JPH05236264A - Color picture device - Google Patents

Abstract

PURPOSE:To reduce the number of times of repetitive learning and to reduce the number of neuron components by providing a subtractor obtaining an error between study data and an output color signal to the device so as to revise the offset. CONSTITUTION:A selection section 4 is switched so that study data being input data of a storage section 3 are inputted to a conversion section 1 and an error between the data converted by the conversion section 1 and study output data from the storage section 3 is obtained by a subtractor 5. Then, each weight and each offset are slightly changed by using the method of steepest descent so that the square error decreases based on the error. In this case, the neural network is structured to be three blocks with respect to an output color signal and an intermediate layer neuron 11h is coupled with only an output layer neuron 11o. Thus, the coupling between final output data and the neuron 11h is strengthened and the coupling with the other neuron 110 is weakened. Thus, the number of times of repetitive learning is reduced and the number of neuron components can be decreased.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a color image device for converting a color image into an electric signal in a color facsimile or a color copying machine.

[0002]

2. Description of the Related Art For example, FIG.
FIG. 3 is a configuration diagram of a signal processing unit of the color copying machine shown in Japanese Patent Publication. A conversion unit 1 receives y (yellow), m (magenta), and c (cyan) as input color signals, corrects colors and gradations using a neural network, and determines toner density. Reference numeral 9 is a host computer that gives learning data to the neural network. FIG. 6 is a block diagram showing a neural network of the conversion unit 1 of the signal processing unit of the color copying machine. The neural network has a three-layer structure, and there is no connection between neurons in the same layer, and each neuron is connected to all neurons in other layers via weights. Here, 11i is an input layer neuron, 11h is an intermediate layer neuron, and 11o is an output layer neuron.

Next, the operation will be described with reference to FIG. Input color signal (ymc) and output color signal (YM
In order to configure the network (FIG. 6) of the conversion unit 1 so as to be suitable for C), learning data (hereinafter, s is attached) representing the relationship between the input color signal and the output color signal is used to calculate each weight of the network and Offset optimization (referred to as learning) is performed first. Each weight and offset are initialized prior to learning. (Step A1). Next, the input color signal (ymc) _s of the learning data is input to the conversion unit 1. (Step A2). The conversion unit 1 calculates using the current weights and offsets to obtain an output.
Conventionally, learning of the neural network of the conversion unit 1 is performed by using back propagation. Neuron 11
Can be represented by a symbol as shown in FIG. Where In is n
The second input, O, represents the output. A weight Wn representing the strength of the coupling is attached to each input. An offset value θ is attached to each neuron, and the output O
Is

[0004]

[Equation 1]

[0005] Here, the function f is generally a sigmoid function f (x) = 1 / (1 + exp (-x)).
Is used. (Step A3) Output data (YMC) _{s of the} learning data and output data (YM) of the conversion unit 1
The error of C) is calculated (step A4). Based on the error obtained in step A4, each weight is reduced so that the squared error is reduced,
Change the offset slightly. The change is made by, for example, the fastest descent method (step A5). Next, the learning data ((ym
c) _s and (YMC) _s ) are updated (step A
6). At this time, it is judged whether all the learning data has finished,
If not completed, steps A2 to A6 are repeated (step A7). When the process is completed for all learning data, it is determined whether the sum of the errors of the output color signals for each learning data is less than or equal to a certain value. If it is larger than a certain value, the learning data is returned to the original step A2
To step A7 are repeated (step A8). If the sum of the errors of the output color signals with respect to the learning data is less than a certain value, it is assumed that the network optimization (learning) is completed.

[0006]

The conversion unit 1 of the conventional color copying machine performs learning by backpropagation and is composed of the neural network having the above structure. Therefore, in order to improve the color reproduction accuracy. There was a problem that the number of neurons in the middle layer had to be increased, and therefore the hardware became large.

