EP4423727A1 - Personenverifikation in porträtgemälden - Google Patents
Personenverifikation in porträtgemäldenInfo
- Publication number
- EP4423727A1 EP4423727A1 EP22803311.4A EP22803311A EP4423727A1 EP 4423727 A1 EP4423727 A1 EP 4423727A1 EP 22803311 A EP22803311 A EP 22803311A EP 4423727 A1 EP4423727 A1 EP 4423727A1
- Authority
- EP
- European Patent Office
- Prior art keywords
- images
- image
- person
- portrait
- similarity
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/168—Feature extraction; Face representation
- G06V40/169—Holistic features and representations, i.e. based on the facial image taken as a whole
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
- G06V10/751—Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/761—Proximity, similarity or dissimilarity measures
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
- G06V10/7747—Organisation of the process, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/84—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using probabilistic graphical models from image or video features, e.g. Markov models or Bayesian networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/172—Classification, e.g. identification
Definitions
- the invention relates to a method and a system for person verification in portrait paintings and a computer program product.
- a portrait typically depicts a person, specifically a person's face or head.
- a portrait can be represented by a painting or a photograph.
- portraits in the form of busts, sculptures, masks, or mosaics are also known.
- a portrait painting can refer to a painting that shows the portrait of a person in the form of a painted or drawn painting.
- the portrait painting usually has a section that shows the face or the head of the person depicted.
- historical portrait paintings depict important people of an epoch and also allow conclusions to be drawn about the lives of historically important people.
- Numerous portrait paintings were created during the Renaissance, partly due to advances in painting techniques.
- the people depicted in historical portrait paintings were usually so relevant that several portrait paintings were made of one person.
- several portraits of a person were made because a person was depicted in portraits at different times in his life.
- a person can also have been depicted by different painters. It is also conceivable that a person has been depicted in several portrait paintings from different perspectives.
- an assumption is made as to the identity of the person depicted.
- Such an assumption can be based in particular on the statement of an art expert or art historian.
- the expert or historian for use a portrait painting for which the person depicted is verified and then compares this with the portrait painting in question to arrive at a statement as to the identity of the person depicted.
- the technical problem thus arises of creating a method and a system for person verification in portrait paintings and a computer program product which verifies the identity of a person depicted with high reliability and accuracy.
- a basic idea of the invention is to verify the identity of a person in a portrait painting to be checked.
- the invention provides for this, with the help of Images of several portraits of a reference person and images of several portraits of one or more contrarians, who are different from the reference person, to determine a technical criterion using a machine learning method, which serves to objectively distinguish the reference person from a contrarian.
- a method for verifying persons is thus made possible without the aforementioned disadvantages, it being verified whether the person depicted in a portrait painting to be checked corresponds to the reference person or not.
- An image designates an image section of a portrait painting, which preferably shows at least part of the face of the person depicted in the portrait painting. However, the image section can also include the entire portrait painting or also a region surrounding the portrait painting.
- the image can be encoded in the form of an image file, which is available, for example, as a RAW, JPEG or PNG file.
- an image can be generated by means of an image acquisition device by imaging a portrait painting or a section of the portrait painting by means of an image acquisition device.
- An image-specific feature or features can be determined as a vector.
- the features determined in the form of a vector cannot be interpreted intuitively or directly for a human user.
- the images of the reference image set differ at least in that at least one image of the reference image set maps to the first of the at least two portrait paintings and another image maps to the second of the at least two portrait paintings.
- the set of reference images can include more than two images, even if these images each only image one of two portrait paintings.
- a portrait painting can be represented by a number of images from, for example, different perspectives, and these images can then be part of the set of reference images.
- the reference person is a person, usually historically relevant, whose identity has been handed down, i.e. previously known.
- the identity of a person can be handed down or previously known, for example, in that a title of a portrait painting names the identity of the person depicted.
- All images of the set of reference images are images of portrait paintings that depict the same reference person. At least two portraits exist of the reference person, preferably four or more portraits. The more portrait paintings of the reference person the reference image set includes, the better the reliability of the proposed method.
- a contrarian is another person, usually also historically relevant, whose identity has been handed down or is known in advance.
- the set of contrarian images includes at least two different images of portrait paintings of the same contrarian, preferably of several different contrarians. Different contrarians are persons whose identities are different from each other. It is also important that the contrarian(s) is/are different from the reference person, i.e. that the reference and contrarian(s) are different persons or identities. There are at least two portraits of each opponent, but preferably three or more portraits, in particular 3 to 5 portraits. Different numbers of portrait paintings can also exist for different contrarians. For example, there may be 100 or more portraits for one opponent, with fewer portraits for another opponent. More preferably, the set of contrarian images includes images of at least two different contrarians.
- the image to be examined relates to a portrait painting depicting a person whose identity has not been handed down, i.e. is unknown. Such a person can also be a historically relevant person.
- the aim of the method described here is to verify whether this unknown person matches the reference person or not.
- the portrait paintings or images of the set of contrasting images are preferably selected in such a way that they are similar to the portrait paintings of the set of reference images with regard to the painting style or the painting technique used.
- the portrait paintings were made by the same painter or in the same or a similar era.
- the gender, age or other person-specific characteristics of the reference person can also be taken into account when choosing the portrait paintings or images of the set of contrarian images.
- a selection criterion when providing the reference and/or contra-image set such as only females or only males, or only portraits by a particular artist.
- a range of values can also be specified that affects a range of ages of the people depicted, such as only likenesses of people who were between 30 and 40 years old at the time a portrait painting was made. In this way, differences between the portrait paintings can be traced back as far as possible to the different people depicted and do not depend solely on different painting styles or, for example, the sex or age of the people depicted or similar.
- all portraits of the set of reference and contra-images are preferably selected in such a way that they also resemble the portrait of the image to be checked with regard to the aforementioned selection criteria. Assumptions regarding the identity of the person to be verified can also influence the selection of the portrait paintings of the reference or contra-image set.
