Abstract: The disclosure describes one or more implementations of a neural network architecture pruning system that automatically and progressively prunes neural networks. For instance, the neural network architecture pruning system can automatically reduce the size of an untrained or previously-trained neural network without reducing the accuracy of the neural network. For example, the neural network architecture pruning system jointly trains portions of a neural network while progressively pruning redundant subsets of the neural network at each training iteration. In many instances, the neural network architecture pruning system increases the accuracy of the neural network by progressively removing excess or redundant portions (e.g., channels or layers) of the neural network. Further, by removing portions of a neural network, the neural network architecture pruning system can increase the efficiency of the neural network.

Abstract: The present disclosure relates to shaping the architecture of a neural network. For example, the disclosed systems can provide a neural network shaping mechanism for at least one sampling layer of a neural network. The neural network shaping mechanism can include a learnable scaling factor between a sampling rate of the at least one sampling layer and an additional sampling function. The disclosed systems can learn the scaling factor based on a dataset while jointly learning the network weights of the neural network. Based on the learned scaling factor, the disclosed systems can shape the architecture of the neural network by modifying the sampling rate of the at least one sampling layer.

Abstract: The technology employs a patch-based multi-scale Transformer (300) that is usable with various imaging applications. This avoids constraints on image fixed input size and predicts the quality effectively on a native resolution image. A native resolution image (304) is transformed into a multi-scale representation (302), enabling the Transformer's self-attention mechanism to capture information on both fine-grained detailed patches and coarse-grained global patches. Spatial embedding (316) is employed to map patch positions to a fixed grid, in which patch locations at each scale are hashed to the same grid. A separate scale embedding (318) is employed to distinguish patches coming from different scales in the multiscale representation. Self-attention (508) is performed to create a final image representation. In some instances, prior to performing self-attention, the system may prepend a learnable classification token (322) to the set of input tokens.

Abstract: The present disclosure relates to systems, non-transitory computer-readable media, and methods that utilize a progressive refinement network to refine alpha mattes generated utilizing a mask-guided matting neural network. In particular, the disclosed systems can use the matting neural network to process a digital image and a coarse guidance mask to generate alpha mattes at discrete neural network layers. In turn, the disclosed systems can use the progressive refinement network to combine alpha mattes and refine areas of uncertainty. For example, the progressive refinement network can combine a core alpha matte corresponding to more certain core regions of a first alpha matte and a boundary alpha matte corresponding to uncertain boundary regions of a second, higher resolution alpha matte. Based on the combination of the core alpha matte and the boundary alpha matte, the disclosed systems can generate a final alpha matte for use in image matting processes.

Abstract: The disclosure describes one or more implementations of a neural network architecture pruning system that automatically and progressively prunes neural networks. For instance, the neural network architecture pruning system can automatically reduce the size of an untrained or previously-trained neural network without reducing the accuracy of the neural network. For example, the neural network architecture pruning system jointly trains portions of a neural network while progressively pruning redundant subsets of the neural network at each training iteration. In many instances, the neural network architecture pruning system increases the accuracy of the neural network by progressively removing excess or redundant portions (e.g., channels or layers) of the neural network. Further, by removing portions of a neural network, the neural network architecture pruning system can increase the efficiency of the neural network.

Abstract: Methods, systems, and non-transitory computer readable storage media are disclosed for utilizing a plurality of neural networks in a multi-branch pipeline to generate image masks for digital images. Specifically, the disclosed system can classify a digital image as a portrait or a non-portrait image. Based on classifying a portrait image, the disclosed system can utilize separate neural networks to generate a first mask portion for a portion of the digital image including a defined boundary region and a second mask portion for a portion of the digital image including a blended boundary region. The disclosed system can generate the mask portion for the blended boundary region by utilizing a trimap generation neural network to automatically generate a trimap segmentation including the blended boundary region. The disclosed system can then merge the first mask portion and the second mask portion to generate an image mask for the digital image.

Abstract: The present disclosure relates to systems, methods, and non-transitory computer-readable media that generate preliminary object masks for objects in an image, surface the preliminary object masks as object mask previews, and on-demand converts preliminary object masks into refined object masks. Indeed, in one or more implementations, an object mask preview and on-demand generation system automatically detects objects in an image. For the detected objects, the object mask preview and on-demand generation system generates preliminary object masks for the detected objects of a first lower resolution. The object mask preview and on-demand generation system surfaces a given preliminary object mask in response to detecting a first input. The object mask preview and on-demand generation system also generates a refined object mask of a second higher resolution in response to detecting a second input.

Abstract: A method includes training a first model to measure the banding artefacts, training a second model to deband the image, and generating a debanded image for the image using the second model. Training the first model can include selecting a first set of first training images, generating a banding edge map for a first training image, where the map includes weights that emphasize banding edges and de-emphasize true edges in the first training image, and using the map and a luminance plane of the first training image as input to the first model. Training the second model can include selecting a second set of second training images, generating a debanded training image for a second training image, generating a banding score for the debanded training image using the first model, and using the banding score in a loss function used in training the second model.

