Image analysis – Learning systems – Neural networks
Reexamination Certificate
1996-02-15
2001-10-09
Mehta, Bhavesh (Department: 2621)
Image analysis
Learning systems
Neural networks
C382S176000, C382S264000
Reexamination Certificate
active
06301381
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates to a neurofilter, which is a non-linear filter implemented as a neural network, and to a method of training a neurofilter to produce an output signal which discriminates between text and picture regions of an image expressed by image data which are supplied to the neurofilter.
The invention further relates to a filter apparatus in which a neurofilter is utilized to compensate for errors in the output signal from a conventional filter.
2. Prior Art
In the prior art, it has been necessary for designers of filters, for such applications as serial signal processing or image data processing, to have extensive experience in that field of work. This is due to the fact that considerable experience is necessary, to enable the designer to set the filter parameters to optimum values. Setting of the filter parameters is difficult, due to the fact that in general the transfer function of a filter is non-linear. Hence, it is necessary to use linear approximation functions to establish a preliminary design of a filter, and for the designer to then try to optimize the design by modifying the values of the filter parameters, on the basis of his experience.
Hence, in the prior art, the degree to which the performance of a filter approached optiumum has been dependent upon the experience of the filter designer, i.e. in general it has not been possible for other individuals to design a filter.
Furthermore even in the case of an experienced filter designer, due to the various forms of non-linearity of operation of an actual filter, it has been impossible to actually achieve optimum results.
SUMMARY OF THE INVENTION
It is an objective of the present invention to overcome the problems of the prior art set out above, by providing a neurofilter which is implemented as a suitably trained neural network, whereby an individual without experience in the field can easily produce a filter which provides optimum performance.
It is a further objective of the invention to provide a neurofilter which produces an output signal that discriminates between text and picture regions of an image expressed by image data which are supplied to the neurofilter, and to provide a suitable method of training such a neurofilter.
It is moreover an objective of the invention to provide a filter apparatus based on a conventional filter, and including a neurofilter for correcting errors in the output signal from the conventional filter, to thereby reduce the severity of design constraints of the conventional filter.
To achieve the above objectives, according to a first aspect of the invention, the invention provides a neurofilter which is a non-linear filter implemented as a neural network that has been trained (by adjustment of the weighting coefficients of the neural network) to have a predetermined filter response to an input signal.
According to a second aspect, the invention provides a method of training a neurofilter which produces an output signal that discriminates between text and picture regions of an image expressed by image data which are supplied to the neurofilter, with the training comprising a process of supplying image data expressing respective local regions of an image to the neurofilter, with said local regions alternately being within text regions and picture regions of the image, while setting a training signal to the ‘1’ logic state when a local region within a text region is being supplied to the neurofilter and setting the training signal to the ‘0’ logic state when a local region within a picture region is being supplied to the neurofilter.
According to another aspect, the invention provides a filter apparatus comprising a parallel combination of a neurofilter which has undergone a suitable training procedure and a conventional filter, with an output signal produced from the neurofilter, in response to an input signal supplied to the filter apparatus, being combined with an output signal produced from the conventional filter in response to that input signal, for thereby obtaining an output signal in which errors in the response characters of the conventional filter have been compensated by the output signal from the neurofilter. As a result, it becomes very much easier to design and manufacture such a conventional filter while obtaining optimum performance from the filter apparatus.
With the present invention, since filter parameters are determined by a neurofilter, i.e. a neural network which has been trained to provide a suitable non-linear response to an input signal, it becomes possible for anyone to construct a filter which provides optimum results.
In the case of a filter apparatus which is a combination of a neurofilter and a conventional filter, only the error component of the output signal from the conventional filter is compensated by the output signal from the neurofilter. It has been found in practice that this enables a highly effective filter apparatus to be achieved.
REFERENCES:
patent: 5148495 (1992-09-01), Imao et al.
patent: 5187592 (1993-02-01), Sugiyama et al.
patent: 5245445 (1993-09-01), Fujisawa
patent: 5339365 (1994-08-01), Kawai et al.
patent: 5608819 (1997-03-01), Ikeuchi
Israel Gopstein Clark & Brody
Matsushita Electric Idustrial Co., Ltd.
Mehta Bhavesh
LandOfFree
Neurofilter, and method of training same to operate on image... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Neurofilter, and method of training same to operate on image..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Neurofilter, and method of training same to operate on image... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2606786