- All images, or the two sets of images described and the image to be checked, can be made available by reading them in via an interface and/or by retrieving them from a database.
- the images can be stored in a memory device, for example in a memory device designed as a ROM or RAM or a memory device comprising a ROM or RAM. This is described in more detail below. It is also conceivable that the images are provided in different ways, i.e. that, for example, an image can be provided by reading it in via an interface and another image can be provided by retrieving it from a database.
- An image-specific feature can represent one or more properties of an image that are objectively and technically analyzable.
- a property may be a geometric property relating to a human face, such as the spacing, shape, or arrangement of a person's eyes, nose, and/or mouth.
- it can also be a property that relates to an artistic aspect with regard to a painting style or painting technique used.
- the thickness of a brush stroke in the image can affect the determination of an image-specific feature.
- the thickness of such a brush stroke or, for example, the distance between a person's eyes can, in particular in an image, over a number of pixels or the distance of pixels from one another or the like can be quantified.
- an image-specific feature can also be a property that is not directly recognizable in the image.
- the machine learning method can be trained to take the described properties into account when determining an image-specific feature. This, in particular the training of a neural network, is explained in more detail below.
- All images of the reference and contra-image set are analyzed using the machine learning method.
- the image to be checked is also analyzed using the machine learning method.
- For each image, at least one image-specific feature is automatically determined using the machine learning method and assigned to the respective image.
- the machine learning method is preferably trained in such a way that the image-specific features are similar for the images that depict the same person. Further, the image-specific features are dissimilar for the images depicting different people. This is due in particular to the fact that the images that depict the same person are based on similar, in particular person-specific, properties.
- a similarity of features preferably depends on a distance between features or can be represented by this, for example if the specific features were determined as vectors.
- the vectors for images that depict the same person can be arranged spatially close to one another, i.e. they can be similar, while on the other hand the vectors for images that depict different people can be arranged spatially far apart in the vector space.
- a specific similarity between two features is preferably referred to as a similarity value.
- Such a similarity value can be obtained, for example, from the reciprocal of the Euclidean
- the similarity value is determined via a cosine similarity of two vectors or features.
- the similarity value is determined via the cosine of the angle between two vectors, with the smaller or acute angle of two congruent intersection angles being used as the angle for determining the cosine.
- the cosine of a zero angle is one, while the cosine of a pi/2 angle is zero. For example, a small angle results in a large cosine value and thus a larger similarity value, while a larger angle results in a smaller cosine value and thus a smaller similarity value.
- the similarity value can correspond to the cosine.
- the determined similarity values are used to determine the first and further distribution.
- the first distribution is determined in particular from those similarity values that result from the difference between features of images that depict the same person, i.e., for example, from the difference between a feature of an image of the reference image set and a feature of another image of the reference image set. Furthermore, similarity values can also be used, which result from the difference between features of images of the contrarian image set that depict the same contrarian. In other words, the first distribution can also be referred to as a first set of values, which includes the explained similarity values.
- the further distribution is then determined in particular from those similarity values that result from the difference between features of images that depict different people, ie for example from the difference between a feature of an image of the set of reference images and an image of the set of contrarian images. It is also possible that similarity values are used, which result from the difference between features of images of different contrarians.
- the second distribution can also be referred to as a second set of values, which includes the similarity values explained.
- the first distribution of similarities is also known as the match distribution or genuine distribution and the further distribution of similarities as non-match distribution or imposter distribution.
- the first distribution can be the value set (4, 5, 5, 6).
- the variable for evaluating a criterion for distinguishing the reference person from the contrarian(s) is determined as a function of the first and/or the further distribution.
- the determined variable can be used in particular as a threshold value when evaluating the criterion.
- the size is preferably determined from a value range of the similarity values of the first and/or the further distribution.
- the range of values can relate to an intersection or an overlapping range of the first and further distribution, with the size being determined, for example, as the mean value, minimum value or maximum value of this intersection/this range.
- the determined quantity could be determined as the similarity value of 4.
- the size can also be determined as the particularly lowest similarity value of the first distribution.
- the size is determined as the mean of the first distribution. In analogy to the example mentioned above, the size could thus be determined as the similarity value of 5. What is essential is that the determined size makes it possible to evaluate a criterion, the result of the evaluation making it possible to distinguish between the person depicted in an image to be checked and the reference person or contrarian(s).
- the size is determined such that a predetermined
- False acceptance rate eg a false acceptance rate of 0.001
- False acceptance rate is not exceeded, i.e. no more than a certain percentage of actually not with the corresponding reference person is verified as identical to the person to be verified according to the method as identical to the reference person.
- the size for example, those image pairs of the contrarian image set that do not depict the same person can be used, with a similarity value being determined for each image pair and the size then being selected in such a way that, using the size, only a maximum of such a percentage of those that actually do not match pairs of images that is less than or equal to the predetermined false acceptance rate.
- the size can also be determined in such a way that a provisional size is selected as the starting value for an iteration, such as a value from the intersection of the first and further distribution, and the size is determined by iteration in such a way that the predetermined false acceptance rate is not exceeded.
- the criterion can be evaluated by comparing the determined size with the measure of similarity.
- the degree of similarity results from a comparison of the features of the image to be checked with each image in the set of reference images. For example, a similarity value can be determined for each pair of images, comprising the image to be checked and an image of the reference image set, with the degree of similarity then being determined, for example, as a minimum value, mean value or maximum value of the large number of similarity values.
- the specific degree of similarity indicates how similar the person to be verified and the reference person are.
- the similarity measure could be determined as an average similarity value of 5. But another way of determining the degree of similarity is also possible.
- the determined variable thus represents a threshold value of a threshold value criterion which, for example, must be exceeded for the verification by the similarity measure.
- the degree of similarity does not meet the criterion, for example because the degree of similarity was determined as an average similarity value of 3, then the result of the verification would be that the person to be verified does not match the reference person.
- the method described here enables person verification in portrait paintings, which is based on objective and technically determinable properties of portrait paintings and which, in particular, is not only traced back to the error-prone opinion of a human being.
