Patent Grant
7149354
1. Technical Field
An embodiment of the present invention generally relates to detecting edges of an image. More particularly, an embodiment of the present invention relates to extracting a scene structure using a gradient runs analysis.
2. Discussion of the Related Art
The concept of coding digital images with runs dates back to the beginning of computer vision. A digital image may comprise sequences of equal intensity pixels adjacent in a row/column, and the edges of objects in the image must somehow be detected in order to properly reproduce the image electronically. For example, a binary runs representation of a binary image, wherein the relationship between binary runs is analyzed, is often used for image compression, enumeration and extraction of connected components, and for generating image structural descriptions using graphs. Building contours of connected components based on a binary runs analysis may enable multilevel image analysis in real-time.
On the other hand, an edge-detecting algorithm may use image gradients to determine the edges of an object within an image. For purposes of illustration, the notion of a binary run in an image may be generalized as a group of adjacent pixels having the same properties. For example, segments of a piecewise-linear approximation of a row/column of a grayscale image are known as a gradient run. According to this technique, steep gradient runs characterize object borders in an image. Consequently, the object borders may be tracked by using the relation of gradient runs in adjacent rows/columns.
In order to track all the edges in an image, it is necessary to be able to combine the row runs information and the column runs information. While the relationship between binary runs may be used to track the borders of an object within a binary image, the complexity of an edge-detecting algorithm required to incorporate the relationship between gradient runs relating to a gray scale image has thus far prevented the development of such an algorithm.
Furthermore, some edge detection techniques require that a neighborhood of pixels be explored in order to obtain image edges. This technique substantially burdens the cache in a processor, for example.
Reference in the specification to “one embodiment”, “an embodiment”, or “another embodiment” of the present invention means that a particular feature, structure or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. Thus, appearances of the phrase “in one embodiment” or “according to an embodiment” appearing in various places throughout the specification are not necessarily all referring to the same embodiment. Likewise, appearances of the phrase “in another embodiment” or “according to another embodiment” appearing in various places throughout the specification are not necessarily referring to different embodiments.
The method of extracting a scene structure may speed up and facilitate image analysis. Within the method, memory may be accessed sequentially during edge extraction, thereby utilizing cache more efficiently than some other techniques. Within the method, features (e.g., straight edges, regions, and junctions) of the image may be conveniently obtained by utilizing a planar graph. The method may be highly adaptive to ambiguities. For example, the method may be robust to noise and illumination variations. According to an embodiment of the method, a gradient runs construction, a gradient runs analysis, and a generation of a planar graph may be processed more quickly because different parts of the image may be processed independently.
Gradient runs may be used to represent an image, for example, by considering a scalar function z=ƒ (x,y). The meaning of the function may depend upon the specific situation for which gradient runs are to be employed. For instance, the function may represent intensity in grayscale or a single channel in a multi-channel image. A horizontal gradient run may be defined by the equation <x1,x2,y,z1,z2>=HRy, where x1 is the coordinate of the beginning of the run, x2 is the coordinate of the end of the run, y is the index of the image row, z1=ƒ (x1,y), and z2=ƒ (x2,y). A vertical gradient run may be defined by the equation <y1,y2,x,z1,z2>=VRx, where y1 is the coordinate of the beginning of the run, y2 is the coordinate of the end of the run, x is the index of the image column, z1=ƒ (x,y1), and z2=ƒ (x,y2).
In an embodiment of the present invention, it may be assumed, for purposes of illustration, that the area below ƒ is the intensity in grayscale. The construction of gradient runs may be reduced to a problem of piecewise-linear approximation of the function ƒ. The type of approximation method to be employed is dependent upon prior information concerning the image to be manipulated. For example, if a number of edges that cross the row/column is known, then an optimization problem may be encountered. If a noise level is known, then an approximation method based on the maximum deviation of pixel values from the approximating segment may be employed. For instance, a Douglas-Peucker approximation may be employed.
