Adobe, Inc.

The Penn State Research Foundation

Xin Lu

Zhe Lin

Xiaohui Shen

Jimei Yang

Jen-Chan Jeff Chien

Jiahui Yu

Zichuan Liu

Hailin Jin

Wentian Zhao

Yumin Jia

Kun Wan

Jianchao Yang

Zhaowen Wang

Radomir Mech

James Z Wang

Jia Li

Chenxi Liu

Poonam Suryanarayan

Reginald B Adams

Michelle Newman

Jianming Zhang

Jonathan Worthen Brandt

Chen Fang

Yilin Wang

Connelly Barnes

Sarah Kong

Yijun Li

Chih-Yao Hsieh

I-Ming Pao

Yeojin Kim

Jinoh Oh

Wei Xiong

Yibing Song

Sheng-Wei Huang

Yingcong Chen

Shitong Wang

He Qin

Kalyan Krishna Sunkavalli

Joon-Young Lee

Haoxiang Li

Gahye Park

Yi-Hsuan Tsai

Jianbo Ye

Ke Wang

Yu Luo

Zejun Huang

Chao Yang

Mingyuan Wu

Rameswar Panda

In implementations of systems for efficient object segmentation, a computing device implements a segment system to receive a user input specifying coordinates of a digital image. The segment system computes receptive fields of a machine learning model based on the coordinates of the digital image. The machine learning model is trained on training data to generate segment masks for objects depicted in digital images. The segment system processes a portion of a feature map of the digital image using the machine learning model based on the receptive fields. A segment mask is generated for an object depicted in the digital image based on processing the portion of the feature map of the digital image using the machine learning model.

Efficient object segmentation

Methods, systems, and non-transitory computer readable media are disclosed for generating artistic images by applying an artistic-effect to one or more frames of a video stream or digital images. In one or more embodiments, the disclosed system captures a video stream utilizing a camera of a computing device. The disclosed system deploys a distilled artistic-effect neural network on the computing device to generate an artistic version of the captured video stream at a first resolution in real time. The disclosed system can provide the artistic video stream for display via the computing device. Based on an indication of a capture event, the disclosed system utilizes the distilled artistic-effect neural network to generate an artistic image at a higher resolution than the artistic video stream. Furthermore, the disclosed system tunes and utilizes an artistic-effect patch generative adversarial neural network to modify parameters for the distilled artistic-effect neural network.

Generating stylized images on mobile devices

A method for performing efficient mixed-precision search for an artificial neural network (ANN) includes training the ANN by sampling selected candidate quantizers of a bank of candidate quantizer and updating network parameters for a next iteration based on outputs of layers of the ANN. The outputs are computed by processing quantized data with operators (e.g., convolution). The quantizers converge to optimal bit-widths that reduce classification losses bounded by complexity constrains.

Efficient mixed-precision search for quantizers in artificial neural networks

Methods and systems are provided for generating enhanced image. A neural network system is trained where the training includes training a first neural network that generates enhanced images conditioned on content of an image undergoing enhancement and training a second neural network that designates realism of the enhanced images generated by the first neural network. The neural network system is trained by determine loss and accordingly adjusting the appropriate neural network(s). The trained neural network system is used to generate an enhanced aesthetic image from a selected image where the output enhanced aesthetic image has increased aesthetics when compared to the selected image.

Aesthetics-guided image enhancement

Methods, systems, and non-transitory computer readable storage media are disclosed for generating image mattes for detected objects in digital images without trimap segmentation via a multi-branch neural network. The disclosed system utilizes a first neural network branch of a generative neural network to extract a coarse semantic mask from a digital image. The disclosed system utilizes a second neural network branch of the generative neural network to extract a detail mask based on the coarse semantic mask. Additionally, the disclosed system utilizes a third neural network branch of the generative neural network to fuse the coarse semantic mask and the detail mask to generate an image matte. In one or more embodiments, the disclosed system also utilizes a refinement neural network to generate a final image matte by refining selected portions of the image matte generated by the generative neural network.

Generating image mattes without trimap segmentations via a multi-branch neural network

The present disclosure describes systems, non-transitory computer-readable media, and methods for utilizing a machine learning model trained to determine subtle pose differentiations to analyze a repository of captured digital images of a particular user to automatically capture digital images portraying the user. For example, the disclosed systems can utilize a convolutional neural network to determine a pose/facial expression similarity metric between a sample digital image from a camera viewfinder stream of a client device and one or more previously captured digital images portraying the user. The disclosed systems can determine that the similarity metric satisfies a similarity threshold, and automatically capture a digital image utilizing a camera device of the client device. Thus, the disclosed systems can automatically and efficiently capture digital images, such as selfies, that accurately match previous digital images portraying a variety of unique facial expressions specific to individual users.

