JPH0451374A - Picture processor - Google Patents

Abstract

PURPOSE:To much more accurately reproduce an original multivalued picture by executing a discrimination processing of a character thinning part or a picture part and respective restoration processings of the character thinning part and the picture part in the restoration processing of the multivalued picture from a binarized picture by means of different neural networks. CONSTITUTION:A latch 102 latches data on seven picture elements at every data of seven lines from a line buffer 101. Data of 49 bits of 7X7 is inputted to an area judgement neuro chip 103. The neuro chip 103 outputs '0' when data is judged to be a halftone picture and output '1' when it is judged to be a thinning part. Data of 25 bits is inputted to both a multivalueing neuro chip for picture 105 and a multivalueing neuro chip for thinning part 106 through a latch 104. The neuro chips 105 and 106 output multivalued data of eight bits. A selector 107 is controlled by the neuro chip 103 and outputs data of the neuro chip 105 when it is the halftone picture and data of the neuro chip 106 when it is the thinning part.

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 for restoring an original multivalued image from a binarized image, a rectangular smoothing filter circuit is used, or the processing results of the filter are stored in a ROM or RAM.
This is done by having it as an LUT. FIG. 7 shows an example of a device that performs multi-value conversion using an LUT obtained from a 3×3 filter.

In the figure, 701 is a line buffer configured with, for example, a FIFO or the like, which receives input of binary data from an image input device (not shown) and stores data for three raster lines. 7.02 is a data latch that latches three lines of data from the line buffer 701 and three pixels of data for each line. Therefore, binary image data of a 3×3 window is obtained from this data latch 702. This 3×3 9-bit data is given, for example, as an address of a ROM-format LUT 703. The LUT 703 is determined by the output of the smoothing filter, and outputs multivalued data of 256 gradations (8 pips) for input data.

[Problems to be Solved by the Invention] However, in the above method, for example, as shown in FIG.
When processed using the 3×3 smoothing filter shown in the figure, halftone image parts are multivalued relatively well, but edge parts where there is a large difference in density, such as thin character lines, become blurred. . FIG. 9 shows how the edge portion is processed using the 3×3 smoothing filter shown in FIG. 8(a). On the other hand, the 8th
If processing is performed using a smoothing filter that focuses on the pixel of interest in order to strengthen thin line areas, as shown in Figure (b), the above-mentioned blurring will be eliminated, but the graininess peculiar to binary images will occur in the halftone areas. remains strong, resulting in differences in density in smooth areas such as the background and human skin. FIG. 1O shows how the halftone portion is processed by the 3×3 smoothing filter shown in FIG. 8(b).

A possible solution to this problem is to distinguish the regions and perform multi-value processing on the thin line part and the image part separately, but the method for separating regions from binary image data has not yet been established. It is.

[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 binary image data, a discriminating means for determining whether the input binary image data is a thin character line portion or an image portion using a neural network, and a multi-value restoration process for the thin character line portion. , a first multi-value conversion means that performs multi-value restoration processing on the image portion using a neural network, and a second multi-value conversion means that performs multi-value restoration processing on the image portion using a neural network, and control means for controlling the second multi-value conversion means to perform multi-value restoration processing.

[Operation] According to the present invention, in order to estimate a multivalued image from an input binarized image, the area between the thin character line part and the image part is discriminated using the neural screw 1 to work, and 'I 'll
Based on the separate results, multi-value conversion is performed by multi-value conversion means using a neural network provided corresponding to the thin character line portion and the image portion, respectively.

[Example code] FIG. 1 is a block diagram showing one embodiment of the present invention.

In FIG. 1, reference numeral 1 denotes a line buffer composed of FIFOs 1a to 1d, which receives a binary data input from an image input device (not shown) and stores data for four raster lines.

2 is a data latch that latches 4 lines of data from the line buffer 101 and 4 pixels of data for each line. Therefore, from this data latch 2, binary image data of a 4×4 window is obtained.

This 4×4 16-bit data is given as an address of the conversion table 3 in ROM format. The conversion table 3 is determined by a neural network, as will be described later, and outputs multi-value data of 256 gradations (8 bits) for input data.

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

First, 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. 2(a) as an example.

In the neural network shown in Figure 2(a),
In the input layer 201, the output (i-ou
t) 204 is input to the intermediate layer 202 consisting of one layer as the number of neurons jj), and the output from the intermediate layer 202 (j-ou
t) 205 is inputted to the output layer 203 as the number kk), and the output layer 203 outputs an output (k-out) 206. In addition, 207 is the ideal output (1deal-
out).

