US 5,491,776 AGranted

Signal processing apparatus and learning method therefor

AssigneeKawasaki Steel Corporation

InventorRyoichi Dangi

Life of this patent

Estimating the term…

Loading patent…

Abstract

Mathematical calculation processing is executed at a high speed for an input signal of image information or the like. Weight coefficients corresponding to a DCT (discrete cosine transform) are set in a neural network NN of an n-layer structure including an input layer I1, (an intermediate layer H1) and an output layer O1. The input layer I1 is constituted from a number of input interfaces equal to the number of picture elements included in a block 12, and the output layer O1 is constituted from a number of neurons equal to the number of DCT coefficients to be outputted. When input information of the block 12 is inputted to the neural network NN, DCT coefficients a(0) to a(63) which are results of DCT processing of the image information are outputted immediately. The weight coefficient is determined by learning using, as teacher data, DCT coefficients obtained by actual mathematical calculation of the same image information as that inputted to the neural network NN by a DCT transform section.

A signal processing device for processing a portion of an input signal, the input signal portion having a predetermined size, the signal processing device comprising:

output selection means for selecting one of a first output signal and a second output signal as a final output signal based on a predetermined threshold;
first output signal generating means for computing the first output signal from the input signal portion based on a predetermined signal transform algorithm;
algorithm selection means for selecting one of a plurality of signal transform algorithms as the predetermined signal transform algorithm in the first output signal generating means;
a neural network operating concurrently with the first output signal generating means and generating the second output signal from the input signal portion based on a current set of learned weighted coefficients, the neural network comprising:
an input layer of neurons having a number of neurons equal to the predetermined size of the input signal portion;
an output layer of neurons generating the second output signal, each neuron of the output layer connected to each neuron of the input layer;
the current set of learned weighted coefficients, wherein each output layer neuron receives an output from each one of the input layer neurons by multiplying an output of each one of the input layer neurons by one of the current set of learned weighted coefficients;
learning means for generating the current set of learned weighted coefficients based on a comparison of the first and second output signals, the learning means operating concurrently with the first output signal generating means;
storage means for storing a plurality of sets of learned weighted coefficients, each set of learned weighted coefficients corresponding to one of the plurality of signal transform algorithms;
and
setting means for setting one of the plurality of sets of learned weighted coefficients as the current set of learned weighted coefficients based on the signal transform algorithm selected by the algorithm selection means.

8 more elements

Facts

Application

08/290,080Filed 12 Aug 1994 · continuation

Priority

JP 3-2191335 Aug 1991

Granted

13 Feb 1996

Term ends

Not estimatedThe source records don’t support an estimate.

Classification9 codes

G06F 17/147
Fourier, Walsh or analogous domain transformations , e.g. Laplace, Hilbert, Karhunen-Loeve, transforms: discrete orthonormal transforms, e.g.…
G06N 3/04
Neural networks: architecture, e.g. interconnection topology
G06N 3/0499
Architecture, e.g. interconnection topology: feedforward networks
G06N 3/09
Learning methods: supervised learning

Abstract

Claims1