Abstract: The present disclosure relates to systems, methods, and non-transitory computer-readable media that upsample and refine segmentation masks. Indeed, in one or more implementations, a segmentation mask refinement and upsampling system upsamples a preliminary segmentation mask utilizing a patch-based refinement process to generate a patch-based refined segmentation mask. The segmentation mask refinement and upsampling system then fuses the patch-based refined segmentation mask with an upsampled version of the preliminary segmentation mask. By fusing the patch-based refined segmentation mask with the upsampled preliminary segmentation mask, the segmentation mask refinement and upsampling system maintains a global perspective and helps avoid artifacts due to the local patch-based refinement process.

Abstract: The present disclosure relates to systems, methods, and non-transitory computer readable media for accurately and flexibly generating harmonized digital images utilizing an object-to-object harmonization neural network. For example, the disclosed systems implement, and learn parameters for, an object-to-object harmonization neural network to combine a style code from a reference object with features extracted from a target object. Indeed, the disclosed systems extract a style code from a reference object utilizing a style encoder neural network. In addition, the disclosed systems generate a harmonized target object by applying the style code of the reference object to a target object utilizing an object-to-object harmonization neural network.

Abstract: Video streams uploaded to a video hosting platform are transcoded using quality-normalized transcoding parameters dynamically selected using a learning model. Video frames of a video stream are processed using the learning model to determine bitrate and quality score pairs for some or all possible transcoding resolutions. The listing of bitrate and quality score pairs determined for each resolution is processed to determine a set of transcoding parameters for transcoding the video stream into each resolution. The bitrate and quality score pairs of a given listing may be processed using one or more predefined thresholds, which may, in some cases, refer to a weighted distribution of resolutions according to watch times of videos of the video hosting platform. The video stream is then transcoded into the various resolutions using the set of transcoding parameters selected for each resolution.

Abstract: 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.

Abstract: The present disclosure relates to systems, non-transitory computer-readable media, and methods that utilize a progressive refinement network to refine alpha mattes generated utilizing a mask-guided matting neural network. In particular, the disclosed systems can use the matting neural network to process a digital image and a coarse guidance mask to generate alpha mattes at discrete neural network layers. In turn, the disclosed systems can use the progressive refinement network to combine alpha mattes and refine areas of uncertainty. For example, the progressive refinement network can combine a core alpha matte corresponding to more certain core regions of a first alpha matte and a boundary alpha matte corresponding to uncertain boundary regions of a second, higher resolution alpha matte. Based on the combination of the core alpha matte and the boundary alpha matte, the disclosed systems can generate a final alpha matte for use in image matting processes.

Abstract: Various embodiments of systems and methods for cross media joint friend and item recommendations are disclosed herein.

Abstract: A system and methods are disclosed for optimal format selection for video players based on visual quality. The method includes generating a plurality of reference transcoded versions of a reference video, obtaining quality scores for frames of the plurality of reference transcoded versions of the reference video, generating a first training input comprising a set of color attributes, spatial attributes, and temporal attributes of the frames of the reference video, and generating a first target output for the first training input, wherein the first target output comprises the quality scores for the frames of the plurality of reference transcoded versions of the reference video. The method further includes providing the training data to train a machine learning model on (i) a set of training inputs comprising the first training input and (ii) a set of target outputs comprising the first target output.

Abstract: 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.

Abstract: A trained model is retrained for video quality assessment and used to identify sets of adaptive compression parameters for transcoding user generated video content. Using transfer learning, the model, which is initially trained for image object detection, is retrained for technical content assessment and then again retrained for video quality assessment. The model is then deployed into a transcoding pipeline and used for transcoding an input video stream of user generated content. The transcoding pipeline may be structured in one of several ways. In one example, a secondary pathway for video content analysis using the model is introduced into the pipeline, which does not interfere with the ultimate output of the transcoding should there be a network or other issue. In another example, the model is introduced as a library within the existing pipeline, which would maintain a single pathway, but ultimately is not expected to introduce significant latency.

Abstract: 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.

Abstract: The present disclosure relates to systems, methods, and non-transitory computer readable media for accurately, efficiently, and flexibly generating harmonized digital images utilizing a self-supervised image harmonization neural network. In particular, the disclosed systems can implement, and learn parameters for, a self-supervised image harmonization neural network to extract content from one digital image (disentangled from its appearance) and appearance from another from another digital image (disentangled from its content). For example, the disclosed systems can utilize a dual data augmentation method to generate diverse triplets for parameter learning (including input digital images, reference digital images, and pseudo ground truth digital images), via cropping a digital image with perturbations using three-dimensional color lookup tables (“LUTs”).

Abstract: The present disclosure relates to systems, methods, and non-transitory computer-readable media that utilizes a neural network having a hierarchy of hierarchical point-wise refining blocks to generate refined segmentation masks for high-resolution digital visual media items. For example, in one or more embodiments, the disclosed systems utilize a segmentation refinement neural network having an encoder and a recursive decoder to generate the refined segmentation masks. The recursive decoder includes a deconvolution branch for generating feature maps and a refinement branch for generating and refining segmentation masks. In particular, in some cases, the refinement branch includes a hierarchy of hierarchical point-wise refining blocks that recursively refine a segmentation mask generated for a digital visual media item.