- the described automated characteristic assignment with regard to the images of a reference person and the images of one or more contrarians, as well as the resulting distributions, can advantageously achieve high reliability and high accuracy in the person verification, which can also be carried out quickly and inexpensively can.
- the images are provided by reading them in via an interface or from a database and/or by capturing the portrait paintings using an image capturing device.
- Reading in refers to the transmission of the images described above from one electronic device to another electronic device, such as from a server to a system which is designed for person verification according to an embodiment described in this disclosure.
- a system can, for example, be or comprise a mobile phone or a computer.
- a server can be external to the system and form or include a database, with images, in particular image files, of various portrait paintings being stored in the database and the person depicted in at least some of the images being known in advance and the corresponding identity information being assigned to the image.
- information about a painter, an epoch, a painting style, etc. or the sex or age of the person depicted can also be assigned to an image.
- the images can then be provided by transferring the images and the images associated information from the database to a storage device of the system.
- the images can be stored on a memory device of the system or temporarily stored until the person verification process is completed.
- Reading in the images via an interface is also possible.
- the interface enables data transmission between two electronic devices, such as between a non-system storage device and the aforementioned system.
- reading in or transferring the images via the interface and/or from the database using different connection standards such as USB, Bluetooth, WLAN, Wi-Fi etc. can be made possible.
- the images or a part of the images can also be provided by capturing a portrait painting with an image capturing device.
- an image capturing device can in particular be part of the system for personal verification, such as a mobile phone camera.
- the image capturing device can have one or more image sensors, such as CCD or CMOS sensors.
- An image sensor captures one or more images of a portrait painting based on rays of light striking the image sensor from the portrait painting or part of the portrait painting.
- An image of a portrait painting generated in this way is preferably encoded in the form of a previously described image file.
- the image-specific features are determined as vectors in a vector space, with the first and/or further distribution and/or the degree of similarity being determined on the basis of specific differences between the vectors.
- the properties of an image explained at the outset are represented by entries in the vector by a feature determined as a vector. These entries can be numerical values.
- An entry can be a component or an element of a vector, with the number of components/elements of a vector specifying the dimension of the vector space.
- the vector space can thus be a multi-dimensional mathematical space, preferably with all the particular features being located in the same vector space.
- the differences between the vectors can be differences between vectors, for example. Such a difference can be the Euclidean distance between two vectors. How the first or further distribution and/or the degree of similarity can be determined from the differences between the vectors or from the absolute differences of the vectors has already been described above.
- a technical analysis of an image using the machine learning method has the advantage that a feature determined as a vector does not directly indicate a property or properties of an image, such as an interpupillary distance, but that the vector uses the machine learning method to can represent such or another property in an abstracted or abstract form.
- This type of representation of properties of an image in a vector results in an objective comparability of different images through a comparison of vectors or features.
- image-specific features are determined as vectors in a vector space, this advantageously results in a simple and objective comparison between the properties of different images.
- the objectivity of the comparison results in particular from the normalization of the vectors.
- a value from a value range of the first and/or further distribution is determined as the variable for evaluating the criterion.
- the criterion such as the initially mentioned threshold value criterion
- a value can therefore be a threshold value from which it can be assumed that the person depicted in the images compared with one another corresponds, such as the mean value of the similarity values in an overlapping region of the first and the further distribution.
- the degree of similarity to be determined for the person verification can then also be determined as a value from the value range of the similarity values of the first and/or further similarity distribution, so that the criterion can be evaluated quickly and easily.
- a first probability distribution is generated from the first distribution and a further probability distribution is generated from the further distribution.
- a similarity value or similarity values from predetermined similarity value ranges are determined several times, for example because the underlying geometric properties such as the interpupillary distance in the images of the reference person and thus the images are the same .
- a corresponding frequency of a similarity value can thus result when determining similarity values.
- a similarity value is determined multiple times for comparisons of features between images of different people, for example between the reference person and the contrarian. This can be the case in particular if there are identical differences between the features of the images of the reference and contrarian, such as differences in the determined interpupillary distance.
- similarity values of the first and further distribution can accumulate, in particular a histogram of the frequencies can be derived for the first and further distribution.
- the set of reference and/or contrarian images preferably includes images of a large number of different portrait paintings, for example images of between 10 and 100 different portrait paintings, so that a corresponding number of images also results in a correspondingly large number of features. It can be the case, in particular, that corresponding similarity values of the first and further distribution accumulate in such a way that an approximately normally distributed histogram for the first and/or the further distribution results. However, it is also possible that the histogram does not have a normal distribution or its properties.
- a probability distribution can be generated from a histogram.
- a particularly continuous probability distribution can be generated from the similarity values using what is known as a kernel density estimation method. These are usually not normally distributed.
- the first probability distribution associates a respective similarity value with a probability for the matched features to match. Matching of the matched features indicates that the images show the same person with respect to these matched features.
- the further probability distribution associates a specific similarity value with a resulting probability of the matched features not matching.
- a mismatch of the matched features indicates that the images show different people with respect to those matched features.
- the highest probability (frequency) that the matched characteristics match could be at a similarity value of 5, for example because a similarity value of 5 corresponds to the expected value of the first probability distribution. Accordingly, the highest probability (frequency) that the matched characteristics do not match could be at a similarity value of 3. Correspondingly lower probabilities of a match or non-match could then be assigned to the remaining similarity values of the first and further probability distributions.
- the similarity value of 5 can have a probability of, for example, 80% for a match are assigned. Due to an overlap of the first and further probability distribution, the similarity value of 5 can also be assigned a probability of 20% for non-matching, for example. Depending on the first and further probability distribution, a similarity value of 3 can, for example, be assigned a probability of 60% for a non-match and a probability of 40% for a match.
- the verification can advantageously be supplemented by stochastic analysis methods known to those skilled in the art. Furthermore, the first and further probability distribution also makes it possible to specify a probability for the correctness of the result of the verification. This is explained in more detail below.
