Abstract: The present disclosure relates to a font recognition system that employs a multi-task learning framework and training to improve font classification and remove negative side effects caused by intra-class variances of glyph content. For example, in one or more embodiments, the font recognition system trains a hybrid font recognition neural network that includes two or more font recognition neural networks and a weight prediction neural network. The hybrid font recognition neural network determines and generates classification weights based on which font recognition neural network within the hybrid font recognition neural network is best suited to classify the font in an input text image. By employing a hybrid trained font classification neural network, the font recognition system can improve overall font recognition as well as remove the negative side effects from diverse glyph content.

Abstract: The present disclosure relates to a font recognition system that employs a multi-task learning framework and adversarial training to improve font classification and remove negative side effects caused by intra-class variances of glyph content. For example, in one or more embodiments, the font recognition system adversarial trains a font recognition neural network by minimizing font classification loss while at the same time maximizing glyph classification loss. By employing an adversarially trained font classification neural network, the font recognition system can improve overall font recognition by removing the negative side effects from diverse glyph content.

Abstract: The present disclosure relates to a font recognition system that employs a multi-task learning framework to jointly improve font classification and remove negative side effects caused by intra-class variances of glyph content. For example, in one or more embodiments, the font recognition system can jointly train a font recognition neural network using a font classification loss model and triplet loss model to generate a deep learning neural network that provides improved font classifications. In addition, the font recognition system can employ the trained font recognition neural network to efficiently recognize fonts within input images as well as provide other suggested fonts.

Abstract: Systems and techniques for converting a low resolution image to a high resolution image include receiving a low resolution image having one or more noise artifacts at a neural network. A noise reduction level is received at the neural network. The neural network determines a network parameter based on the noise reduction level. The neural network converts the low resolution image to a high resolution image and removes one or more of the noise artifacts from the low resolution image during the converting by the using the network parameter. The neural network outputs the high resolution image.

Abstract: The present disclosure relates to an asymmetric font pairing system that efficiently pairs digital fonts. For example, in one or more embodiments, the asymmetric font pairing system automatically identifies and provides users with visually aesthetic font pairs for use in different sections of an electronic document. In particular, the asymmetric font pairing system learns visually aesthetic font pairs using joint symmetric and asymmetric compatibility metric learning. In addition, the asymmetric font pairing system provides compact compatibility spaces (e.g., a symmetric compatibility space and an asymmetric compatibility space) to computing devices (e.g., client devices and server devices), which enable the computing devices to quickly and efficiently provide font pairs to users.

Abstract: Constraining memory use for overlapping virtual memory operations is described. The memory use is constrained to prevent memory from exceeding an operational threshold, e.g., in relation to operations for modifying content. These operations are implemented according to algorithms having a plurality of instructions. Before the instructions are performed in relation to the content, virtual memory is allocated to the content data, which is then loaded into the virtual memory and is also partitioned into data portions. In the context of the described techniques, at least one of the instructions affects multiple portions of the content data loaded in virtual memory. When this occurs, the instruction is carried out, in part, by transferring the multiple portions of content data between the virtual memory and a memory such that a number of portions of the content data in the memory is constrained to the memory that is reserved for the operation.

Abstract: Systems and techniques that synthesize an image with similar texture to a selected style image. A generator network is trained to synthesize texture images depending on a selection unit input. The training configures the generator network to synthesize texture images that are similar to individual style images of multiple style images based on which is selected by the selection unit input. The generator network can be configured to minimize a covariance matrix-based style loss and/or a diversity loss in synthesizing the texture images. After training the generator network, the generator network is used to synthesize texture images for selected style images. For example, this can involve receiving user input selecting a selected style image, determining the selection unit input based on the selected style image, and synthesizing texture images using the generator network with the selection unit input and noise input.

Abstract: The present disclosure includes methods and systems for generating captions for digital images. In particular, the disclosed systems and methods can train an image encoder neural network and a sentence decoder neural network to generate a caption from an input digital image. For instance, in one or more embodiments, the disclosed systems and methods train an image encoder neural network (e.g., a character-level convolutional neural network) utilizing a semantic similarity constraint, training images, and training captions. Moreover, the disclosed systems and methods can train a sentence decoder neural network (e.g., a character-level recurrent neural network) utilizing training sentences and an adversarial classifier.

Abstract: Systems and techniques for identification of fonts include receiving a selection of an area of an image including text, where the selection is received from within an application. The selected area of the image is input to a font matching module within the application. The font matching module identifies one or more fonts similar to the text in the selected area using a convolutional neural network. The one or more fonts similar to the text are displayed within the application and the selection and use of the one or more fonts is enabled within the application.

Abstract: Photometric stabilization for time-compressed video is described. Initially, video content captured by a video capturing device is time-compressed by selecting a subset of frames from the video content according to a frame sampling technique. Photometric characteristics are then stabilized across the frames of the time-compressed video. This involves determining correspondences of pixels in adjacent frames of the time-compressed video. Photometric transformations are then determined that describe how photometric characteristics (e.g., one or both of luminance and chrominance) change between the adjacent frames, given movement of objects through the captured scene. Based on the determined photometric transformations, filters are computed for smoothing photometric characteristic changes across the time-compressed video. Photometrically stabilized time-compressed video is generated from the time-compressed video by using the filters to smooth the photometric characteristic changes.

