JPH0451376A - Picture processor - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an image processing device that restores an original multivalued image from a binarized image.

[Prior Art] Conventionally, as a method of restoring an original multivalued image from a binary image, there is a method of applying a rectangular filter to a binary image and converting it into multivalued images.

Furthermore, a method has been proposed in which multi-value data is estimated from a binary pattern of a rectangular area in a binarized image using a neural network.

[Problems to be Solved by the Invention] The method using the neural network described above achieves high-quality multi-value restoration through learning.

However, multi-value restoration using a neural network depends on the binarization method of the binarized data used for learning, so in the current situation where there are various binarization methods,
For example, even if a neural network trained on an image that has been binarized using the error diffusion method is used to restore an image that has been binarized using the dither method, good multivalued data may not necessarily be obtained. was there.

Therefore, it is desirable to perform multilevel conversion using a neural network that has been trained to correspond to the binarization method of the binarized image that you want to restore. It is not always clear whether the

[Means for Solving the Problems] In order to solve the above problems, the image processing device of the present invention has the following features:
an input means for inputting a binarized image; a plurality of multi-value conversion means each corresponding to a specific binarization method and performing multi-value restoration processing of the binarized image using a neural network; and the input means The binarization method of the binarized image input from
A discriminating means for discriminating by a neural network, a selecting means for selecting one of the plurality of multi-value encoding means based on the discrimination result by the discriminating means, and a multi-value encoding means selected by the selecting means, and control means for controlling the binarized image inputted by the input means to be multivalued.

[Operation] According to the present invention, the binarization method of the input binarized image is determined by the discriminating means, and the binarized image is multiplied by the multi-value converting means corresponding to the determined binarization method. Value.

[Embodiment] FIG. 1 is a block diagram showing one embodiment of the present invention,
FIG. 2 is a flowchart 1 showing the processing procedure of the image processing apparatus of the present invention. Steps 201 to 207 written on the left side of FIG.
08 to 212 are the processing procedures for multi-value quantization. In the present invention,
First, a binarization method is estimated, and then multi-value processing is performed.

Hereinafter, the functions of each part shown in FIG. 1 will be explained while explaining the processing procedure shown in FIG. 2. In Figure 1, 101 is a FIFO
It is assumed that the line buffer is composed of a line buffer that receives input of binary data and stores data for four raster lines. Reference numeral 102 denotes a data latch that latches data of four pixels for each line of four lines from the line buffer 101. Therefore, binary image data of a 4×4 window is obtained from this data latch 102 (
Step 202).

This 4×4 16-bit data is given as an address of a ROM format conversion table 103. As will be described later, the conversion table 103 is determined by a neural network, and converts input data into data corresponding to one of predetermined binarization methods, into a binary value estimated from the input data. output as a converted format (step 203). The present embodiment is configured to correspond to the error diffusion method, the Bayer dither method, the Fat-Jung dither method, and the simple threshold individuality method, as shown in FIG. The counter 109 independently counts the four types of data corresponding to the four binarization methods. For example, when it is determined in the table 103 that the input image is obtained by the error diffusion method, A portion counter is incremented (step 204). Step 20
In step 5, it is checked whether steps 202 to 204 have been completed for all images, and if they have not been completed, the process is further repeated.

When the above processing is completed for all images, selector 1
At step 10, the one with the largest count value by the counter 109 is estimated to be the binarization method for this image, is selected, and is set in the latch 111. The set data is used as a selection signal for the selector 108, which will be described later (step 206). The above is the estimation process of the binarization method.

Next, the purpose of multivalue processing is performed.

First, in step 209, 4×4 image data is extracted. This is the same process as in step 202 described above.

The extracted image data is input to each of multi-value conversion tables 104 to 107 stored in ROM. Tables 104 to 107 each correspond to the four binarization methods described above, and are tables for multi-value processing using a neural network, as will be described later.

The output of each table 104 to 107 is input to a selector 108. The selector 108 selects one of the outputs from the tables 104 to 107 in response to the selection signal from the latch 111, and outputs it as final multivalued data (step 210). The output selected here is the selection unit 11
This is the output from the table corresponding to the estimated binarization method selected by 0.