- At least one plausibility value is determined when determining the measure of similarity, the plausibility value being determined as a ratio between the probabilities assigned to a (similarity) value in the first and further probability distribution.
- the plausibility value can then be the quotient of the probability of a match and the probability of a non-match of the features, with the probability of a match resulting from the first probability distribution and the probability of a non-match resulting from the further probability distribution.
- a plausibility value can thus be assigned to each similarity value.
- the two probabilities assigned to a similarity value are relativized by the plausibility value.
- a plausibility value can then be determined for each determined similarity value in order to determine the degree of similarity.
- the first plausibility value is, for example, 8/2. If the probabilities for a second similarity value of 3 have been assigned 40% for a match and 60% for a non-match, then this results the second plausibility value is 2/3.
- the reciprocal values of the quotients just described can also be used as plausibility values.
- all plausibility values can then be multiplied with one another or their mean value formed, which, as explained above, is determined for each similarity value that is determined when comparing the image to be checked and each image of the reference image set.
- This product or the mean value forms an overall plausibility value.
- the degree of similarity can then be determined as the similarity value to which the overall plausibility value is assigned, in particular as a function of the previously explained assignment of plausibility values to similarity values.
- the degree of similarity can thus be determined with increased accuracy and reliability, since the described relativizing of the probability values results in a weighting when determining similarity values and outliers have less of an influence on the degree of similarity.
- a reliability variable is determined, this variable indicating a probability of the correctness of the result.
- the reliability variable can indicate the probability of the person identified as the reference person actually not corresponding to the reference person and thus corresponding to a false acceptance rate (false match rate). If the person to be verified was identified as not corresponding to the reference person (i.e. the similarity measure does not meet the criterion), the reliability variable can indicate the probability of the person not identified as a reference person actually corresponding to the reference person and thus a false non-acceptance rate (false non match rate) match.
- the threshold-specific false acceptance rate and the threshold-specific false non-acceptance rate can be determined from the determined probability distributions for each threshold value (similarity threshold value), whereby this relationship can be represented, for example, in the form of an ROC curve (receiver-operating characteristic curve). This relationship can then be evaluated to determine the reliability variable. Creating a Receiver Operating The characteristic curve is known to those skilled in the art. As a function of this relationship, the variable for evaluating the criterion can also—as explained above—be determined, in particular as the threshold value at which the false acceptance rate is less than or equal to a predetermined threshold value, eg 0.001.
- the reliability variable and the result of the verification are output via an output device, in particular to a user of a system for personal verification. In this way, it can be conveyed to the user how correct the result determined with the proposed method is.
- the reliability variable can also be compared with a predetermined threshold value, with a reliability variable that is smaller than the threshold value indicating a particularly reliable correctness of the result.
- the verification result is advantageously quantified by a probability. In this way, the similarity between a person to be verified and the reference person can be assessed quickly and objectively.
- the machine learning method is trained as follows
- Pre-training a neural network with the first set of training images providing a further set of training images with images from portrait paintings, with each image being assigned at least one painting-specific basic truth,
- the method of machine learning trained in this way is designed to determine image-specific features for images of portrait paintings.
- An untrained machine learning method must learn to determine at least one image-specific feature for an image of a portrait painting with a quality that enables the person verification method described at the beginning to be carried out with sufficient accuracy.
- a neural network that has not yet been trained can be adapted by training with a first and a further set of training images.
- a neural network can comprise a number of levels, with a large number of neurons being able to be arranged in each level, with the arrangement and links between the various levels being adaptable by training.
- a neural network suitable for this purpose can be a residual neural network, for example.
- a first set of training images is provided with portrait photographs that form images of the first set of training images. If portrait photographs are available as digital image files, the images can be provided by them or in the form of these image files. Of course, it is also possible to generate the images of the first training image set by imaging portrait photographs, e.g. with an image acquisition device.
- This first set of training images preferably includes a sufficiently high number of images, such as 10,000 or more images, in particular up to 1,000,000 images.
- the images of the first set of training images are photographs of people or faces, in particular portrait photographs.
- the first training image set preferably includes images of both males and females. The images also preferably depict people with different phenotypes, ie in particular people with different visual appearances.
- At least two images of each person There are preferably at least two images of each person, with the at least two images preferably imaging the person from different perspectives.
- At least one photo-specific basic truth is assigned to each image.
- the basic truth corresponds to the identity of a person depicted in an image.
- the large number of images of photographs can be provided quickly and inexpensively, for example, by retrieving them from a database.
- a portrait photograph and in a portrait painting at least the full face of a person is preferably depicted.
- At least part of the images of the first training image set is generated by properties of images that are already part of the first Training image set are changed, for example by methods of image processing known to those skilled in the art. Such a change can take place, for example, by changing the orientation or mirroring or by changing the contrast strength of an image.
- at least part of the images or all images of the first training image set is/are provided by properties of portrait photographs being changed in such a way that these simulate images of portrait paintings.
- image processing methods known to those skilled in the art can be used to change portrait photographs or images of portrait photographs in such a way that these images look like portrait paintings, ie the images simulate portrait paintings. This can be referred to as data augmentation of the first training image set.
- the neural network is then pre-trained by analyzing the images of the first training image set using the neural network and adapting the neural network depending on the result of the analysis, i.e. in particular depending on at least one image-specific feature determined with the neural network.
- Such an adaptation of the neural network can take place in particular by an iterative approximation to an optimization criterion.
- Analysis of an image by the neural network can mean, for example, that the individual pixels, in particular their intensity values, of an image encoded as an image file are analyzed by the neural network.
- the neural network can detect edge, curve and/or color gradients in an image from the differences, in particular of intensities, between pixels.
- At least one image is an input variable and at least one feature determined by the analysis is an output variable of the neural network.
- a first subset of, for example, three images (e.g. image A, image B and image C) of the first set of training images is analyzed in a first run of the training, with two images A, B being assigned the identity of the same person as the basic truth , i.e. two of the images A, B depict the same person.
