JPH0451375A - Image processing device - Google Patents

Abstract

PURPOSE:To improve the precision of a multivalue processing by providing a multivalue means by means of a neural network, which inputs picture data in an area where a picture element being the object of multivaluing becomes asymmetric in a picture processor estimating multivalued picture data from binarized picture data. CONSTITUTION:Windows are set to be asymmetric as against a notable picture element. A line buffer 1 receives binary data for four lines. A data latch 2 obtains binary picture data of the windows of 4X4. Data is given to the conversion table 3 of a ROM system as an address. Thus, the conversion table 3 outputs multivalued data of 256 layers to input data. The conversion table 3 is decided by the neural network 201 and the like. The neural network 201 and the like learn based on a back propagation method, obtains outputs as against all input patterns of 4X4 and set them to be a lookup table by ROM.

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 binary image, a rectangular area window centered on the target pixel is provided, and a neural network is applied from the binary pattern inside the window. A method has also been proposed for estimating multivalued data using this method.

[Problems to be Solved by the Invention] However, with the method of providing a window as described above, good results cannot be obtained unless a large window is provided.
In that case, the number of reference pixels would be too large, making it difficult to implement hardware for multi-value conversion using a neural network.

By the way, in the error diffusion method and the average density approximation method, which are binarization methods that preserve the density, the density error caused by the binarization of a certain pixel is distributed to the pixels immediately to the right and directly below and saved.

Therefore, the density of the pixel to be processed can be expressed as D=(error from the left)+(error from above)+(own error) 10. Here, α is an error received from other pixels due to the diffusion matrix in the case of the error diffusion method.

That is, in the binarization method that preserves density, the pixel of interest is strongly influenced from the left and above.

Therefore, when estimating a multivalued image from binarized image data, it is sufficient to place emphasis on the left side and above the pixel of interest.

[Means and effects for solving the problems] In order to solve the above problems, the image processing apparatus of the present invention includes an input means for inputting binary image data consisting of a plurality of pixels including a pixel to be multivalued; a multi-value conversion means for performing multi-value restoration processing of the multi-value conversion target pixel from input binary image data using a neural network; Input the image data of the area.

[Example] In this example, a window as shown in FIG. 3 is provided. The window in FIG. 3 includes many pixels above and to the left of the pixel of interest.

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

In FIG. 1, l is a line buffer composed of FIFOla to ld, which receives 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 latches 4 pixels of data for each line. Therefore, from this data ledge 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. As will be described later, the conversion table 3 is determined by a neural network, and outputs multivalued data of 256 gradations (8 bits) in response to input data.

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. 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 input to the output layer 203 as the number of neurons kk), and from the output layer 203, the output (k - out) 206
is output. Further, 207 is an ideal output (ideal output).
out).

In a neural network, input data and an ideal output for it (1 deal-out, ) are prepared,
By comparing this with the output (k-out) 206, the coupling strength in the intermediate layer WJI [jj, ii] (208 in the figure) and the coupling strength in the output layer WkJ [kk, j
, j] (209 in the figure) is determined.

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
, +: [jj, ii], Wk, [kk, jj]
Give the initial value of . Here, in consideration of convergence in the learning process, a value in the range of -0.5 to +0.5 is selected.

Next, input data 1-ou for learning is inputted in the sub 5402.
t, (i), and in step 5403 this data 1
- Set out(i) to the input layer. Further, in step 5404, an ideal output (ideal-out) for input data 1-out(i) is prepared.

Therefore, in step 5405, the output of the intermediate layer, j-ou
Find t(j).

