JP6763637B2 - Image restoration method - Google Patents

Description

The present invention relates to an image restoration method for restoring an observed image (deteriorated image) in which a function representing an image deterioration process is a Gaussian function to obtain a restored image as close as possible to the original image without blurring.

The main causes of deterioration of the image obtained through the camera lens are the relative movement of the camera and the subject, what is called out-of-focus or out-of-focus, the random fluctuation of the medium such as a thick air layer, and the aberration of the lens. And so on. The function that expresses the deterioration process is different for each. Here, the case of restoring an image deteriorated due to lens aberration will be taken up. The function representing the deterioration process of the image due to the aberration of the lens is a Gaussian function.

Patent Document 1 describes an image restoration device that generates a restored image with suppressed ringing and noise by restoring a true image based on a known PSF (point spread function) and one deteriorated image. Are listed.

The image restoration device described in Patent Document 1 includes a base image generation processing unit that generates a base image by performing a base image generation processing based on a deteriorated image, and a base image generated from the base image generation processing unit and a PSF. A blurred image generation processing unit that generates a blurred image by performing convolution integration with (point spreading function), and a degraded residual image by calculating the difference between the generated blurred image and the degraded image. The restored residual image is restored by performing image restoration processing on the generated deterioration residual image and the generated deterioration residual image based on the generated base image so as to change the strength of the constraint. It is provided with an image restoration processing unit for generating a restored image and a restoration image generation processing unit for generating a restored image by adding a generated base image to the generated restoration residual image.

In addition, Non-Patent Document 1 describes the engineering application of wavelets with respect to image processing / signal processing.

JP-A-2009-271725

"Signal processing and image processing by wavelet" by Hiroki Nakano, Shizuo Yamamoto, Yasuo Yoshida, published in August 1999, Kyoritsu Shuppan Open CV 2.2 documentation

However, with the technique disclosed in Patent Document 1, even if a restored image with suppressed ringing and noise can be generated, the image deteriorated due to the aberration of the lens is restored and the restored image is as close as possible to the original image without blurring. It's hard to get.

The present invention has been made in view of such problems of the prior art, and restores an observed image (deteriorated image) in which the function representing the image deterioration process is a Gaussian type function without blurring. It is an object of the present invention to provide an image restoration method for obtaining a restored image as close as possible to the original image.

In order to solve the above problems, the present invention is an image restoration method for restoring a deteriorated image to obtain a restored image as close as possible to the original image without blurring, and a preparatory processing step for converting the deteriorated image into a YUV color space. A decomposition processing step of decomposing a Y image (Y plane) in the YUV color space obtained by this preparatory processing step into a plurality of decomposed images, and a Y image (Y image) by reconstructing the decomposed image obtained by this decomposition processing step. It is provided with a reconstruction processing step of forming a Y plane) and combining it with a UV image (UV plane) converted in the preparatory processing step to obtain a restored image.

In the image restoration method, the preparatory processing steps are, for example, 1 to 7 below.
1. 1. The deteriorated image (input image) f (x, y) is converted into the YUV color space.
2. 2. The Y image f in the YUV color space is converted into 32 bits.
3. 3. Create a diffusion value array of σ ₀ to _J from the input parameters (restoration depth t, ratio k).
4. Creating a Gaussian kernel Gσ ₀ ² diffusion value sigma ₀ ^2.
5. To create a LOG kernel LOGσ ₁ ² ~LOGσ _J ² diffusion value σ ₁ ² ~σ _J ^2.
6. A restoration depth array with σ as −t is created by simply using the array contents of 3 above.
7. Perform the same operations as 4 to 5 above with the abdominal depth array of 6 above to create a restored kernel.
The decomposition processing steps are, for example, the following 8 to 10.
8. Image Gσ ₀ ² * f ₀ of applying the Gaussian kernel Gσ ₀ ² diffusion value sigma ₀ ² of the 4 in the Y image f ₀ is subtracted from the Y image f ₀ to generate the decomposed image f ₁ of the decomposed once ( f ₁ = f ₀ −f ₀ * Gσ ₀ ² ).
The separation image f ₁ after 9.1 times exploded subtracted from the image f ₁ * LOGshiguma ₁ ² a separation image f ₁ after one decomposed according to the LOG kernel LOGshiguma ₁ ² diffusion values sigma ₁ ² of the 5 2 The decomposed image f ₂ after the decomposition is generated (f ₂ = f ₁ −f ₁ * LOGσ ₁ ² ).
10. The same operation as in 9 above is repeated to generate decomposed images f _J , and J decomposed images are obtained.
Further, the reconstruction processing steps are, for example, the following 11 to 14.
11. The restored LOG kernels LOGσ ₁ ² _{-t to} LOGσ _J ² _-t created in 7 above are applied to the decomposed images f _{1 to} f _J , respectively, and the sum of all the images is obtained to create a total image.
12. The restored Gaussian kernel Gσ ₀ ² _-t created in 7 above is applied to the Y image, and the sum image created in 11 above is added.
13. The Y image is returned to the bit depth of the input image f (x, y).
14. The image obtained in 13 above is combined with the UV plane of the input image f (x, y) as a Y plane and output as a restored image I'(x, y).