- the identity of another person is assigned to the remaining image C as basic truth, so that this image C thus depicts a different person than images A and B.
- the neural network has generated three image-specific features/feature vectors which, after the first pass of the training usually do not have a high quality with regard to possible conclusions about the depicted persons or the basic truths.
- An adaptation/training of the neural network can now be carried out in such a way that a feature-dependent evaluation variable is minimized or maximized or reduced to the extent that the evaluation variable corresponds to a predetermined threshold value or is smaller than this predetermined threshold value, or increased to the extent that the evaluation variable corresponds to a (further) predetermined threshold value or is greater than this (further) predetermined threshold value.
- an evaluation variable-dependent cost function can be minimized or maximized by the training, in particular up to a predetermined level.
- An evaluation variable can be a difference between two features/feature vectors determined by the analysis.
- a threshold value can be, for example, a predetermined difference between the two features/feature vectors for the images that depict the same person, with the evaluation variable being minimized by means of iterative adjustments, i.e. training, of the neural network or being minimized to the extent that it is smaller than the predetermined one difference is.
- Another threshold may be a predetermined further difference between the two features/feature vectors determined for the images depicting different people.
- the evaluation variable can be maximized or maximized to the extent that it is greater than the predetermined further difference by means of the iterative adaptations of the neural network. This procedure can be described as an approximation to an optimization criterion.
- the procedure described is also carried out for further subsets of the first training image set that are different from the first and from one another, so that the neural network is optimized for determining features from images, in particular of different people.
- a triplet loss function can be used for the training explained, in particular as a cost function. The person skilled in the art is familiar with the use of such a function.
- a subset of the first training image set cannot be used for the training, but the images of the subset can be retained for assessing the quality of the neural network.
- the retained images without the photo-specific basic truth are analyzed by the pre-trained neural network and the basic truths assigned to the images are used, to check whether the features determined in this way allow conclusions to be drawn about the people depicted in the images or whether certain features for images of the same person do not differ from one another by more than a predetermined amount and/or certain features for images of different people by more than a predetermined amount dimensions differ from each other.
- a difference between the characteristics of such pairs of images that do not depict the same person can be determined, for example.
- the identities of the people depicted are known from the assigned basic truths.
- the quality can be sufficient if it is determined with the help of the determined difference that the predetermined level is exceeded.
- a minimum value for the difference between two features of images that do not depict the same person can be defined as a predetermined measure.
- a further training image set is provided.
- Such a specialization of a pre-trained neural network is also referred to as transfer learning.
- the further set of training images includes images of different portrait paintings, preferably 300 or more images, with each image being assigned at least one painting-specific basic truth. Such a painting-specific basic truth can in particular concern or be the identity of a person depicted.
- the pre-training of the neural network makes it possible for the additional training image set to include a smaller number of images than the first training image set. This is particularly advantageous because images of portrait paintings are not usually available in as large a number as there are photographs.
- the first and the further training image set preferably include at least two images per identity, ie at least two images of the same person.
- the images or all images of the further training image set are provided by properties of portrait photographs, in particular portrait photographs of the first training image set or portrait photographs that are not part of the first training image set, being changed in such a way that these images are portrait paintings simulate.
- image processing methods known to those skilled in the art can be used to change portrait photographs or images of portrait photographs in such a way that the appearance of these images Portrait paintings are modeled on, so the images simulate portrait paintings.
- other properties of the images can also be changed, such as a contrast strength or an orientation of the images.
- This procedure can also be referred to as data augmentation.
- the number of images of the first and further training image set can be increased in a simple manner.
- the pre-trained neural network is then adapted/trained by means of the further training image set, preferably analogously to the described pre-training procedure, in such a way that features for images of portrait paintings of the same person do not deviate from one another by more than a predetermined amount and/or differ more than a predetermined amount for images of portraits of different people.
- the further training of the neural network can now be carried out in such a way that feature-dependent evaluation variables are minimized or maximized or reduced to the extent that the evaluation variable corresponds to a predetermined threshold value or is smaller than this predetermined threshold value, or increased to the extent that the evaluation variable corresponds to or is greater than this (further) predetermined threshold value.
- an evaluation variable-dependent cost function can be minimized or maximized by the training, in particular up to a predetermined level.
- This procedure can also be referred to as an approximation to an optimization criterion.
- a subset of the further training image set for example 30 images of the 300 images, are not used when specializing the pre-trained network and these images are then used analogously to the above-mentioned procedure for assessing the quality of the specialized neural network.
- the described training of the method of machine learning advantageously means that features suitable for reliable verification can also be determined for images from portrait paintings to which no basic truth is assigned.
- the trained method of machine learning uses in particular the specialized neural network.
- the machine learning method to be trained only with the first set of training images.
- a first training image set with portrait photographs can be provided, with each image being assigned at least one photo-specific basic truth.
- a neural network is trained with the first training image set, the machine learning method trained in this way being designed to determine image-specific features for images of portrait paintings.
- the first training image set can include portrait photographs and/or portrait paintings simulated in accordance with the preceding explanations and/or images of portrait paintings.
- the first training image set in this case comprises only simulated portraitures or only images of portraitures or only simulated portraitures and images of portraitures.
- At least one image-specific feature can be determined for an image of a portrait painting, as explained. Such an image-specific feature can then be used, as described above, for a method for person verification in portrait paintings.
- the training of the machine learning method described here ensures that person verification in portrait paintings is made possible with sufficient quality or accuracy and reliability.
- the trained neural network can comprise a plurality of levels, in particular levels connected in series, with intermediate results in the analysis of an image becoming more abstract as the level progresses.
- the result of the analysis or the output variable is then at least one image-specific feature. Due to the abstraction described during the analysis of an image, the features determined in this way can also relate to properties of an imaged face that remain hidden from the subjective human view. This advantageously results in more differentiated features compared to an analysis of an image by a human that increase reliability and accuracy in person verification in portrait paintings.
- a system for personal verification comprises an image acquisition device and/or an interface, a storage device, a data processing device and an output device, the system being designed or configured to carry out a method according to one of the previously disclosed embodiments.
