JPH07121695A - Image gradation conversion method and image processing device - Google Patents

Abstract

PURPOSE:To obtain a satisfactory gradation conversion image without performing the area division of digital image data by judging either a first or second weighted mean value is near to the value of a remarked picture element. CONSTITUTION:A first temporary value:P(1) is selected from gradation (n values) after conversion by a selector 20, and the first weighted mean value B1 can be found by adding a value in which a weighted mean coefficient is multiplied by the value P(1) on SIGMA. Also, a second temporary value:P(2) is selected by the selector 20, and the second weighted mean value B2 can be found by adding a value in which the weighted mean coefficient is multiplied by the value P(2) on SIGMA. When it is judged that ¦A-B2¦ is smaller than ¦A-Bl¦ as a result of comparison of an error, the content of an error holding circuit 70 is updated by the value ¦A-B2¦, and also, the content of a conversion candidate hold circuit 30 is updated. Also, when it is judged that ¦A-B2¦ is larger than or equal to ¦A-B1¦, the error hold circuit 70 and the conversion candidate hold circuit 30 are not updated.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image gradation conversion method for converting the gradation of digital image data in which characters and images are mixed, and an image processing apparatus using this method.

[0002]

2. Description of the Related Art An image in which each pixel forming an image is displayed in white or black is called a binary image.
On the other hand, an image with several levels of brightness between white and black is called a multi-level gray scale image. The process of converting a multi-tone image into a monochrome binary image is called binarization. For example, each pixel is from 0 (black) to 255 (white)
An image of 256 gradations having values up to 1 requires an information capacity of 8 bits for each pixel, but when binarized, each pixel is compressed to an information capacity of 1 bit, and the data amount is 1
/ 8.

Conventionally, in converting the gradation of digital image data in which characters and images are mixed, binary storage is required in the character area, and halftone reproducibility is required in the image area. However, there has been no single image gradation conversion method that satisfies such conflicting requirements. Therefore, as a pre-step of gradation conversion, the character area and the image area of the input digital image data are first detected, and the digital image data is divided into the character area and the image area based on the detection result. It was
Then, in the character area, it is possible to use white and black 2 with a certain threshold.
Using a simple binarization method that normalizes the pixel value to the level,
In the image area, an intermediate reproduction method by area gradation using a pattern matrix of 4 × 4 pixels in the vertical and horizontal directions is used, and different processing is performed for each area.

[0004]

However, in the above-mentioned conventional method, it is necessary to divide the input digital image data into the character area and the image area as a pre-step of gradation conversion, and the gradation conversion method is also used. Since different methods are required for the character area and the image area, the processing speed becomes slow, the circuit scale becomes large and complicated, and as a result, the image processing apparatus becomes large and costly. there were. In view of the above problems, an object of the present invention is to provide an image gradation conversion method for obtaining a favorable gradation conversion image for both characters and images without dividing the character area and the image area of the digital image data. That is. Another object of the present invention is to provide a high-speed image processing apparatus for converting gradation without dividing a character area and an image area of digital image data.

[0005]

In order to solve the above problems, the present invention has the following constitution. That is, the present invention is 2
Image gradation conversion for binarizing each pixel of digital image data in which character pattern data represented by a value image and multi-gradation image data represented by a plurality of gradations included in the binary section are mixed In the method, a step of selecting a target pixel from a set of pixels to be converted, and a step of setting a first value from the two values as a first temporary value of the target pixel,
Obtaining a first weighted average value from the first temporary value and the values of a plurality of pixels in the vicinity of the target pixel; and selecting a second value from the two values that is different from the first value. Setting the pixel as a second temporary value, determining a second weighted average value from the second temporary value and the values of a plurality of pixels in the vicinity of the pixel of interest, and the first weighted average value. And a second weighted average value that is closer to the value of the pixel of interest, and if the first weighted average value is closer to the value of the pixel of interest than the second weighted average value. If the first value is the updated value of the pixel of interest, and the second weighted average value is closer to the value of the pixel of interest than the first weighted average value, then the second value is the updated value of the pixel of interest. And the step of repeating all the steps until all the target pixels have been processed once, and the predetermined convergence condition is satisfied. An image gradation conversion method characterized by comprising the a step of repeating the whole process until.

