Deep learning based adaptive arithmetic coding and codelength regularization
AssigneeWaveOne Inc.
InventorsOren Rippel, Lubomir Bourdev
Life of this patent
Estimating the term…
Abstract
A deep learning based compression (DLBC) system applies trained models to compress binary code of an input image to a target codelength. For a set of binary codes representing the quantized coefficents of an input image, the DLBC system applies a first model that is trained to predict feature probabilities based on the context of each bit of the binary codes. The DLBC system compresses the binary code via adaptive arithmetic coding based on the determined probability of each bit. The compressed binary code represents a balance between a reconstruction quality of a reconstruction of the input image and a target compression ratio of the compressed binary code.
Claims16
A computer-implemented method comprising:
- receiving, from a first trained model, quantized coefficients of a representation of an input image, wherein the first trained model is iteratively trained on a plurality of training data, the training data comprising quantized coefficients of a corpus of training images, wherein one or more of the training iterations of the first trained model comprises:
- receiving a training image,
- extracting quantized coefficients from the training image based on a current set of training parameters of the first trained model,
- computing a penalty for one or more extracted quantized coefficients, and
- adjusting the current set of training parameters to reduce the computed penalties for the extracted quantized coefficients;
- converting the received quantized coefficients to one or more processing units, each processing unit having one or more bits;
- for each processing unit, computing a probability of its value based on a context of the processing unit and one or more pre-computed feature probabilities received from a second trained model;
- and
- generating compressed binary codes of the input image using entropy coding based on the computed probability of the value of each processing unit.
A non-transitory computer-readable storage medium including instructions that, when executed by a processor, cause the processor to perform steps including:
- receiving, from a first trained model, quantized coefficients of a representation of an input image, wherein the first trained model is iteratively trained on a plurality of training data, the training data comprising quantized coefficients of a corpus of training images, wherein one or more of the training iterations of the first trained model comprises:
- receiving a training image,
- extracting quantized coefficients from the training image based on a current set of training parameters of the first trained model,
- computing a penalty for one or more extracted quantized coefficients, and
- adjusting the current set of training parameters to reduce the computed penalties for the extracted quantized coefficients:
- converting the received quantized coefficients to one or more processing units, each processing unit having one or more bits;
- for each processing unit, computing a probability of its value based on a context of the processing unit and one or more pre-computed feature probabilities received from a second trained model;
- and
- generating compressed binary codes of the input image using entropy coding based on the computed probability of the value of each processing unit.
Dependencies are inferred from the claim text.