- the system is embodied as a portable terminal, e.g., a mobile phone, or is provided by the terminal or the elements of the terminal.
- the exemplary embodiments for a corresponding image acquisition device, interface or storage device have already been explained above.
- the data processing device can be embodied as an integrated circuit on a microchip or can comprise one.
- the data processing device can also be a microcontroller or include one.
- the data processing device is designed to analyze images of portrait paintings or image files using the machine learning method and to carry out the claimed person verification method according to an embodiment described in this disclosure.
- An output device can be designed as a display, for example, and can be used to output a result variable or a reliability variable of the personal verification to a user.
- a computer program product with a computer program comprising software means for executing one, several or all steps of the method for person verification in portrait paintings according to one of the embodiments described in this disclosure, when the computer program is executed by or in a computer or an automation system .
- a program which, when running on a computer or in an automation system, causes the computer or the automation system to carry out one or more or all steps of the method for person verification in portrait paintings according to one of those described in this disclosure to carry out embodiments, and/or a program storage medium on which the program is stored (in particular in a non-transitory form), and/or to a computer, which comprises the program storage medium, and/or a (physical, e.g. electrical, e.g. technically produced ) signal wave, e.g. a digital signal wave, carrying information representing the program, e.g. the aforesaid program, comprising e.g. code means suitable for carrying out any or all of the method steps described herein.
- a program storage medium on which the program is stored in particular in a non-transitory form
- a computer which comprises the program storage medium, and/or a (physical, e.g. electrical, e.g. technically produced ) signal wave, e.g. a digital signal wave
- the method according to the invention is, for example, a computer-implemented method.
- all of the steps or only some of the steps (i.e. fewer than the total number of steps) of the method of the invention can be performed by a computer.
- One embodiment of the computer-implemented method is using the computer to perform a data processing method.
- the computer comprises, for example, at least one microcontroller or processor and, for example, at least one memory in order to (technically) process the data, for example electronically and/or optically.
- the processor consists, for example, of a substance or composition which is a semiconductor, for example at least partially n- and/or p-doped semiconductors, for example at least one II, III, IV, V, VI semiconductor material, for example (doped ) silicon and/or gallium arsenide.
- the steps described, in particular the determination steps and the verification step, are carried out by a computer, for example. Determination steps or calculation steps are, for example, steps for determining data as part of the technical process, for example as part of a program.
- a computer is, for example, any type of data processing device, e.g. electronic data processing device.
- a computer can be a device that is commonly regarded as such, e.g.
- a computer may consist of a system (network) of "sub-computers", each sub-computer representing a computer in its own right. Steps that are executed or carried out by a computer or an automation system can in particular be the determination steps and/or the verification step.
- the computer program product advantageously enables a method for person verification in portrait paintings to be carried out in accordance with one of the Disclosure-described embodiments for which technical advantages have been described above.
- FIG. 2 shows a flow chart of an embodiment of a method according to the invention for person verification in portrait paintings
- FIG. 4 shows an embodiment according to the invention of a system for person verification in a portrait painting
- the reference image set RM shows the reference image set RM with three different images of the reference person R1.
- the images can depict portrait paintings of the reference person R1, which were made by different painters and/or at different points in time in the life of the reference person R1.
- What is essential is that it is already known for all images of the reference image set RM that they image the reference person R1.
- the contrarian image set KM comprises a total of six images, two images depicting different portraits of a first contrarian K1, two images depicting different portraits of a second contrarian K2 and two images depicting different portraits of a third Show opponent K3.
- Each image of the contra-image set KM forms a portrait painting of a historically relevant person K1, K2, K3, whose identity is also previously known.
- the contrarians K1, K2, K3 are different from each other, so they refer to different people.
- all contrarians K1, K2, K3 are different from the reference person R1.
- FIG. 1 Also shown in FIG. 1 is an image PB to be checked of a portrait painting of a person P1 to be verified.
- the images of the reference image set RM were selected in such a way that the person verification method (see FIG. 2) can be used to verify whether the person P1 shown in the image PB to be checked corresponds to the reference person R1.
- FIG. 2 shows a schematic flow chart of a method according to the invention for person verification in portrait paintings.
- the method comprises six steps S1 to S6, which are explained below.
- step S1 the images shown in FIG. 1 are provided.
- the reference image set RM and the contra-image set KM can be made available by retrieving images encoded as image files from a database DB (see FIG. 4).
- the image PB to be checked can be provided by capturing a portrait painting P using an image capturing device BE (see FIG. 4). It is also conceivable that the image to be checked is also retrieved from the database DB.
- the images provided in this way can be temporarily stored in a storage device SE (see FIG. 4).
- step S2 the images of the reference image set RM are supplied as an input variable to a method of machine learning ML by means of a data processing device DV (see FIG. 4), the data processing device DV carrying out this method.
- the machine learning process ML analyzes the images using a specialized neural network SNN (see Figure 3), so that at least one image-specific feature is determined for each image and this is assigned to the respective image as the output variable of the machine learning process.
- a multiplicity of image-specific features MR thus result for the images of the reference image set RM.
- the same procedure is used for the images of the contra-image set KM and for the image PB to be checked, so that a large number of image-specific features MK result for the images of the contra-image set KM and at least for the image PB to be checked a feature MP.
- the features can be defined as vectors and arranged in a vector space.
- step S3 From the image-specific features MR, which are assigned to the images of the reference image set RM, and the features MK, which are assigned to the images of the contrasting image set, in step S3, by comparing the features of images of each image pair, whose images depict the same person, similarity values AW for Determination of a first distribution determined. Similarity values for determining a further distribution are determined by comparing the features of images of each pair of images whose images depict different people. A similarity value AW is determined in particular from the difference between two features or from the Euclidean distance between two vectors. In this case, decreasing difference represents increasing similarity.
- the probability distributions WV1, WV2 generated from the first and further distribution, e.g. with a kernel density estimator, are shown schematically in FIG.
