Abstract: Digital image selection techniques are described that employ machine learning to select a digital image of an object from a plurality of digital images of the object. The plurality of digital images each capture the object for inclusion as part of generating digital content, e.g., a webpage, a thumbnail to represent a digital video, and so on. In one example, digital image selection techniques are described that employ machine learning to select a digital image of an object from a plurality of digital images of the object. As a result, the service provider system may select a digital image of an object from a plurality of digital images of the object that has an increased likelihood of achieving a desired outcome and may address the multitude of different ways in which an object may be presented to a user.

Abstract: An image search system uses a multi-modal model to determine relevance of images to a spoken query. The multi-modal model includes a spoken language model that extracts features from spoken query and a language processing model that extract features from an image. The multi-model model determines a relevance score for the image and the spoken query based on the extracted features. The multi-modal model is trained using a curriculum approach that includes training the spoken language model using audio data. Subsequently, a training dataset comprising a plurality of spoken queries and one or more images associated with each spoken query is used to jointly train the spoken language model and an image processing model to provide a trained multi-modal model.

Abstract: Methods, systems, and non-transitory computer readable media are disclosed for generating a composite image comprising objects in positions from two or more different digital images. In one or more embodiments, the disclosed system receives a sequence of images and identifies objects within the sequence of images. In one example, the disclosed system determines a target position for a first object based on detecting user selection of the first object in the target position from a first image. The disclosed system can generate a fixed object image comprising the first object in the target position. The disclosed system can generate preview images comprising the fixed object image with the second object sequencing through a plurality of positions as seen in the sequence of images. Based on a second user selection of a desired preview image, the disclosed system can generate the composite image.

Abstract: Certain embodiments involve visually augmenting images of three-dimensional containers with virtual elements that fill one or more empty regions of the three-dimensional containers. For instance, a computing system receives a first image that depicts a storage container and identify sub-containers within the storage container. The computing system selects, from a virtual object library, a plurality of virtual objects that are semantically related to the sub-container. The computing system determines an arrangement of the virtual objects within the sub-container based on semantics associated with the sub-container and the plurality of virtual objects. The computing system generates a second image that depicts the arrangement of the plurality of virtual objects within the storage container and sub-containers. The computing system generates, for display, the second image depicting the storage container and the arrangement of the virtual objects.

Abstract: An image differentiation system receives input feature vectors for multiple input images and reference feature vectors for multiple reference images. In some cases, the feature vectors are extracted by an image feature extraction module trained based on training image triplets. A differentiability scoring module determines a differentiability score for each input image based on a distance between the input feature vectors and the reference feature vectors. The distance for each reference feature vector is modified by a weighting factor based on interaction metrics associated with the corresponding reference image. In some cases, an input image is identified as a differentiated image based on the corresponding differentiability score. Additionally or alternatively, an image modification module determines an image modification that increases the differentiability score of the input image. The image modification module generates a recommended image by applying the image modification to the input image.

Abstract: A framework generates a thumbnail to represent a video on a webpage based on a prominent individual appearing in both the video and content of the webpage. Content of a webpage on which a video is to be posted is analyzed to identify individuals represented in the webpage content. Frames of the video are also analyzed to identify individuals in the video. A first individual that appears in both the webpage content and the video is selected based on a score for the first individual determined based on the webpage content. Subsequent to selecting the first individual, frames of the video that include the first individual are analyzed to select a first frame of the video that includes the first individual. A thumbnail to represent the video on the webpage is generated from the first frame, and the thumbnail is provided for presentation on the webpage to represent the video.

Abstract: An image differentiation system receives input feature vectors for multiple input images and reference feature vectors for multiple reference images. In some cases, the feature vectors are extracted by an image feature extraction module trained based on training image triplets. A differentiability scoring module determines a differentiability score for each input image based on a distance between the input feature vectors and the reference feature vectors. The distance for each reference feature vector is modified by a weighting factor based on interaction metrics associated with the corresponding reference image. In some cases, an input image is identified as a differentiated image based on the corresponding differentiability score. Additionally or alternatively, an image modification module determines an image modification that increases the differentiability score of the input image. The image modification module generates a recommended image by applying the image modification to the input image.

Abstract: The present disclosure relates to systems, non-transitory computer-readable media, and methods that implement an image filter for enhancing light text and removing document shadows. In particular embodiments, the disclosed systems use a modified adaptive thresholding approach the relies on image gradients to efficiently guide the thresholding process. In addition, the disclosed systems use a machine-learning model to generate a document shadow map. The document shadow map can include text reflections. Accordingly, the disclosed systems remove text reflections from the document shadow map (e.g., by using an interpolated shadow intensity value of neighboring shadow map pixels). In turn, the disclosed systems use the document text mask and the document shadow map cleaned of text reflections to remove shadows from the digital image. Further, the disclosed systems enhance text in the shadow-removed digital image based on contrast stretching.

Abstract: In implementations of systems for role classification, a computing device implements a role system to receive data describing a corpus of text that is associated with a user ID. Feature values of features are generated by a first machine learning model by processing the corpus of text, the features representing questions with respect to the corpus of text and the feature values representing answers to the questions included in the corpus of text. A classification of a role is generated by a second machine learning model by processing the feature values, the classification of the role indicating a relationship of the user ID with respect to a product or service. The role system outputs an indication of the classification of the role for display in a user interface of a display device.

Abstract: Methods and systems are provided for providing directional assistance to guide a user to position a camera for centering a person's face within the camera's field of view. A neural network system is trained to determine the position of the user's face relative to the center of the field of view as captured by an input image. The neural network system is trained using training input images that are generated by cropping different regions of initial training images. Each initial image is used to create a plurality of different training input images, and directional assistance labels used to train the network may be assigned to each training input image based on how the image is cropped. Once trained, the neural network system determines a position of the user's face, and automatically provides a non-visual prompt indicating how to center the face within the field of view.

