Google Inc.

Sarah Parker

Debargha Mukherjee

Yue Chen

Urvang Joshi

Yuxin Liu

James Bankoski

Michele Mae Covell

Chen Wang

Ramin D Zabih

Elliott Karpilovsky

Emil Keyder

Guided restoration is used to restore video data degraded from a video frame. The video frame is divided into restoration units (RUs) which each correspond to one or more blocks of the video frame. Restoration schemes are selected for each RU. The restoration schemes may indicate to use one of a plurality of neural networks trained for the guided restoration. Alternatively, the restoration schemes may indicate to use a neural network and a filter-based restoration tool. The video frame is then restored by processing each RU according to the respective selected restoration scheme. During encoding, the restored video frame is encoded to an output bitstream, and the use of the selected restoration schemes may be signaled within the output bitstream. During decoding, the restored video frame is output to an output video stream.

Guided restoration of video data using neural networks

Multiple global motion models associated with respective segments of a current frame are decoded from a compressed bitstream. Each global motion model is based on a segmentation of the current frame and represents a respective underlying motion of blocks within a respective segment. Blocks of the current frame are decoded by: for each inter-predicted block of a segment, decoding, form the compressed bitstream, an indication of whether to decode the each inter-predicted block based on a global motion model of the multiple global motion models and associated with the segment, or whether to decode the each inter-predicted block based on a motion vector that is different from the global motion model; and decoding the each inter-predicted block based on the indication.

Segmentation-based parameterized motion models

Image coding using guided machine learning restoration may include obtaining reconstructed frame data by decoding, obtaining a restored frame by restoring the reconstructed frame, and outputting the restored frame. Obtaining the restored frame may include obtaining a reconstructed block, obtaining guide parameter values, obtaining a restored block, and including the restored block in the restored frame. Obtaining the restored block may include inputting the reconstructed block to an input layer of a trained guided convolutional neural network, wherein the neural network is constrained such that an output layer has a defined cardinality of channels, obtaining, from the output layer, neural network output channel predictions, obtaining a guided neural network prediction as a linear combination of the guide parameter values and the neural network output channel predictions, and generating the restored block using the guided neural network prediction.

Video coding with guided machine learning restoration

Generating a compound predictor block for a current block of video includes generating, for the current block, a first predictor block using one of inter-prediction or intra-prediction and generating a second predictor block. The first predictor block includes a first pixel and the second predictor block includes a second pixel that is co-located with the first pixel. A first weight is determined for the first pixel using a difference between a value of the first pixel and a value of the second pixel. A second weight is determined for the second pixel using the first weight. The compound predictor block is generated by combining the first predictor block and the second predictor block. The compound predictor block includes a weighted pixel that is determined using a weighted sum of the first pixel and the second pixel using the first weight and the second weight.

Compound prediction for video coding

A hybrid apparatus for coding a video stream includes a first encoder. The first encoder includes a neural network having at least one hidden layer, and the neural network receives source data from the video stream at a first hidden layer of the at least one hidden layer, receives side information correlated with the source data at the first hidden layer, and generates guided information using the source data and the side information. The first encoder outputs the guided information and the side information for a decoder to reconstruct the source data.

Hybrid motion-compensated neural network with side-information based video coding

Video coding may include generating, by a processor, a decoded frame by decoding a current frame from an encoded bitstream and outputting a reconstructed frame based on the decoded frame. Decoding includes identifying a current encoded block from the current frame, identifying a prediction coding model for the current block, wherein the prediction coding model is a machine learning prediction coding model from a plurality of machine learning prediction coding models, identifying reference values for decoding the current block based on the prediction coding model, obtaining prediction values based on the prediction coding model and the reference values, generating a decoded block corresponding to the current encoded block based on the prediction values, and including the decoded block in the decoded frame.

Image and video coding using machine learning prediction coding models

Transform-level partitioning of a prediction residual block is performed to improve compression efficiency of video data. During encoding, a prediction residual block is generated responsive to prediction-level partitioning performed against a video block, a transform block partition type to use is determined based on the prediction residual block, a non-recursive transform-level partitioning is performed against the prediction residual block according to the transform block partition type, and transform blocks generated as a result of the transform-level partitioning are encoded to a bitstream. During decoding, a symbol representative of the transform block partition type used to encode transform blocks is derived from the bitstream, inverse transformed blocks are produced by inverse transforming encoded video data associated with the prediction residual block, and the prediction residual block is reproduced according to the transform block partition type and used to reconstruct the video block, which is output within an output video stream.

