Image analysis – Image enhancement or restoration – Variable threshold – gain – or slice level
Reexamination Certificate
2001-05-21
2002-12-03
Boudreau, Leo (Department: 2621)
Image analysis
Image enhancement or restoration
Variable threshold, gain, or slice level
C382S108000, C382S172000, C382S190000, C382S205000, C382S218000, C382S239000, C382S258000, C382S264000, C382S270000, C382S274000, C348S356000, C358S406000, C708S207000
Reexamination Certificate
active
06490375
ABSTRACT:
FIELD OF THE INVENTION
This invention relates to machine vision, and particularly to identifying a good reference site in an image using a machine-vision system.
BACKGROUND
In the field of machine vision, a common need is to identify and subsequently repeatably find a reference site within an image. The image is captured by the machine-vision system, where the image is a digital representation of an object within the field of view. The reference site is used to orient the machine to the object. For the reference site to be useful, it must be reliably found and is preferably unique within the field of view of the image.
Uses for reference sites are in the inspection field, where probes must be positioned for test, and in computer control of manufacturing processes, where sequences of operations must be precisely aligned. In the field of inspection of semiconductor devices, reference sites often are flats or notches on an edge of a wafer.
In such applications, the reference site is chosen typically by an operator. Then, the machine-vision system captures an image of the reference site, which is then stored as a template or a model. Thereafter during production, also known in the industry as run-time, the machine-vision system acquires an image of the object, and attempts to find a feature in the image that matches the template or model of the reference site. When the template or model is located, the machine-vision system performs further operations on the run-time images, such as positioning, searching for features or deviation analysis.
In printing applications, frequently reference sites are fiducial marks on the periphery of the object in view. For instance, crosses positioned at edges allow documents to be properly positioned for scanning. Typically, fiducials positioned at edges also orient automatic devices prior to printing.
Many newly designed applications, however, have removed fiducial marks from the objects being viewed. Increasingly unique characteristics from the natural patterns of the image are used as reference sites instead of the fiducial marks.
Without fidicials, operators must select the reference sites based on the “feel” of the image. To choose a viable reference site, which result in commercially good operation, the operator interacting with the selection program must be highly trained. Otherwise, the reference site is found inconsistently during production, if at all. Moreover, as the patterns within the images become more complex, the operator's job becomes even more difficult. Consequently, locating the reference site at run-time has become even less reliable.
In addition to unreliability introduced by the lack of operator training, the use of patterns in the image as the reference site has inherent difficulties. When patterns are used as the reference site, the reference site is a portion of the image that is segmented by a window. A unique pattern within a window that has minimal information in a particular direction may not be a suitable reference site for positioning in that direction. Such a reference site cannot indicate whether the image is upside down. Further, the pattern may not be unique within the field of view such as feature
140
and feature
140
′ illustrated in image
100
of FIG.
1
. Without further processing, the machine cannot choose which of the features
140
or
140
′ is the reference site. Therefore, the machine may not be oriented properly with respect to the image. Thus, any non-unique feature is not an optimum reference site.
One attempt to replace operators with machines has had limited success owing to the inability to identify the appropriate criteria that repeatably represent a good reference site while producing a fast result. The attempt was embodied in AUTOTRAIN, a software tool sold by Cognex Corporation. AUTOTRAIN measured the standard deviation of two one-dimensional projections and uniqueness of windows to determine if a window was a good reference site candidate. Although AUTOTRAIN was fast, it did not always choose the appropriate reference site, and, at times, would propose a reference site of random noise.
SUMMARY
A method is disclosed for identifying a good reference site in an image which uses at least one measurable parameter to filter windows in the image and present an ordered set of the filtered windows as possible good reference sites.
The at least one measurable parameter indicates the extent of “redundant” information within the possible reference sites. A reference site with “redundant” information can be found repeatably even when portions of the reference site are omitted, skewed, obstructed, or altered owing to production conditions. The method recognizes that selecting proposed reference sites with “redundant” information greatly enhances the quality of proposed reference sites. Preferably “uniqueness” is measured in conjunction with “redundancy.”
A value for the parameter is computed for each window in the image, where the windows contain the possible reference sites. Using the value, an ordered set of the windows is produced, where the order depends upon the value of the parameter. Optionally, a portion of the windows are not ordered and instead removed from consideration as reference sites if they fall below a threshold associated with the parameter. The threshold can leave only a single window in the ordered set.
In a preferred embodiment, a plurality of parameters filters the windows in sequence. In this embodiment, each window must pass a threshold of the first parameter before the second parameter is measured. The method can use any combination of parameter thresholds or lack thereof
Further, when a plurality of parameters are used, the value of each parameter is scaled relative to the other parameters, such as normalized for instance, and then combined to produce the ordered set of windows. Accurate methods of scaling the parameters are disclosed. Optionally, the parameters are weighted according to their relevance before the values of the parameters are combined. At the end, the window having the highest value is expected to be used as the reference site.
In a preferred embodiment, to improve the set of possible reference sites, the conditions for finding a reference site are adjusted for each application and for each image in which a reference site is sought. The conditions can include window size, sampling rate, and weights of the parameters, for example. It is recognized by the invention that providing versatility for these values and states improves the proposed reference sites for each image. Thus, by customization, the embodiment improves selection of the reference site. The method may also be applied iteratively to find which conditions are best for each application and image by evaluating the scores of the proposed reference site from each implementation of the method.
A specific implementation of this method is illustrated and claimed which efficiently extracts information from the image using the parameters of “symmetry”, “orthogonality”, and “uniqueness,” where “symmetry” and “orthogonality” both indicate the extent of “redundancy” of a possible reference site. In a preferred implementation, “symmetry” is the first filter, because it is recognized by the invention that symmetry can be measured on low-resolution images without skewing information, i.e. it is linear, and, therefore, computationally less time-intensive.
The invention solves some of the problems of the prior art. In particular, the invention recognizes the need to customize the method to each application.
The customization compensates for the inherent difficulties of using portions of the image for reference sites. In such an application, the role of the operator is reduced, and therefore the reliability of the proposed reference sites is enhanced. Further, the invention measures certain possible reference sites to increase the likelihood that the proposed reference sites are found repeatably.
REFERENCES:
patent: 4731863 (1988-03-01), Sezan et al.
patent: 4819197 (1989-04-01), Blais
patent: 50
Garakani Arman
Taycher Leonid
Boudreau Leo
Calabresi Tracy
Cognex Corporation
Mariam Daniel G.
LandOfFree
Methods for finding peaks in a characteristic surface of an... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Methods for finding peaks in a characteristic surface of an..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Methods for finding peaks in a characteristic surface of an... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2972187