Image analysis – Pattern recognition
Reexamination Certificate
2000-01-27
2003-05-06
Boudreau, Leo (Department: 2621)
Image analysis
Pattern recognition
C382S230000, C382S229000
Reexamination Certificate
active
06560360
ABSTRACT:
BACKGROUND OF THE INVENTION
The present invention relates generally to automated recognition systems for classifying and identifying patterns as objects within a library of objects, and more specifically to a recognition system including feed forward, feed back multiple neural networks with context driven recognition.
As it is generally known, preprocessing forms an integral part of many existing artificial recognition systems. In such systems, preprocessing converts or transforms an input signal into a more suitable form for further processing. Some common steps performed by existing recognition systems during preprocessing of an input signal include normalization, noise reduction (filtering), and feature extraction. There are various reasons for using preprocessing, including the fact that input signals often contain noise, and that preprocessing can sometimes effectively eliminate irrelevant information. Furthermore, performing data reduction during preprocessing may result in at least two advantages for neural network based recognition systems: a) dimensionality of the input vectors is reduced, affording better generalization properties, and b) training is generally much simpler and faster when smaller data sets are used. However, obtaining these advantages by preprocessing in some cases may introduce certain potential problems, since preprocessing may sometimes result in the loss of important information from the input signal.
For example, a standard procedure for noise removal in two-dimensional images is generally known as “smoothing” of the input signal. One of the simplest ways to smooth an image includes convolving it with a mask of a fixed size and setting the value of every pixel within the mask to the average value of the pixels within the mask. The smoothing process can be controlled by a set of parameters, one of which is the size of the mask. The size of the mask may vary from one pixel (when there is no smoothing) to the size of the whole image. Varying the value of this mask size parameter from low to high results in changing the resolution of the image from sharp to very blurry. It is, however, very difficult to correctly determine the value of the mask size parameter prior to recognition. This is because image recognition may be impaired if the resulting image resolution is too poor. For example, in the case of handwriting recognition, omission of a noise-like structure, such as a dot above “i” or “j”, due to poor image resolution, could impair recognition.
One reason why some of the problems encountered at the preprocessing level are not possible to resolve during preprocessing is that some information which is necessary for choosing optimal preprocessing parameters is not yet available. Accordingly, it would be desirable to have a recognition system that provides interaction between higher (cognitive) level processing and the preprocessor in order to change the parameters controlling the preprocessor according to dynamically determined, higher level expectations.
U.S. Pat. No. 4,760,604 of Cooper et al. discloses a recognition system that may use multiple preprocessors. In that system, each of a number of adaptive modules can facilitate a different preprocessing scheme, but the preprocessing is done “globally” on the whole input pattern. It would be desirable, however, to use feedback information for a) locating a section or portion of the input pattern that needs additional preprocessing, and b) changing the preprocessing of only such a section or portion. This would be especially useful during recognition of complex objects where a specific type of “re-preprocessing” appropriate for one section could improve recognition of that particular section, but may have an adverse effect on recognition of the rest of the object if applied globally. Moreover, in order to selectively apply different preprocessing techniques to different regions of an object, a recognition system should be able to appropriately segment an object into parts. It would be desirable, therefore, to have a system which provides an appropriate, segmented representation of an object for the purpose of applying different preprocessing techniques at different regions of the object.
An effective recognition system should further be capable of recognizing a pattern regardless of its position and/or size within the overall context in which it is received. For example, the position of an address written on an envelope or a package can vary significantly, and the sizes of individual letters and numbers within the address are also not fixed. In many existing recognition systems, this problem is addressed at the preprocessing stage. In such existing systems, prior to a recognition stage, the received image is “cleaned”, in that the text is located, and surrounded by a rectangle that is then re-scaled or normalized. However, this approach suffers from significant drawbacks since at the level of preprocessing it is not yet known which sections of the input signal represent text or speech, and which sections represent background. Therefore, the output of such existing systems often consists of numerous false segmentations. It would be desirable to have a recognition system that does not rely on (prior to recognition) pre-segmentation of the input signal.
In addition, it would be desirable to provide a translationally invariant representation of an input object, within a system which also provides scale invariant recognition. The translationally invariant representation would permit an object to be described with the same values regardless of its position or location within the input signal. Such scale invariant recognition would allow recognition of an object that may appear in different scales or sizes, while the system need only be trained on one scale.
Another challenging problem in the field of pattern recognition is the identification and classification of objects that are only partially present within the input signal, or that are partially occluded for some reason. It would further be desirable to provide a recognition system that allows for the recognition of incomplete or partially occluded patterns as parts of a recognizable object.
One of the problems in the field of sequence analysis, as occurs in speech or cursive writing recognition systems, is referred to as the segmentation/binding dilemma. This problem stems from the fact that in order to unambiguously segment an input word into letters, the input word must be known, but in order to recognize the word, its constituent letters must be known. An existing approach to solving this problem (e.g. in cursive recognition applications), is to first make all possible segmentations, and then to choose an “optimal” one, with respect to the set of potentially recognizable objects, according to some predetermined criterion or cost function. Since this method is often computationally intensive, it is desirable to use an algorithm that can efficiently search the space of potentially recognizable objects. To this end, researchers have often used some variation of conventional dynamic programming optimization techniques. However, in some cases, dynamic programming techniques do not decrease the computational complexity of selecting an optimal segmentation. Accordingly, it would be desirable to have a method for recognizing an object which permits features of the object to be individually recognized and associated with the object, based on context information of some kind. Such a method should advantageously lend itself to a high degree of parallelism in its implementation, thus resulting in fast recognition results obtained in relatively few cycles. Such a method could, in some cases, advantageously provide an alternative to the dynamic programming based post-processing employed in existing systems to find an optimal segmentation of an input pattern.
Another problem related to sequence analysis in recognition systems is selecting a convenient representation of an input pattern that captures the sequential nature of the input signal. One technique used
Cooper Leon N
Neskovic Predrag
Reilly Douglas L.
Akhavannik Hussein
Boudreau Leo
Nestor, Inc.
Weingarten Schurgin, Gagnebin & Lebovici LLP
LandOfFree
Feed forward feed back multiple neural network with context... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Feed forward feed back multiple neural network with context..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Feed forward feed back multiple neural network with context... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3055995