ULTRA-HIGH COMPRESSION OF IMAGES BASED ON DEEP LEARNING

• US 20160292589A1
• Filed: 04/03/2015
• Published: 10/06/2016
• Est. Priority Date: 04/03/2015
• Status: Active Grant

• Current Assignee: Mitre Corporation
• Original Assignee: Mitre Corporation
• Inventors: TAYLOR, Robert M. Jr., NECIOGLU, Burhan
• Granted Patent: US 11,221,990 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 16/1744 using compression, e.g. sparse files
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/047 Probabilistic or stochastic networks
• CPC Class Codes: G06T 9/002 using neural networks

Abstract

A system for machine learning model parameters for image compression, including partitioning image files into a first set of regions, determining a first set of machine learned model parameters based on the regions, the first set of machine learned model parameters representing a first level of patterns in the image files, constructing a representation of each of the regions based on the first set of machine learned model parameters, constructing representations of the image files by combining the representations of the regions in the first set of regions, partitioning the representations of the image files into a second set of regions, and determining a second set of machine learned model parameters based on the second set of regions, the second set of machine learned model parameters representing a second level of patterns in the image files.

Claims

1. A computer implemented method for machine learning model parameters for image compression, comprising:

partitioning a plurality of image files stored on a first computer memory into a first set of regions;

determining a first set of machine learned model parameters based on the first set of regions, the first set of machine learned model parameters representing a first level of patterns in the plurality of image files;

constructing a representation of each region in the first set of regions based on the first set of machine learned model parameters;

constructing representations of the plurality of image files by combining the representations of the regions in the first set of regions;

partitioning the representations of the plurality of image files into a second set of regions;

determining a second set of machine learned model parameters based on the second set of regions, the second set of machine learned model parameters representing a second level of patterns in the plurality of image files; and

storing the first set of machine learned model parameters and the second set of machine learned model parameters on one or more computer memories.

2. The computer implemented method of claim 1, wherein the first set of machine learned model parameters and the second set of machine learned model parameters are used to compress and decompress a digital image file stored on a second computer memory.

3. The computer implemented method of claim 1, wherein:

each of the image files comprises pixel values for rendering an image on a display and each of the regions in the first set of regions comprises a subset of the pixel values for rendering a corresponding region of the image on the display.

4. The computer implemented method of claim 1, wherein:

each region in the first set of regions overlaps at least one other region in the first set of regions; and

each region in the second set of regions overlaps at least one other region in the second set of regions.

5. The computer implemented method of claim 4, wherein each of the first set of regions comprises an M by M array of pixel values, wherein M is an integer.

6. The computer implemented method of claim 5, wherein each of the second set of regions comprises an N by N array of pixel values, wherein N is greater than M.

7. The computer implemented method of claim 1, wherein:

the first set of machine learned model parameters and the second set of machine learned model parameters are determined based on a statistical method for estimating a structure underlying a set of data.

8. The computer implemented method of claim 1, wherein:

determining a first set of machine learned model parameters comprises machine learning a first union of subspaces that estimates each of the regions in the first set of regions, wherein the first set of machine learned model parameters includes basis vectors and offsets for subspaces in the first union of subspaces; and

determining a second set of machine learned model parameters comprises machine learning a second union of subspaces that estimates each of the regions in the second set of region, wherein the second set of machine learned model parameters includes basis vectors and offsets for subspaces in the second union of subspaces.

9. The computer implemented method of claim 8, wherein:

the first union of subspaces and the second union of subspaces are machine learned based on a Mixture of Factor Analyzers.

10. The computer implemented method of claim 8, wherein:

a quantity and a dimension of subspaces in the first union of subspaces are determined based on the plurality of image files; and

a quantity and a dimension of subspaces in the second union of subspaces is determined based on the representations of the plurality of image files.

11. A computer implemented method for compressing an image file, comprising:

partitioning an image file into a first set of regions;

constructing a representation of each region in the first set of regions based on a first set of machine learned model parameters, the first set of machine learned model parameters representing a first level of image patterns;

constructing a first representation of the image file by combining the representations of the regions in the first set of regions;

partitioning the first representation of the image file into a second set of regions;

constructing a representation of each region in the second set of regions based on a second set of machine learned model parameters, the second set of machine learned model parameters representing a second level of image patterns;

constructing a second representation of the image file by combining the representations of the regions in the second set of regions, wherein the second representation comprises coefficients of the machine learned model parameters in the second set of machine learned model parameters;

selecting a plurality of elements from a predetermined list of elements to represent the model parameter coefficients; and

storing a plurality of indices to the plurality of elements in a memory.