The reason will be described below with reference to FIG. Focusing on one of the intermediate layer neurons 11h, 3
It is connected to two output layer neurons 11o. Therefore,
The output of the intermediate layer neuron 11h is input to the three output layer neurons 11o via the weight of each connection. In the case of (a), it is accelerated for the Y output layer neuron 11o, and also for the M output layer neuron 11o.
The output layer neuron 11o of C works suppressively.
In the case of (b), the Y output layer neurons 11o act proactively and the M and C output layer neurons 11o act suppressively. In addition, there are many connection weights, from small to large.

When the weight of the connection of one intermediate layer neuron 11h is determined, the output layer neuron 11o is output.
The degree of influence on is fixed. Since the weight of the connection is determined by the relationship with the learning data of the three connected output layer neurons 11o, it becomes an intermediate value of the three neurons. Therefore, in order to improve the accuracy of color reproduction, it is necessary to increase the number of intermediate layer neurons 11h and prepare a large number of intermediate layer neurons 11h having different degrees of coupling with the three output layer neurons 11o.

Three input layer neurons 11i, nine intermediate layer neurons 11h, and three output layer neurons 11o
When an experiment was performed using a neural network composed of, the average error between the converted value of the input signal and the expected output value was 5%, and the maximum error was 10%.

As described above, in the conventional color / gradation correction using the neural network, in order to improve the color reproduction accuracy, it is necessary to increase the number of neurons in the intermediate layer, so that the hardware becomes large. was there.

The present invention has been made in order to solve the above problems, and an object thereof is to obtain a color image device using a neural network having a high color reproduction accuracy with a small number of intermediate layer neurons.

[0012]

A color image device according to the present invention is based on a conversion unit for converting an input signal into an output signal and a weight of the conversion unit based on an error between a color signal in a storage unit and an output color signal. A weight changing unit to be changed, a storage unit for storing learning data indicating a relationship between an input signal and an output signal, and a subtractor for obtaining an error between the color signal and the output color signal in the storage unit are provided.

Further, in the color image device for correcting the color image signal, the color image signal is provided with a correcting means for correcting the color of the color image signal by the neural network and a learning means for changing the weight and offset of the neural network to optimize the network. The configuration of the neural network of the conversion unit forming the main part of the correction means is divided into n (n ≧ 3) blocks for the output color signal, and the intermediate layer neuron is connected to only one output layer neuron to obtain the color image signal. Is to correct.

[0014]

In the color image device according to the present invention, since the intermediate layer neurons of the neural network formed by the conversion unit 1 are connected to only one output layer neuron, the final output data and the intermediate layer neurons are strongly connected. Therefore, the number of iterations of learning for obtaining a given input color signal and an expected output color signal corresponding thereto is reduced, and the number of constituent neurons can be reduced.

[0015]

EXAMPLES Example 1. An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a configuration of a signal processing unit of the present invention. In FIG. 1, 1 is a conversion unit that converts an input signal into an output signal, 2 is a weight changing unit that changes the weight of the conversion unit 1 based on the error between the color signal of the storage unit 3 and the output color signal, and 3 is an input unit. A storage unit 4 for storing learning data representing the relationship between the color signal and the output color signal, and a selector 5 for switching the input signal between the color signal from the storage unit 3 and the color image signal from a sensor (not shown). Is a subtracter for obtaining an error between the color signal of the storage unit 3 and the output color signal.

FIG. 2 is a diagram showing a neural network of the conversion unit 1. In FIG. 2, 11i is an input layer neuron, 11h is an intermediate layer neuron, and 11o is an output layer neuron. Image data r, g, and b are input to the three input layer neurons 11i, and their outputs are input to all the intermediate layer neurons 11h. Also, one output layer neuron 11o is not all neurons in the intermediate layer, but three predetermined intermediate layer neurons 11h.
The output of is input. Each output layer neuron 11o is
It outputs signals corresponding to R, G, and B. As described above, the conversion unit 1 is composed of a neural network including a combination of neurons.

