Color clustering and segmentation using sigma filtering

Claim 2

A fast method of image segmentation that retains edge information comprising the steps of: applying a sigma filter to generate a smoothed image; dividing the filtered image into a plurality of non-overlapping windows; applying a sequential clustering algorithm to filtered pixels of each window to generate potential clusters; validating generated potential clusters while rejecting clusters representing pixels on boundaries; and generating a list of cluster centroids in a chosen color space; wherein the step of applying a sigma filter comprises the steps of: estimating a threshold value as 2σ; computing a distance, d, from the center pixel in the window in image ƒ to a next unvisited pixel, p, this distance being measured in terms of a RGB (red, green, blue) color space difference between the unvisited pixel p in a filter mask, M, the filter mask M having a value of “1” to mark an included pixel and “0” otherwise; and determining whether processing of the window is complete and, if not, repeating the steps of computing a distance, d, and determining whether a computed distance d is greater than the threshold value.

Claim 3

The fast method of image segmentation recited in claim 1 , wherein the sigma filter is based on a model of an image over a local area as a summation of a constant, k, and a noise component, n(x,y), wherein the noise is approximated as a Gaussian distribution, which is completely characterized by its mean, μ, and standard deviation, σ, and that, for a Gaussian distribution, 95% of pixel values in a window are located within 2σ of the mean, μ, where the “mean” refers to an average of color values within a window, The sigma filter identifying those pixels that belong to a same class as a center pixel of the window.

Claim 4

A fast method of image segmentation that retains edge information comprising the steps of: applying a sigma filter to generate a smoothed image; dividing the filtered image into a plurality of non-overlapping windows; applying a sequential clustering algorithm to filtered pixels of each window to generate potential clusters; validating generated potential clusters while rejecting clusters representing pixels on boundaries; and generating a list of cluster centroids in a chosen color space; wherein the step of applying a sigma filter comprises the steps of: estimating a threshold value; computing a distance, d, from the center pixel in the window in image ƒ to a next unvisited pixel, p; determining whether a computed distance d is greater than the threshold value and if not, including the pixel position of the unvisited pixel p in a filter mask, M, the filter mask M having a value of “1” to mark an included pixel and “0” otherwise; and determining whether processing of the window is complete and, if not, repeating the steps of computing a distance, d, and determining whether a computed distance d is greater than the threshold value.

Claim 5

A fast method of image segmentation that retains edge information comprising the steps of: applying a sigma filter to generate a smoothed image; dividing the filtered image into a plurality of non-overlapping windows; applying a sequential clustering algorithm to filtered pixels of each window to generate potential clusters; validating generated potential clusters while rejecting clusters representing pixels on boundaries; and generating a list of cluster centroids in a chosen color space; wherein the step of applying a sigma filter comprises the steps of: estimating a threshold value; computing a distance, d, from the center pixel in the window in image ƒ to a next unvisited pixel, p; determining whether a computed distance d is greater than the threshold value and if not, including the pixel position of the unvisited pixel p in a filter mask, M, the filter mask M having a value of “1” to mark an included pixel and “0” otherwise; and determining whether processing of the window is complete and, if not, repeating the steps of computing a distance, d, and determining whether a computed distance d is greater than the threshold value.

Claim 6

The fast method of image segmentation recited in claim 3 , wherein the step of applying a sigma filter comprises the steps of: modifying the window if the number or included mask positions is significantly less than the number of pixels within the window by replacing the center pixel by the mean of its surrounding pixels; and processing with the new value for the center pixel.

Claim 7

The fast method of image segmentation recited in claim 3 , wherein the step of applying a sigma filter further comprises the steps of: modifying the window if the number of included mask positions is less than N, where N=(┌1.5σ┐)+1, by replacing the center pixel by the mean of its surrounding pixels; and processing with the new value for the center pixel.

Claim 8

The fast method of image segmentation recited in claim 3 , wherein the step of estimating a threshold value comprises the steps of: locating edge points in the image ƒ; calculating a local standard deviation δ n in terms of color (RGB) values for pixels surrounding each edge point, where n indexes the edge points, and repeating this process for each edge point in an image ƒ; sorting a sequence of standard deviations {δ n } for all edge points; and calculating δ by the following rule: δ=min{δ i |δ i ≧0.1×max{δ i }} to capture statistics of regions that surround edge pixels in the image so that smoothing across edges can be subsequently avoided.

Claim 9

The fast method of image segmentation recited in claim 3 , wherein the sigma filter is based on a model of an image over a local area as a summation of a constant, k, and a noise component, n(x,y), wherein the noise is approximated as a Gaussian distribution, which is completely characterized by its mean, μ, and standard deviation, σ, and that, for a Gaussian distribution, 95% of pixel values in a window are located within 2σ of the mean, μ, where the “mean” refers to an average of color values within a window, The sigma filter identifying those pixels that belong to a same class as a center pixel of the window.