Capturing digital images utilizing a machine learning model trained to determine subtle pose differentiations

The present disclosure describes systems, non-transitory computer-readable media, and methods for accurately and efficiently removing objects from digital images taken from a camera viewfinder stream. For example, the disclosed systems access digital images from a camera viewfinder stream in connection with an undesired moving object depicted in the digital images. The disclosed systems generate a temporal window of the digital images concatenated with binary masks indicating the undesired moving object in each digital image. The disclosed systems further utilizes a generator as part of a 3D to 2D generative adversarial neural network in connection with the temporal window to generate a target digital image with the region associated with the undesired moving object in-painted. In at least one embodiment, the disclosed systems provide the target digital image to a camera viewfinder display to show a user how a future digital photograph will look without the undesired moving object.

Automatically removing moving objects from video streams

The architectural complexity of a neural network is reduced by selectively pruning channels. A cost metric for a convolution layer is determined. The cost metric indicates a resource cost per channel for the channels of the layer. Training the neural network includes, for channels of the layer, updating a channel-scaling coefficient based on the cost metric. The channel-scaling coefficient linearly scales the output of the channel. A constant channel is identified based on the channel-scaling coefficients. The neural network is updated by pruning the constant channel. Model weights are updated via a stochastic gradient descent of a training loss function evaluated on training data. The channel-scaling coefficients are updated via an iterative-thresholding algorithm that penalizes a batch normalization loss function based on the cost metric for the layer and a norm of the channel-scaling coefficients. When the layer is batch normalized, the channel-scaling coefficients are batch normalization scaling coefficients.

Reducing architectural complexity of convolutional neural networks via channel pruning

Generating stylized images in real time on mobile devices

The present disclosure relates to systems, methods, and non-transitory computer-readable media that generate refined segmentation masks for digital visual media items. For example, in one or more embodiments, the disclosed systems utilize a segmentation refinement neural network to generate an initial segmentation mask for a digital visual media item. The disclosed systems further utilize the segmentation refinement neural network to generate one or more refined segmentation masks based on uncertainly classified pixels identified from the initial segmentation mask. To illustrate, in some implementations, the disclosed systems utilize the segmentation refinement neural network to redetermine whether a set of uncertain pixels corresponds to one or more objects depicted in the digital visual media item based on low-level (e.g., local) feature values extracted from feature maps generated for the digital visual media item.

Iteratively refining segmentation masks

Utilizing a machine learning model trained to determine subtle pose differentiations to automatically capture digital images

The present disclosure describes systems, non-transitory computer-readable media, and methods for accurately and efficiently removing objects from digital images taken from a camera viewfinder stream. For example, the disclosed systems access digital images from a camera viewfinder stream in connection with an undesired moving object depicted in the digital images. The disclosed systems generate a temporal window of the digital images concatenated with binary masks indicating the undesired moving object in each digital image. The disclosed systems further utilizes a 3D to 2D generator as part of a 3D to 2D generative adversarial neural network in connection with the temporal window to generate a target digital image with the region associated with the undesired moving object in-painted. In at least one embodiment, the disclosed systems provide the target digital image to a camera viewfinder display to show a user how a future digital photograph will look without the undesired moving object.

The present disclosure relates to training and utilizing an image exposure transformation network to generate a long-exposure image from a single short-exposure image (e.g., still image). In various embodiments, the image exposure transformation network is trained using adversarial learning, long-exposure ground truth images, and a multi-term loss function. In some embodiments, the image exposure transformation network includes an optical flow prediction network and/or an appearance guided attention network. Trained embodiments of the image exposure transformation network generate realistic long-exposure images from single short-exposure images without additional information.

Utilizing an image exposure transformation neural network to generate a long-exposure image from a single short-exposure image

Predicting patch displacement maps using a neural network is described. Initially, a digital image on which an image editing operation is to be performed is provided as input to a patch matcher having an offset prediction neural network. From this image and based on the image editing operation for which this network is trained, the offset prediction neural network generates an offset prediction formed as a displacement map, which has offset vectors that represent a displacement of pixels of the digital image to different locations for performing the image editing operation. Pixel values of the digital image are copied to the image pixels affected by the operation.

Predicting patch displacement maps using a neural network

Digital image completion by learning generation and patch matching jointly is described. Initially, a digital image having at least one hole is received. This holey digital image is provided as input to an image completer formed with a dual-stage framework that combines a coarse image neural network and an image refinement network. The coarse image neural network generates a coarse prediction of imagery for filling the holes of the holey digital image. The image refinement network receives the coarse prediction as input, refines the coarse prediction, and outputs a filled digital image having refined imagery that fills these holes. The image refinement network generates refined imagery using a patch matching technique, which includes leveraging information corresponding to patches of known pixels for filtering patches generated based on the coarse prediction. Based on this, the image completer outputs the filled digital image with the refined imagery.

Digital image completion by learning generation and patch matching jointly

Growing community of inventors

Mountain View, CA, United States of America

Xin Lu