JPH0447474A - Picture processing system using neural network - Google Patents

Abstract

PURPOSE:To improve the precision of picture processing by coupling partial neural elements in the output layer to only partial neural elements in the intermediate layer. CONSTITUTION:Binary picture data of a restoration object inputted from the external is inputted to a line memory 100 having the capacity corresponding to 5 lines. Picture data of 5X5 areas is segmented from picture data in the line memory by a segmenting circuit 101 and is sent to a neural network 102. In the network 102, picture data is inputted to the input layer and is subjected to prescribed picture processing and is outputted from the output layer, and partial neural elements in the output layer are coupled to only partial elements in the intermediate layer. Thus, the quality of picture processing and the learning precision are improved.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to an image processing method using a neural network. related to image processing methods.

[Prior Art] Conventionally, there have been techniques for converting multivalued images into binarized images, or for restoring multivalued image data from reverse binary image data,
Many types of image processing are being performed, including techniques for applying spatial filtering.

However, when conventional binarization methods such as the dither method are used for binarization, the resolution and edge components of character portions tend to be lost, causing deterioration. In addition, the ED (error diffusion) method generates a unique striped pattern during binarization, and the image quality is not necessarily considered to be good, and the deterioration of text is also improved compared to the dither method. Although it has been done, it is still not perfect. Regarding the ED method, for example, see the literature (Floyd Sternl
oerg:SID, International
SYMP, Syp, Digest of T
ech-nical Papers 4-3 Apri
1'75) for a sufficient understanding.

Furthermore, a technique for restoring a multilevel halftone image from a binary image using a neural network has never been attempted.
Conventionally, at best, a multilevel halftone image is restored by applying a smoothing filter to the binary image.

[Problems to be Solved by the Invention] In multivalued image restoration using this conventional smoothing filter, the resolution of character areas deteriorates, and a sufficient number of gradations cannot be obtained for image areas with smooth density gradients. A problem occurred. Therefore, there are expectations for a technique for restoring a multivalued image from a binary image, for example, by image processing using a neural network. A feature of neural networks is that their processing accuracy increases dramatically through learning. However, it is not simple to use a neural network for binarization processing or multi-value restoration processing.

In particular, when simultaneously processing images that contain a mixture of binary images such as characters and halftone images such as photographs, a problem arises in which the coupling constant diverges during learning. It is no exaggeration to say that no consideration had been given to what kind of network configuration should be used to form a neural network for image processing.

Therefore, the present invention simultaneously processes an image containing a mixture of binary images such as characters and halftone images such as photographs, and performs image processing with high precision, such as restoring a multivalued image from a binary image. This paper proposes an image processing method using a neural network that can perform the following tasks.

[Means for Solving the Problems] The configuration of the present invention for achieving the above problems uses a neural network consisting of an input layer, at least one intermediate layer, and an output layer, and inputs image data to the input layer. Input,
An image processing method that performs predetermined image processing within this network and outputs the processed image data from the output layer, in which some neural elements in the output layer are coupled only to some neural elements in the intermediate layer. It seems like it is.

When performing block processing on images, depending on the image quality,
The application area of block processing should be changed. In the image processing device of the present invention, the neural elements in the output layer that are given to output an image of a certain quality are connected only to some neural elements in the intermediate layer that fit into blocks of appropriate size. This improves the quality of image processing,
Learning accuracy can be improved.

[Embodiments] Two image processing apparatuses (first and second embodiments) to which the image processing method of the present invention is applied will be described below with reference to the accompanying drawings. These embodiments disclose an image processing method for restoring a binary image to a multivalued image. That is,
Generally speaking, as shown in Figure 2, a 5x5 pixel block is input, the block is converted into multi-value edge image data and multi-value halftone image data, and the data in the block is The image quality is determined, and based on the determination result, the optimal one of the multivalued edge image data and the multivalued halftone image data is selected and output.

It goes without saying that the size of the blocks shown in FIG. 2 is merely an example.

Image processing method characteristic of these two embodiments ■: Use of a network; ■: When learning the network, a 5×5
The blocks are input as training data, but
For learning edge-like images such as characters, use learning data included in a 3x3 area of the 5x5 area, and for learning halftone images, use image data from all 5x5 blocks; (2): Learning of images with edge characteristics such as characters is performed first, and neural connection constants (synaptic parameters) closely related to image processing with strong edge characteristics are determined. When learning a halftone image, the neural coupling constant determined first is fixed, and the neural coupling constant for that halftone image is determined later (the second learning method). ).

