Google Inc.

Elliott Karpilovsky

Debargha Mukherjee

Lester Lu

Yue Chen

Rohit Shankar Singh

Sarah Parker

Urvang Joshi

David Charles Minnen

Sung Jin Hwang

Methods, systems, and apparatus, including computer programs, for compression and decompression of video data using an ensemble of machine learning models. Methods can include defining for each frame in a video, a plurality of blocks in the frame. Methods can further include processing the frames of video in sequential sets, wherein each sequential set is at least a current frame () of video and a prior frame () of video in the ordered sequence. Each respective prediction of a block in the frame of the video includes providing, as input to a prediction model a first and the second border () of a current block () of the current frame, a first and a second border () for a respective current block () of the prior frame and the respective current block () of the prior frame.

Video inter/intra compression using mixture of experts

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

Residual coding using vector quantization (VQ) is described. A flag indicating whether a residual block for the current block is encoded using VQ. In response to the flag indicating that the residual block is encoded using VQ, a parameter indicating an entry in a codebook is decoded, and the residual block is decoded using the entry. In response to the flag indicating that the residual block is not encoded using VQ, the residual block is decoded based on a skip flag indicating whether the current block is encoded using transform skip. The current block is reconstructed using the residual block.

Vector quantization for prediction residual coding

Transform modes are derived for inter-predicted blocks using side information. A prediction residual is generated for a current video block using a reference frame. Side information associated with one or both of the current video block or the reference frame is identified. A trained transform is determined from amongst multiple trained transforms based on the side information, in which each of the trained transforms is determined using individual side information types and combinations of the individual side information types and the side information represents values of one of the individual side information types or one of the combinations of the individual side information types. The prediction residual is transformed according to the trained transform, and data associated with the transformed prediction residual and the side information are encoded to a bitstream.

Inter-prediction mode-dependent transforms for video coding

Decoding video data includes, for a block encoded using a prediction mode, determining a transform mode for the block using the prediction mode. The transform mode is a first mode when the prediction mode is an inter-prediction mode and is a second mode when the prediction mode is an intra-prediction mode. The first mode is an available first transform type that is a combination of transforms selected from first fixed transforms and first learned transforms that each comprise a respective transformation matrix generated iteratively using blocks predicted using the inter-prediction mode. The second mode is an available second transform type that is a combination of transforms selected from second fixed transforms, which is a proper subset of the first fixed transforms, and a second learned transform comprising a transformation matrix that is generated iteratively using blocks predicted using the intra-prediction mode. Decoding the block uses the prediction and transform modes.

Combination of mode-dependent and fixed transform types in video coding

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

Coding a block of video data includes determining a prediction mode for the block, which is an inter-prediction or intra-prediction mode, determining a transform type for the block, and coding the block using the prediction mode and the transform type. The transform type is one of a first plurality of transform types when the prediction mode is the inter-prediction mode, and is one of a second plurality of transform types when the prediction mode is the intra-prediction mode. The first plurality of transform types includes first fixed transform types and first mode-dependent transform types that are based on a first learned transform generated using inter-predicted blocks. The second plurality of transform types includes second fixed transform types and second mode-dependent transform types that are based on a second learned transform generated using intra-predicted blocks. The first and second fixed transform types have at least some fixed transform types in common.

Transform modes are derived for inter-predicted blocks using side information available within a bitstream. An inter-predicted encoded video block and side information are identified within a bitstream. Based on the side information, a trained transform is determined for inverse transforming transform coefficients of the inter-predicted encoded video block from amongst multiple trained transforms. The transform coefficients of the inter-predicted encoded video block are inverse transformed according to the trained transform to produce a prediction residual. A video block is reconstructed using the prediction residual and the reference frame. The video block is then output within an output video stream for storage or display. To determine the trained transforms, a learning model uses individual side information types and combinations of the individual side information types processed against a training data set.

Growing community of inventors

Santa Clara, CA, United States of America

Elliott Karpilovsky