IMPORTANCE FILTERING FOR IMAGE RETARGETING

• US 20120121204A1
• Filed: 05/02/2011
• Published: 05/17/2012
• Est. Priority Date: 11/11/2010
• Status: Active Grant

• Current Assignee: Seiko Epson Corporation (Seiko Group)
• Original Assignee: Seiko Epson Corporation (Seiko Group)
• Inventors: Xiao, Jing, Ding, Yuanyuan
• Granted Patent: US 8,494,302 B2
• Time in Patent Office: Days
• US Class Current: 382/260
• CPC Class Codes: G06T 3/04 Context-preserving transformations, e.g. by using an importance map panospheric to cylindrical image transformations G06T3/12
• CPC Class Codes: G06V 10/462 Salient features, e.g. scale invariant feature transforms [SIFT]

Abstract

A content-aware image retargeting technique uses an “importance filtering” technique to preserve important information in the resizing of an image. The image saliency is first filtered, guided by the image itself to achieve a structure-consistent importance map. The pixel importance is then used as the key constraint in computing the gradient map of pixel shifts from the original resolution to the target resolution. Finally the shift gradient is integrated across the image by a weighted filtering process to construct a smooth pixel shift-map and render the target image. The weight is again controlled by the pixel importance. The two filtering processes enforce the maintaining of structural consistency while preserving the important contents in the target image. The simple nature of the present filter operations allow for real-time applications and easy extension to video retargeting, as the structural constraints from the original image naturally convey the temporal coherence between frames.

Claims

1. A method of retargeting a digital original image to a digital target image, said method comprising:

(a) accessing said original image, wherein the size of at least one dimension of said original image is different than the size of the corresponding one dimension in said target image, and a scaling factor relates the size of said one dimension of said original image to the size of said corresponding one dimension in said target image;

(b) constructing an image saliency of said original image;

(c) filtering the image saliency under the guidance of said original image to produce an importance map, S^Imp, that matches pixels to salient structures in said original image;

(d) estimating a shift-map gradient G according to said scaling factor and under the constraint of said importance map S^Imp, wherein pixels that are matched to salient structures are assigned a consistently distributed gradient;

(e) defining a shift-map M from said shift-map gradient G, wherein said shift-map M estimates pixel shifts from said original image to said target image; and

(f) applying said shift-map M to said original image to construct said target image.

2. The method of claim 1, wherein in step (b), said image saliency defines a saliency measure for each pixels of said original image, said saliency measure being computed from local image features within said original image according to the saliency of said features.

3. The method of claim 1, wherein in step (c), said image saliency is filtered under the guidance of the whole of said original image by using a guided filter.

4. The method of claim 1, wherein in step (c), a face detector is incorporated into the filtering of the image saliency to identify the pixels matched to salient structures, and salient structures are determined from saliency measures in said image saliency.

5. The method of claim 1, wherein in step (c):

said importance map S^Imp defines salient structures, said salient structures being important structures bet preserved in said target image; and

for each defined salient structure, the pixels match to the defined salient structure constitute a group of matched pixels, and the matched pixels within each group are assigned smoothly distributed salient measures.

6. The method of claim 5, wherein said matched pixels within each group are assigned substantially similar salient measures.

7. The method of claim 1, wherein in step (d), said shift-map gradient G defines each pixel'"'"'s shift relative to its neighboring pixels.

8. The method of claim 1, wherein in step (d), said consistently distributed gradient is substantially the same gradient value indicative of a minimal shift relative to neighboring pixels.

9. The method of claim 1, wherein in step (d), said consistently distributed gradient is nonlinearly related to said scaling factor.

10. The method of claim 9, wherein said consistently distributed gradient is lower than a gradient linearly related to said scaling factor.

11. The method of claim 1, wherein in step (d), said consistently distributed gradient is effective for maintaining distortion of said salient structure lower than distortion of unimportant regions of said original image, as defined by said importance map S^Imp.

12. The method of claim 1, wherein in step (d):

said importance map S^Imp is normalized so that S^Imp(x, y)ε

[0,1];

said shift-map gradient G renders bigger gradients for pixels whose importance value as defined by said normalized importance map S^Imp approach 0, and renders smaller gradients for pixels whose importance value as defined by said normalized importance map S^Imp approach 1.

13. The method of claim 12, wherein said shift-map gradient G is non-linear and gradient values drop faster as importance values approach 1.

14. The method of claim 13, wherein gradient values drops faster when importance values are above 0.75 than when gradient values are not above 0.75.

15. The method of claim 12, wherein the shift-map gradient G is defined as:

16. The method of claim 12, wherein the shift-map gradient G is defined as:

17. The method of claim 12, wherein the shift-map gradient G is defined as:

18. The method of claim 12, wherein the shift-map gradient G is defined as:

19. The method of claim 1, wherein in step (e), the defining of said shift-map M from said shift-map gradient G, includes integrating the shift-map gradient G by a weighted filtering process, wherein pixels matched to salient structures are weighted higher than pixels not matched to salient structures.

20. The method of claim 1, wherein shift-map M is defined by the following optimization:

M=M_opt=min_M′

|∇

_xM′

−

G|

s.t.M′

(0,y)=0,M′

(X,y)=|X′

−

X|

where M′

is a pixel shift that when optimized renders the final shift-map M, X is dimensional coordinate of a pixel in said original image along said one dimension of the original image, X′

is dimensional coordinate of the corresponding pixel in said target image along the corresponding one dimension in said target image, and x and y are coordinate in a two dimensional system including said one dimension of the original image.

21. The method of claim 1, wherein shift-map M is defined by following relationship:

22. The method of claim 21, wherein weight w is defined as w(x, y)=e^SImp(x,y).