Further, according to the present invention, a digital pattern in which character pattern data represented by a binary image and multi-tone image data represented by a first gradation (m value) included in the binary section are mixed. Each pixel of the image data is set to the first gradation (m
Image gradation conversion method for converting the gradation into digital image data represented by a second gradation (n value; m> n) that is smaller than the value) from a set of pixels forming the digital image data to be converted. A step of selecting the pixel of interest, and a value indicating the first brightness from the second gradation; P
Setting ₍₁₎ as a first provisional value of the pixel of interest; determining a first weighted average value from the first provisional value and values of a plurality of pixels in the vicinity of the pixel of interest; A step of setting a value indicating the second brightness from the second gradation; P ₍₂₎ as a second tentative value of the target pixel, the second tentative value and the vicinity of the target pixel The step of obtaining a second weighted average value from the values of a plurality of pixels, and a value of the k-th (n ≧ k ≧ 3); P _(k) from the second gradation (n value); Of the k-th temporary average value, a step of obtaining the k-th temporary average value from the k-th temporary value and the values of a plurality of pixels near the target pixel, Determining which is the closest to the value of the pixel of interest; setting the temporary value that yields the weighted average value closest to the pixel of interest as the updated value of the pixel of interest; A step of repeating the whole process until the processing of the elephants pixel is completed, an image gradation conversion method characterized by comprising the a step of repeating the whole process until a predetermined convergence condition is satisfied.

According to the present invention, a digital pattern in which character pattern data represented by a binary image and multi-tone image data represented by a first gradation (m value) included in the binary section are mixed. Each pixel of the image data is set to the first gradation (m
Value), the value set as the temporary value of the pixel of interest in the image gradation conversion method for gradation conversion into digital image data represented by a second gradation (n value; m> n)

[Equation 2] The image gradation conversion method is characterized by being determined by the formula (1).

Further, according to the present invention, in any one of the image gradation conversion methods described above, when the weighted average value is obtained from the provisional value of the target pixel and the values of a plurality of pixels in the vicinity of the target pixel, The larger the distance to the pixel, the smaller the weighted average coefficient of the neighboring pixels can be set.

The present invention is also an image processing apparatus using a planar neural network in which units arranged in a plane are connected to units in the vicinity thereof so as to be able to exchange information with each other. , Means for calculating a weighted linear sum of inputs from units in the vicinity of the unit, means for holding the pixel value before gradation conversion assigned to the unit, and possible gradation values after conversion And a means for generating.
An image processing device characterized by executing the image gradation conversion using the image gradation conversion method described in (1).

[0010]

According to the present invention, with the above configuration, digital image data in which a character area and an image area are mixed is divided into a binary area or a first gradation without dividing the digital image data into the character area and the image area. From (m value) to the second gradation (n value: m>
n) can be gradation converted. Further, the image gradation conversion can be executed by using a neural network in which each unit is arranged in a plane.

The neural network is a network in which units, which are engineering models of neurons (also called nerve units or neurons), which are considered to be the basic constituent units of the brain, are interconnected. Neural networks are classified into hierarchical networks and interconnected networks according to their connection form. The hierarchical network is also called a perceptron type network and has a structure in which a large number of units are connected in multiple layers. The first layer of the hierarchical network is an input layer for inputting information from the outside, the final layer is an output layer for outputting information to the outside, and the layer between them is called an intermediate layer. The characteristic of the hierarchical connection is that there is no connection within the layer,
The coupling between layers is that there is only unidirectional coupling from the input layer to the output layer.

