Perceptual segmentation of images

1. A perceptual-segmentation system, comprising:

a computer; and

one or more perceptual-segmentation programs, stored within the computer and executed by the computer, that

identify regions within an image;

for each identified region,

compute or access each of one or more pixel-feature values for pixels within region,

using the one or more pixel-feature values computed or accessed for pixels within the region, compute one or more region-feature values, and

using the one or more region-feature values, determine a region type for the region and associating the region type with the region; and

store the associations of region types with regions for use in subsequent image-processing tasks;

wherein the one or more perceptual-segmentation programs, stored within the computer and executed by the computer, identify regions within the image by:

for each of a number of specific object-identifying methods,

applying the object-identifying method to locate occurrences of objects of a specific type in the image; and

disambiguating any regions of overlap of two or more objects identified by the object-identifying methods.

2. The perceptual-segmentation system of claim 1 wherein pixel features associated with a pixel include one or more of:

one or more intrinsic pixel features, including

a grayscale value, and

three values that specify a color of the pixel according to a color model;

one or more numeric values that specify a spatial location of the pixel within an image;

a numeric value computed from pixel-feature values of the pixel and additional pixels that are located in a neighborhood of the pixel; and

a numeric value computed by comparing one or more intrinsic pixel-values with a model or distribution of pixel-feature values.

3. The perceptual-segmentation system of claim 1 wherein segment features associated with a segment or region include one or more of:

an average, median, or other mathematical value computed from pixel-feature values of pixels within the segment or region;

an area computed for the segment or region;

a relative area computed for the segment or region;

numeric values that specify a position of the segment or region within an image;

one or more numeric values that indicate an orientation of the region with an image;

a morphological characteristic computed for the segment or region;

a value computed by comparing a value computed from pixel-feature values of pixels within the segment or region to a model or distribution;

a numeric value computed form pixel-feature values of the pixel and additional pixels that are located in a neighborhood of the pixel;

a numeric value computed by comparing one or more intrinsic pixel-values with a model or distribution of pixel-feature values; and

a numeric value from segment-feature values of the segment or region and additional segments or regions that are located in a neighborhood of the segment or region.

4. A perceptual-segmentation system, comprising:

a computer; and

one or more perceptual-segmentation programs, stored within the computer and executed by the computer, that

identify regions within an image;

for each identified region,

compute or access each of one or more pixel-feature values for pixels within region,

using the one or more pixel-feature values computed or accessed for pixels within the region, compute one or more region-feature values, and

using the one or more region-feature values, determine a region type for the region and associating the region type with the region; and

store the associations of region types with regions for use in subsequent image-processing tasks;

wherein the one or more perceptual-segmentation programs, stored within the computer and executed by the computer, compute global statistics for the image;

wherein the one or more perceptual-segmentation programs, stored within the computer and executed by the computer, compute, for each segment or region of an image:

per-segment statistics,

a probability that the segment corresponds to a blue-sky region of the image, P(blue_sky),

a probability that the segment corresponds to a gray-sky region of the image, P(gray_sky), and

a probability that the segment corresponds to a snow region of the image, P(snow); and

wherein the one or more perceptual-segmentation programs, stored within the computer and executed by the computer, identify and tag segments corresponding to regions of sky in the image using the computed probabilities P(blue_sky), P(gray_sky), and P(snow).

5. The perceptual-segmentation system of claim 4

wherein the probability P(blue_sky) computed for a segment is related to a location of the segment within the image, colors within the segment, a size of the segment, a texture computed for the segment, an intensity computed for the segment, and a correspondence of color changes along two orthogonal traces within the segment to expected color changes for blue sky;

wherein the probability P(gray_sky) computed for a segment is related to a location of the segment within the image, colors within the segment, a size of the segment, a texture computed for the segment, and an intensity computed for the segment; and

wherein the probability P(snow) computed for a segment is related to a location of the segment within the image, colors within the segment, a size of the segment, a texture computed for the segment, and an intensity computed for the segment.

6. The perceptual-segmentation system of claim 5 wherein the one or more perceptual-segmentation programs, stored within the computer and executed by the computer, identify and tag segments corresponding to regions of sky in the image using the computed probabilities P(blue_sky), P(gray_sky), and P(snow) by:

for each segment in the image,

initially tagging the segment as “blue sky” when P(blue_sky) computed for the segment is greater than a threshold and greater than both P(gray_sky) and P(snow),

otherwise initially tagging the segment as “gray sky” when P(gray_sky) computed for the segment is greater than a threshold and greater than P(snow),

otherwise initially tagging the segment as “snow” when P(snow) computed for the segment is greater than a threshold, and

otherwise leaving the segment untagged or, equivalently, tagged as “unknown;” and

selecting a dominant sky color and tagging those segments with a likelihood of representing sky of the dominant sky color greater than a threshold value as sky, while untagging previously tagged segments with a likelihood of representing sky of the dominant sky color less than the threshold value.