Extended transform partitions for video compression

Generating a compound predictor block of a current block of video can include generating, for the current block, predictor blocks comprising a first predictor block including first predictor pixels and a second predictor block including second predictor pixels; using at least a subset of the first predictor pixels to determine a first weight for a first predictor pixel of the first predictor pixels; obtaining a second weight for a second predictor pixel of the second predictor pixels, where the second predictor pixel is co-located with the first predictor pixel; and generating the compound predictor block by combining the first predictor block and the second predictor block, where the predictor block includes a weighted pixel that is determined using a weighted sum of the first predictor pixel and the second predictor pixel using the first weight and the second weight, respectively.

An apparatus for encoding a current frame of a video. The apparatus includes a memory and a processor. The processor is configured to execute instructions stored in the memory to generate, for each reference frame of a subset of available reference frames, at least one respective candidate global motion model (GMM); partition the current frame into blocks; generate an aggregated residual frame for the current frame; and encode the respective residual blocks in a compressed bitstream. To generate the aggregated residual frame includes to select, for predicting each block of the blocks, a respective selected GMM, where the respective selected GMM corresponds to the one of the at least one respective candidate GMMs that minimizes a total error associated with the aggregated residual frame; and obtain respective residual blocks for the block.

Diversified motion using multiple global motion models

Generating a compound predictor block of a current block includes generating, for the current block, predictor blocks including a first predictor block formed of first predictor pixels. Using at least a subset of the first predictor pixels, a first modulation value for modulating a first weight to be applied to a first predictor pixel of the first predictor pixels is determined. The compound predictor block is generated using the first predictor pixel, the first weight, and the first modulation value.

Restoring a degraded tile of a degraded frame resulting from reconstruction is disclosed. A method includes, for a scaling factor of at least some scaling factors, recursively filtering the degraded tile using the scaling factor to generate a respective restored tile, and determining a respective error for the respective restored tile with respect to the source tile. The method also includes selecting an optimal scaling factor from the at least some scaling factors and encoding, in an encoded bitstream, a scaling parameter based on the optimal scaling factor. The optimal scaling factor corresponding to a smallest respective error. An apparatus includes a processor and non-transitory memory storing instructions. The instructions cause the processor to determine, from an encoded bitstream, a scaling factor, which determines how strongly edges in the degraded tile affect filtering operations, and recursively filter, resulting in a restored tile, the degraded tile using the scaling factor.

Restoration in video coding using domain transform recursive filters

A method for encoding a current frame of a video includes jointly determining respective motion models for reference frames and encoding the current frame using the respective motion models. The reference frames are used for encoding the current frame. Jointly determining respective motion models for reference frames includes determining respective aggregated residuals for combinations of candidate motion models and selecting the combination of candidate motion models that corresponds to the smallest aggregated residual. The respective motion models correspond to the candidate motion models of the selected combination.

A method includes segmenting the video frame with respect to a reference frame resulting in a segmentation, determining a first motion vector for the current block based on the segmentation, determining a second motion vector for the current block using translational motion compensation, and encoding, for the current block, the one of the first motion vector and the second motion vector corresponding to a smaller error. The segmentation includes a segment containing the current block and a parameterized motion model for the segment. Another method includes identifying a parameterized motion model corresponding to a motion model type, associating the parameterized motion model with a segment of a reference frame, and, in response to determining that the current block is encoded using the parameterized motion model, decoding the current block using the parameterized motion model.

A current block of a video frame can be encoded or decoded using parameterized motion models. First and second parameterized motion models are identified. The first parameterized motion model corresponds to a first motion model type, and the second parameterized motion model corresponds to a second motion model type. The first and second parameterized motion models are associated with one or more reference frames. One of the first or second parameterized motion models is selected along with an associated reference frame, such as based on a lowest prediction error. A motion vector is generated between the current block and the selected reference frame by warping pixels of the current block to a warped patch of the selected reference frame according to the selected parameterized motion model. A prediction block is generated using the motion vector, and the current block is encoded or decoded using the prediction block.

Growing community of inventors

San Francisco, CA, United States of America

Sarah Parker