Abstract: Font replacement based on visual similarity is described. In one or more embodiments, a font descriptor includes multiple font features derived from a visual appearance of a font by a font visual similarity model. The font visual similarity model can be trained using a machine learning system that recognizes similarity between visual appearances of two different fonts. A source computing device embeds a font descriptor in a document, which is transmitted to a destination computing device. The destination compares the embedded font descriptor to font descriptors corresponding to local fonts. Based on distances between the embedded and the local font descriptors, at least one matching font descriptor is determined. The local font corresponding to the matching font descriptor is deemed similar to the original font. The destination computing device controls presentations of the document using the similar local font. Computation of font descriptors can be outsourced to a remote location.

Abstract: Font recognition and similarity determination techniques and systems are described. In a first example, localization techniques are described to train a model using machine learning (e.g., a convolutional neural network) using training images. The model is then used to localize text in a subsequently received image, and may do so automatically and without user intervention, e.g., without specifying any of the edges of a bounding box. In a second example, a deep neural network is directly learned as an embedding function of a model that is usable to determine font similarity. In a third example, techniques are described that leverage attributes described in metadata associated with fonts as part of font recognition and similarity determinations.

Abstract: Photometric stabilization for time-compressed video is described. Initially, video content captured by a video capturing device is time-compressed by selecting a subset of frames from the video content according to a frame sampling technique. Photometric characteristics are then stabilized across the frames of the time-compressed video. This involves determining correspondences of pixels in adjacent frames of the time-compressed video. Photometric transformations are then determined that describe how photometric characteristics (e.g., one or both of luminance and chrominance) change between the adjacent frames, given movement of objects through the captured scene. Based on the determined photometric transformations, filters are computed for smoothing photometric characteristic changes across the time-compressed video. Photometrically stabilized time-compressed video is generated from the time-compressed video by using the filters to smooth the photometric characteristic changes.

Abstract: Font recognition and similarity determination techniques and systems are described. In a first example, localization techniques are described to train a model using machine learning (e.g., a convolutional neural network) using training images. The model is then used to localize text in a subsequently received image, and may do so automatically and without user intervention, e.g., without specifying any of the edges of a bounding box. In a second example, a deep neural network is directly learned as an embedding function of a model that is usable to determine font similarity. In a third example, techniques are described that leverage attributes described in metadata associated with fonts as part of font recognition and similarity determinations.

Abstract: In some embodiments, techniques for synthesizing an image style based on a plurality of neural networks are described. A computer system selects a style image based on user input that identifies the style image. The computer system generates an image based on a generator neural network and a loss neural network. The generator neural network outputs the synthesized image based on a noise vector and the style image and is trained based on style features generated from the loss neural network. The loss neural network outputs the style features based on a training image. The training image and the style image have a same resolution. The style features are generated at different resolutions of the training image. The computer system provides the synthesized image to a user device in response to the user input.

Abstract: Systems and techniques that synthesize an image with similar texture to a selected style image. A generator network is trained to synthesize texture images depending on a selection unit input. The training configures the generator network to synthesize texture images that are similar to individual style images of multiple style images based on which is selected by the selection unit input. The generator network can be configured to minimize a covariance matrix-based style loss and/or a diversity loss in synthesizing the texture images. After training the generator network, the generator network is used to synthesize texture images for selected style images. For example, this can involve receiving user input selecting a selected style image, determining the selection unit input based on the selected style image, and synthesizing texture images using the generator network with the selection unit input and noise input.

Abstract: Methods and systems are provided for ranking images against queries. A visual modality ranking of visual features of a digital image against a query is generated. A language modality ranking of text features of text associated with the digital image against the query is also generated. A multi-modal neural network determines importance weightings of the language modality ranking and the visual modality ranking against the query. The visual modality ranking and the language modality ranking are combined into a multi-modal ranking of the digital image against the query based on the importance weightings. The digital image is provided as a search result of the query based on the multi-modal ranking.

Abstract: Font replacement based on visual similarity is described. In one or more embodiments, a font descriptor includes multiple font features derived from a visual appearance of a font by a font visual similarity model. The font visual similarity model can be trained using a machine learning system that recognizes similarity between visual appearances of two different fonts. A source computing device embeds a font descriptor in a document, which is transmitted to a destination computing device. The destination compares the embedded font descriptor to font descriptors corresponding to local fonts. Based on distances between the embedded and the local font descriptors, at least one matching font descriptor is determined. The local font corresponding to the matching font descriptor is deemed similar to the original font. The destination computing device controls presentations of the document using the similar local font. Computation of font descriptors can be outsourced to a remote location.

Abstract: Techniques and systems are described to generate a compact video feature representation for sequences of frames in a video. In one example, values of features are extracted from each frame of a plurality of frames of a video using machine learning, e.g., through use of a convolutional neural network. A video feature representation is generated of temporal order dynamics of the video, e.g., through use of a recurrent neural network. For example, a maximum value is maintained of each feature of the plurality of features that has been reached for the plurality of frames in the video. A timestamp is also maintained as indicative of when the maximum value is reached for each feature of the plurality of features. The video feature representation is then output as a basis to determine similarity of the video with at least one other video based on the video feature representation.

Abstract: The present disclosure is directed towards systems and methods for generating a new aligned image from a plurality of burst image. The systems and methods subdivide a reference image into a plurality of local regions and a subsequent image into a plurality of corresponding local regions. Additionally, the systems and methods detect a plurality of feature points in each of the reference image and the subsequent image and determine matching feature point pairs between the reference image and the subsequent image. Based on the matching feature point pairs, the systems and methods determine at least one homography of the reference image to the subsequent image. Based on the homography, the systems and methods generate a new aligned image that is that is pixel-wise aligned to the reference image. Furthermore, the systems and methods refines boundaries between local regions of the new aligned image.