Abstract: Systems and methods are disclosed for estimating aesthetic quality of digital images using deep learning. In particular, the disclosed systems and methods describe training a neural network to generate an aesthetic quality score digital images. In particular, the neural network includes a training structure that compares relative rankings of pairs of training images to accurately predict a relative ranking of a digital image. Additionally, in training the neural network, an image rating system can utilize content-aware and user-aware sampling techniques to identify pairs of training images that have similar content and/or that have been rated by the same or different users. Using content-aware and user-aware sampling techniques, the neural network can be trained to accurately predict aesthetic quality ratings that reflect subjective opinions of most users as well as provide aesthetic scores for digital images that represent the wide spectrum of aesthetic preferences of various users.

Abstract: The invention is directed towards segmenting images based on natural language phrases. An image and an n-gram, including a sequence of tokens, are received. An encoding of image features and a sequence of token vectors are generated. A fully convolutional neural network identifies and encodes the image features. A word embedding model generates the token vectors. A recurrent neural network (RNN) iteratively updates a segmentation map based on combinations of the image feature encoding and the token vectors. The segmentation map identifies which pixels are included in an image region referenced by the n-gram. A segmented image is generated based on the segmentation map. The RNN may be a convolutional multimodal RNN. A separate RNN, such as a long short-term memory network, may iteratively update an encoding of semantic features based on the order of tokens. The first RNN may update the segmentation map based on the semantic feature encoding.

Abstract: Digital image defect identification and correction techniques are described. In one example, a digital medium environment is configured to identify and correct a digital image defect through identification of a defect in a digital image using machine learning. The identification includes generating a plurality of defect type scores using a plurality of defect type identification models, as part of machine learning, for a plurality of different defect types and determining the digital image includes the defect based on the generated plurality of defect type scores. A correction is generated for the identified defect and the digital image is output as included the generated correction.

Abstract: Methods and systems are provided for generating harmonized images for input composite images. A neural network system can be trained, where the training includes training a neural network that generates harmonized images for input composite images. This training is performed based on a comparison of a training harmonized image and a reference image, where the reference image is modified to generate a training input composite image used to generate the training harmonized image. In addition, a mask of a region can be input to limit the area of the input image that is to be modified. Such a trained neural network system can be used to input a composite image and mask pair for which the trained system will output a harmonized image.

Abstract: Provided are systems and techniques that provide an output phrase describing an image. An example method includes creating, with a convolutional neural network, feature maps describing image features in locations in the image. The method also includes providing a skeletal phrase for the image by processing the feature maps with a first long short-term memory (LSTM) neural network trained based on a first set of ground truth phrases which exclude attribute words. Then, attribute words are provided by processing the skeletal phrase and the feature maps with a second LSTM neural network trained based on a second set of ground truth phrases including words for attributes. Then, the method combines the skeletal phrase and the attribute words to form the output phrase.

Abstract: Methods and systems are provided for using a single image of an indoor scene to estimate illumination of an environment that includes the portion captured in the image. A neural network system may be trained to estimate illumination by generating recovery light masks indicating a probability of each pixel within the larger environment being a light source. Additionally, low-frequency RGB images may be generated that indicating low-frequency information for the environment. The neural network system may be trained using training input images that are extracted from known panoramic images. Once trained, the neural network system infers plausible illumination information from a single image to realistically illumination images and objects being manipulated in graphics applications, such as with image compositing, modeling, and reconstruction.

Abstract: Embodiments of the present disclosure relate to a sky editing system and related processes for sky editing. The sky editing system includes a composition detector to determine the composition of a target image. A sky search engine in the sky editing system is configured to find a reference image with similar composition with the target image. Subsequently, a sky editor replaces content of the sky in the target image with content of the sky in the reference image. As such, the sky editing system transforms the target image into a new image with a preferred sky background.

Abstract: Methods and systems for recognizing people in images with increased accuracy are disclosed. In particular, the methods and systems divide images into a plurality of clusters based on common characteristics of the images. The methods and systems also determine an image cluster to which an image with an unknown person instance most corresponds. One or more embodiments determine a probability that the unknown person instance is each known person instance in the image cluster using a trained cluster classifier of the image cluster. Optionally, the methods and systems determine context weights for each combination of an unknown person instance and each known person instance using a conditional random field algorithm based on a plurality of context cues associated with the unknown person instance and the known person instances. The methods and systems calculate a contextual probability based on the cluster-based probabilities and context weights to identify the unknown person instance.

Abstract: Systems and methods are disclosed for segmenting a digital image to identify an object portrayed in the digital image from background pixels in the digital image. In particular, in one or more embodiments, the disclosed systems and methods use a first neural network and a second neural network to generate image information used to generate a segmentation mask that corresponds to the object portrayed in the digital image. Specifically, in one or more embodiments, the disclosed systems and methods optimize a fit between a mask boundary of the segmentation mask to edges of the object portrayed in the digital image to accurately segment the object within the digital image.

Abstract: In various implementations, a personal asset management application is configured to perform operations that facilitate the ability to search multiple images, irrespective of the images having characterizing tags associated therewith or without, based on a simple text-based query. A first search is conducted by processing a text-based query to produce a first set of result images used to further generate a visually-based query based on the first set of result images. A second search is conducted employing the visually-based query that was based on the first set of result images received in accordance with the first search conducted and based on the text-based query. The second search can generate a second set of result images, each having visual similarity to at least one of the images generated for the first set of result images.