In particular, point (■) will be explained. Image data with strong edge characteristics, such as characters, has strong locality. On the other hand, if the current image data is a halftone image and this halftone image is binarized by, for example, an error diffusion method or a dither method, locality is weak and the original image tone is dispersed over a wide area. do. Therefore, in order to train a neural network to restore multilevel image data from binary image data whose original image is halftone, it is preferable to refer to a larger area. In order to train a neural network to restore multi-valued image data from binary image data when the image is a strong one, it is preferable to refer to a smaller area. That is, especially in the case of restoring an image with strong edge characteristics, if all 25 pixels of 5×5 are referred to, the binary data of the 25 pixels will be complicatedly intertwined, and it will be difficult to match the ideal signal. For learning edge-like images such as characters, 3x out of a 5x5 area is required.
This is the reason why the learning data included in the area No. 3 is used, and the image data of all 5×5 blocks are used for learning the halftone image.

Configuration of Image Processing Apparatus> FIG. 3 shows the configuration of an image processing apparatus common to the first and second embodiments. In the figure, binary image data to be restored that is input from the outside is input to a line memory having a capacity for five lines. The image data of a 5×5 area is extracted from the image data in the line memory by the extraction circuit 101 and sent to the neural network 102 . In the first embodiment, the network 102 has a configuration as shown in FIG.
It has the configuration shown in the figure. That is, the image processing apparatus of the first embodiment and the image processing apparatus of the second embodiment differ mainly in the connections between the neural networks in the network 102.

The neural coupling constants of this network 102 are learned in advance by the procedure described below, and the results of determining from this network 102 the image signal G suitable for a halftone image, the image signal E suitable for an edge image, and their image tones. These three signals are input to the selector 103. In accordance with the region separation signal S, the selector 103 selects and outputs either an image signal G in which a halftone image is better restored, or an image signal E in which an edge image is better restored.

The concept of neural networks themselves is explained in detail in, for example, ``Using neural networks for pattern recognition, signal processing, and knowledge processing'' (Nikkei Electronics; August 10, 1987).

First Embodiment> The configuration of a network related to an image processing apparatus of a first embodiment will be described with reference to FIG.

This network 102 has 5×5=25 neural elements in the input layer to input 5×5 blocks. Further, in order to output the halftone image signal G, the character image signal E, and the area separation signal S, three neural elements were provided in the output layer. Further, in the first embodiment, only one layer is used as the intermediate layer, and the total number of elements is 5×5=25.

The connections between neural elements in each layer will be explained. Each of the 25 elements of the input layer is coupled to each of the 25 elements of the intermediate layer.

The coupling between the elements of the intermediate layer and the output layer will be explained. An output layer neural outputting a halftone image signal G is coupled to each of all 25 elements of the intermediate layer. That is, as explained in connection with FIG. 2, the halftone image signal G is 5×5
It is determined by referring to all pixels of . In other words, the coupling constant between the a-layer neural outputting the halftone image signal G and each of all 25 elements of the intermediate layer is also 5X5=2.
It is determined by referring to all five pixels.

The output layer neural, which outputs the region separation signal S, is also coupled to each of all 25 elements of the intermediate layer. That is, both this separation signal S and the coupling constant between the output layer neural that outputs this S and each of all 25 elements of the intermediate layer are:
It is determined by referring to all pixels in a 5×5 block.

On the other hand, in contrast to G and S, the output layer neural outputting the character image signal E is coupled only to 3×3=9 elements of the intermediate layer. That is, as explained in connection with FIG. 2, the character image signal E is determined with reference to 3×3=9 pixels. The multivalued character image signal E is converted into binary data “
0”, 0.4.0.5.0 when restored from “1”
．． Since it has a characteristic much closer to "0" or "1" than the value 6..., it is simplified by reducing the coupling with the output element E.

The position of the 3x3 block in the 5x5 block will be explained. In principle, 5X5 in Figure 2
There is no strong restriction on the positional relationship between the block and the 3×3 block. However, it should be determined according to what kind of binarization process was used to generate the binary image data from the original image. In this example, multi-value data for one pixel of interest is restored from a 5x5 block, so as shown in Figure 2, the center of the 5x5 block is set to the center of the 3x3 block. It is preferable to match the

Next, the mode of learning will be explained using the following two examples.