In the neural network, input data and an ideal output (1deal-out) for the input data are prepared, and by comparing this with the output (k-out) 206, the connection strength WJI [j, Liil (in the figure) in the intermediate layer is determined. 2
08), coupling strength W8゜[kk, ju
Determine l (209 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
JI [jj, iil, WkJ[kk, jjl
Give the initial value of . Here, a value in the range of -0.5 to +0.5 is selected in consideration of convergence during the learning process.

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, data 1out(i) from the input layer is multiplied by the weighting coefficient W J I of the intermediate layer, and the summation SumvJ is expressed as Su
mFJ=ΣW J l [jj, il * 1-o
ut(i), and then add si to this Sumrt.
By applying the gmoid function, the output j of the j-th hidden layer is
−out(j) is calculated by

Next, in step 8406, -out(
Find k). This procedure is similar to step 8406.

That is, the output j-out(j) from the intermediate layer is multiplied by the weighting coefficient W kJ of the output layer, and the sum Sum- is expressed as Su
mrk=ΣW, [kk, jjl * j-out(
j), and then add si to this 5Lllllrk.
By applying the gmoid function, −out (k) is calculated on the output of the k-th hidden layer as follows. 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-ou
-out(k) umbrella (1-k-out(k
) Find l. Here, k-out(k)*(1-k-
out(k)) is the sigmoid function −out(k
It has the meaning of differentiation of l.

Next, in step 5408, the variation width of the weighting coefficient of the output layer ΔWkJ[kk, jjl, ΔW, [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 coefficients WkJ[kk, jul and WkJ[kk, jul
=WkJ[kk, jul+ΔW, [kk, jjl] Update. In other words, learn.

Next, in step 5410, the intermediate layer teacher signal teac
Calculate h−j (j). For that purpose, first, Sum
++J=Σteach−k(k)*wk, [jj,
Based on ii, calculate the inverse contribution from the output layer to each element of the intermediate layer. Next, from this SumllJ,
The intermediate layer teaching signal teach-j (j) is calculated using the following equation.

teach-j (j) = j-out(j)*(1
-j-out(j))*Sums,.

Next, in step 5411, the change width △WJI [jj+H] of the weighting coefficient of the intermediate layer is calculated as △WJ+[jj, i
il = β*1-out(i)*teach-j (j
)+α*△WJI [calculated by jj+ iil.

In step 5412, the weighting coefficient WJI [jj, iil is changed to WJI [jj, 11
] = WJl [jj+ 111+△WJI [jj,
il and update. In other words, learn.

In this way, in steps 8401 to 412, the weighting coefficient W,
1 and Wl were learned once.

In step 5413, it is checked whether the above learning has been performed on all the data, and in step 5414, it is checked whether learning has been performed a predetermined number of times to sufficiently converge the weighting coefficients. 8401-4
Repeat step 12.

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 will be used.

Next, a case where the above-mentioned "learning" is performed on a neural network for region separation between thin character line parts and image parts in this embodiment will be explained.

First, the input data is the value of a pixel (0 or 1, respectively) within a 7×7 window of two pieces of image data centered on the pixel of interest.

Therefore, the number of neurons in the input layer is 49, the number of neurons in the output layer is 1 when the area is a thin line part, and 1 because the flag is O when it is an image area, and the number of neurons in the intermediate layer is:
Although it is arbitrary, the number is set to 25 in this embodiment.

On the other hand, as the ideal output, a map is prepared.

This map is, for example, a map obtained by applying an edge detection filter to the original multivalued image as shown in Figure 3(b), or a map obtained by artificially specifying areas as shown in Figure 3(c). Possible examples include maps.

Furthermore, as for how to select input data, a learning pixel is selected at random, and a 7×7 window centered on that pixel is provided.

Using the above, the above-described learning procedure is repeated until the values converge, and the weighting coefficient at the time of convergence is determined as the region separation parameter.

When performing the separation process, a threshold value of 05 is set for a value between 0 and 1.0, and anything above that is considered a thin line part, and anything below that is considered an image.

Next, in order to obtain weighting coefficients for multi-value conversion, two images are prepared as input data, a multi-value image is used as ideal output, and a map as described above is prepared.

In this embodiment, 25 pieces of data within a 5×5 window centered on the pixel of interest are used for multi-value conversion both for the thin line portion and for the image portion.

Therefore, the number of neurons in the input layer is 25, the number of neurons in the output layer is 1 because there is one halftone value, and the number of neurons in the intermediate layer is arbitrary, but in this embodiment, it is 15.

First, the weighting coefficients for the thin line part and the image part are -
〇, a random value other than O from 5 to 0.5 is given as an initial value.

As described above, the input data is given by randomly selecting a pixel of interest and setting a 5×5 window centered on that pixel.

For the ideal output, the value D (0 to 255) of the multivalued image located at the position corresponding to the pixel of interest is given as 0 and 1, so as shown in Fig. 6, i, deal-output
Convert and set by = D / 255 X O, 8 + 0.1.

