JPH04142877A - Color image forming device - Google Patents

Abstract

PURPOSE:To prevent the deterioration in the color reproducibility by revising configuration of learning data based on the degree of achievement of learning for each preset conduction in the case of implementing the learning by the back propagation method. CONSTITUTION:A printer control signal generated by a learning system 20 is given to a recording system 40 to output a color hard copy and its output is read photoelectrically by a color separation system 30. Then the learning system 20 uses a printer control signal given to the recording system 40 as a teacher signal, uses a color separation signal outputted from the color separation system 30 as an input signal and uses the input signal and the teacher signal as learning data so as to apply learning by the back propagation method and the configuration of the learning data is revised based on the degree of achievement of learning for each preset condition during the continuity of learning. Thus, an error as the entire learning data is reduced and the deterioration in the reproducibility around a specific color is improved.

Description

Detailed Description of the Invention "Field of Industrial Application" The present invention applies a back propagation method to a color correction system that photoelectrically reads a color image and converts the obtained color separation signals into control signals given to a printer. The present invention relates to a color image forming apparatus using a neural network that performs learning using a neural network.

[Prior Art] Conventionally, in order to obtain a hard copy of a photoelectrically read color image using a color printer, a color correction system is indispensable because the spectral reflection characteristics of ink are insufficient. A color masking method is widely used as a color correction method. To determine the matrix for the color masking method, first create evenly distributed printer control signals and determine the print result, as described in Journal of the Institute of Image Electronics Engineers, Vol. 14, No. 5, pages 298 to 307. The printed results are read photoelectrically to obtain color separation signals. A matrix is determined by finding the nonlinear correspondence between these pairs of printer control signals and color separation signals using the method of least squares.

Neural network is a journal of the Institute of Image Electronics Engineers, Proceedings 89-0
5-04, pages 19 to 24, it is possible to replace the color masking method as a color correction method, and the printer control number and color separation signal for matrix determination in the color masking method are , it can be used as an input signal and a teacher signal in the backpropagation method.

The backpropagation method is based on the rPDP model”D.
Written by E. Ramelhart et al., Translated and supervised by Toshi Keri, Industry Map;! F
1989, Chapter 8. The method is described below.

Each component that makes up a neural network is called a unit or neuron. As shown in Figure 4, neurons have long, slender projections that branch into several branches, with the tips serving as synapses, which are connectors to other neurons.

The excitation output P of neuron i is P, = f(E,)... (1) However, f(E,)... (2) 1+exp(-El) ΣW, P, from neuron j to neuron 1 The synaptic load on O○ < W,, < OO.

When the synaptic load WlI is negative, it means an inhibitory connection, and when it is positive, it means an excitatory connection.

This method deals with networks constructed by hierarchically connecting neurons.

Consider a network consisting of L layers. The first layer is called the input layer, the second layer (L is the largest layer) is the output layer, and the second layer to the first layer is called the input layer.
The layers up to 7-1 are called middle layers.

The process in which the excitatory output of a neuron acts as a force for the next neuron, causing the stimulus to propagate one after another, is called network firing. Firing begins by applying an external stimulus to neurons in the input layer.

One firing ends when the stimulus reaches the output layer neuron and the excitation output of the output layer neuron is determined.

The excitatory output P of a neuron is a function of the input stimulus and the synaptic load. The purpose of network learning is to change the synaptic load so that the excitation output P of the output layer neuron approaches the ideal value when a set of input stimulus and ideal excitation output is given. During learning, the error between the output of the output layer neuron and the ideal excitation output is propagated in the opposite direction of the stimulus transmission path to obtain the error of the excitation output of the middle layer neuron, and the synaptic loads are updated to reduce these errors.

The ideal excitement output is called teacher confidence.

The training data consists of a pair of input signals of human layer neurons and teacher signals. During learning, large amounts of data are repeatedly given. First, prepare a piece of training data and apply a human stimulus to the human layer neurons to start firing. When firing ends, synaptic loads are updated according to the following formula.

W + + (t +1) −W + + (t)
1ηδ1+P,... (4)η The fluctuation rate δ represents the error in the excitation output of neuron i, and for the output layer neuron, δ1= (T P f' (E,), and for the middle layer neuron, δ1=Σ Wkl δk f '(El).

however, : Ideal excitation output (teacher signal) df(E) of neuron i f.

(E) E The above is one cycle of learning, and learning progresses by repeating this cycle.

The learning termination conditions are: 1. When the amount of change in synaptic load falls below a certain threshold.

2. When the correct answer rate exceeds a certain threshold.

3. When a preset number of learning cycles are completed.

and so on.

[Problem to be solved by the invention] However, if learning is performed using the backpropagation method on all of the large amount of prepared learning data,
The error between the excitement output (output signal) obtained from the input data and the teacher signal varies depending on the individual learning data. When viewed as a color correction system in a color image forming apparatus, when an input signal close to learning data I<] with a large error is received, the color reproducibility is degraded.

The present invention has been made in order to solve the above-mentioned problems, and reduces the error of the learning data as a whole when performing learning using the backpropagation method, and improves the decrease in reproducibility near a specific color. The purpose is to

[Means for Solving the Problem] In order to achieve this object, the present invention includes a color separation system that photoelectrically reads a document, a color correction system that uses a neural network, and a color hard copy that outputs a color hard copy from a printer control signal. In a color image forming apparatus that has a recording system and a learning system that performs neural network learning using the backpropagation method, when performing learning using the backpropagation method, the degree of learning achievement is determined based on preset conditions. The structure of the learning data is changed as follows.