1. During learning, a binary image is given as a learning input image, and a multivalued image is given as an ideal output to the neural network 102.

However, it is impossible to obtain a multi-value image, which is the ideal output, by restoring a binary image, so a multi-value image whose data is clearly known (this is assumed to be the ideal signal) is converted into a binary image. , given as a learning input signal. or,
Halftone image signal G.

Learning is performed by giving exactly the same ideal signal to the character image signal E. That is, as shown in FIG. 4, multivalued image data 110 as an original image is prepared, and this image data 110 is
Binarization is performed using a 5×5 dither matrix 111 and a 3×3 fixed threshold matrix 112. Also, image data 1
10 is also applied to the region separation filter 113 to obtain the signal S. The selector 114 selects the binarizer 11 according to the signal S.
At step 1112, one of the binarized signals G and H is selected. The binary image data 116 for learning thus obtained is cut out into 5×5 blocks and input into the input layer of the network shown in FIG. Furthermore, the multilevel signal 110 of the original image and the region separation signal 115 are used as ideal output signals when calculating the coupling constant.

It should be noted that the region separation signal S for learning may be obtained by performing edge detection or correction from a multivalued ideal signal.

Alternatively, an image may be provided in which the character area and the halftone area are clearly divided into, for example, rectangular shapes from the beginning. Furthermore, if a character image and a halftone image are prepared and an image is created by combining them, the ideal signal of the region separation signal can be generated naturally.

The binary image data 116 is cut out into 5×5 blocks and input into the input layer of the network shown in FIG. Further, the multi-level signal 110 and the region separation signal 115 of the original image are given to the output layer as ideal outputs.

In this case, halftone multivalued image data may also be input for learning, but since the number of connections in the middle layer is small for the character image signal E, the inputs in the input layer are "1" and "0".
” and the weight of the center pixel (pixel of interest) remain strongly, and characters and edges with high frequency components tend to remain, so there is no problem even if halftone multi-valued image data is input for learning.

- State Part 2 This second learning method is divided into two stages (Step I, Step ■) as shown in Figures 5A and 5B. however,
Regarding the generation of learning data, as in the first learning method, the data shown in FIG. 4 is used.

Figure 5A shows the coupling constant between the input layer and the intermediate layer, some neural elements in the intermediate layer, and the neural elements E in the output layer while inputting a learning signal that is optimal for outputting the character image signal E. Step I of determining the coupling constant between. That is, in step (2), the binary character image signal prepared for 5×5 learning is input to all 25 elements of the input layer, and the coupling constant indicated by * in FIG. 5B is determined.

Thus, the coupling constant (denoted by .) determined for the character image is fixed during the learning for the halftone image signal G and the region separation signal S in step H of FIG. 5B. . That is, in step (2), the binary halftone image signal prepared for 5×5 learning is input to all 25 elements of the input layer, and the remaining coupling constants indicated by O in FIG. 5B are It is determined.

9 of the intermediate layer where the neural elements (E, G) of the output layer of the character image signal E and the halftone image signal G are commonly connected.
For (=3×3) neural elements, a coupling constant that matches both image characteristics and the region separation signal S can be obtained. Since the 16 unshared intermediate layer elements may have coupling constants that satisfy only two types, the halftone image signal G and the region separation signal S, the convergence of the coupling constants during learning is very good.

In addition, even if the coupling constant of the 9 elements of the shared intermediate layer described above is determined to better output the character image signal E, but is not sufficient as the coupling constant for G and S, the coupling constant with the 16 elements that are not shared is determined to better output the character image signal E. Depending on the bond strength (coupling constant), G, S
The correction works to bring the signal closer to the ideal signal.

Note that if the image data for learning in step I includes a halftone image, or if the image data for learning in step H includes a character image,
Learning is performed only when the ideal signal of the region separation signal S indicates a character image and a halftone image, respectively.
All you have to do is update the coupling constant and not perform learning in other cases.

We propose the following modification to the second learning method. That is, in step (2) of learning method 2, both characters and halftone images may be provided as the ideal signal, input signal, and region separation signal, and learning may be performed in all cases.

In addition, in step E, in order to ideally determine the coupling constant for the character image signal E, in the coupling between the output layer or intermediate layer and the layer in the previous stage, the coupling constants other than those that share all the couplings with the previous stage element are set to 1. It is also possible to perform learning by assuming a state in which the bond is completely disconnected from time to time, determine the remaining coupling constant, and fix this.