7. A perceptual-segmentation system, comprising:

a computer; and

one or more perceptual-segmentation programs, stored within the computer and executed by the computer, that

identify regions within an image;

for each identified region,

compute or access each of one or more pixel-feature values for pixels within region,

using the one or more pixel-feature values computed or accessed for pixels within the region, compute one or more region-feature values, and

using the one or more region-feature values, determine a region type for the region and associating the region type with the region; and

store the associations of region types with regions for use in subsequent image-processing tasks;

wherein following associating region types with the identified regions, the one or more perceptual-segmentation programs, for each of one or more particular region types, further:

compute a color model from the regions associated with a the particular region type; and

apply the color model to other regions of the image to identify any additional pixels or regions of the image corresponding to the particular region type as determined by comparing between the color of the other segments and the computed color model.

8. A perceptual-segmentation system, comprising:

a computer; and

one or more perceptual-segmentation programs, stored within the computer and executed by the computer, that

identify regions within an image;

for each identified region,

compute or access each of one or more pixel-feature values for pixels within region,

using the one or more pixel-feature values computed or accessed for pixels within the region, compute one or more region-feature values, and

using the one or more region-feature values, determine a region type for the region and associating the region type with the region; and

store the associations of region types with regions for use in subsequent image-processing tasks;

wherein the one or more perceptual-segmentation programs, stored within the computer and executed by the computer, identify segments within an image corresponding to regions in the image that represent human skin and tag the segments by:

detecting faces within the image;

for each detected face,

determining a rectangular region of the image that includes the face,

partitioning the rectangular region into face and non-face partitions based on a general skin-color model and pixel position,

determining a face-specific color model for the face, based on the face partition,

generating a face-skin map for the rectangular region,

generating a body-skin model based on pixels within the rectangular region identified by the face-skin map as corresponding to face skin, and

identifying additional segments within the image corresponding to body-skin regions based on the body-skin model.

9. A perceptual-segmentation system, comprising:

a computer; and

one or more perceptual-segmentation programs, stored within the computer and executed by the computer, that

identify regions within an image;

for each identified region,

compute or access each of one or more pixel-feature values for pixels within region,

using the one or more pixel-feature values computed or accessed for pixels within the region, compute one or more region-feature values, and

using the one or more region-feature values, determine a region type for the region and associating the region type with the region; and

store the associations of region types with regions for use in subsequent image-processing tasks;

wherein the one or more perceptual-segmentation programs, stored within the computer and executed by the computer, tag segments within an image corresponding to regions in the image that represent foliage by:

for each untagged segment in the image,

computing a probability that the segment corresponds to a region of foliage in the image, P(foliage), based on a color model for foliage,

computing a probability that the segment corresponds to a region of forest in the image, P(forest), based on a color model for forest,

computing a probability that the segment corresponds to a region of ground in the image, P(ground), based on a color model for forest,

when P(foliage) is greater than one of P(forest) and P(ground) as well as greater than a threshold value, tagging the segment as foliage, and

otherwise tagging the segment based on a computed probability that the segment corresponds to a natural region of the image and not to an artificial region of the image, P(natural), by

when P(natural) is greater than a threshold value, P(foliage) is greater than a threshold value, and a computed intensity for the segment is less than a threshold intensity, tagging the segment as foliage, and

when P(naturel) is greater than a threshold value, P(forest) is greater than a threshold value, and a computed intensity for the segment is less than a threshold intensity, tagging the segment as foliage.

10. A method, executed by a computer, for perceptual segmentation of an image, comprising:

identifying regions within an image;

for each identified region,

computing or accessing each of one or more pixel-feature values for pixels within the region,

using the one or more pixel-feature values computed or accessed for pixels within the region to compute one or more region-feature values, and

using the one or more region-feature values to determine a region type for the region and associating the region type with the region; and

storing, on a non-transitory, computer-readable medium, the associations of region types with regions for use in subsequent image-processing tasks;

wherein identifying the regions within the image is accomplished by:

for each of a number of specific object-identifying methods,

applying the object-identifying method to locate occurrences of objects of a specific type in the image; and

disambiguating any regions of overlap of two or more objects identified by the object-identifying methods.

11. The method of claim 10 wherein pixel features associated with a pixel include one or more of:

one or more intrinsic pixel features, including:

a grayscale value, and

three values that specify a color of the pixel according to a color model,

one or more numeric values that specify a spatial location of the pixel within an image,

a numeric value computed from pixel-feature values of the pixel and additional pixels that are located in a neighborhood of the pixel, and

a numeric value computed by comparing one or more intrinsic pixel-values with a model or distribution of pixel-feature values;

and wherein segment features associated with a segment or region include one or more of:

an average, median, or other mathematical value computed from pixel-feature values of pixels within the segment or region,

an area computed for the segment or region,

a relative area computed for the segment or region,

numeric values that specify a position of the segment or region within an image,

one or more numeric values that indicate an orientation of the region within an image,

a morphological characteristic computed for the segment or region,

a value computed by comparing a value computed from pixel-feature values of pixels within the segment or region to a model or distribution,

a numeric value computed from pixel-feature values of the pixel and additional pixels that are located in a neighborhood of the pixel,

a numeric value computed by comparing one or more intrinsic pixel-values with a model or distribution of pixel-feature values, and

a numeric value from segment-feature values of the segment or region and additional segments or regions that are located in a neighborhood of the segment or region.