12. The computer implemented method of claim 11, wherein

each region in the first set of regions overlaps at least one other region in the first set of regions;

and

each region in the second set of regions overlaps at least one other region in the second set of regions.

13. The computer implemented method of claim 11, wherein the first set of machine learned model parameters and the second set of machine learned model parameters are based on a statistical model for estimating a structure underlying a set of training data.

14. The computer implemented method of claim 12, wherein the statistical model comprises a Mixture of Factor Analyzers.

15. The computer implemented method of claim 11, wherein the first representation comprises coefficients of the machine learned model parameters in the first set of machine learned model parameters;

the method further comprising;

selecting a second plurality of elements from a second predetermined list of elements to represent the coefficients of the machine learned model parameters in the first set of machine learned model parameters; and

storing a second plurality of indices to the second plurality of elements in the memory.

16. A computer implemented method for decompressing an image file comprising:

opening a compressed image file comprising a plurality of indices to a plurality of elements from a predetermined list of elements, wherein the predetermined list of elements is based on a first set of machine learned model parameters representing a first level of image patterns and a second set of machine learned model parameters representing a second level of image patterns;

retrieving the plurality of elements from the predetermined list of elements based on the plurality of indices;

constructing a plurality of regions of an image file by combining the plurality of elements with the second set of machine learned model parameters; and

blending the plurality of regions to generate a decompressed image file.

17. The computer implemented method of claim 16, wherein the first set of machine learned model parameters and the second set of machine learned model parameters are based on a statistical model for estimating a structure underlying a set of training data.

18. The computer implemented method of claim 16, wherein the statistical model comprises a Mixture of Factor Analyzers.

19. A computer implemented method for decompressing an image file comprising:

opening a compressed image file comprising;

a first plurality of indices to a first plurality of elements from a first predetermined list of elements, wherein the first predetermined list of elements is based on a first set of machine learned model parameters representing a first level of image patterns, and

a second plurality of indices to a second plurality of elements from a second predetermined list of elements, wherein the second predetermined list of elements is based on a second set of machine learned model parameters representing a second level of image patterns;

retrieving the first plurality of elements from the first predetermined list of elements based on the first plurality of indices;

constructing a first plurality of regions of an image file by combining the first plurality of elements with the first set of machine learned model parameters;

retrieving the second plurality of elements from the second predetermined list of elements based on the second plurality of indices;

constructing a second plurality of regions of an image file by combining the second plurality of elements with the second set of machine learned model parameters and the first plurality of regions; and

blending the second plurality of regions to generate a decompressed image file.

20. A system for compressing an image file comprising:

a training module for learning model parameters comprising a processor and a memory and one or more applications stored in the memory that include instructions for;

partitioning a plurality of image files into a first set of regions;

determining a first set of machine learned model parameters based on the first set of regions, the first set of machine learned model parameters representing a first level of patterns in the plurality of image files;

constructing a representation of each region in the first set of regions based on the first set of machine learned model parameters;

constructing representations of the plurality of image files by combining the representations of the regions in the first set of regions;

partitioning the representations of the plurality of image files into a second set of regions; and

determining a second set of machine learned model parameters based on the second set of regions, the second set of machine learned model parameters representing a second level of patterns in the plurality of image files; and

a compression module configured to generate a compressed image file based on the first set of machine learned model parameters and the second set of machine learned model parameters.

21. The system of claim 20, comprising a decompression module configured to decompress the compressed image file based on the second set of machine learned model parameters.

22. A non-transitory computer readable storage medium storing one or more programs, the one or more programs comprising instructions, which when executed by an electronic device, cause the device to:

partition an image file into a first set of regions;

construct a representation of each region in the first set of regions based on a first set of machine learned model parameters, the first set of machine learned model parameters representing a first level of image patterns;

construct a first representation of the image file by combining the representations of the regions in the first set of regions;

partition the first representation of the image file into a second set of regions;

construct a representation of each region in the second set of regions based on a second set of machine learned model parameters, the second set of machine learned model parameters representing a second level of image patterns;

construct a second representation of the image file by combining the representations of the regions in the second set of regions, wherein the second representation comprises coefficients of the machine learned model parameters in the second set of machine learned model parameters;

select a plurality of elements from a predetermined list of elements to represent the model parameter coefficients; and

store a plurality of indices to the plurality of elements in a memory.