The above steps 209 to 210 are repeated until it is determined in step 211 that all images have been completed.

Control of each section described above is performed by a CPU (not shown).

Next, estimation of a binarization method using a neural network and multi-value processing using a neural network will be explained.

First, a general learning procedure in a pack propagation neural network will be explained using FIG. 4(a) as an example.

In the neural network shown in Figure 4(a),
In the input layer 401, the output (i-ou
t) 404 is input to the intermediate layer 402 consisting of one layer as the number of neurons jj), and the output from the intermediate layer 402 (j-ou
t) 405 is inputted to the output layer 403 as the number kk), and the output layer 403 outputs an output (k-out) 406. In addition, 407 is the ideal output (1deal-
out).

In the neural network, input data and an ideal output (1 deal-out) for the input data are prepared, and by comparing this with the output (k-out) 406, the connection strength WJI [jj, iil (in the figure) in the middle layer is determined. 4
08), the coupling strength WkJ [kk, jj
Determine l (409 in the figure).

The learning procedure using the neural network described above is
This will be explained in more detail using the flowchart shown in FIG.

First, in step 5401, the weighting coefficient (coupling strength) W
Give the initial values of Jr[jj+iil, WkJ[kk+jul. Here, in consideration of convergence in the learning process, a value in the range of -0.5 to +0.5 is selected.

Next, in step 5402, the learning input data 1-ou
t(i), and in step 5403 this data 1-
Set out(i) to the input layer. Also, step 5
At 404, an ideal output (ideaL-out) for the input data 1-out(i) is prepared.

Therefore, in step 5405, the intermediate layer output j-out
Find (j).

First, the data Lout(i) from the input layer is multiplied by the weighting coefficients W and l of the intermediate layer, and the summation Sumx is expressed as SumpJ
=ΣW J l [jj, il * 1-out(
i) and then add sigmo to this SumtJ
By applying the id function, the output j-ou of the j-th intermediate layer is
Calculate t(j) by:

Next, in step 5406, -out(
Find k). This procedure is similar to step 5406.

That is, the output j-out(j) from the intermediate layer is multiplied by the weighting coefficient W0 of the output layer, and the total sum Sumrk is calculated as Sumrk.
rk=ΣW, [kk, jjl * j-out(j
) and then add sigmoi to this SumFk
Apply the d function to the output of the k-th hidden layer -out
Calculate (k) by Note that this output value has been normalized.

Next, in step 5407, the output obtained above -out(k) is compared with the ideal output 1deal-out(k) prepared in step 5404, and the teacher signal teach-k(k) of the output layer is As, teach-
k(k) = (ideal-out(k)-k-o
ut(k))* to -out(k)*(1-k-out(
Find k)). Here, k-out(k)*(l −
k-out(k)) is the sigmoid function -out
It has the significance of the differentiation of (k).

Next, in step 8408, the variation width of the weighting coefficient of the output layer △WkJ[kk+ jul is calculated as △WkJ[kk, jj
l = β*j-out(j)*teach-k(k)+
Calculate by α*ΔWllJ[kk, jj]. Here, α is a stabilizing constant, and β is a constant called a learning constant, which plays the role of suppressing sudden changes.

In step 5409, the weighting coefficient WkJ [kk, jj] is changed to Wb, + [kk,
jj] =WkJ[kk, jul+△Wk, [kk,
jj] and update. In other words, study.

Next, in step 5410, the intermediate layer teacher signal teac
h−,j (calculate l. For that purpose, first, SLl
mBj"Σteach-k(k)* W l, J[jj
, ii, calculate the inverse contribution from the output layer to each element of the intermediate layer. Next, from this SllmBj, the intermediate layer teaching signal teach-j (j) is calculated using the following formula.

teach-j(j) = j-out(j)*(1-
j-out(j)l*suma4 Then, step 541
1, the change width of the weight coefficient of the middle layer △W,z [, +j,
+i], △WJ+ [jj, fil =β*1−o
ut(i)*teach-j(j)+a*△WJI [
Calculate by jj, fil.