On the other hand, the mutual connection type network has a structure in which a large number of units are mutually connected without forming a hierarchical structure. Unlike a hierarchical network in which information flows from an input layer to an output layer only in one direction, a feedback loop exists in an interconnected network, so that information repeatedly travels inside the network. The state of the network starts from a certain initial state, and while each unit repeatedly changes the internal state and the output, the state of the entire network reaches a certain stable equilibrium state or goes around several states. Fall into a patrol state. As a typical network of this kind, a Hopfield type neural network (Reference 1: "Neural Computing-Theory and Practice-" PHILIP D.W.
By ASSERMAN Naohiro Ishii / Minoru Tsukada Co-translation Morikita Publishing, pp 81-96 1993, Reference 2: Hopfie
ld JJ and Tank DW: "Neural Computation of Decis
ions in Optimization Problems ", Biol.Cybern., 52, p
p.141-152 1985) and the Boltzmann machine.

According to the present invention, image gradation conversion is performed by a configuration based on the Hopfield neural network. FIG. 5 shows a general configuration diagram of the Hopfield neural network. In the figure, each unit 6 constituting the Hopfield neural network is mutually connected to all other units through a path 7 capable of bidirectionally transmitting information, and the output of one unit is different from that of the other unit. A feedback loop path is configured to return to the unit via the.

Further, in the Hopfield type neural network, there is a concept of a coupling coefficient as the ease of transmitting information between the units. When receiving information from other units, each unit of the Hopfield neural network is supposed to receive a value obtained by multiplying the information value by this coupling coefficient.

In the original Hopfield type neural network, if there are N units, all N units are connected to each other. However, in the image processing of the present invention, the units arranged in a plane are , It is sufficient if it is connected to a nearby unit within a predetermined distance between units. Even if there are more connections, 0 is assigned to the coupling coefficient, and information transmission as a neural network does not occur between units distant from each other. In order to apply the above neural network to image gradation conversion, each pixel to be converted is assigned to each unit of the neural network. Then, assuming that the coupling coefficient between a certain unit and its neighboring unit is the weighted average coefficient at the time of gradation conversion, the unit to which the pixel of interest is assigned calculates the weighted average coefficient from the unit to which the neighboring pixel is assigned to the value of the neighboring pixel. You can receive the multiplied value.

The gradation conversion algorithm is as follows. When gradation conversion of a pixel is performed, the weighted average value of the pixel of interest and the pixel of interest after gradation conversion is taken into consideration as the brightness of the pixel of interest before gradation conversion in consideration of this pixel of interest and its neighboring pixels. If the gradation conversion is performed so as to be closest to that, a high quality gradation conversion result can be obtained. Pixels that are separated from the target pixel by a certain amount or more need not be considered at all in the gradation conversion of the target pixel.

Next, FIG. 6 shows an example of how to take neighboring pixels. In FIG. 6, the star symbol is the target pixel 8, and the ⊚ mark and the ◯ mark are neighboring pixels 9 that are considered during the gradation conversion of the target pixel 8.
Indicates. The reason why the pixels in the vicinity of 12 are selected is as follows. First, in the image area, it is considered that only pixels close to the pixel of interest to be subjected to gradation conversion are involved. Secondly, since there are many vertical and horizontal lines in the character area, it is considered that the information on the vertical and horizontal pixels with respect to the pixel of interest is more involved than the information on the diagonal pixels. If neighboring pixels are selected in consideration of such a situation, the number of neighboring pixels does not have to be 12.