Abstract: Techniques and systems are described for performing semantic text searches. A semantic text-searching solution uses a machine learning system (such as a deep learning system) to determine associations between the semantic meanings of words. These associations are not limited by the spelling, syntax, grammar, or even definition of words. Instead, the associations can be based on the context in which characters, words, and/or phrases are used in relation to one another. In response to detecting a request to locate text within an electronic document associated with a keyword, the semantic text-searching solution can return strings within the document that have matching and/or related semantic meanings or contexts, in addition to exact matches (e.g., string matches) within the document. The semantic text-searching solution can then output an indication of the matching strings.

Abstract: Techniques are provided herein for generating improved document summaries that consider the amount of time that has passed since the user last accessed the document. The length of time that has passed since the user has accessed each previous portion of the document is used as a variable to determine how much the summary should focus on each of the previously read sections of the document. When a document is accessed by a user, a relevance score is assigned to content from previously accessed sections of that document, where the relevance score is weighted based on how long ago each of the sections was accessed by the user. Once the various content items of previous sections have been provided relevance scores, selected sentences with the highest relevance scores are fed to a deep learning sequence-to-sequence model is used to build the document summary.

Abstract: A method of generating accessible content is described. Embodiments of the method identifies a plurality of channels for a multimedia communication session, generate a master timeline for the communication session, wherein the master timeline comprises a chronological ordering of events from each of the channels, and wherein each of the events is associated with event-specific audio data, and present the multimedia communication session to a user to enable the user to transition among the channels based on the master timeline.

Abstract: Techniques and systems are described for performing semantic text searches. A semantic text-searching solution uses a machine learning system (such as a deep learning system) to determine associations between the semantic meanings of words. These associations are not limited by the spelling, syntax, grammar, or even definition of words. Instead, the associations can be based on the context in which characters, words, and/or phrases are used in relation to one another. In response to detecting a request to locate text within an electronic document associated with a keyword, the semantic text-searching solution can return strings within the document that have matching and/or related semantic meanings or contexts, in addition to exact matches (e.g., string matches) within the document. The semantic text-searching solution can then output an indication of the matching strings.

Abstract: Systems and methods for removing objects from images are disclosed. An image processing application identifies a boundary of each object of a set of objects in an image. The image processing application identifies a completed boundary for each object of the set of objects by providing the object to a trained model. The image processing application determines a set of masks. Each mask corresponds to an object of the set of objects and represents a region of the image defined by an intersection of the boundary of the object and the boundary of a target object to be removed from the image. The image processing application updates each mask by separately performing content filling on the corresponding region. The image processing application creates an output image by merging each of the updated masks with portions of the image.

Abstract: Techniques are disclosed for generating image recommendations to facilitate the sale of a product. An example methodology includes identifying a product category associated with an image of the product provided by the seller, and a product sub-category associated with the product image. The method further includes retrieving one or more images of for-sale items. The retrieval is based on a search of for-sale listings using the identified product category and the identified product sub-category. The method further includes clustering the retrieved images of for-sale items into groups, each group associated with a perspective viewpoint of the for-sale item. The method further includes providing a selected image from each group as an image recommendation. The selection is based on a value score associated with each of the images of the for-sale items. A graphical status indicating completeness of the seller's image set is updated in response to recommended images being adopted.

Abstract: Methods, apparatus, and systems are provided for obfuscating facial identity in images by synthesizing a new facial image for an input image. A base face is detected from or selected for an input image. Facial images that are similar to the base face are selected and combined to create a new facial image. The new facial image is added to the input image such that the input image includes a combination of the base face and the new facial image. Where no base face is detected in the input image, a base face is selected from reference facial images based at least on pose keypoints identified in the input image. After a new facial image is generated based on the selected base face, a combination of the new facial image and the base facial image are added to the input image by aligning one or more pose keypoints.

Abstract: The present disclosure relates to a cloud-based system including a server for generating and nurturing leads using within-a document lead nurturing. The server enables a marketer to generate initial content and contextual content, specify a consumption condition with respect to the initial content, and generate an electronic file including the initial content and executable code configured to monitor consumption of the initial content. A user (e.g., possible lead) uses a reader program of their device to open the electronic file, which presents the initial content and triggers execution of an instruction of the executable code to monitor interactions of the user with the initial content to determine whether the consumption condition has been satisfied. The reader program executes another instruction of the executable code to present the contextual content and notify the server that the user is a possible lead, when the consumption condition is satisfied.

Abstract: The present disclosure is directed towards methods and systems for providing relevant video scenes in response to a video search query. The systems and methods identify a plurality of key frames of a media object and detect one or more content features represented in the plurality of key frames. Based on the one or more detect content features, the systems and methods associate tags indicating the detected content features with the plurality of key frames of the media object. The systems and methods, in response to receiving a search query including search terms, compare the search terms with the tags of the selected key frames, identify a selected key frame that depicts at least one content feature related to the search terms, and provide a preview image of the media item depicting the at least one content feature.

Abstract: Techniques are disclosed for generating image recommendations to facilitate the sale of a product. An example methodology includes identifying a product category associated with an image of the product provided by the seller, and a product sub-category associated with the product image. The method further includes retrieving one or more images of for-sale items. The retrieval is based on a search of for-sale listings using the identified product category and the identified product sub-category. The method further includes clustering the retrieved images of for-sale items into groups, each group associated with a perspective viewpoint of the for-sale item. The method further includes providing a selected image from each group as an image recommendation. The selection is based on a value score associated with each of the images of the for-sale items. A graphical status indicating completeness of the seller's image set is updated in response to recommended images being adopted.