Generator networks for generating images with predetermined counts of objects

A method for manufacturing a generator network that is configured to generate images with multiple objects, comprising the following steps:

• providing a generator network that is configured to generate images with at least one instance of at least one type of object, wherein the generator network includes at least one residual layer, and/or convolutional layer, and/or combination of a convolutional layer and a sampling layer, and/or fully connected layer;
• modifying the generator network by replacing the residual layer, and/or convolutional layer, and/or combination of convolutional layer and sampling layer, and/or fully connected layer, with a dense block that includes a sequence of multiple convolutional layers, wherein an outcome of at least one of the sequence of multiple convolutional layers is inputted into multiple succeeding convolutional layers in the sequence;
• and
• training the modified generator network, wherein the modified generator network is configured to map a combination of a noise sample and at least one target count of objects to a generated image by:
• providing a set of training images and, for each respective training image of the training images, at least one actual count of objects that the respective training image contains,
• providing a discriminator network that is configured to map an image to a combination of:
• a decision of whether the image is a training image or a generated image produced by the modified generator network, and at least one predicted count of objects in the image,
• drawing noise samples and target counts of objects,
• mapping, by the modified generator network, the noise samples and target counts of objects to generated images,
• randomly drawing images from a pool including the generated images and the training images,
• supplying the randomly drawn images to the discriminator network, the discriminator network mapping each respective drawn image of the randomly drawn images to a combination of:
• a decision whether the respective drawn image is a training image or a generated image, and at least one predicted count of objects in the respective drawn image;
• optimizing discriminator parameters that characterize a behavior of the discriminator network with a goal of improving an accuracy with which the discriminator network distinguishes between generated images and training images;
• optimizing generator parameters that characterize a behavior of the modified generator network with a goal of deteriorating the accuracy;
• and
• further optimizing both the generator parameters and the discriminator parameters with a goal of improving a match between the predicted count of objects on the one hand, and the actual count or the target count of objects on the other hand.

A method for training an image classifier, comprising the following steps:

• providing a trained generator network by:
• providing a set of training images and, for each respective training image of the training images, at least one actual count of objects that the respective training image contains;
• providing a generator network that is configured to map a combination of a noise sample and at least one target count of objects to a generated image;
• providing a discriminator network that is configured to map an image to a combination of:
• a decision of whether the image is a training image or a generated image produced by the generator network, and at least one predicted count of objects in the image;
• drawing noise samples and target counts of objects;
• mapping, by the generator network, the noise samples and target counts of objects to generated images;
• randomly drawing images from a pool including the generated images and the training images;
• supplying the randomly drawn images to the discriminator network, the discriminator network mapping each respective drawn image of the randomly drawn images to a combination of:
• a decision whether the respective drawn image is a training image or a generated image, and at least one predicted count of objects in the respective drawn image;
• optimizing discriminator parameters that characterize a behavior of the discriminator network with a goal of improving an accuracy with which the discriminator network distinguishes between generated images and training images;
• optimizing generator parameters that characterize a behavior of the generator network with a goal of deteriorating the accuracy;
• and
• further optimizing both the generator parameters and the discriminator parameters with a goal of improving a match between the predicted count of objects on the one hand, and the actual count or the target count of objects on the other hand;
• generating, using the trained generator network, classifier training images that contain predetermined numbers of instances of objects, wherein the objects are of predetermined types;
• obtaining, for each classifier training image of the classifier training images, using the predetermined types of objects and the predetermined numbers of instances, a label that is indicative of one or more classes to which the image classifier shall map the classifier training image;
• providing the classifier training images to the image classifier, thereby mapping the classifier training images to classes;
• rating a difference between the mapped classes and the labels associated with the training images with a predetermined classifier cost function;
• and
• optimizing classifier parameters that characterize a behavior of the image classifier with the goal of improving the rating by the classifier cost function.