The weighted average coefficient of the target pixel 8 and the neighboring pixel 9 becomes a coupling coefficient between the units forming the neural network, and the neighboring 8 which is one pixel away from the target pixel.
The weighted average coefficient from the pixel (⊚) is set to a, and the weighted average coefficient b from four neighboring pixels (∘) separated by 2 pixels is set.
b. In addition, the sum of the weighted average coefficients of the neighboring pixels and the weighted average coefficient z of the provisional value of the own unit, that is, z
As long as + 8a + 4b = 1, any value may be set as the weighted average coefficient. If neighboring eight pixels values x ₁ ~x ₈ apart by one pixel from the target pixel, the values of four neighboring pixels separated by two pixels from the target pixel and x ₉ ~x _12, a weighted sum of neighboring pixels Σ Becomes equation (2).

Σ = a (x ₁ + x ₂ + ... + x ₈ ) + b (x ₉ + z ₁₀ + ... + x ₁₂ ) ... (2)

From the above, when the weighted average coefficient of gradation conversion is used as the coupling coefficient between units and the state of each unit is made to correspond to the value of each pixel, it is represented by a Hopfield neural network as a binary image. It is possible to perform gradation conversion of digital image data in which character pattern data and multi-gradation image data are mixed without dividing the character pattern data and the image data. Moreover, due to the parallelism of the Hopfield type neural network, gradation conversion can be executed at high speed by highly parallel processing.

[0020]

Embodiments of the present invention will now be described in detail with reference to the drawings. FIG. 1 is a configuration diagram of an embodiment of an image processing apparatus according to the present invention. In the figure, an image processing apparatus of the present invention comprises an image input section 1 for converting a document or the like in which characters and images are mixed into digital image data for input, an image storage section 2 for storing digital image data, and a digital image. Image gradation conversion unit 3 for gradation conversion of data
And an image output unit 4 for outputting digital image data after gradation conversion, and a CPU 5 for controlling the entire apparatus. In FIG. 1, thick arrows indicate the flow of image data, and thin arrows indicate the flow of control.

FIG. 2 is a flow chart of an embodiment of the image gradation conversion method according to the present invention. In this embodiment, digital image data in which character pattern data represented by a binary image and multi-tone image data represented by a plurality of gradations (m values) included in the binary section are mixed. Is an example in which the present invention is applied to image gradation conversion for binarizing each pixel of.
In FIG. 2, it is assumed that the digital image data to be subjected to gradation conversion is loaded in advance in the Hopfield type neural network.
In this state, as described in the item of action, each unit of the Hopfield type neural network is read with digital image data before gradation conversion pixel by pixel.

In the flowchart of FIG. 2, first, a target pixel which is a gradation conversion target is randomly selected (step S10). Next, the outputs of the twelve units that carry the pixels in the vicinity of the pixel of interest are read as the inputs (x1, x2, ..., X12) of the unit that carries the pixel of interest (step S20). Next, the values of the 12 neighboring pixels are multiplied by the weights of the weighted average coefficients and added to obtain the weighted sum Σ of the neighboring pixels (step S3).
0). It is desirable that the sum of products (step S30) for calculating Σ be performed at the same time as the reading of the neighboring pixels in step S20.

Then, the value obtained by multiplying the black value by the weighted average coefficient as the first temporary value of the two values is added to the above Σ,
A first weighted average value is calculated (step S40). Next, the value obtained by multiplying the white value by the weighted average coefficient as the second temporary value of the two values is added to the above Σ, and the second weighted average value is calculated (step S41). Next, the value of the pixel of interest (pixel of its own unit) is read (step S
42). At this time, steps S40 to S42
It is desirable to be processed in parallel.

Next, the two weighted average values are compared with the value of the pixel of interest, and the weighted average value closer to the value of the pixel of interest is selected (step S50). Then, a provisional value that produces a weighted average value closer to the value of the target pixel, that is, black or white is selected (step S60), and the target pixel is updated with that value (step S70). Next, it is determined whether the gradation conversion of all the pixels has been completed (step S80). If NO, the process returns to step S10.
The next pixel of interest is selected. If YES in the determination in step S80, the convergence condition is determined (step S90). If the convergence condition is satisfied, the gradation conversion is ended, but if the convergence condition is not satisfied, another process is performed. Since the gradation conversion is performed for each pixel, the process returns to step S10.

