Visual Perception Model For Hi-Fidelity Image Watermarking

• US 20080069396A1
• Filed: 09/14/2006
• Published: 03/20/2008
• Est. Priority Date: 09/14/2006
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Kirovski, Darko, He, Shan
• Granted Patent: US 7,720,253 B2
• Time in Patent Office: Days
• US Class Current: 382/100
• CPC Class Codes: G06T 1/0028 Adaptive watermarking, e.g. Human Visual System [HVS]-based watermarking
• CPC Class Codes: G06T 2201/0202 whereby the quality of watermarked images is measured; Measuring quality or performance of watermarking methods; Balancing between quality and robustness

Abstract

Exemplary systems and methods implement an exemplary visual perception model. This model can be implemented in such applications as imaging, compression, denoising, annotation of hi-fidelity images, etc. In one implementation a perceptual filter estimates an entropy statistic and a standard deviation statistic associated with each pixel of an image, then mixes the entropy statistic and the standard deviation statistic to map an estimation of human perceptual tolerance to noise or alteration in parts of the image. In one application, metadata may be added to an image via a system of synergistic watermarks, in which the watermarks strengthen each other and are placed according to the noise tolerance of the pixels in order to maintain hi-fidelity of the image.

Claims

1. A method, comprising:

scanning an image pixel by pixel;

for each pixel being scanned, evaluating the pixel with a perceptual filter, including;

computing a differential standard deviation statistic and an entropy statistic for a group of pixels centered on the pixel; and

mixing the differential standard deviation statistic and the entropy statistic to model a noise tolerance associated with the pixel.

2. The method as recited in claim 1, wherein the noise tolerance denotes a degree of the change in the pixel attribute that can be achieved without being perceived by the human visual system when the pixel attribute is changed for one or more of the pixels.

3. The method as recited in claim 2, wherein the pixel attribute is luminance.

4. The method as recited in claim 1, further comprising creating a noise tolerance map for the image from the noise tolerances of each pixel and using the noise tolerance map to perform denoising, medical imaging, or image compression.

5. The method as recited in claim 1, further comprising embedding data in the image by changing the pixel attribute for the pixels that possess a value of the noise tolerance above a threshold.

6. The method as recited in claim 5, wherein the data comprise a spread spectrum watermark or the data comprise bits of metadata for annotating the image.

7. The method as recited in claim 5, wherein the threshold is based on a human visual perception of an amount of the noise that can be added to an image without degrading a perceived fidelity of the image.

8. The method as recited in claim 1, further comprising embedding a spread spectrum watermark in the image to signal the presence of a metadata watermark and to provide registration or alignment of the image for blind watermark detection that does not have access to the original image;

wherein the data of the spread spectrum watermark is placed in the image based on the noise tolerance of the pixels.

9. The method as recited in claim 8, wherein the embedding further comprising creating the spread spectrum watermark in a block of the image and then tiling the same block in the image via a lapped biorthogonal transform.

10. The method as recited in claim 8, further comprising embedding the metadata watermark into the image based on the noise tolerance of the pixels by:

partitioning each block of the spread spectrum watermark into smaller blocks;

embedding each bit of the metadata by quantizing first order statistics of the coefficients of the smaller blocks, wherein the quantizing uses an even number of a quantization step size if the bit is a 0 bit and uses an odd number of the quantization step size if the bit is a 1 bit;

allocating changes to the attribute of the pixels in each smaller block such that the change is proportional to the noise tolerance of the pixel.

11. The method as recited in claim 8, wherein embedding the metadata watermark strengthens a robustness of the spread spectrum watermark to image compression and image cropping while maintaining a fidelity of the image.

12. A system, comprising:

a pixel-wise noise tolerance estimator for evaluating sites within an image for tolerance to alteration of the pixels of the image;

a perceptual filter in the pixel-wise noise tolerance estimator, including;

an entropy filter to compute an entropy statistic for a group of pixels centered on the pixel and including the pixel;

a standard deviation filter to compute a differential standard deviation statistic for the group of pixels;

a mixer for combining the entropy statistic and the differential standard deviation statistic to model a noise tolerance associated with the pixel;

wherein the noise tolerance denotes a degree of change in a pixel attribute that can be achieved without being perceived by the human visual system when the pixel attribute is changed for one or more of the pixels.

13. The system as recited in claim 12, further comprising a noise tolerance map of the image for deciding which sites in the image can tolerate alteration of an attribute of the pixels in the sites without being significantly perceived by a human visual system, wherein the alteration of the attribute of the pixels is to be performed for adding data to the image, for denoising, for medical imaging, or for image compression.

14. The system as recited in claim 13, further comprising an annotation engine, including:

a spread spectrum watermarker to embed a first watermark signaling the presence of a second watermark, wherein the first watermark comprises a spread spectrum watermark tiled over the image via a lapped biorthogonal transform; and

a metadata watermarker to embed the second watermark for storing metadata in the image, including;

a regional statistic quantifier to partition each block of the spread spectrum watermark into smaller blocks and embed each bit of the metadata by quantizing first order statistics of the coefficients of the smaller blocks, wherein the quantizing uses an even number of a quantization step size if the bit is a 0 bit and uses an odd number of the quantization step size if the bit is a 1 bit.

15. The system as recited in claim 14, further comprising an embedding site discriminator to allocate changes to the attribute of the pixels in each smaller block such that the change is proportional to the noise tolerance of the pixel.

16. The system as recited in claim 15, wherein the embedding site discriminator embeds the metadata into the image by changing the pixel attribute for the pixels that possess a value of the noise tolerance above a threshold.

17. The system as recited in claim 16, wherein the threshold is based on a human visual perception of an amount of the noise that can be added to an image without degrading a perceived fidelity of the image.

18. The system as recited in method as recited in claim 14, further comprising a multi-watermark enhancer to embed the second watermark such that the first watermark is strengthened in robustness to image compression and image cropping while maintaining hi-fidelity of the image.

19. A system, comprising:

means for visual modeling, including;

means for estimating an entropy statistic associated with each pixel of an image;

means for estimating a standard deviation statistic associated with each pixel of the image; and

means for mixing the entropy statistic and the standard deviation statistic to create an image-wide estimation map of human perceptual tolerance to alteration of the image in each part of the image.

20. The system as recited in claim 19, further comprising one of:

means for applying image compression by determining quantization noise based on the image-wide estimation map;

means for applying error correction by using less resilient coding in image areas that tolerate more noise based on the image-wide estimation map;

means for medical imaging based on the image-wide estimation map;

or

means for denoising the image based on the image-wide estimation map.