Back to results
US 12,720,055 B2Granted
Systems and methods for encoding a deep neural network
AssigneeInterDigital Madison Patent Holdings, SAS
InventorsFabien Racape, Shahab Hamidi-Rad, Swayambhoo Jain
Life of this patent
Estimating the term…
Abstract
The disclosure relates to a method for compression including codebook-based quantization of a data set and corresponding decompression method, signal; bitstream, and encoder and/or decoder device.
Claims20
A method for decoding a bitstream comprising compressed video data, the method comprising:
- receiving the bitstream comprising a flag, wherein the flag is configured to indicate codebook-based quantization or uniform quantization, and wherein the bitstream comprises a tensor of a neural network;
- and
- if the flag indicates a first value, determining that the flag indicates that codebook-based quantization was used to encode the bitstream, receiving an indication of a codebook size, and decoding the tensor from the bitstream using inverse quantization in accordance with the codebook size;
- or
- if the flag indicates a second value, determining that the flag indicates that uniform quantization was used to encode the bitstream and decoding the tensor from the bitstream using uniform quantization.
2 more elements
A decoder for decoding a bitstream comprising compressed video data, the decoder comprising:
- a processor and memory configured to:
- receive the bitstream comprising a flag, wherein the flag is configured to indicate codebook-based quantization or uniform quantization, and wherein the bitstream comprises a tensor of a neural network;
- if the flag indicates a first value, determine that the flag indicates that codebook-based quantization was used to encode the bitstream, receive an indication of a codebook size, and decode the tensor from the bitstream using inverse quantization in accordance with the codebook size;
- and
- if the flag indicates a second value, determine that the flag indicates that uniform quantization was used to decode the bitstream and decode the tensor from the bitstream using uniform quantization.
2 more elements
A method for encoding a bitstream, the method comprising:
- receiving video data, the video data comprising a tensor of a neural network;
- encoding the video data according to a Context-based Adaptive Binary Adaptive Coding (CABAC) method to generate the bitstream, wherein the bitstream comprises a flag, wherein the flag is configured to indicate codebook-based quantization or uniform quantization to encode the bitstream;
- wherein encoding the video data according the CABAC method comprises encoding the video data using codebook-based quantization in accordance with a codebook size if the flag indicates a first value or wherein encoding the video data according the CABAC method comprises encoding the video data using uniform quantization if the flag indicates a second value;
- and
- transmitting the encoded bitstream.
2 more elements
Dependencies are inferred from the claim text.