In the above embodiment, the target pixel which is the gradation conversion target is randomly selected (step S10).
However, the selection may be performed sequentially in the raster scan direction of the input image, or the gradation conversion of all pixels may be simultaneously executed by maximizing the characteristics of the neural network.

Next, FIG. 3 shows a block diagram of an embodiment of the image gradation converting section 3 which constitutes the image processing apparatus according to the present invention. The image gradation conversion unit 3 is a neural network in which a Hopfield neural network is arranged in a plane and has mutual information transmission paths only between nearby units. Then, each neuron model constituting this neural network, that is, the unit Q (i, j)
Are interconnected with a unit having an inter-unit distance of 1 with a coupling coefficient a, and are interconnected with a unit having an inter-unit distance of 2 with a coupling coefficient b.

FIG. 4 is a block diagram showing the internal structure of each unit Q (i, j) forming the neural network. In FIG. 4, a unit (neuron model)
Are ₁₂ input terminals x _{1 to} x 12 that receive outputs from the 12 nearby units and terminals for distributing the outputs to the 12 nearby units as terminals for exchanging signals with the outside of the unit. There are output terminals y _{1 to} y _{12 of the} book.

Inside the unit, a pixel holding circuit 10 for holding the value A of the pixel of interest and gradations (P (1) to P (P) to be converted).
A selector 20 for selecting a temporary value of (n value consisting of (n)), a conversion candidate holding circuit 30 for holding the conversion candidate (temporary value) output from the selector 20, and an input from 12 neighboring units First sum-of-products circuit 40 for weighting and adding the values of neighboring pixels to be calculated and calculating Σ;
A second product-sum circuit 50 for weighting a temporary value selected by 0 and adding it to Σ to obtain a weighted average value B, and an error for calculating the absolute value of the difference between the value A of the target pixel and the weighted average value B Calculation circuit 6
0, an error holding circuit 70 that holds the output of the error calculation circuit 60, an error comparison circuit 80 that compares the minimum error up to the previous time and the current error, and if the current error is smaller based on the error comparison, A control circuit (not shown) that controls the content of the conversion candidate holding circuit 30 to be updated with the provisional value of this time at the same time as updating the content of the error holding circuit 70 with the error of this time, and performs other control is provided. ing.

Next, the operation of this embodiment will be described. First, the digital image data before conversion is loaded. At this time, by a control signal (not shown), the register circuit composed of the input terminal x ₁ , the pixel holding circuit 10, and the output terminal y ₂ is connected in a potato-like manner between the units adjacent to each other, and the initial load path is connected. Constitute. Then, the pre-conversion digital image data for each pixel is loaded from the image storage unit 2 into each unit via the initial load path. The initial load path may be one in the neural network or may be divided into several paths and parallelized.

When the initial loading of the digital image data before conversion is completed, each neuron constituting each neural network, that is, each unit is fired. The firing of neurons may be randomly selected one by one to fire, selected to scan a network, or all neurons may fire at the same time.

When the neuron fires, the output of each neighboring unit is taken in by the input terminals x1 to x12, and the first sum of products circuit 40 weights each input signal by the respective coupling coefficient. Multiply the average coefficient to find the sum Σ. This Σ is a weighted average value of neighboring pixels. Next, the selector 20 converts the converted gradation (n
The first tentative value: P (1) is selected from the values) and held in the conversion candidate holding circuit 30, and P (1) is multiplied by the weighted average coefficient by the second product-sum circuit 50. Value and Σ
And are added to obtain the first weighted average value B1. Next, | A−B1 |, which is the absolute value of the difference between B1 and the value A of the pixel before conversion held in the pixel holding circuit 10, is obtained, and this value is held in the error holding circuit 70.