11 Standing Ima1 Next, regarding the learning of the neural network 102,
An outline of the procedure will be explained using FIG.

First, as learning data, for example, binary image data 116 as shown in FIG. 4 and multi-value image data 110 as an ideal output are prepared. This ideal output is the teacher signal (tea
channel signal). Then, after initially appropriately determining the coupling strength W 4 between the input layer and the hidden layer and the coupling strength W0 between the middle layer and the pressure layer, the above-mentioned above is added to the network connected based on these initial values as learning data. Binary image data 116 is input, and eight multivalued image data (kout) are obtained. This output is compared with the atomic value image data as the ideal output, and a teaching signal (teach) is generated from this comparison. This process is shown in FIG. From this teacher signal, the above W j l I W +1 j is corrected while learning by the pack propagation method. This process is shown as ■ in Figure 1. By repeating this learning, the coupling strength W and 1°Wl are modified to be appropriate for restoring a binary image to a multivalued image. Connection strength W J i 1 obtained as a result of learning
W M j is maintained within network 102 .

The learning procedure will be explained in detail using FIG. 6A and FIG. 6B. The steps shown in this flowchart are as follows:
Although the above-described first learning method and second learning method are shown as being common, the differences will be explained each time.

In step 5402 in the figure, initial values are given to the weighting coefficient WJ8 indicating the strength of the connection between the input layer and the intermediate layer, and the weighting coefficient Wk indicating the strength of the connection between the intermediate layer and the output layer. As an initial value, −
〇, select a value in the range of 5 to +0.5. Step 54
In 04, select any 5× from the input data for learning.
Select block 5. Steps 8404 to 8406 are procedures for setting input and output data to the neural network. That is, in step 5406,
The image data 1out(i) (here, i=1 to 16) within this 5×5 area is set to the neural of the input layer.

Note that the signal given to the input layer is image data of 8 bits 0 to 2.
If it is 55, a value between 0 and 1.0 divided by 255 is given.

Next, in step 5408, Ri,! ! Multivalued image data 110 (ideal out) and region separation signal 115 as four outputs are prepared.

Step 5410 shows a procedure for calculating the output kout (k) based on the given transform coefficients.

That is, data 1out(i) from the input layer is multiplied by the coefficients W and I of the intermediate layer, and the sum SumyJ is calculated as Sumr=
=Σ, WJI fj, i) * 1out(i)
・・＋・−(1) Calculate, then output j of the first intermediate layer
51gl1 to normalize out(j) to 0/1
Calculate using the +oid function. Here, F in Sumri is Forward.
ard). Similarly, the output value kout from the intermediate layer to the output layer is determined as follows. First, using the coefficients W and J of the output layer and the output value jout(j) from the intermediate layer, the sum of the products is calculated and the value Su++
Get □.

5ullyi+ = ,X: WIIJ (k, J)
This jout(j)...” ” (3) Next, 0/
Using the 8Ngmoid function to normalize to 1, we obtain the output kout (k) to the output layer according to . Here, k=1 to 3, k=1 means halftone image signal output G, k=2 means character image signal output E, and k=3 means area separation signal output S.

Thus, FORWA for a set of sample data
The calculation in the RD direction is completed. Below is BACKWARD
In other words, this is a procedure for correcting the coupling strength by learning from sample data consisting of a pair of the above input and ideal output.

In step 5412, the output value kout calculated from the initial value (set in step 5402) WJllwilJ and the ideal output 1deal out(k) prepared in advance are calculated.
Compare with. That is, by this comparison, the teacher signal teach k(k) of the output layer is expressed as teac3k(k)=(ideal-out(k)
-kout(k))*kout(k)*(1-kout
(k))... Calculate according to (5). (5) kout(k)*(1-
kout (k) ) has the significance of a differential value of a sigmoid function. Next, in step 5414, the variation width ΔWm7(k, j) of the degree of coupling of the output layer is determined. That is, △L
J(k, J) = β umbrella jout(j)*teach k
(k) 1α*△LJ (k, j) ・・・・・・(6
). Note that α and β are fixed value coefficients, and in this example, they are set to 0.5. In step 5415, the degree of coupling WkJ (k, j) between the intermediate layer and the output correlation is updated, that is, learned, based on this change width ΔWkJ (k, j).