First, data 1out(i) from the input layer is multiplied by the weighting coefficient W J of the intermediate layer, and the sum SumrJ is expressed as Su
mFJ=Σw,, [jj, il*1-out
(i), and then add sigm to this SumrJ.
By applying the oid function, the output j-o of the j-th intermediate layer is
Calculate ut(j) by:

Next, in step 8406, -out is applied to the output of the output layer.
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 WkJ of the output layer, and the total sum 5ullFk is
and then add sigmoid to this Sumr*
Apply the function to the output of the kth hidden layer -out
(10 is calculated by. Note that this output value is 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(kD(l −k−out(k
)). Here, k-out(k)*(1-k-
out(k)) is the sigmoid function −out(k))
k) has the meaning of differentiation.

Next, in step 5408, change width △Wk, [kk, Jjl of the output layer weighting coefficient, △Wt+, (kk,
Jjl = β*j-out(j)*teach-k(
k)+α*△Wk, +[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, based on this variation width, the weighting coefficients WkJ[kk, jj and Wu; [kk, Jj
l=W+B [kk, jul+△WkJ[kk,
Jjl and update. In other words, learn.

Next, in step 5410, the intermediate layer teacher signal teac
Calculate h−j (j). For that purpose, first, Sum
aJ=Σteach-k(k)*W+J[jj+i
Based on il, calculate the inverse contribution from the output layer to each element of the intermediate layer. Next, from this SumBJ, a teacher signal teach-j (j) for the intermediate layer is calculated using the following formula.

teach-j(j)=j-out(j)*(1-j-
out(j))*Sumn J Then step 5411
Then, the change width of the weighting coefficient of the middle layer △WJI [jj, i
il, △WJI [jj+ til =β*1-ou
t(i)*teach-j(j)+α*△WJI [j
Calculate by j, il.

In step 5412, based on this variation range, the weighting coefficients W J I [jj+ iil, WJl[jj,
iil =WJ, [jj, ii]+△WJI l,
il and update. 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
l and Wk 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 8401 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 a binary image in this embodiment will be explained.

First, 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 the error diffusion method.

Therefore, the number of neurons in the input layer is 16, the number of neurons in the output layer is 1 because the multilevel output is for one pixel, and the number of neurons in the middle layer is ζ, which is arbitrary, but in this example, 1.
Two pieces.

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

Furthermore, as for how to select input data, a learning pixel is selected at random, and a 4×4 window is provided for that pixel as shown in FIG. 3 to provide this.

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
6 patterns)), find the output (multivalue 256 gradations),
This is stored in a lookup table (LUT) using ROM, etc.
) can be obtained by

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.

Also, the window size should be set to 21 to make it a LUT.
The limit is 7 patterns, and only up to 17 pixels can be referenced, but in the case of a neurochip, more pixels can be referenced, so it is not limited to 4x4, for example, as shown in Figure 4 (a). It may be 5×5.

Furthermore, the window need not be limited to a rectangular shape, as shown in Figure 4 (
1) may also be used.

[Effects of the Invention] As described above, according to the present invention, in order to estimate a multi-valued image from an input binary image, image data of an asymmetric region of a pixel to be multi-valued is collected. Input,
Since we now refer to an image that includes many pixels in directions where the pixels to be multivalued are more susceptible, it is now possible to perform multivalued processing with the same level of precision with fewer reference pixels than before. Hardware has become easier.

[Brief explanation of drawings]

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 1 of the learning procedure of the neural network. FIG. 3 is a diagram showing an example of a window, and FIGS. 4(a) and (b) are diagrams showing other examples of windows. 1...Line buffer 2...Data latch 3...Conversion table ZGI

Claims (3)

[Claims] (1) An image processing device for estimating multi-valued image data from binary image data, comprising an input means for inputting binary image data consisting of a plurality of pixels including a pixel to be multi-valued; a multi-value conversion means for performing multi-value restoration processing of the multi-value conversion target pixel from binary image data using a neural network; An image processing device characterized by inputting image data. (2) The image processing apparatus according to claim 1, wherein the inputted image data includes more pixels to the left than to the right of the multilevel conversion target pixel. (3) The image processing apparatus according to claim 1, wherein the inputted image data includes more pixels above the multivalue quantization target pixel than below.