Image analysis – Image transformation or preprocessing – Changing the image coordinates
Reexamination Certificate
2002-04-02
2004-02-10
Johnson, Timothy M. (Department: 2625)
Image analysis
Image transformation or preprocessing
Changing the image coordinates
C382S199000
Reexamination Certificate
active
06690842
ABSTRACT:
FIELD OF THE INVENTION
This invention relates generally to digital image processing, and particularly to edge detection in digital images.
BACKGROUND OF THE INVENTION
Digital images can be formed by many devices and can be used for many practical purposes. Digital image formation devices include TV cameras operating on visible or infrared light, line-scan sensors, flying spot scanners, electron microscopes, X-ray devices, such as CT scanners, and magnetic resonance imagers, for example. Practical applications of digital image formation devices include industrial automation, medical diagnosis, satellite imaging, photographic processing, surveillance, traffic monitoring, document processing, and many others.
To serve these applications, images formed by a digital image formation device are processed by a digital information processing device, such as a general purpose computer executing image processing software, to extract useful information. One very common form of digital image processing, well known in the art, is edge detection. Edge detection can be defined informally as a process for determining the location of boundaries between image regions that are of different and roughly uniform brightness. To be more precise, edge detection can be defined as a process for locating edges in an image, where an edge can be usefully defined as a point in an image where the image gradient magnitude reaches a local maximum in the image gradient direction, or equivalently, where the second derivative of brightness crosses zero in the image gradient direction. It can also be useful to define an edge as a point where the image best fits a one- or two-dimensional ideal step boundary, within a small neighborhood of the point. (Some authors define an edge to be a set of such points, and define the points themselves as edge elements. For present purposes, the distinction is unnecessary; the points will simply be called edges.)
It is recognized in the art that many local maxima of gradient magnitude (or zero-crossings in the second derivative) may exist in an image that are not the result of boundaries between regions, but rather are the result of image noise or other image artifacts. Thus, it is conventional to discard edge points whose gradient magnitude is less than some noise threshold, which threshold can be predetermined, or can be adaptively computed based on image characteristics, and which can vary from point to point in the image, or can be constant over the entire image. Other more sophisticated edge point filtering techniques are known in the art, such as the hysteresis thresholding method of Canny.
It is also recognized in the art that the existence and characteristics of a boundary between regions of different and roughly uniform brightness depends on the scale (resolution) at which the image is processed. Boundaries between small, high spatial frequency regions may not be evident in a coarse, low resolution examination of the image, while boundaries between much larger, low spatial frequency regions may not be evident in a fine, high resolution view (i.e., not seeing the forest for the trees). Thus it is known in the art to perform edge detection at a plurality of spatial frequencies or length scales as appropriate to the application.
The above definition of an edge based on gradient magnitude and direction, while precise, is based on the impractical assumption that an image can be treated as a function of two continuous variables. In practice, however, an image acquired by an image formation device as discussed above is discontinuous and quantized, consisting of an array of pixels, each pixel being disposed at an integer-valued image coordinate, and each pixel having an integer brightness values. Consequently, in practice one can only estimate gradient magnitude and gradient direction, and one can only estimate the position of a gradient maximum or a zero-crossing. Furthermore, in practice, the computational cost and speed of such an estimation must be considered, so that it is desirable to use methods of gradient estimation that are accurate, and yet at the same time computationally inexpensive and fast. However, higher accuracy gradient determination, and edge location based thereon, is typically associated with high cost and/or low speed. Also, low cost and/or high speed gradient determination, and edge location based thereon, is typically associated with low accuracy. Many estimators of gradient magnitude and direction are known in the art, which attempt to strike a reasonable balance between accuracy, computational cost, and speed.
To provide low cost and/or high speed, most known gradient estimators provide very crude estimates of gradient magnitude and direction. In this case, the gradient magnitude accuracy tends to be less for gradient directions not substantially parallel to the axes of the pixel grid, as compared with gradient magnitude accuracy for gradient directions substantially parallel to the axes of the pixel grid. Gradient direction is usually computed to only three bits of precision (i.e., approximated to one of eight discrete directions, e.g., N, NE, E, . . . ) because three-bit accuracy is relatively inexpensive; cost increases significantly beyond three bits.
Although carrying out edge detection to the nearest whole pixel using image gradients is generally straightforward and efficient, it is challenging to achieve fast, inexpensive, and accurate subpixel edge detection using image gradients. Alternatively, edge detection based on locally fitting ideal step boundaries can directly provide accurate subpixel edge positions, without requiring intermediate calculations based on an assumption of two continuous variables. Therefore, such local fitting methods dominate the art in applications requiring accurate, subpixel edge detection. However, local fitting methods are relatively expensive and/or slow, and therefore are not practical in high-accuracy applications that also require low cost and/or high speed. Consequently, there is a need for an inexpensive and/or fast method of high-accuracy subpixel edge detection.
SUMMARY OF THE INVENTION
The invention provides an apparatus and method for accurate subpixel edge detection, based on fast and inexpensive estimates of image gradient magnitude and direction. Any method of forming an image may be used, based on either image acquisition using an image formation device, or image synthesis. In either case, the image may optionally be transformed by one or more processing steps of any nature, including but not limited to optical and/or electronic image processing.
The invention provides a method and apparatus for edge detection using an array of gradient magnitude and gradient direction estimates to determine accurate subpixel positions of the edges detected.
Image gradient magnitude and direction are estimated at regularly spaced positions in the image using any gradient estimation method that provides more than three bits of gradient direction accuracy, including but not limited to the gradient estimation method described herein. In a preferred embodiment, an inexpensive gradient estimator providing around seven bits or more of magnitude and direction is used. The estimates are made at a scale (spatial resolution) appropriate to the application. The edge detection process can be performed at a plurality of scales on the same image, if appropriate. The points at which gradient is estimated at a given scale may or may not correspond to the locations of the image pixels.
In an optional step, a gradient estimate is discarded when it is determined not to correspond to a real image feature, using methods known in the art, such as discarding the gradient estimate when the magnitude of the gradient estimate falls below a noise threshold. This optional step may be performed separately, or may be combined with subsequent processing steps.
For each gradient estimate G, having magnitude G
0
and direction G
&thgr;
, one or more neighboring estimates are chosen on each side of G, approximately along the direction G
&thgr;
. In a
Garakani Arman
Silver William
Wallack Aaron
Chawan Sheela
Cognex Corporation
Johnson Timothy M.
Weinzimmer Russ
LandOfFree
Apparatus and method for detection and sub-pixel location of... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Apparatus and method for detection and sub-pixel location of..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Apparatus and method for detection and sub-pixel location of... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3290998