Next, the selector 20 selects a second tentative value: P (2), and a value obtained by multiplying P (2) by a weighted average coefficient and Σ are added to obtain a second weighted average value B2. To be Next, | A−B2 |, which is the absolute value of the difference between B2 and the value A of the pixel before conversion held in the pixel holding circuit 10, is obtained, and | A−B2 | is held in the error holding circuit 70. The error comparing circuit 80 compares | A-B1 |

As a result of error comparison, | AB2 |
If it is smaller than A-B1 |, the content of the error holding circuit 70 is updated with the value of | A-B2 |, and the content of the conversion candidate holding circuit 30 is updated with P (2). If | A−B2 | is greater than or equal to | A−B1 |, the error holding circuit 70 and the conversion candidate holding circuit 30 are not updated.

Similarly, the selector 20 sequentially outputs new temporary values.
, The weighted average value is calculated with the temporary value, the error from the value of the pixel before conversion is calculated, and the error is compared with the minimum error up to the temporarily selected temporary value held in the error holding circuit 70. .
In this way, when the provisional value: P (n) is repeated, the value finally held in the conversion candidate holding circuit 30 becomes the value obtained by gradation conversion of the pixel with the minimum error, and with this value, the pixel holding circuit 10 The content of is updated and the processing of the pixel concerned is once terminated. The order of selecting the temporary value by the selector 20 is based on the above equation (1). If n = 4, P (1) = 0, P (2) = 1, P (3) = 1 /
3, P (4) = 7/9.

Next, another pixel is selected, different pixels are processed one after another, and the processing is continued until all pixels are processed. Alternatively, if all pixels are processed simultaneously, the convergence condition is determined. In determining the convergence condition,
If the convergence condition is not satisfied yet, the process for all pixels is repeated again. At first, most of the pixels are updated with pixel values by gradation conversion, but as the number of calculations is increased, the number of pixels that are updated gradually decreases, and eventually it is almost not updated and settles into a certain state. . This is a feature of the Hopfield neural network. The Hopfield neural network is characterized in that when the energy of the network is considered and the network is operated, the energy of the network decreases and the energy reaches the minimum point.

As for the judgment of the convergence condition, it is judged that the convergence has occurred when the ratio of the number of updated pixels to the total number of pixels becomes a fixed ratio, for example, 5% or less. When the pixel update due to the gradation conversion is converged, the contents of the pixel holding circuit 10 of each unit are output to the image storage unit 2 in a potato-like manner via the initial load path, and the digital image data after the gradation conversion is imaged. It is output by the output unit 4.

[0037]

As described above, according to the present invention, 2
Each pixel of digital image data in which character pattern data represented by a value image and multi-gradation image data represented by a plurality of gradations (m values) included in the binary interval are mixed or binarized In an image gradation conversion method for performing n-value conversion (m> n), an image gradation conversion method for obtaining a favorable gradation conversion image for both characters and images without dividing the character area and the image area of digital image data Can be provided. Further, according to the present invention, there is an effect that it is possible to provide a high-speed image processing apparatus which performs gradation conversion without dividing a character area and an image area of digital image data.

[Brief description of drawings]

FIG. 1 is a configuration diagram showing an entire image processing apparatus of the present invention.

FIG. 2 is a flowchart of an embodiment of the image gradation conversion method of the present invention.

FIG. 3 is a block diagram showing a neural network that constitutes an image gradation conversion unit of the image processing apparatus of the present invention.

FIG. 4 is a block diagram showing each neuron model (unit) of the neural network that constitutes the image gradation conversion unit of the image processing apparatus of the present invention.

FIG. 5 is an explanatory diagram of neighboring pixels at the time of image gradation conversion.

