. Device manufacture As illustrated in Figure 7, a method 300 of producing the device 15 configured to determine whether an object of interest is present in an inspection image generated using penetrating radiation, may comprise: obtaining, at S31, a deep learning algorithm 1 generated by the method 200 according to any aspects of the disclosure; and storing, at S32, the obtained deep learning algorithm 1 in the memory 151 of the device 15. The deep learning algorithm 1 may be created and stored using any suitable representation, for example as a data description comprising data elements specifying calculation conditions and their calculation outputs. Such a data description could be encoded e.g. using XML or using a bespoke binary representation. The data description is then interpreted by the processor 152 running on the device 15 when applying the deep learning algorithm 1. Alternatively, the machine learning algorithm may generate the deep learning algorithm 1 directly as executable code (e.g. machine code, virtual machine byte code or interpretable script). This may be in the form of a code routine that the device 15 can invoke to apply the deep learning algorithm 1. Regardless of the representation of the deep learning algorithm 1, the deep learning algorithm 1 effectively defines a dimmed vector calculation algorithm (comprising a set of rules) based on input data (i.e., detected objects in the inspection image). The device 15 may be connected temporarily to the system 110 to transfer the generated deep learning algorithm (e.g. as a data file or executable code) or the transfer may occur using a storage medium (e.g. memory card). In a preferred approach, the deep learning algorithm 1 is transferred to the device 15 from the system 110 over the network connection 30 (this could include transmission over the Internet from a central location of the system 110 to a local network where the device 15 is located). The deep learning algorithm 1 is then installed at the device 15. The deep learning algorithm 1 could be installed as part of a firmware update of device software, or independently. Installation of the deep learning algorithm 1 may be performed once (e.g. at time of manufacture or installation) or repeatedly (e.g. as a regular update). The latter approach can allow the detection performance of the deep learning algorithm to be improved over time, as new training images become available. Further details and examples The disclosure may be advantageous but is not limited to customs and / or security applications. The disclosure typically applies to cargo inspection systems (e.g., land, sea or air cargo). The inspection system 154 of Figure 1, acting as an inspection system, may be configured to inspect the container of the cargo 155, e.g., by transmission of inspection radiation through the container of the cargo 155. The container of the cargo 155 configured to contain the cargo may be, as a non-limiting example, placed on a vehicle. In some examples, the vehicle may comprise a trailer configured to carry the container of the cargo 155. The inspection system 154 of Figure 1 may comprises a source configured to generate the inspection radiation. The radiation source is configured to cause the inspection of the cargo through the material (usually steel) of walls of the container of the cargo 155, e.g. for detection and / or identification of the cargo. Alternatively or additionally, a part of the inspection radiation may be transmitted through the container of the cargo 155 (the material of the container of the cargo 155 being thus transparent to the radiation), while another part of the radiation may, at least partly, be reflected by the container of the cargo 155 (called “back scatter”). In some examples, the inspection system 154 may be mobile and may be transported from a location to another location (the inspection system 154 may comprise an automotive vehicle). In the source, electrons are generally accelerated under a voltage comprised between 100keV and 15MeV. In mobile inspection systems, the power of the X-ray source may be e.g., between 100keV and 9.0MeV, typically e.g., 300keV, 2MeV, 3.5MeV, 4MeV, or6MeV, for a steel penetration capacity e.g., between 40mm to 400mm, typically e.g., 300mm (12in). In static inspection systems, the power of the X-ray source may be e.g., between 1MeV and 10MeV, typically e.g., 9MeV, for a steel penetration capacity e.g., between 300mm to 450mm, typically e.g., 410mm (16.1 in). In some examples, the source may emit successive X-ray pulses. The pulses may be emitted at a given frequency, comprised between 50 Hz and 1000 Hz, for example approximately 200 Hz. According to some examples, the detectors may be mounted on a gantry. The gantry for example forms an inverted “L”. In mobile inspection systems, the gantry may comprise an electro-hydraulic boom (not shown on the Figures) which can operate in a retracted position in a transport mode and in an inspection position. The boom may be operated by hydraulic actuators (such as hydraulic cylinders). In static inspection systems, the gantry may comprise a static structure. It should be understood that the inspection radiation source may comprise sources of other penetrating radiation, such as, as non-limiting examples, sources of ionizing radiation, for example gamma rays or neutrons. The inspection radiation source may also comprise sources which are not adapted to be activated by a power supply, such as radioactive sources, such as using Co60 or Cs137. In some examples, the inspection system comprises detectors, such as X-ray detectors, optional gamma and / or neutrons detectors, e.g., adapted to detect the presence of radioactive gamma and / or neutrons emitting materials within the cargo, e.g., simultaneously to the X-ray inspection. In some examples, detectors may be placed to receive the radiation reflected by the container. In the context of the present disclosure, the container may be any type of container, such as a holder or a box, etc. The container may thus be, as non-limiting examples a palette (for example a palette of European standard, of US standard or of any other standard) and / or a train wagon and / or a tank and / or a boot of the vehicle and / or a “shipping container” (such as a tank or an ISO container or a non-ISO container or a Unit Load Device (ULD) container). In some examples, one or more memory elements (e.g., the memory of one of the processors) can store data used for the operations described herein. This includes the memory element being able to store software, logic, code, or processor instructions that are executed to carry out the activities described in the disclosure. A processor can execute any type of instructions associated with the data to achieve the operations detailed herein in the disclosure. In one example, the processor could transform an element or an article (e.g., data) from one state or thing to another state or thing. In another example, the activities outlined herein may be implemented with fixed logic or programmable logic (e.g., software / computer instructions executed by a processor) and the elements identified herein could be some type of a programmable processor, programmable digital logic (e.g., a field programmable gate array (FPGA), an erasable programmable read only memory (EPROM), an electrically erasable programmable read only memory (EEPROM)), an ASIC that includes digital logic, software, code, electronic instructions, flash memory, optical disks, CD-ROMs, DVD ROMs, magnetic or optical cards, other types of machine-readable mediums suitable for storing electronic instructions, or any suitable combination thereof. As one possibility, there is provided a computer program, computer program product, or computer readable medium, comprising computer program instructions to cause a programmable computer to carry out any one or more of the methods described herein. In example implementations, at least some portions of the activities related to the processors may be implemented in software. It is appreciated that software components of the present disclosure may, if desired, be implemented in ROM (read only memory) form. The software components may, generally, be implemented in hardware, if desired, using conventional techniques. Other variations and modifications of the system will be apparent to the skilled in the art in the context of the present disclosure, and various features described above may have advantages with or without other features described above. The above embodiments are to be understood as illustrative examples, and further embodiments are envisaged. It is to be understood that any feature described in relation to any one embodiment may be used alone, or in combination with other features described, and may also be used in combination with one or more features of any other of the embodiments, or any combination of any other of the embodiments. Furthermore, equivalents and modifications not described above may also be employed without departing from the scope of the invention, which is defined in the accompanying claims.