WkJ (k, j) = △LJ (k, j) + LJ
(k, J) (7) Next, in step 5416, the intermediate layer teacher signal teac
Calculate hj (j). That is, first, 5"°"""i
teach-k(k)1N...(kJ) ,,
< 8), calculate the contribution from the output layer in the BACKWARD direction to each neural in the intermediate layer. Then, by normalizing this contribution using the differential value of the sigmoid function according to the following equation, the intermediate layer teaching signal teach j (j) is calculated.

teachj (j) = jout(j) umbrella (1-jo
ut (j)) Umbrella Sums.

(9) Note that teachj (j) in equation (9) has a meaning as an ideal signal in the intermediate layer.

Next, in step 5418, this teacher signal teach
Using j (j), change width of the degree of coupling of the intermediate layer △W J
Calculate l (j, i). That is, △WJl(j, i
)=β*1out(i)*teach-j(j)+αinΔwJI(j,i) (10) Then, in step 5420, the degree of connectivity WJ+(j,i) is updated.

That is, WJ l(j, i) = WJ + (, L i)+
ΔWJIF, 1).

In this way, based on the pack propagation method, the degree of connectivity WJl is calculated from one set (3x3 multivalued image data) as Lout and the corresponding dithered binary data 1deal out as ideal output data. W
k, has been learned - times. In step 5422, it is checked whether such operations have been performed on all sets of sampling input data, and the procedures from step 8404 to step 5420 are repeated until they have been performed on all sets of sampling data.

If the learning for all sets of sampling data is performed only once, the accuracy is considered to be low, so step 5
The processing of steps 3404 to 5422 is repeated until the accuracy is considered to be high (according to the judgment in step 424).

Note that when learning is performed using the second learning method described above (FIGS. 5A and 5B), step 5415 and step 5 are performed.
The updating process of the coupling constant W at 420 is performed only for necessary ones. That is, when learning the coupling constant indicated by . in FIG. 5B (step I), updating of the coupling constant W indicated by 0 in steps 5415 and 5420 is omitted. Conversely, when learning the coupling constants indicated by ○ (step ■), step 5415 and step 5420 of the coupling constant W indicated by .
Updates in are omitted.

Note that the positions of the blocks specified in step 5404 should not be sequential (it is better to specify them randomly).

The above is the explanation of the principle of the first embodiment and the learning procedure based on the pack propagation method in the embodiment.

Second Embodiment FIG. 7 is a block diagram showing the configuration of a neural network according to a second embodiment. In the first embodiment described above, as shown in FIG. 1, the number of neural elements related to the output of the character image signal E was 25 in the input layer and 9 in the intermediate layer. In this second embodiment, the number of neural elements related to the output of the single character image signal E is nine in both the input layer and the output layer, and in particular, the nine elements in the input layer correspond only to the pixel of interest and the nine pixels adjacent to it. are doing. For this reason, it is possible to learn the coupling constant most suitable for the process of restoring a binary image from a character image into a multivalued image without being affected by the "1" and "O" of surrounding pixels.

The configuration of the device of this second embodiment is that of the first embodiment (third embodiment).
Figure) can be used. Furthermore, the control procedure for learning is the same as that of the first embodiment (Figures 6A and 6B).
It is possible to use Also, the learning method is the first
It is possible to refer to the examples. Furthermore, the data of the first embodiment (FIG. 4) can also be used to create data for learning.

8A and 8B are FIG. 5A of the first embodiment.

This corresponds to FIG. 5B.

<Modifications> The present invention can be modified in various ways without departing from the spirit thereof.

In the two embodiments described above, the block size is 5×
5, the present invention is not limited thereto.

In addition, in these embodiments, for example, 3×3=
9 elements share output with all output layer elements, but
This number does not need to be 9.

Furthermore, it is not necessary that the number of elements in the input layer and the intermediate layer of the neural network match, and there is no particular restriction that the number be 25.

Further, in the embodiment, there is only one intermediate layer, but this number may be freely set to two, three, etc.

Further, although the above-described embodiment describes restoration from binary to multi-value, the present invention is not limited thereto, and can be applied very easily to conversion processing from multi-value to binary.

[Effects of the Invention] As explained above, the image processing method of the present invention uses a neural network consisting of an input layer, at least one intermediate layer, and an output layer, inputs image data to the input layer, and processes this network. An image processing method that performs predetermined image processing in the output layer and outputs the processed image data from the output layer, in which some neural elements in the output layer are connected only to some neural elements in the intermediate layer. It seems like there are.