Then, referring to the value of the map at the position corresponding to the pixel of interest, if the value is 1, the weighting coefficient of the thin line part is used, and if it is 0, the weighting coefficient of the halftone part is used to obtain one output.

A teacher signal is obtained from the difference between the output and the ideal output, and the weighting coefficients are updated.

By repeating the above procedure, weighting coefficients (parameters) for multi-value conversion for each of the thin line part and the halftone part are obtained.

Using the above, the above-described learning procedure is repeated until the values converge, and the weighting coefficient at the time of convergence is determined as the parameter for region separation.

FIG. 1 is a block diagram showing one embodiment of the present invention.

In FIG. 1, reference numeral 101 denotes a line buffer composed of, for example, a FIFO or the like, which receives input of binary data from an image input device (not shown) and stores data for seven raster lines.

A data latch 102 latches seven pixel data for each seven lines of data from the line buffer 101. Therefore, from this data latch 102,
Binary image data of a 7×7 window is obtained.

This 7×7 49-bit data is sent to the area determination circuit 10.
3 will be done manually. The area determination circuit 103 outputs 1 if it determines that the image is a halftone image and if it determines that it is an O1 thin line portion. This output value 113 becomes a selection control signal for the selector 107.

Further, among the outputs of the line buffer 101, the central five lines are 25-bit data centered on the pixel of interest, and the 25-bit data is transferred to the halftone image multi-value conversion circuit 105, It is input to both the multi-value conversion circuit 106 for images for thin line portions, and each of the multi-value conversion circuits 114 and 115
Outputs one 8-bit multivalue deck. Both outputs are inputted manually to the selector 107, and the above-mentioned area determination circuit 1
If the result of area determination is a halftone image, the output data of the multi-value conversion circuit 114 is output, and if the result of area determination is a thin line part, the output data of the multi-value conversion circuit 115 is output.

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

Furthermore, in binary multivalue conversion, learning of weighting coefficients for region separation may be performed using a 5×5 window centered on the pixel of interest. According to this, the accuracy decreases a little, but the processing speed can be increased. Conversely, it is also possible to increase the window size. Also,
Even in multi-value conversion, the window size may be changed for each target area (for example, it may be 7x7 for halftones and 5x5 for thin line areas).

In addition, in the above embodiment, the weighting coefficients for both the thin line part and the halftone image part are learned at once using the map, but the image data of only the thin line part and only the halftone image part are learned separately. You may also prepare and study it.

In addition, in order to eliminate the unnaturalness of multilevel conversion at the boundary between two areas, in order to obtain the weighting coefficients used for multilevel conversion of thin line areas, not only the thin line areas but also the intermediate tones are trained at the same time. Weighting factors may also be used.

[Effects of the Invention] As explained above, according to the present invention, in order to estimate a multivalued image from an input binarized image, neural network software is used to discriminate between thin character lines and image areas. Based on the discrimination results, the thin character line parts are converted into multi-value data while preserving the edges. In the halftone image, graininess caused by binarization is removed and the image is multi-valued.

Therefore, the original multivalued image can be reproduced fairly accurately.

In addition, because the area discrimination between thin character lines and image areas is performed using a neural network, unlike processing using a filter, the area discrimination includes information such as the positional relationship of dots and connections, so it is quite accurate. be able to.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram of an image processing apparatus according to an embodiment, and FIG.
) is a conceptual diagram of the neural network, and FIG. 2(b) is a flowchart of the learning procedure of the neural network. Figures 3 (a) to (C) are diagrams explaining the learning of region separation, Figure 4 is a diagram showing an example of binary data, and Figure 5 (a) to (c) are diagrams illustrating the learning of region separation. FIG. 6 is a diagram showing the relationship between multi-value data and ideal output data. FIG. 7 is a diagram showing an example of a conventional device for binary-to-multi-value conversion. Figures (a) and (b) are diagrams showing an example of a smoothing filter, Figure 9 is a diagram showing an example of multi-value conversion of the edge part, and Figure 10 is an example of multi-value conversion of the halftone part. FIG. ...Line buffer...Data ledge...Region judgment neurodeep...Multi-level neurochip for images...Multi-level neurodeep for thin line areas...Selector (α) Cb) (C) Border N mouth (HJ-Water 7 Gyakuzutsutsu Figure (of) <b>

Claims (1)

[Scope of Claim] An image processing device for restoring multivalued image data from binary image data, comprising: an input means for inputting binary image data; a first multi-value conversion means that performs multi-value restoration processing on the character fine line portion using a neural network; A second multi-value conversion means performed by a network, and a control means for controlling the first or second multi-value restoration processing to be performed by the first or second multi-value conversion means based on the determination result of the determination means. Characteristic image processing device.