Claims

1. A device configured to detect an object of interest in an inspection image of cargo, the inspection image being generated using penetrating radiation, the device comprising:a memory storing a trained deep learning algorithm configured to calculate a dimming vector indicative of features of objects in an inspection image of cargo;a controller coupled to the memory and configured to:obtain the inspection image of the cargo;detect one or more objects in the obtained inspection image;use the trained deep learning algorithm to calcuiate a dimming vector for each of the one or more detected objects,wherein the dimming vector comprises one or more coefficients indicative of features of each of the one or more detected objects;wherein each coefficient is calculated as a value dimmed with respect to a corresponding reference value learned by the deep learning algorithm during training, andwherein the reference value corresponds to a value of a coefficient of a reference vector calculated for a reference object of interest as detected in a reference inspection image by an inspection system when the reference object of interest is positioned along a reference axis and in a reference orientation and is exposed, without any obstruction, to penetrating radiation, perpendicularly to the reference axis and the reference orientation; and determine whether the one or more detected objects correspond to an object of interest, based on the calculated dimming vector.

2. The device of claim 1, wherein the trained deep learning algorithm is configured to calculate the one or more coefficients comprising at least one of:one or more statistical measures; and / orone or more descriptors of grey levels; and / orone or more values indicative of entropy; and / orone or more descriptors of texture; and / orone or more geometrical characteristics; and / orone or more values indicative of Hu moments; and / orone or more features of Fast Fourier Transforms, FFT.

3. The device of claim 2, wherein the one or more statistical measures comprise at least one of:an average of image intensity; and / ora median indicative a central intensity value; and / orone or more quartiles, indicative of a distribution of intensity; and / ora standard variation, indicative of dispersion of intensity; and / ora skewness, indicative of asymmetrical distribution of intensity; and / ora kurtosis, indicative of flattening of distribution of intensity; and / ora range, indicative of internal contrast within intensity; and / or a variation coefficient, indicative of normalised value of dispersion.

4. The device of claim 2 or claim 3, wherein the one or more descriptors of grey levels are indicative of a co-occurrence matrix of the grey levels and comprise at least one of:a value indicative of energy, such as an angular second moment, ASM; and / ora value indicative of contrast; and / ora value indicative of homogeneity; and / ora value indicative of correlation; and / ora value indicative of dissimilarity5. The device of any claims 2 to 4, wherein the one or more values which are configured to be indicative of entropy are indicative of texture complexity.

6. The device of any claims 2 to 5, wherein the one or more descriptors of texture comprise at least one of:one or more values indicative of local binary patterns, LBP; and / or one or more values indicative of Gabor energy.

