Image analysis – Pattern recognition – Feature extraction
Reexamination Certificate
1997-10-09
2001-03-27
Coles, Edward L. (Department: 2722)
Image analysis
Pattern recognition
Feature extraction
C382S157000, C382S159000, C382S165000, C382S197000, C382S203000, C382S282000, 36, C706S020000, C706S029000
Reexamination Certificate
active
06208758
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
This invention relates to a method for recognizing the presence or absence of a predetermined object image in an image. This invention particularly relates to a method for recognizing an object image wherein, during image information processing, a judgment is made as to whether a candidate for a predetermined object image, which candidate has been extracted from an image, is or is not the predetermined object image. This invention also relates to a learning method for a neural network, wherein a target object image, for which the learning operations are to be carried out, is extracted from an image, and the learning operations of a neural network for carrying out recognition of a predetermined object image are carried out with respect to the extracted target object image. This invention further relates to a method for discriminating an image wherein, during image information processing, a judgment as to whether a given image is or is not a predetermined image is made accurately without being adversely affected by a change in the angle of the image, rotation of the image and a background of the image.
2. Description of the Prior Art
A human being views an image and recognizes what the thing embedded in the image is. It is known that this action can be divided into two steps. A first step is to carry out “discovery and extraction” by moving the viewpoint, setting a target of recognition at the center point of the visual field, and at the same time finding the size of the object. A second step is to make a judgment from a memory and a knowledge of the human being as to what the object present at the viewpoint is. Ordinarily, human beings iterate the two steps and thereby acquire information about the outer world.
On the other hand, in conventional techniques for recognizing a pattern by carrying out image processing, typically in pattern matching techniques, importance is attached only to the second step. Therefore, various limitations are imposed on the first step for “discovery and extraction.” For example, it is necessary for a human being to intervene in order to cut out a target and normalize the size of the target. Also, as in the cases of automatic reading machines for postal code numbers, it is necessary for a target object to be placed at a predetermined position. As pattern recognizing techniques unaffected by a change in size and position of a target, various techniques have been proposed wherein a judgment is made from an invariable quantity. For example, a method utilizing a central moment, a method utilizing a Fourier description element, and a method utilizing a mean square error have been proposed. With such methods, for the purposes of recognition, it is necessary to carry out complicated integrating operations or coordinate transformation. Therefore, extremely large amounts of calculations are necessary in cases where it is unknown where a target object is located or in cases where a large image is processed. Also, with these methods, in cases where a plurality of object images are embedded in an image, there is the risk that their coexistence causes a noise to occur and causes errors to occur in recognizing the object images. Thus these methods are not satisfactory in practice.
As a model, which is unaffected by the size of a target object or by a shift in position of a target object and which can accurately recognize the target object, a model utilizing a neocognitron, which is one of techniques for neural networks, has been proposed. The neocognitron is described by Fukushima in “Neocognitron: A Neural Network Model for a Mechanism of Pattern Recognition Unaffected by Shift in Position,” Collected Papers of The Institute of Electronics and Communication Engineers of Japan, A, J62-A(10), pp. 658-665, October 1979. Neural networks constitute one of research techniques for neural information processing, which is referred to as the constructive method and which aims at clarifying the information processing principle of a brain by constructing an appropriate neural circuitry model with full consideration given to the facts known physiologically and results of research, investigating the actions and performance of the model, and comparing the actions and performance of the model with those of the actual human brain. Research has been conducted to develop various models, such as visual models, learning models, and associative memory models. In particular, the neocognitron model is tolerant of a shift in position of an object image embedded in an image. The neocognitron carries out pattern matching and self-organizing learning operations on a small part of a target object image, assimilates a shift in position at several stages with a layered architecture, and thereby tolerates the shift in position.
In the neocognitron, the operation for tolerating a shift in position of a feature little by little at several stages plays an important role in eliminating adverse effects of a shift in position of an input pattern and carrying out pattern recognition tolerant of a deformation of the input pattern. Specifically, adverse effects of shifts in position between local features of an input pattern, which shifts are due to various deformations, such as enlargement and reduction, of the input pattern, are assimilated little by little during the process for putting the features together. Ultimately, an output can be obtained which is free of adverse effects of comparatively large deformation of the input pattern.
As described above, the neocognitron is based on the principle that the pattern matching is carried out on a small part of a target object, and a shift in its position is assimilated at several stages through a layered architecture. However, with such a principle, a limitation is naturally imposed on achievement of both the accurate recognition and the assimilation of the shift in position. It has been reported, for example, by Nagano in “Neural Net for Extracting Size Invariant Features,” Computrol, No. 29, pp. 26-31, that the neocognitron can ordinarily tolerate only approximately four times of fluctuation in size. As for the shift in position, the neocognitron can tolerate only approximately two or three times the size of a target object. The tolerance capacity remains the same also in a recently proposed neocognitron model which is provided with a selective attention mechanism.
How the visual function of a human being carries out the first step has not yet been clarified. On the other hand, how the viewpoint moves has been clarified to some extent as described, for example, by Okewatari in “Visual and Auditory Information Processing in Living Body System,” Information Processing, Vol. 23, No. 5, pp. 451-459, 1982, or by Sotoyama in “Structure and Function of Visual System”, Information Processing, Vol. 26, No. 2, pp. 108-116, 1985. It is known that eyeball movements include a saccadic movement, a follow-up movement, and involuntary movement. Several models that simulate these eye movements have been proposed. For example, a model in which the viewpoint is moved to the side of a larger differential value of an image is proposed, for example, by Nakano in “Pattern Recognition Learning System,” Image Information (I), 1987/1, pp. 31-37, or by Shiratori, et al. in “Simulation of Saccadic Movement by Pseudo-Retina Mask,” ITEJ Tec. Rep. (Technical Report of The Institute of Television Engineers of Japan), Vol. 14, No. 36, pp. 25-30, ICS′ 90-54, AIPS′ 90-46, June 1990. Also, a model in which the viewpoint is moved to the side of a higher lightness is proposed, for example, by Hirahara, et al. in “Neural Net for Specifying a Viewpoint,” ITEJ Tec. Rep., Vol. 14, No. 33, pp. 25-30, VAI′ 90-28, June 1990. Additionally, a model in which the viewpoint is moved to a point of a contour having a large curvature is proposed, for example, by Inui, et al. in Japanese Unexamined Patent Publication No. 2(1990)-138677. However, these proposed models are rather simple and do not well simulate the human visual function
Ono Shuji
Osawa Akira
Coles Edward L.
Fuji Photo Film Co. , Ltd.
Rosenblum David
Sughrue Mion Zinn Macpeak & Seas, PLLC
LandOfFree
Method for learning by a neural network including extracting... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method for learning by a neural network including extracting..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method for learning by a neural network including extracting... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2493912