FIG. 6 is a general configuration diagram of a Hopfield neural network.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 image input section 2 image storage section 3 image gradation conversion section 4 image output section 5 CPU 6 unit 7 bidirectional path 8 target pixel 9 neighboring pixel 10 pixel holding circuit 20 selector 30 conversion candidate holding circuit 40 first sum-of-products circuit 50 Second product-sum circuit 60 Error calculation circuit 70 Error holding circuit 80 Error comparison circuit x _{1 to} x ₁₂ Input terminal from neighboring unit y _{1 to} y ₁₂ Output terminal to neighboring unit

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification number Office reference number FI Technical indication location H04N 1/407 4226-5C H04N 1/40 101 E

Claims (5)

[Claims] 1. Each pixel of digital image data in which character pattern data represented by a binary image and multi-gradation image data represented by a plurality of gradations included in the binary interval are mixed is represented by a binary value. In the image gradation conversion method for converting, a step of selecting a target pixel from a set of pixels to be converted, and a step of setting a first value from the two values as a first temporary value of the target pixel A step of obtaining a first weighted average value from the first temporary value and the values of a plurality of pixels in the vicinity of the target pixel; and a second value different from the first value from the two values, Setting a second provisional value of the pixel of interest, obtaining a second weighted average value from the second provisional value and the values of a plurality of pixels in the vicinity of the target pixel, and the first weighted average A process for determining which of the value and the second weighted average value is closer to the value of the pixel of interest. When the first weighted average value is closer to the value of the target pixel than the second weighted average value, the first value is set as the updated value of the target pixel, and the second weighted average value is If the value of the pixel of interest is closer to the value of the pixel of interest than the first weighted average value, then the step of setting the second value as the updated value of the pixel of interest, and repeating all the steps until the processing of all the target pixels is completed. And a step of repeating all the above steps until a predetermined convergence condition is satisfied, the image gradation conversion method. 2. Digital image data in which character pattern data represented by a binary image and multi-gradation image data represented by a first gradation (m value) included in the binary interval are mixed. An image gradation conversion method for converting each pixel into digital image data represented by a second gradation (n value; m> n) smaller than the first gradation (m value), A step of selecting a target pixel from a set of pixels forming the target digital image data, and a value indicating the first brightness from the second gradation; P ₍₁₎
Is set as a first provisional value of the target pixel, a first weighted average value is obtained from the first provisional value and the values of a plurality of pixels in the vicinity of the target pixel, and the second A value indicating the second brightness from the gradation; P ₍₂₎
Is set as a second temporary value of the pixel of interest, a second weighted average value is obtained from the second temporary value and the values of a plurality of pixels in the vicinity of the pixel of interest, A step of setting a k-th (n ≧ k ≧ 3) value; P _(k) as a k-th temporary value of the pixel of interest from the gradation (n-value); Determining a k-th weighted average value from the values of a plurality of pixels in the vicinity of the target pixel; determining which of the k weighted average values is closest to the value of the target pixel; A step of setting the temporary value that has generated the closest weighted average value as an updated value of the pixel of interest, a step of repeating all the steps until processing of all target pixels is completed, and a predetermined convergence condition is satisfied And a step of repeating all the above steps up to the step of converting an image gradation. 3. The image gradation conversion method according to claim 2, wherein the value set as the temporary value of the pixel of interest is An image gradation conversion method characterized by being determined by the formula (1). 4. The image gradation conversion method according to claim 1, wherein a weighted average value is obtained from a temporary value of the pixel of interest and values of a plurality of pixels near the pixel of interest. An image gradation conversion method, wherein the weighted average coefficient of pixels in the vicinity is set to be smaller as the distance from the pixel of interest to the pixel in the vicinity increases. 5. An image processing apparatus using a planar neural network in which units arranged in a plane are connected to units in the vicinity thereof so as to be capable of exchanging information with each other. Means for calculating the weighted linear sum of the inputs from the units in the vicinity of, the means for holding the pixel value before the gradation conversion assigned to the unit, and the possible value of the converted gradation. An image processing apparatus comprising: a means for performing image gradation conversion using the image gradation conversion method according to claim 1.