Technologies are described for retrieving documents using image representations in the documents and is based on intra-image features. The identification of elements within an image representation can allow for deeper understanding of the image representation and for better relating image representations based on their intra-image features. The intra-image features present in image representations can be used in searches. Search results can further be reranked to improve search results. For example, reranking can allow search results to conform to intra-image dominant image features.

A machine learning system may automatically produce classifier algorithms and configuration parameters by selecting them into a set of predetermined unitary algorithms and associated parametrization values. Multiple digital representations of input object items may be produced by varying the position and orientation of the object to be classified and/or of the sensor to capture a digital representation of the object, and/or by varying a physical environment parameter which changes the digital representation capture of the object by the sensor. A robot arm or a conveyor may vary the object and/or the sensor positions and orientations. The machine learning system may employ genetic programming to facilitate the production of classifiers suitable for the classification of multiple digital representations of input object items. The machine learning system may automatically generate reference template signals as configuration parameters for the unitary algorithms to facilitate the production of classifiers suitable for the classification of multiple digital representations of input object items.

A panoptic labeling system includes a modified panoptic labeling neural network (“modified PLNN”) that is trained to generate labels for pixels in an input image. The panoptic labeling system generates modified training images by combining training images with mask instances from annotated images. The modified PLNN determines a set of labels representing categories of objects depicted in the modified training images. The modified PLNN also determines a subset of the labels representing categories of objects depicted in the input image. For each mask pixel in a modified training image, the modified PLNN calculates a probability indicating whether the mask pixel has the same label as an object pixel. The modified PLNN generates a mask label for each mask pixel, based on the probability. The panoptic labeling system provides the mask label to, for example, a digital graphics editing system that uses the labels to complete an infill operation.

The invention provides a model training method and system that uses pretrained features of a teacher neural network trained on a billion-size dataset to train a student neural network. The model training method leverages the teacher neural network to design a more stable loss function that works well with more sophisticated learning rate schedules to reduce training time and make the augmentation designing process more natural.

The invention provides a model training method and system that uses pretrained features of a teacher neural network trained on a billion-size dataset to train a student neural network. The model training method leverages the teacher neural network to design a more stable loss function that works well with more sophisticated learning rate schedules to reduce training time and make the augmentation designing process more natural.

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for objection classification. One of the methods includes: obtaining a first set of images of objects that have a likelihood of being at a property that satisfies a likelihood threshold; generating, for each object, a binary classifier from a set of images of the respective object; determining, using at least one of the binary classifiers, that an image of an unknown object was classified as an object from the objects; in response to determining, using the binary classifiers, that the image of the unknown object was classified as an object from the objects, selecting a second set of images of unknown objects that does not include the image; and generating a multiclass classifier for use classifying objects using i) the first set as respective classes and ii) the second set that does not include the image.

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for objection classification. One of the methods includes: obtaining a first set of images of objects that have a likelihood of being at a property that satisfies a likelihood threshold; generating, for each object, a binary classifier from a set of images of the respective object; determining, using at least one of the binary classifiers, that an image of an unknown object was classified as an object from the objects; in response to determining, using the binary classifiers, that the image of the unknown object was classified as an object from the objects, selecting a second set of images of unknown objects that does not include the image; and generating a multiclass classifier for use classifying objects using i) the first set as respective classes and ii) the second set that does not include the image.

A method for detecting from images correct placement and function, or incorrect placement and function, of a clip of a transportation box of wafers in sterile or similar conditions obtains an image template comprising features of clip and obtains a first detection image of a working clip. An object region focusing on the imaged clip in the first detection image is determined according to the image template. Part of the working image is selected as a first preset location. The part of the image is masked to obtain a second detection image, the masking obscures the background region of the part of the image but not the clip-object region, and displays the unobscured clip-object region. The second detection image is input into a trained neural network model to determine whether the clip is in sealed or unsealed state. An electronic device and a non-transitory storage medium are also disclosed.

A method for detecting from images correct placement and function, or incorrect placement and function, of a clip of a transportation box of wafers in sterile or similar conditions obtains an image template comprising features of clip and obtains a first detection image of a working clip. An object region focusing on the imaged clip in the first detection image is determined according to the image template. Part of the working image is selected as a first preset location. The part of the image is masked to obtain a second detection image, the masking obscures the background region of the part of the image but not the clip-object region, and displays the unobscured clip-object region. The second detection image is input into a trained neural network model to determine whether the clip is in sealed or unsealed state. An electronic device and a non-transitory storage medium are also disclosed.

Methods, systems, and computer programs are presented for adding new features to a network service. A method includes receiving an image depicting an object of interest. A category set is determined for the object of interest and an image signature is generated for the image. Using the category set and the image signature, the method identifies a set of publications within a publication database and assigns a rank to each publication. The method causes presentation of the ranked list of publications at a computing device from which the image was received.