[Operation] In the present invention having the above configuration, a printer control signal generated by the learning system is applied to the recording system to output a color hard copy, and the output is photoelectrically read by the color separation system. The learning system uses a printer control signal directed to the recording system as a teacher signal, uses a color separation signal output from the color separation system as an input signal, and performs learning by a backpropagation method using the input signal and the teacher signal as learning data. However, during the continuation of learning, the configuration of the learning data is changed based on the learning achievement level for each preset condition.

[Example] Hereinafter, an example embodying the present invention will be described with reference to the drawings.

FIG. 1 shows a color correction system using a neural network in a color image forming apparatus such as a digital color copying machine. - As an example, an input layer (1
1), an intermediate layer (12) having three or more units coupled to each unit of the input layer (11), and an intermediate layer (12) having three or more units coupled to each unit of the input layer (11).
It is constituted by a three-layer neural network including an output layer (13) which is connected to each unit of (2) and outputs, for example, a CMY three-value printer control signal.

As shown in Figure 2, the overall configuration of the color image forming apparatus includes a color correction system (10), a learning system (20), and a color separation system (3).
0) and a recording system (40).

The learning system (20) performs learning processing on the neural network in the color correction system (10) according to the backpropagation method, and generates a printer control signal as a "teacher signal" and sends it to the recording system (40). The color hard copy printed by the recording system (40) is read from the color separation system (30) and used as an "input signal".

For example, the recording system (40) is cyan/magenta/yellow 3
If 256 levels can be expressed for each color using colored ink, the total number of colors will be approximately 16.7 million colors, but the learning system (2o) will be able to express cyan, magenta, yellow from the expressible levels. A learning data set consisting of a plurality of pairs of "one teacher signal" and "input signal" is created by sampling at equal intervals. For example, if 12 samples are sampled from each color, CMYS values are obtained as 12 cubes, that is, 2744 "teacher signals", and pairs with the same number of "input signals" RGBa values read from the color separation system are obtained.

The learning system (20) extracts each pair of "teacher signal" and "input signal" in the learning data set and converts them into "human signals".
That is, the RGB values read from the color separation system are converted to the color correction system (10
) to the input layer. Then, the neural network calculates C as an output signal according to this RGBS value.
Output the MYa value. The backpropagation method is applied by comparing this CMYa value and the CMYS value of the "teacher signal."

For example, in the second example, the learning count is counted once after the backpropagation method is applied to the learning data set consisting of 2744 pairs. In repeating learning, the learning data is reconstructed under appropriate conditions, for example, every 100 learnings.

FIG. 3 shows a flowchart for reconstructing learning data.

In the reference value calculation (Sl), the error E between the output signal generated by the correction system (10) and the "teacher signal" for the "input signal" is calculated for all learning data, and the average value is calculated. Let it be the reference value M.

△E- (c.

C,)2+(M.

M,)2+(Y.

Yl)2 c.

I M.

Y.

Output value of each neuron in the output layer by the neural network Teacher signal (printer control signal given to the recording system) Σ ΔE ・(9) n Number of data in the learning dataset In data extraction (S2), reference value calculation (Sl) The error E for each pair of learning data obtained in step 1 is compared with the reference value M, and learning data pairs having an error greater than or equal to the reference value are extracted. In the learning data reconstruction (S3), the learning data selected in the data extraction (S2) is added to the already existing learning data set. Although the training dataset reconstructed by (S3) contains two or more pieces of the same training data,
By continuing learning using this learning data set, the backpropagation method is applied more often to learning data that exceeds the reference value, contributing to overall accuracy improvement.

In addition, as a second example of reconstructing learning data, the third example
The learning data is sorted in ascending order by the value of E to the error calculated in the standard value calculation (Sl) in the figure, and the error value of the appropriate rank from the binary values is used as the standard value, and in (S2),
In (S1), the learning data that is equal to or greater than the reference value obtained is extracted, and in (S3), the existing learning data set is discarded and (S1) is extracted.
The learning data extracted in 2) is used as a new learning data set. In this second embodiment, the learning data that is less than the reference value calculated in (Sl) is reduced, which is also effective in shortening the learning time.

After the learning is completed, the color original is read by the color separation system, and this is used as an input signal, and the color correction system neural network outputs an output signal.This output signal is given to the recording system as a printer control signal, so that the color original is read. You can obtain color hard copies with more faithful color reproduction.

In this embodiment, if the characteristics of the color separation system or recording system change due to aging or the like, it is possible to always obtain highly accurate color hard copies by redoing the learning.

[Effects of the Invention] As is clear from the detailed description above, according to the present invention, when performing learning by the backpropagation method, learning data is calculated based on the learning achievement level for each preset condition. By changing the configuration, it is possible to reduce the error between the output signal generated during learning and the teacher signal, which has the excellent effect of preventing a decline in color reproducibility near a specific color. There is.

[Brief explanation of drawings]

1 to 4 show an embodiment embodying the present invention. FIG. 1 is a block diagram of a color correction system using neural network software, and FIG. 2 is a block diagram of a color image forming apparatus. FIG. 3 is a flowchart showing the procedure for reconstructing learning data, and FIG. 4 is a schematic diagram of a neuron. In the figure, 10 is a color correction system, 11 is an input layer of the neural network, 12 is an intermediate layer of the neural network, 13 is an output layer of the neural network, 20 is a learning system, 30 is a color separation system, and 40 is a recording system.

Claims (1)

[Claims] 1. A color separation system that photoelectrically reads an original image, a color correction system that uses a neural network, a recording system that generates a color hard copy from a printer control signal, and a neural network that uses a backpropagation method. A color image forming apparatus having a learning system that performs learning is characterized in that when performing learning by a backpropagation method, the configuration of learning data is changed based on the learning achievement level for each preset condition. Color image forming apparatus.