- the first and further probability distributions WV1, WV2 indicate for each specific similarity value AW of the first or further distribution which probability W for a match or non-match is associated with this similarity value AW (see FIG. 5).
- a variable GK is determined for evaluating a criterion.
- the variable GK is determined as an average of the similarity values AW in the value range of an overlapping range UB of the first and further similarity distribution (see FIG. 5).
- the variable GK is therefore also a similarity value.
- variable GK can also be determined as a similarity value AW for which a predetermined false acceptance rate is not exceeded.
- step S5 a similarity measure AM is determined from a comparison between the features MR, which are assigned to the images of the reference image set RM, and the feature/features MP, which are assigned to the image PB to be checked.
- a similarity value AW is thus determined in each case by comparing the test image-specific feature(s) MP with the reference image-specific feature(s) of each reference image RB.
- the similarity measure AM can then be determined as an average of the similarity values AW determined in this way become.
- a plausibility value is determined for each specific similarity value AW and the similarity measure AM is then determined as the similarity value AW of the first or further distribution to which the product or the mean value of the determined plausibility values is assigned as a plausibility value.
- the degree of similarity AM is then compared with the variable GK to evaluate the criterion.
- the criterion for differentiation can then be used to verify the person P1, with the person P1 to be verified being verified as the same person as the reference person R1 if the determined degree of similarity AM exceeds a threshold value corresponding to the variable GK. If the degree of similarity is less than or equal to the size, then the person P1 to be verified is not verified as a person who matches the reference person R1.
- a result of the verification VE determined in this way can then be output via an output device AE (see FIG. 4).
- FIG. 3 shows a flowchart for training a machine learning method ML.
- a first training image set TM1 is provided in a training step T1.
- the first training data set TM1 comprises 10,000 to 1,000,000 portrait photographs, ie photographs of people or their faces.
- the identities of the people depicted are assigned to the respective images as photo-specific basic truth GW1, so the identity of the people depicted is known in advance.
- a neural network NN is pre-trained using the first training image set TM1.
- a first run two images of the same person and an image of another person are analyzed by the neural network NN.
- each image is an input variable and an image-specific feature set determined by the analysis is an output variable of the neural network NN, the feature set being able to include or encode precisely one or more than one feature.
- a first difference between the two feature sets of images that are intended to represent the same person is determined, and a further difference is determined for the feature sets of images that are intended to represent different persons.
- the neural network NN can then be adjusted in an optimization OK in such a way that the first Difference is minimized and/or the further difference is maximized.
- the neural network NN can be adjusted in the optimization OK by adjusting the arrangement of the neurons in a level of the neural network NN or by adjusting a link between the levels of the neural network NN.
- a triplet loss function can be used to perform the optimization OK.
- the remaining images of the first training image set TM1 can be used for further optimization OK of the neural network NN.
- a corresponding quality is achieved in the feature determination by the neural network NN, for example because features for images of portraits of the same person do not differ from one another by more than a predetermined amount and/or features for images of portraits of different people differ by more than a predetermined amount, so the pre-training of the neural network NN is complete.
- a further training image set TM2 is also provided in a training step T3.
- the further training image set TM2 includes, for example, 300 images of portrait paintings. At least one painting-specific basic truth GW2 is assigned to each image, which in particular assigns the identity of a person depicted to an image.
- the two sets of training images TM1, TM2 can be provided via a database, as already explained for the other sets of images.
- the pre-trained neural network VNN is then specialized in determining features in images of portrait paintings in a training step T4 using the further training image set TM2.
- the pre-trained neural network VNN is trained by the specialization in such a way that differences in images, which are due to a painting style, for example, are to be taken into account when determining image-specific features.
- the result of such training is a specialized neural network SNN, which can be used in a machine learning process ML.
- the method of machine learning ML is then designed to determine image-specific features MR, MK, MP for images of portrait paintings in such a way that features for images of portrait paintings of the same person do not deviate from one another by more than a predetermined amount and/or Features for portraiture images of different people differ more than a predetermined amount.
- a database DB can be stored on a server external to the system, which the system SY can access via an interface SN, for example using an Internet connection.
- a reference image set RM with images of a reference person R1 and a contrarian image set KM with images of several contrarians K1, K2, K3 are retrieved from the database and transmitted via the interface SN to a memory device SE of the system SY, e.g. designed as a RAM and thus provided.
- a portrait painting P is recorded by means of an image recording device BE.
- the portrait painting P shows a person to be verified P1.
- the image PB of the person P1 generated in this way is also temporarily stored in the storage device SE and is thus made available.
- the data processing device DV designed as a microcontroller is now used to carry out the method according to the invention as described in FIG.
- An output device AE designed as a display can, in particular, output the result VE of a personal verification with regard to the image PB or the portrait painting P to be checked. This can advantageously make it easier for a user to quickly and inexpensively verify a person P1 depicted in a portrait painting P against a reference person R1.
- the fifth shows a value range of the similarity values AW determined for the first and further distribution and the first and further probability distribution WV1, WV2 generated therefrom.
- the first probability distribution WV1 indicates the probability W with which a specific similarity value AW indicates a match.
- the further probability distribution WV2 indicates the probability W with which a specific similarity value AW indicates a non-match.
- the probability distributions WV1, WV2 are determined in particular in such a way that the probability is greater than zero for all similarity values AW.
- the similarity values AW are in this case on the abscissa of the coordinate system shown entered, while the probabilities W are entered on the ordinate as a function of the similarity values AW.
- the first and further probability distributions WV1, WV2 have a range UB in which the probability distributions overlap, a value range of similarity values AW of the first and the further distribution being assigned to the range UB.