When performing block processing on images, depending on the image quality,
Although the application area of block processing should be changed, in the image processing method of the present invention, the neural elements of the output layer given to output an image of a certain quality are Since it is connected only to some neural elements in the intermediate layer that are compatible with the above, the quality of image processing and learning accuracy can be improved.

Additionally, as a side effect, there is no need to use multiple neural networks to obtain final outputs with different properties (and the circuit size can be reduced).

[Brief explanation of the drawing]

Figure 1 is a diagram showing the configuration of the neural network used in the first embodiment, and Figure 2 shows neural coupling constants learned when the input image is a binary image and when it is a halftone image. FIG. 3 is a diagram showing the range of pixels to be referred to when the present invention is applied. FIG. 4 is a diagram showing the configuration of the image processing apparatus of the second embodiment. A system diagram when creating learning data for a neural network in the second embodiment, FIGS. 5A and 5B explain two steps I and II in the second learning method of the first embodiment. Figures 6A and 6B are 1. A flowchart of the control procedure for learning in the second embodiment. Figure 7 is a diagram showing the configuration of the neural network used in the second embodiment. Figures 8A and 8B are the second learning of the second embodiment. How-
FIG. 2 is a diagram illustrating two steps I and II in FIG. In the figure, 100...line memory, 101...block extraction circuit, 102...neural network, 10
3...Selector. Patent applicant: Canon Co., Ltd. - Agent: Patent attorney: Yasunori Otsuka (1 person) Manpower: 1 cP threshold: 1st? figure poka layer

Claims (12)

[Claims] (1) Using a neural network consisting of an input layer, at least one intermediate layer, and an output layer, image data is input to the input layer, predetermined image processing is performed within this network, and data is output from the output layer. An image processing method that outputs processed image data using a neural network characterized in that some neural elements in an output layer are connected only to some neural elements in an intermediate layer. Processing method. (2) The input image data includes a mixture of images of different image quality, and each of the plurality of neural elements of the output layer simultaneously outputs output image data of different image quality. Image processing method using the described neural network. (3) Each of the plurality of neural elements of the output layer has:
3. The image processing method using a neural network according to claim 2, wherein the image processing method is a halftone image signal, an edge portion or character portion image signal, and a region separation signal. (4) The image processing method using a neural network according to claim 1, wherein some neural elements in the intermediate layer are connected to all neural elements in the input layer. . (5) An image using a neural network according to claim 1, wherein some of the neural elements of the intermediate layer are connected only to some of the neural elements of the input layer. Processing method. (6) When learning the network, first learn and determine a coupling constant between some neural elements in the output layer and some neural elements in the intermediate layer that are connected to this neural element. . The image processing method using a neural network according to claim 2, wherein the remaining coupling constants are determined by re-learning while fixing the previously determined coupling constants. (7) When first determining the coupling constants of some neural elements, as an ideal signal to be given to the network,
5. The image processing method using a neural network according to claim 3, wherein an image having a stronger locality among the images of different image quality is used. (8) The image processing method using a neural network according to claim 7, wherein the image with stronger locality is an image with stronger edge characteristics, such as a character image. (9) The coupling constant determined during relearning is for neural elements other than some of the neural elements in the intermediate layer;
5. The image processing method using a neural network according to claim 4, wherein the coupling constant is a coupling constant between neural elements of the output layer other than some of the neural elements of the output layer. (10) The coupling constant determined during relearning is for neural elements other than some neural elements in the intermediate layer, neural elements in the output layer other than some neural elements in the output layer, and one neural element in the input layer. 2. The image processing method using a neural network according to claim 1, wherein the coupling constant is a coupling constant between a neural element other than the neural element of the neural network. (11) A final image signal is obtained by switching between a halftone image signal and a text image signal from a neural element in the output layer using a region separation signal from a neural element in the output layer. An image processing method using the neural network described in Section 3 of (12) Using a neural network consisting of an input layer, at least one intermediate layer, and an output layer, input an image containing strong and weak local image quality into blocks to the input layer, and An image processing method that performs predetermined image processing on this block image within a work and outputs the processed image data from the output layer, wherein neural elements of the output layer are provided for each of the image qualities, The output layer neural elements for image quality with strong locality are coupled only between some neural elements of the input layer and some neural elements of the intermediate layer, and are connected only between some neural elements of the input layer and some neural elements of the intermediate layer. An image processing method using a neural network, characterized in that part of the image data of the block image is input to the neural element.