Image analysis – Learning systems – Neural networks
Reexamination Certificate
1997-11-13
2001-06-05
Chang, Jon (Department: 2623)
Image analysis
Learning systems
Neural networks
C706S026000
Reexamination Certificate
active
06243490
ABSTRACT:
BACKGROUND OF THE INVENTION
This invention relates to neural networks well suited for various types of data processing and, more particularly, to techniques for reducing the size of neural-network circuitry and raising the speed at which arithmetic operations are performed in a neural network. Furthermore, the invention relates to image processing using a neural network.
The fact that various processing operations are performed by neural networks has been the subject of research in recent years. In order to raise the processing accuracy of processing using a neural network, each element (neuron) constituting the neural network receives a large number of items of input data, subjects these items of input data to processing and outputs the results. Specifically, a neural network generally comprises an input layer, an intermediate layer and an output layer. For example, in a case where the number of neurons in the input layer is i and the number of neurons in the intermediate layer is j, i-number of multiplications and i-number of additions are performed by one neuron of the intermediate layer, and these multiplications and additions are performed j times by the intermediate layer as a whole. Furthermore, j-number of multiplications and j-number of additions are performed by one neuron of the output layer. These arithmetic operations, which are enormous in number, are performed whenever there is an input. In other words, the multiplication and addition operations have the form of a network.
Accordingly, when a neural network is merely constructed of hardware, the scale of the circuitry is very large and processing speed is reduced as a result.
Next, consider image processing by a neural network. In examples of such processing, image compression and character recognition of Chinese characters, numerals and the like have been attempted. However, studies relating to the processing of natural images by neural network, spatial filtering processing and image conversion have not yet been carried out. A review of masking processing is reported in the paper of a research group of the
Image Electronics Engineers of Japan
(1989).
FIG. 1
is a simplified block diagram showing such a neural network for masking processing disclosed in the papers of the aforesaid research group. As shown in
FIG. 1
, an input layer comprises input elements for inputs Y, M, C, and an output layer similarly comprises output elements for outputs Y′, M′, C′ obtained as the result of masking processing. Two layers are provided as intermediate layers, each of which comprises 10 or 20 nodes (neural elements, or neurons). In
FIG. 1
, only one intermediate layer is shown for the sake of simplicity.
In the arrangement of
FIG. 1
, the intermediate layer has a large number of elements, namely ten, and the intermediate layer is a double layer. For this reason, the length of time needed for learning and execution is very long.
The present inventors have given careful consideration to the foregoing points. Specifically, in the prior-art example of
FIG. 1
, the neural elements of the preceding intermediate layer and the neural elements of the succeeding intermediate layer are coupled on an equivalent basis. Similarly, the neural elements of the succeeding intermediate layer and the neural elements of the output layer are coupled on an equivalent basis.
Consequently, when the coupling constants are learned, a sense of direction cannot be given to the convergence of the learning of the coupling constants because there is no difference between the neural elements of the intermediate layers.
The Y, M and C image signals are mutually independent in nature. At learning, image signals for learning Y, M and C are applied to respective ones of the Y, M and C input elements of the input layer, and ideal signals Y′, M′ and C′ resulting from masking are applied to respective ones of the Y′, M′ and C′ output elements of the output layer.
In other words, when the Y, M and C learning signals having mutually independent properties enter the neural network, convergence at learning of the coupling coefficients is in a state of disorder with regard to the coupling coefficients between the input layer and the intermediate layers and the coupling coefficients between the preceding stage of the intermediate layer and the succeeding stage of the intermediate layer. Accordingly, convergence requires a great deal of time. In addition, as the result of convergence, only the learned results regarding solely the coupling coefficients between the elements of the succeeding intermediate layer and the elements of the output layer contribute to the formation of a neural network for masking processing while maintaining the aforementioned mutually independent properties. More specifically, coefficients other than the coupling coefficients between the elements of the intermediate layers and the three elements of the output layer are decided so as to be favorable for all of the outputs Y′, M′, C′ and Bk′. Therefore, even if, say, the ideal output of Y′ has excellent accuracy, the other ideal outputs exhibit poor accuracy, or vice versa. Even if learning is repeated, a situation arises in which the coupling coefficients oscillate with a certain fixed range.
Accordingly, when it is attempted to suppress the range of oscillation and execute highly precise masking, a dilemma is encountered in which the intermediate layer must be made a double layer, for example, and the number of elements in each intermediate layer must be increased. That is, since learning for masking processing is substantially decided only between the elements of the succeeding intermediate layer and the elements of the output layer, the number of intermediate-layer elements must be increased and, as a result, convergence of the learning process takes considerable time. In addition, since the result of the foregoing is that the coupling coefficients between the input and intermediate layers do not participate in masking processing, a high precision cannot be obtained in such masking processing. Greater accuracy cannot be hoped for even if the number of intermediate-layer elements is increased further.
The foregoing problem essentially arises between image signals having little mutual participation, as between like ones of the image signals. Therefore, the problem does not arise only in masking processing and only when using the image signals Y, M and C; it can arise also in many types of image processing in general.
Considered next will be image processing such as a technique for binarizing a multivalued image, a technique for restoring multivalued image data from binary image data, and a technique for spatial filtering.
In a case where a multivalued image is binarized using a conventional binarizing method, such as the dither method, the resolution of character portions declines and edge components are lost and tend to deteriorate. With the ED (error diffusion) method, stripe patterns peculiar to binarization are produced, and the picture characteristic is not always good. Though the deterioration of character portions is less than with the dither method, the improvement is only partial. for a sufficient understanding of the ED method, see Floyd Stelnloerg; SID, International SYMP. Syp. Digest of Technical Papers 4-3 April '75.
A technique for restoring a multivalued half-tone image from a binary image using a neural network has not yet been realized. In the prior art, the best that can be done is to restore a multivalued half-tone image by applying a smoothing filter to a binary image.
In such conventional restoration of a multivalued image by a smoothing filter, the resolution of character portions declines and portions of the image having a smooth density gradient do not develop a sufficient number of tones. Accordingly, a technique for restoring a multivalued image from, say, a binary image by image processing using a neural network is eagerly awaited. The characterizing feature of a neural network
Canon Kabushiki Kaisha
Chang Jon
Fitzpatrick ,Cella, Harper & Scinto
LandOfFree
Data processing using neural networks having conversion... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Data processing using neural networks having conversion..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Data processing using neural networks having conversion... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2532179