- the mean of the value range of the similarity values AW in the overlapping area UB can be determined as variable GK for evaluating the criterion.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- Databases & Information Systems (AREA)
- Artificial Intelligence (AREA)
- Human Computer Interaction (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
Description
Claims
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| EP21205625.3A EP4174791A1 (de) | 2021-10-29 | 2021-10-29 | Verfahren und system zur personenverifikation in porträtgemälden und computerprogrammprodukt |
| PCT/EP2022/079450 WO2023072775A1 (de) | 2021-10-29 | 2022-10-21 | Personenverifikation in porträtgemälden |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| EP4423727A1 true EP4423727A1 (de) | 2024-09-04 |
Family
ID=78414543
Family Applications (2)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| EP21205625.3A Withdrawn EP4174791A1 (de) | 2021-10-29 | 2021-10-29 | Verfahren und system zur personenverifikation in porträtgemälden und computerprogrammprodukt |
| EP22803311.4A Pending EP4423727A1 (de) | 2021-10-29 | 2022-10-21 | Personenverifikation in porträtgemälden |
Family Applications Before (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| EP21205625.3A Withdrawn EP4174791A1 (de) | 2021-10-29 | 2021-10-29 | Verfahren und system zur personenverifikation in porträtgemälden und computerprogrammprodukt |
Country Status (3)
| Country | Link |
|---|---|
| US (1) | US20240273862A1 (de) |
| EP (2) | EP4174791A1 (de) |
| WO (1) | WO2023072775A1 (de) |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR100745981B1 (ko) * | 2006-01-13 | 2007-08-06 | 삼성전자주식회사 | 보상적 특징에 기반한 확장형 얼굴 인식 방법 및 장치 |
| US10839493B2 (en) * | 2019-01-11 | 2020-11-17 | Adobe Inc. | Transferring image style to content of a digital image |
-
2021
- 2021-10-29 EP EP21205625.3A patent/EP4174791A1/de not_active Withdrawn
-
2022
- 2022-10-21 WO PCT/EP2022/079450 patent/WO2023072775A1/de not_active Ceased
- 2022-10-21 EP EP22803311.4A patent/EP4423727A1/de active Pending
-
2024
- 2024-04-29 US US18/648,680 patent/US20240273862A1/en active Pending
Also Published As
| Publication number | Publication date |
|---|---|
| WO2023072775A1 (de) | 2023-05-04 |
| US20240273862A1 (en) | 2024-08-15 |
| EP4174791A1 (de) | 2023-05-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| DE69332265T2 (de) | Verfahren zur maskenerzeugung | |
| DE68928895T2 (de) | Verfahren und Gerät für universelle adaptiv lernende Bildmessung und -erkennung | |
| DE60208223T2 (de) | Anordnung und verfahren zur gesichtserkennung unter verwendung von teilen des gelernten modells | |
| DE69309300T2 (de) | Verfahren und vorrichtung zur artikelklassifizierung | |
| DE112019002848T5 (de) | System und verfahren zum auffinden und klassifizieren von mustern in einem bild mit einem bildverarbeitungssystem | |
| DE112019000093T5 (de) | Diskriminierungsvorrichtung und Maschinenlernverfahren | |
| DE69322095T2 (de) | Verfahren und gerät zur identifizierung eines objekts mittels eine geordneten folge von grenz-pixel-parametern | |
| DE60118606T2 (de) | Auswahlverfahren für wimperntusche, auswahlsystem für wimperntusche und gerät für die wimperntuscheberatung | |
| DE102018108780A1 (de) | Bearbeitungsflächenqualitätsbewertungsvorrichtung | |
| DE19521346A1 (de) | Bilduntersuchungs/Erkennungsverfahren, darin verwendetes Verfahren zur Erzeugung von Referenzdaten und Vorrichtungen dafür | |
| WO2000010119A1 (de) | Verfahren zum erkennen von objekten in digitalisierten abbildungen | |
| DE102017220307A1 (de) | Vorrichtung und Verfahren zum Erkennen von Verkehrszeichen | |
| DE102022208083A1 (de) | Trainieren eines neuronalen Netzwerks mit Hilfe von Wissensgraphen | |
| DE102018222797A1 (de) | Informationsverarbeitungsvorrichtung, Bildverarbeitungsverfahren und Aufzeichnungsmedium, auf dem das Bildverarbeitungsprogramm gespeichert ist | |
| WO2000079471A2 (de) | Verfahren und vorrichtung zur segmentierung einer punkteverteilung | |
| DE102018205561A1 (de) | Vorrichtung zur Klassifizierung von Signalen | |
| DE112020007472T5 (de) | Lernnutzungssystem, nutzungsvorrichtung, lernvorrichtung, programm und lernnutzungsverfahren | |
| DE69802372T4 (de) | Klassifizierungssystem und -verfahren mit N-Tuple- oder RAM-basiertem neuronalem Netzwerk | |
| WO2023072775A1 (de) | Personenverifikation in porträtgemälden | |
| DE102014108924B4 (de) | Ein halbüberwachtes Verfahren zum Trainiern eines Hilfsmodells zur Erkennung und Erfassung von Mehrfachmustern | |
| DE102023130646A1 (de) | Verfahren zur Dimensionsreduzierung | |
| DE19834718C2 (de) | Digitale Bildverarbeitung für ein Qualitätskontrollsystem | |
| DE202022107266U1 (de) | Ein Gesichtserkennungssystem zur Unterscheidung zwischen ähnlichen Gesichtern | |
| DE112022001991T5 (de) | Tumorzelen-isolinien | |
| EP1170678B1 (de) | Verfahren und Vorrichtung zur automatischen Suche relevanter Bilddatensätze |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: UNKNOWN |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE INTERNATIONAL PUBLICATION HAS BEEN MADE |
|
| PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: REQUEST FOR EXAMINATION WAS MADE |
|
| 17P | Request for examination filed |
Effective date: 20240523 |
|
| AK | Designated contracting states |
Kind code of ref document: A1 Designated state(s): AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC ME MK MT NL NO PL PT RO RS SE SI SK SM TR |
|
| DAV | Request for validation of the european patent (deleted) | ||
| DAX | Request for extension of the european patent (deleted) | ||
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: EXAMINATION IS IN PROGRESS |
|
| 17Q | First examination report despatched |
Effective date: 20250923 |