In an embodiment, only image rows are considered. For example, a run [x1,x2] of pixels that satisfies the following conditions may be chosen for each row:
∀×ε[x1,x2], (I′(x)>T1 & {overscore (I′)}d(x)>T2) or (I′(x)<−T1 & {overscore (I′)}d(x)<−T2), where I′(x)=ƒx′(x,y),
and Īd(x) is the mean value of the intensity. The condition T1≧0 depends on the image contrast, and the conditions T2,d≧0 depend on the noise.
Referring to
xbar(R)=argmin(||/(x)−z2|−|/(x)−z1||).
For example, a runs construction based on the preceding algorithm may take 10–15 ms (non-optimized code) for an image with 320×240 pixel resolution on a personal computer having a Pentium® III 700 MHz processor.
Referring again to
According to an embodiment of the present invention, localization of the image edges may involve edge detection, which is generally based on a smoothing and differentiation of the image. The first derivatives ƒx′,ƒy′ of the image at point (x,y) may be estimated by utilizing coefficients, kR, which indicate the pitch of the gradient runs that include point (x,y). If R is a horizontal run, then kR may be represented by the following equation:
kR=(z2−z1)/(x2−x1).
If R is a vertical run, then kR may be represented by the following equation:
kR=(z2−z1)/(y2−y1).
Depending on the image contrast, Kbar>0 may be chosen such that runs having |kR|<Kbar may be considered to represent uniform regions in the image, and runs having |kR|≧Kbar may be considered to represent peculiar image domains.
For purposes of illustration, gradient runs generally may be classified into two categories: (1) growing and (2) falling. In growing gradient runs, z2−z1>0. In falling gradient runs, z2−z1<0. To simplify the discussion, only horizontal runs will be considered, though the discussion is generally the same for vertical runs. For example, two gradient runs, R1 and R2, may be connected if they lie in adjacent rows and the following condition is satisfied:
(x2(R1)≧x1(R2))&(x2(R2)≧x1(R1)).
In an embodiment, two runs, R1 and R2, of the same type (e.g., both horizontal runs or both vertical runs) may be edge-forming if the following conditions are satisfied:
According to an embodiment, all edge-forming runs may be detected during a single pass. The line that connects the points, xbar(R), of the edge-forming runs may approximate a border between uniform regions. An edge in an image may be represented as a polyline that connects the points, xbar(R), of the edge-forming runs.
For example,
Returning to
In an embodiment, intervals of polylines obtained from vertical runs that represent a single image edge may be eliminated. A graph may be obtained that has vertices corresponding to the points, xbar(R), of linked runs. The degree of the vertices may not be greater than two.
In an image, the color of a region may be estimated, for example, by the intensity of runs endpoints that lie inside the region. In an embodiment, region environs may be estimated by the intensity of runs endpoints that lie outside the region.
Referring to
Returning to
According to an embodiment of the present invention, the method may further include eliminating 230 a redundant edge from a plurality of image edges. In an embodiment, the method may further include linking 240 a plurality of image edges into a planar graph. A broken edge may be restored 250 by a forced extension.
According to an embodiment of the present invention, the processor 1120 may eliminate a redundant edge. In an embodiment, the processor 1120 may restore a broken edge.
While the description above refers to particular embodiments of the present invention, it will be understood that many modifications may be made without departing from the spirit thereof. The accompanying claims are intended to cover such modifications as would fall within the true scope and spirit of an embodiment of the present invention. The presently disclosed embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of an embodiment of the invention being indicated by the appended claims, rather than the foregoing description, and all changes that come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein.
| Number | Name | Date | Kind |
|---|---|---|---|
| 4656507 | Greaves et al. | Apr 1987 | A |
| 4817174 | Nakatani | Mar 1989 | A |
| 4995087 | Rathi et al. | Feb 1991 | A |
| 5532843 | Otsuki | Jul 1996 | A |
| 6392759 | Kuwata et al. | May 2002 | B1 |
| 6628833 | Horie | Sep 2003 | B1 |
| 7003160 | Horie | Feb 2006 | B1 |
| Number | Date | Country | |
|---|---|---|---|
| 20040114805 A1 | Jun 2004 | US |
