Methods and apparatus for model-based visual descriptors compression

Patent metadata

Field	Value
Publication number	US-10021395-B2
Application number	US-201514953124-A
Country	US
Kind code	B2
Filing date	Nov 27, 2015
Priority date	Nov 27, 2014
Publication date	Jul 10, 2018
Grant date	Jul 10, 2018

Abstract

A particular implementation determines parameters of a generative probabilistic model from visual descriptors extracted from at least one image. The extracted visual descriptors are quantized and encoded using the model-based arithmetic encoding to be stored or for transmission to a decoder. The model parameters are also stored to be available to a decoder, or transmitted directly to a decoder. A decoder uses the stored, or received, model parameters to reconstruct the generative probabilistic model and then to decode the visual descriptors. The visual descriptors are used for image analysis tasks, such as image retrieval or object detection. A particular implementation uses a Gaussian mixture model as a generative probabilistic model.

First claim

The invention claimed is:

1. A method of compressing visual descriptors from at least one image by exploiting redundancy of natural image descriptors, comprising:
extracting the visual descriptors from at least one image, said visual descriptors describing key points in images;
creating model parameters of a generative probabilistic model from the extracted visual descriptors in a maximum likelihood sense;
quantizing and encoding said model parameters;
quantizing said extracted visual descriptors; and,
applying a model-based arithmetic encoding to said quantized extracted visual descriptors using said encoded model parameters exploiting redundancy of the visual descriptors within the at least one image for compression of the visual descriptors.

2. The method of claim 1 , comprising storing at least one of said encoded model parameters and said encoded visual descriptors.

3. The method of claim 1 , comprising transmitting at least one of said encoded model parameters and said encoded visual descriptors to a decoder.

4. The method of claim 1 , said encoding of said quantized extracted visual descriptors comprising:
associating each said visual descriptor with a corresponding Gaussian mixture model component for which the likelihood of said visual descriptor is maximum;
rearranging said visual descriptors by order of Gaussian mixture model component indices so that the visual descriptors are non-decreasing;
encoding Gaussian mixture model component indices using a predictive entropy coding scheme; and
encoding each said visual descriptor using a multivariate Gaussian-based arithmetic coding.

5. The method of claim 1 , wherein the generative probabilistic model is a Gaussian mixture model.

6. An apparatus for compressing visual descriptors from at least one image by exploiting redundancy of natural image descriptors, comprising:
a receiver of visual descriptors extracted from at least one image, said visual descriptors describing key points in images;
a processor, configured to determine model parameters of a generative probabilistic model from the extracted visual descriptors in a maximum likelihood sense;
a quantizer of said model parameters;
an encoder of said quantized model parameters;
a quantizer of said extracted visual descriptors; and
an encoder for encoding said quantized extracted visual descriptors using said model parameters by applying a model based arithmetic encoding to exploit redundancy of the visual descriptors within the at least one image for compression of the visual descriptors.

7. The apparatus of claim 6 , wherein at least one output of the model parameters and encoded visual descriptors encoders are stored.

8. The apparatus of claim 6 , wherein at least one output of the model parameters and encoded visual descriptors encoders are transmitted to a decoder.

9. The apparatus of claim 6 , said encoder of said quantized extracted visual descriptors comprising:
a processor that associates each visual descriptor with a corresponding Gaussian mixture model component for which the likelihood of said visual descriptor is maximum;
a second processor that rearranges visual descriptors by order of Gaussian mixture model component indices so that the visual descriptors are non-decreasing;
an encoder of Gaussian mixture model component indices that uses a predictive entropy coding scheme; and
an encoder of each visual descriptor using a multivariate Gaussian-based arithmetic coding.

10. The apparatus of claim 6 , wherein the generative probabilistic model is a Gaussian mixture model.

11. A method of decoding compressed encoded visual descriptors from at least one image, comprising:
receiving a bit stream comprising quantized generative probabilistic model parameters determined from visual descriptors extracted from the at least one image in a maximum likelihood sense;
reconstructing a generative probabilistic model using said quantized generative probabilistic model parameters;
receiving a bit stream comprising the compressed encoded visual descriptors, the compressed encoded visual descriptors describing key points in the at least one image;
decoding said compressed encoded visual descriptors using said reconstructed generative probabilistic model thereby exploiting redundancy of the compressed encoded visual descriptors within the at least one image; and,
performing an image analysis using said decoded visual descriptors.

12. The method of claim 11 , wherein the generative probabilistic model is a Gaussian mixture model.

13. An apparatus for decoding compressed encoded visual descriptors from at least one image, comprising:
a receiver of a bit stream comprising quantized generative probabilistic model parameters determined from visual descriptors extracted from the at least one image in a maximum likelihood sense;
a processor to reconstruct a generative probabilistic model using said quantized generative probabilistic model parameters;
a receiver of a bit stream comprising the compressed encoded visual descriptors, the compressed encoded visual descriptors describing key points in the at least one image;
a processor to decode the compressed encoded visual descriptors using said reconstructed generative probabilistic model thereby exploiting redundancy of the compressed encoded visual descriptors within the at least one image; and,
a processor to perform image analysis using said decoded visual descriptors.

14. The apparatus of claim 13 , wherein the generative probabilistic model is a Gaussian mixture model.

Assignees

• Thomson Licensing

Inventors

• Ozerov Alexey
• Vigouroux Jean-Ronan
• Lefebvre Frederic

Classifications

• H04N19/50

  using predictive coding (H04N19/61 takes precedence) · CPC title

• H04N19/91

  Entropy coding, e.g. variable length coding [VLC] or arithmetic coding · CPC title

• H04N19/94 Primary

  Vector quantisation · CPC title

• H04N19/44

  Decoders specially adapted therefor, e.g. video decoders which are asymmetric with respect to the encoder · CPC title

• H04N19/136 Primary

  Incoming video signal characteristics or properties · CPC title