A method for obtaining a spatial layout of a machine that is to comprise a predetermined number of instances of one or more predetermined components, comprising the following steps:

• training a generator network by:
• providing a set of training images and, for each respective training image of the training images, at least one actual count of objects that the respective training image contains,
• providing the generator network, the generator network being configured to map a combination of a noise sample and at least one target count of objects to a generated image,
• providing a discriminator network that is configured to map an image to a combination of:
• a decision of whether the image is a training image or a generated image produced by the generator network, and at least one predicted count of objects in the image;
• drawing noise samples and target counts of objects,
• mapping, by the generator network, the noise samples and target counts of objects to generated images,
• randomly drawing images from a pool including the generated images and the training images,
• supplying the randomly drawn images to the discriminator network, the discriminator network mapping each respective drawn image of the randomly drawn images to a combination of:
• a decision whether the respective drawn image is a training image or a generated image, and at least one predicted count of objects in the respective drawn image,
• optimizing discriminator parameters that characterize a behavior of the discriminator network with a goal of improving an accuracy with which the discriminator network distinguishes between generated images and training images,
• optimizing generator parameters that characterize a behavior of the generator network with a goal of deteriorating the accuracy, and
• further optimizing both the generator parameters and the discriminator parameters with a goal of improving a match between the predicted count of objects on the one hand, and the actual count or the target count of objects on the other hand;
• wherein the training images include images that show various counts of instances of the predetermined components, and wherein the at least one actual count associated with each of the training images is a count of instances of a predetermined component in the respective training image;
• drawing a noise sample;
• and
• supplying the drawn noise sample and the predetermined number of instances of the predetermined components to the generator network, thereby obtaining a generated image with the spatial layout.

A non-transitory computer-readable storage medium on which is stored a computer program for training a generator network that is configured to generate images with multiple objects, the computer program, when executed by one or more computer, causing the one or more computers to perform the following steps:

• providing a set of training images and, for each respective training image of the training images, at least one actual count of objects that the respective training image contains;
• providing a generator network that is configured to map a combination of a noise sample and at least one target count of objects to a generated image;
• providing a discriminator network that is configured to map an image to a combination of:
• a decision of whether the image is a training image or a generated image produced by the generator network, and at least one predicted count of objects in the image;
• drawing noise samples and target counts of objects;
• mapping, by the generator network, the noise samples and target counts of objects to generated images;
• randomly drawing images from a pool including the generated images and the training images;
• supplying the randomly drawn images to the discriminator network, the discriminator network mapping each respective drawn image of the randomly drawn images to a combination of:
• a decision whether the respective drawn image is a training image or a generated image, and
• at least one predicted count of objects in the respective drawn image;
• optimizing discriminator parameters that characterize a behavior of the discriminator network with a goal of improving an accuracy with which the discriminator network distinguishes between generated images and training images;
• optimizing generator parameters that characterize a behavior of the generator network with a goal of deteriorating the accuracy;
• and
• further optimizing both the generator parameters and the discriminator parameters with a goal of improving a match between the predicted count of objects on the one hand, and the actual count or the target count of objects on the other hand.

One or more computers configured to train a generator network that is configured to generate images with multiple objects, the one or more computers configured to:

• provide a set of training images and, for each respective training image of the training images, at least one actual count of objects that the respective training image contains;
• provide a generator network that is configured to map a combination of a noise sample and at least one target count of objects to a generated image;
• provide a discriminator network that is configured to map an image to a combination of:
• a decision of whether the image is a training image or a generated image produced by the generator network, and at least one predicted count of objects in the image;
• draw noise samples and target counts of objects;
• map, by the generator network, the noise samples and target counts of objects to generated images;
• randomly draw images from a pool including the generated images and the training images;
• supply the randomly drawn images to the discriminator network, the discriminator network mapping each respective drawn image of the randomly drawn images to a combination of:
• a decision whether the respective drawn image is a training image or a generated image, and at least one predicted count of objects in the respective drawn image;
• optimize discriminator parameters that characterize a behavior of the discriminator network with a goal of improving an accuracy with which the discriminator network distinguishes between generated images and training images;
• optimize generator parameters that characterize a behavior of the generator network with a goal of deteriorating the accuracy;
• and
• further optimize both the generator parameters and the discriminator parameters with a goal of improving a match between the predicted count of objects on the one hand, and the actual count or the target count of objects on the other hand.