In step 5412, based on this change width, the weighting coefficient WJI [jj, fil, W, [jj, f
il = W, 1+ [jj, ij] △W J +
[,+J,i]] is updated. In other words, learn.

In this way, in steps 8401 to 412, the weighting coefficient WJ is calculated from one set of input data and the ideal output for this
, and W 3. , and were learned once.

In step 5413, it is checked whether the weighting coefficients have sufficiently converged through the learning described above, and if not, steps 3401 to 412 are repeated.

The above is an explanation of the neural network learning procedure based on the pack propagation method.

The learning described above is a preparatory stage for processing, and in actual processing, only the determined weighting coefficients and a table of processing results for all possible inputs using the weighting coefficients will be used. Become.

Next, a case where the above-mentioned "learning" is performed on a neural network for estimating a multivalued image from two images in this embodiment will be explained.

First, the input data is the value of a pixel (0 or 1, respectively) within a 4×4 window of image data that has been binarized using a binarization method and is a learning target.

Therefore, the number of neurons in the input layer is 16, the number of neurons in the output layer is 12 because the multi-value output is for one pixel, and the number of neurons in the middle layer is arbitrary, but in this example, it is 12.
Individual.

On the other hand, the ideal output is multivalued image data, which is the original image of the input data that has been binarized.

In addition, as for how to select input data, a learning pixel is selected at random, and two 4×4 windows containing that pixel are provided.

Using the above, weighting coefficients are determined by the learning procedure described above. (Therefore, the connection of the neural network is determined.) In this embodiment, the processing of this neural network is tabulated. To do this, we need to generate all 4×4 input patterns (21
This can be obtained by determining the output for 6 patterns) and storing it in a ROM.

In this embodiment, in order to support four binarization methods, the above processing (learning and ROMization) is performed independently for each method, and four conversion tables 104 to 10 are created.
Prepare 7.

On the other hand, learning for estimating the binarization method is also performed in the same way.

First, the input data is the value of a pixel (0 or 1, respectively) within a 4×4 window of image data that has been binarized using four binarization methods to be studied.

Therefore, the number of neurons in the input layer is 16, the number of neurons in the output layer is 4 depending on which of the four binarization methods is used, and the number of neurons in the middle layer is arbitrary, but in this example, the number of neurons is as described above. As in the case of , there are 12 pieces.

On the other hand, as an ideal output, the input data binarization method is 4
This is a logical value indicating which one of the two.

In addition, as for how to select input data, as in the case described above, a learning pixel is randomly selected and a 4×4 window containing that pixel is provided.
One neural network is trained by randomly selecting which binarization method to provide data.

In the above embodiments, the neural network processing was performed using a conversion table, but a neurochip having the determined weighting coefficients may also be used.

The window size is not limited to 4x4, but also 5x5.3x.
It may be 4th grade.

Further, the window size may be different between the binarization method estimation and the multi-value image estimation, and furthermore, the window size may be different depending on the binarization method in the multi-value image estimation.

[Effects of the Invention] As explained above, according to the present invention, first, the binarization method of the input binarized image is estimated, and multi-value conversion is performed according to the binarization method. High-quality multivalued images can be obtained.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram of an image processing apparatus according to an embodiment, and FIG.
Flowchart of the overall process. Figure 3 is a diagram showing the relationship between the conversion table and the binarization method. Figure 4 (a) is a conceptual diagram of the neural network. Figure 4 (b) is the learning of the neural network. Flowchart of Procedures FIG. 5 is a diagram showing an example of a window. 101... Line buffer 102... Data latches 103-107... Conversion table If) 8... Selector 109... Counter 110... Selection section 111... Latch Nanato 4ogo ob

Claims (1)

[Claims] An input means for inputting a binarized image, and a plurality of multi-value restoration processes each of which performs multi-value restoration processing of the binarized image using a neural network, each corresponding to a specific binarization method. means for determining the binarization method of the binarized image input from the input means using a neural network; and one of the plurality of multi-value converting means based on the determination result by the discriminating means. and a control means for controlling the binarized image inputted by the input means to be multivalued by the multivalue conversion means selected by the selection means. image processing device.