Abstract: Accelerating object detection techniques are described. In one or more implementations, adaptive sampling techniques are used to extract features from an image. Coarse features are extracted from the image and used to generate an object probability map. Then, dense features are extracted from high-probability object regions of the image identified in the object probability map to enable detection of an object in the image. In one or more implementations, cascade object detection techniques are used to detect an object in an image. In a first stage, exemplars in a first subset of exemplars are applied to features extracted from the multiple regions of the image to detect object candidate regions. Then, in one or more validation stages, the object candidate regions are validated by applying exemplars from the first subset of exemplars and one or more additional subsets of exemplars.

Abstract: In embodiments of event image curation, a computing device includes memory that stores a collection of digital images associated with a type of event, such as a digital photo album of digital photos associated with the event, or a video of image frames and the video is associated with the event. A curation application implements a convolutional neural network, which receives the digital images and a designation of the type of event. The convolutional neural network can then determine an importance rating of each digital image within the collection of the digital images based on the type of the event. The importance rating of a digital image is representative of an importance of the digital image to a person in context of the type of the event. The convolutional neural network generates an output of representative digital images from the collection based on the importance rating of each digital image.

Abstract: Joint depth estimation and semantic labeling techniques usable for processing of a single image are described. In one or more implementations, global semantic and depth layouts are estimated of a scene of the image through machine learning by the one or more computing devices. Local semantic and depth layouts are also estimated for respective ones of a plurality of segments of the scene of the image through machine learning by the one or more computing devices. The estimated global semantic and depth layouts are merged with the local semantic and depth layouts by the one or more computing devices to semantically label and assign a depth value to individual pixels in the image.

Abstract: Systems and methods are disclosed for estimating aesthetic quality of digital images using deep learning. In particular, the disclosed systems and methods describe training a neural network to generate an aesthetic quality score digital images. In particular, the neural network includes a training structure that compares relative rankings of pairs of training images to accurately predict a relative ranking of a digital image. Additionally, in training the neural network, an image rating system can utilize content-aware and user-aware sampling techniques to identify pairs of training images that have similar content and/or that have been rated by the same or different users. Using content-aware and user-aware sampling techniques, the neural network can be trained to accurately predict aesthetic quality ratings that reflect subjective opinions of most users as well as provide aesthetic scores for digital images that represent the wide spectrum of aesthetic preferences of various users.

Abstract: Neural network patch aggregation and statistical techniques are described. In one or more implementations, patches are generated from an image, e.g., randomly, and used to train a neural network. An aggregation of outputs of patches processed by the neural network may be used to label an image using an image descriptor, such as to label aesthetics of the image, classify the image, and so on. In another example, the patches may be used by the neural network to calculate statistics describing the patches, such as to describe statistics such as minimum, maximum, median, and average of activations of image characteristics of the individual patches. These statistics may also be used to support a variety of functionality, such as to label the image as described above.

Abstract: Techniques for planar region-guided estimates of 3D geometry of objects depicted in a single 2D image. The techniques estimate regions of an image that are part of planar regions (i.e., flat surfaces) and use those planar region estimates to estimate the 3D geometry of the objects in the image. The planar regions and resulting 3D geometry are estimated using only a single 2D image of the objects. Training data from images of other objects is used to train a CNN with a model that is then used to make planar region estimates using a single 2D image. The planar region estimates, in one example, are based on estimates of planarity (surface plane information) and estimates of edges (depth discontinuities and edges between surface planes) that are estimated using models trained using images of other scenes.

Abstract: Techniques for increasing robustness of a convolutional neural network based on training that uses multiple datasets and multiple tasks are described. For example, a computer system trains the convolutional neural network across multiple datasets and multiple tasks. The convolutional neural network is configured for learning features from images and accordingly generating feature vectors. By using multiple datasets and multiple tasks, the robustness of the convolutional neural network is increased. A feature vector of an image is used to apply an image-related operation to the image. For example, the image is classified, indexed, or objects in the image are tagged based on the feature vector. Because the robustness is increased, the accuracy of the generating feature vectors is also increased. Hence, the overall quality of an image service is enhanced, where the image service relies on the image-related operation.

Abstract: A digital medium environment includes an image processing application that performs object segmentation on an input image. An improved object segmentation method implemented by the image processing application comprises receiving an input image that includes an object region to be segmented by a segmentation process, processing the input image to provide a first segmentation that defines the object region, and processing the first segmentation to provide a second segmentation that provides pixel-wise label assignments for the object region. In some implementations, the image processing application performs improved sky segmentation on an input image containing a depiction of a sky.

Abstract: The present disclosure is directed toward systems and methods for tracking objects in videos. For example, one or more embodiments described herein utilize various tracking methods in combination with an image search index made up of still video frames indexed from a video. One or more embodiments described herein utilize a backward and forward tracking method that is anchored by one or more key frames in order to accurately track an object through the frames of a video, even when the video is long and may include challenging conditions.

Abstract: Systems and methods are disclosed for segmenting a digital image to identify an object portrayed in the digital image from background pixels in the digital image. In particular, in one or more embodiments, the disclosed systems and methods use a first neural network and a second neural network to generate image information used to generate a segmentation mask that corresponds to the object portrayed in the digital image. Specifically, in one or more embodiments, the disclosed systems and methods optimize a fit between a mask boundary of the segmentation mask to edges of the object portrayed in the digital image to accurately segment the object within the digital image.