7. The device of any claims 2 to 6, wherein the one or more geometrical characteristics comprise at least one of:one or more values indicative of an area; and / orone or more values indicative of a perimeter; and / orone or more values indicative a mass centre; and / orone or more values indicative of an aspect ratio; and / orone or more values indicative of a solidity; and / or one or more values indicative of circularity.

8. The device of any claims 2 to 7, wherein the one or more values indicative of Hu moments comprise seven constant values indicative of the shape of the one or more objects.

9. The device of any claims 2 to 8, wherein the one or more features of the FFT comprise at least one of:a value indicative of an average of the FFT; and / or a value indicative a standard deviation of the FFT.

10. The device of any of the preceding claims, wherein the controller is configured to build a mask over the one or more detected objects.

11. The device of any of the preceding claims, wherein the controller Is further configured to determine whether the one or more detected objects correspond to an object of interest by comparing the one or more calculated dimming vectors to one or more reference vectors of objects of interest.

12. The device of any of claims 1 to 11, wherein the controller is further configured to calculate a dimming score of the one or more detected objects, based on the calculated dimming vector.

13. The device of claim 12, further comprising a user interface comprising a display and wherein the controller is further configured to cause the display to display the calculated dimming score.

14. A method for generating a trained deep learning algorithm configured to calculate a dimming vector indicative of features of an object detected in an inspection image of cargo, the inspection image being generated using penetrating radiation, the method comprising: obtaining a plurality of training inspection images; andtraining the deep learning algorithm using the obtained training inspection images, wherein the training comprises:detecting one or more objects in the obtained training inspection images;for each detected object, predicting one or more coefficients of a dimming vector;determining a difference between each of the predicted coefficients and a corresponding training dimmed coefficient,wherein each training dimmed coefficient is calculated by dividing an observed coefficient by a corresponding reference value, andwherein the reference value corresponds to a value of a coefficient of a reference vector calculated for a reference object of interest as detected in a reference inspection image by an inspection system when the reference object of interest is positioned along a reference axis and in a reference orientation and is exposed, without any obstruction, to penetrating radiation, perpendicularly to the reference axis and the reference orientation; and minimising the determined difference by minimising a loss function.

15. The method of the previous claim, wherein the one or more observed coefficients, the one or more coefficients of the reference vector and the one or more predicted coefficients comprise at least one of:one or more statistical measures; and / orone or more descriptors of grey levels; and / orone or more values indicative of entropy; and / orone or more descriptors of texture; and / orone or more geometrical characteristics; and / orone or more values indicative of Hu moments; and / orone or more features of Fast Fourier Transforms, FFT.

16. The method of the previous claim, wherein the one or more observed coefficients, the one or more coefficients of the reference vector and the one or more predicted coefficients correspond to the coefficients of any of claims 3 to 9.

17. The method of any of claims 14 to 16, wherein the deep learning algorithm is trained to calculate a dimming score of the one or more detected objects, based on the predicted dimming vector, orwherein the loss function comprises a mean square error.

18. The method of any of claims 14 to 17, wherein one or more training inspection images are obtained by altering the inspection image of the reference object of interest or by altering one or more inspection images of objects of Interest.

19. The method of the preceding claim, wherein altering comprises at least one of:adding a Gaussian noise, and / oradding one or more grey or black areas.

20. The method of any of claims 14 to 19, wherein the deep learning algorithm comprises a convolutional neural network.

21. A method of producing a device configured to determine whether an object of interest is present in an inspection image generated using penetrating radiation, the method comprising:obtaining a deep learning algorithm generated by the method according to any one of claims 14 to 20; andstoring the obtained deep learning algorithm in a memory of the device, optionally wherein the storing comprises transmitting the generated deep learning algorithm to the device via a network, the device receiving and storing the deep learning algorithm, optionally wherein the deep learning algorithm is generated, stored and / or transmitted in the form of one or more of: a data representation of the deep learning algorithm; executable code for applying the deep learning algorithm to one or more inspection images.

22. A method for determining whether an object of interest is present in an inspection image generated using penetrating radiation, the method comprising:obtaining an inspection image;applying, to the obtained image, a deep learning algorithm generated by the method according to any one of claims 14 to 20; anddetermining whether an object of interest is present in the inspection image, based on the applying.

23. The method of the preceding claim, wherein obtaining the inspection image comprises: irradiating, using penetrating radiation, one or more real containers configured to contain cargo; anddetecting radiation from the irradiated one or more real containers.

24. The device of any of claims 1 to 13 or the method according to any of claims 14 to 23, wherein the object of interest comprises at least one of:a threat object, such as a weapon; and / ora contraband product, such as a drug product and / or a cigarette product and / or an alcohol bottle.

25. A computer program or a computer program product comprising instructions which, when executed by a processor, enable the processor to perform the method according to any one of claims 14 to 23 or to control the device according to any of claims 1 to 13.

Images