Method for detecting objects in digital images

1. A method for detecting objects in a digital image, comprising the steps of:

a) generating a first segmentation map of the digital image according to a non-object specific criterion;

b) generating a second segmentation map of the digital image according to a object specific criterion; and

c) detecting objects in the digital image using both the first and second segmentation maps.

2. The method claimed in claim 1 , further comprising the step of merging the first and second segmentation maps prior to the detecting step and detecting the objects in the merged map using pattern matching.

3. The method claimed in claim 1 , further comprising the step of detecting objects using pattern matching in the first and second segmentation maps respectively and merging the detected objects.

4. The method claimed in claim 1 , wherein the non-object specific criterion is a color homogeneity criterion and the object specific criterion is an object specific color similarity.

5. The method claimed in claim 4 for detecting skin regions, wherein at least one of the color homogeneity criterion and the object specific color similarity relates to a skin color, and the detected objects are detected skin color regions.

6. The method claimed in claim 5 further comprising the step of detecting red-eye in the detected skin color regions.

7. The method claimed in claim 6 , wherein the step of detecting red eye comprises the steps of:

a) identifying candidate redeye defects in skin colored regions having a characteristic of a face and having groups of pixels with a color characteristic of a redeye defect; and

(b) evaluating the candidate redeye defects relative to one or more properties of a face to determine whether a pair of the candidate redeye defects is consistent with a size expected of an eye based on a distance between the pair of candidate redeye defects.

8. The method claimed in claim 7 , further comprising the step of correcting color of pixels based on a location of the redeye defects.

9. The method claimed in claim 1 , wherein the step of generating a first segmentation map comprises segmenting the digital image into continuous regions of homogeneous color and assigning a score indicating a probability that a region corresponds to objects being detected.

10. The method claimed in claim 1 , wherein the step of generating a second segmentation map comprises the steps of:

a) detecting object pixels in the digital image by assigning an object belief value to each pixel;

b) applying a threshold to the belief values to identify object pixels;

c) extracting continuous regions of object pixels;

d) growing the extracted continuous regions according to a continuity criterion for belief values; and

e) decomposing the grown regions into parts of simple shapes to generate the second segmentation map.

11. The method claimed in claim 10 wherein the step b) of applying a threshold to the belief values to identify object pixels comprises the steps of:

a) forming a probability distribution of at least some of the belief values of the pixels in the digital image;

b) determining an adaptive threshold from the probability distribution; and

c) applying the adaptive threshold to the belief values.

12. The method claimed in claim 11 wherein the adaptive threshold is determined from a probability distribution of belief values for pixels that are on or around a boundary between one or more objects and a background in the image.

13. The method claimed in claim 10 wherein the step b) of applying a threshold to the belief values to identify object pixels comprises the steps of:

a) forming a probability map of the belief values of the pixels in the digital image;

b) determining a gradient of the probability map for each pixel;

c) constructing a histogram using only high-gradient pixels;

d) determining the adaptive threshold from one or more valleys located in the histogram; and

e) applying the adaptive threshold to the belief values.

14. The method claimed in claim 10 wherein the step e) of decomposing the grown regions comprises the steps of:

a) following a contour of each region;

b) smoothing the contour to remove noisy structures;

c) finding critical points of high curvature along the smoothed contour;

d) identifying negative curvature critical points and, for one or more of the negative curvature critical points, choosing another contour point to form a cut; and

e) splitting regions at the cuts to produce the simple shapes.

15. A computer storage medium having instructions stored therein for causing a computer to perform the method of claim 1 .

16. A system for detecting objects in a digital image, said system comprising:

a) a first stage for generating a first segmentation map of the digital image according to a non-object specific criterion;

b) a second stage for generating a second segmentation map of the digital image according to a object specific criterion; and

c) a detector for detecting objects in the digital image using both the first and second segmentation maps.

17. The system claimed in claim 16 , further comprising a stage for merging the first and second segmentation maps, wherein the detector detects the objects in the merged map using pattern matching.

18. The system claimed in claim 16 , wherein the non-object specific criterion is a color homogeneity criterion and the object specific criterion is an object specific color similarity.

19. The system claimed in claim 18 wherein at least one of the color homogeneity criterion and the object specific color similarity relates to a skin color, and the detected objects are detected skin color regions.

20. The system claimed in claim 19 wherein the detector further detects red-eye in the detected skin color regions.

21. The system claimed in claim 20 further comprising a correction stage for correcting color of pixels based on a location of the redeye defects.

22. The system claimed in claim 16 wherein the second stage includes an algorithm for:

a) detecting object pixels in the digital image by assigning an object belief value to each pixel;

b) applying a threshold to the belief values to identify object pixels;

c) extracting continuous regions of object pixels;

d) growing the extracted continuous regions according to a continuity criterion for belief values; and

e) decomposing the grown regions into parts of simple shapes to generate the second segmentation map.

23. The system claimed in claim 22 wherein applying a threshold to the belief values to identify object pixels comprises:

a) forming a probability distribution of at least some of the belief values of the pixels in the digital image;

b) determining an adaptive threshold from the probability distribution; and

c) applying the adaptive threshold to the belief values.

24. The system claimed in claim 23 wherein the adaptive threshold is determined from a probability distribution of belief values for pixels that are on or around a boundary between one or more objects and a background in the image.

25. The system claimed in claim 22 wherein applying a threshold to the belief values to identify object pixels comprises:

a) forming a probability map of the belief values of the pixels in the digital image;

b) determining a gradient of the probability map for each pixel;

c) constructing a histogram using only high-gradient pixels;

d) determining the adaptive threshold from one or more valleys located in the histogram; and

e) applying the adaptive threshold to the belief values.

26. The system claimed in claim 22 wherein decomposing the grown regions comprises:

a) following a contour of each region;

b) smoothing the contour to remove noisy structures;

c) finding critical points of high curvature along the smoothed contour;

d) identifying negative curvature critical points and, for one or more of the negative curvature critical points, choosing another contour point to form a cut; and

e) splitting regions at the cuts to produce the simple shapes.