The operation of the present invention will be described below. Learning of the neural network shown in FIG. 2 is performed by backpropagation. That is, the selector 4 is switched so that the learning data of the input data of the storage unit 3 is input to the conversion unit 1, and the error between the data converted by the conversion unit 1 and the learning data of the output data of the storage unit 3 is subtracted. The weight and offset are changed little by little using the fastest descent method so that the squared error is reduced based on the error obtained by the device 5. On this occasion,
As shown by the broken line and the chain line in FIG. 2, the structure of the neural network is divided into three blocks for the output color signal, and therefore the intermediate layer neuron 11h is connected to only one output layer neuron 11o. , The final output data and the intermediate layer neuron 11h are strongly coupled, and the other output layer neurons 11o are weakly coupled. This is suitable for data that is theoretically separated into three primary colors, such as image data. for that reason,
The number of learning iterations can be reduced and the number of constituent neurons can be reduced.

The relationship between the input data and the output data of the network thus optimized is, for example, as shown in FIG. Here, ● indicates a set (x, y) of the input color signal and the output color signal of the learning data. As shown in FIG.
The feature is that a curve faithful to the learning data can be obtained as an optimal solution.

As a result of the experiment, in the optimized network, the average error between the signal value converted by inputting the input color signal (rgb) of the learning data into the conversion unit 1 and the output color signal (RGB) of the learning data is 3%, maximum error is 5
It was suppressed to%.

In order to convert the input color signal from the sensor, the selector 4 is turned to the input color signal side from the sensor, and conversion is performed by using the optimized network weight and offset.

Example 2. In the first embodiment, the conversion unit 1 and the weight changing unit 2 are provided so that learning is possible. However, as shown in FIG. 4, the conversion unit 1 is composed only of the ROM 8 and learning is performed by an external computer. The color correction may be performed by storing the optimized relationship between the input color signal and the output color signal of the network in the ROM 8 in the form of a table.

[0022]

As described above, according to the present invention, the color correction is made into the structure of the neural network having the structure divided into n blocks for the output layer neuron 11o, and the intermediate layer neuron of the neural network outputs one output. Since it is connected only to the layer neurons, the correspondence between the input color signal and the output color signal, which have a complicated relationship, is realized by a neural network with a small number of the same constituent neurons, and further, the error of each color is reduced to reproduce the color. There is an effect that a highly accurate color image device can be obtained.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a signal processing unit of a color image device according to an embodiment of the present invention.

FIG. 2 is a diagram showing a neural network of a conversion unit 1 of a color image device according to an embodiment of the present invention.

FIG. 3 is a diagram showing a relationship between input data and output data of an optimized network of a color image device according to an embodiment of the present invention.

FIG. 4 is a configuration diagram of a signal processing unit of an image device according to another embodiment of the present invention.

FIG. 5 is a configuration diagram of a signal processing unit of a conventional color copying machine.

FIG. 6 is a configuration diagram of a neural network of a conversion unit 1 of a signal processing unit of a conventional color copying machine.

FIG. 7 is a diagram of symbols showing neurons.

FIG. 8 is a flowchart showing a color correction operation of a conventional color copying machine.

FIG. 9 is a diagram showing a connection between an intermediate layer neuron and an output layer neuron of a signal processing unit of a conventional color copying machine.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Transform section 2 Weight changing section 3 Storage section 5 Subtractor 11h Intermediate layer neuron 11o Output layer neuron

Claims (2)

[Claims] 1. A color image apparatus for correcting a color image signal, comprising: a correction means for correcting the color of the color image signal by a neural network; and a learning means for changing the weight and offset of the neural network to optimize the network. A conversion unit that converts an input signal into an output signal, a weight changing unit that changes the weight of this conversion unit based on the error between the color signal and the output color signal in the storage unit, and learning data that shows the relationship between the input signal and the output signal. A color image device comprising a storage unit for storing and a subtracter for obtaining an error between a color signal of the storage unit and an output color signal. 2. A color image apparatus for correcting a color image signal, comprising: a correction means for correcting the color of the color image signal by a neural network; and a learning means for changing the weight and offset of the neural network to optimize the network. The configuration of the neural network of the conversion unit forming the main part of the correction means is divided into n (n ≧ 3) blocks for the output color signal, and the intermediate layer neuron is connected to only one output layer neuron to obtain the color image signal. A color image device characterized in that