According to the image restoration method according to the present invention, an observed image in which a function representing an image deterioration process, such as a deteriorated image due to lens aberration or a deteriorated image through random fluctuations of a medium such as a thick air layer, is a Gaussian function. Degraded image) can be restored to obtain a restored image as close as possible to the original image without blur.

Configuration diagram of an image restoration device that implements the image restoration method according to the present invention. A flowchart showing a processing procedure of the image restoration method according to the present invention. Explanatory drawing of line buffer operation method in disassembly processing / restoration processing (A) is an image deteriorated due to lens aberration, and (b) is an image after restoration. Is the original picture of the test chart Is a blurred image of the original image Is a restored image of a blurred image

Embodiments of the present invention will be described below with reference to the accompanying drawings. As shown in FIG. 1, the image restoration device 1 that implements the image restoration method according to the present invention includes a preparatory processing means 2 that converts a deteriorated image f (x, y) into a YUV color space, and the preparatory processing means 2 The decomposition processing means 3 that decomposes the Y image (Y plane) f ₀ in the color space of YUV obtained by the above into a plurality of decomposed images f _{1 to} f _J , and the decomposed images f ₁ to obtained by the decomposition processing means 3. Reconstruction processing means for reconstructing f _J to obtain a Y image (Y plane) f and combining it with a UV image (UV plane) fuv converted by the preparation processing means 2 to obtain a restored image I'(x, y). 4 is provided.

The image produced by the lens and the image sensor (imaging sensor) is generally deteriorated by Gaussian blur due to lens aberration and noise due to high frequency. Originally, the light that should be collected in one pixel is blurred by the Gaussian distribution and affects the surrounding pixels. An image with a lens that has no aberration at all does not cause blurring, but the actual image is a set of blurred light with a Gaussian distribution.

To express an image with a mathematical formula, let the two-dimensional Gaussian function be Gt (x, y).
Gt (x, y) = 1 / 2πt * exp (-(x ² + y ² ) / 2t) (Equation 1)
t is a diffusion value indicating the spread of the Gaussian function.

Assuming that the image (original image) without aberration (without blur) is I (x, y), the deterioration process f (x, y) of the blurred image (deteriorated image) is expressed by the following equation.
f (x, y) = K ^t I (x, y) + ω (x, y) = Gt (x, y) * I (x, y) + ω (x, y) (Equation 2)
* Is an operator that represents convolution and is defined by the following equation.

Gt (x, y) * I (x, y) = ∬Gt (x', y') * I (x ・ x', y ・ y') dx'dy' (Equation 3)
In order to calculate this with finite pixels, it is converted to the following equation.
Gt (x, y) * I (x, y) = Σy'Σx' (Gt (x', y') * I (x ・ x', y ・ y')) (Equation 4)
Note that ω (x, y) is omitted because it is considered to be high-frequency noise during observation.

Here, K ^t is an operator that convolves a function with a two-dimensional Gaussian function Gt (x, y). Since t represents the spread of the Gaussian function, if ^t is considered as a time, K ^t can be considered to represent, for example, how the ink at one point gradually diffuses. Of course, the diffusion time is t ≧ 0, but if this is t <0, it can be regarded as the reverse of Gaussian function diffusion, that is, the process of restoration.

If the operator that performs such an operation is expressed as K ^-t , the approximate solution I'(x, y) of the original image I (x, y) can be expressed as the following equation.
I'(x, y) = K ^-t f (x, y) = I (x, y) + K ^-t ω (x, y) (t> 0) (Equation 5)
The relationship between K ^t and K ^{-t is} as follows.
K ^t K ^-t f (x, y) = f (x, y) (Equation 6)

Since it cannot be used as it is, the following relational expression is used.
One is the relational expression K ^-t Gσ ₂ * f = Gσ _2-t * f (σ ² > t) for the convolution of the Gaussian function (Equation 7).
The other is the relational expression K ^-t LOGσ ₂ * f = LOGσ _2-t * f (σ ² > t) for the convolution of the LOG function (Equation 8).

The LOG function is a Laplacian of Gaussian filter often used in image processing, and is obtained by applying the Laplacian operator L to a two-dimensional Gaussian function.
Laplacian operator L = ∂ ² / ∂ x ² + ∂ ² / ∂y ² (Equation 9)

When the Laplacian operator L is applied to the Gaussian function, the following equation is obtained.
LOGσ ₂ (x, y) = 1 / (2πσ ⁴ ) * (2- (x ² + y ² ) / σ ² ) * exp (− (x ² + y ² ) / 2σ ² ) (Equation 10)
If this function (Equation 10) is regarded as a filter, it becomes a bandpass filter that passes an intermediate frequency.

When the deteriorated image f (x, y) is decomposed using Equations 7 and 8,
f = Gσ ₀ ² * f ₁ + Σ ^J _{j = 1} LOGσ _j ² * f _j + f _{J + 1} (Equation 11)
In this proposal, the deteriorated image f (x, y) is converted into the following equation in order to calculate with finite pixels.
f (x, y) = Σ y 'Σ x' (Gσ 0 2 -t (x ', y') * f 1 (x · x ', y · y')) +
Σ ^J _{j = 1} (Σ y'Σ x'(LOG σ _j ² _-t (x', y') * f _j (x ・ x', y ・ y'))) + f _{J + 1} (Equation 12)
Here, σ ² is a scale parameter of the Gaussian function and the LOG function.

f _j is a residual image after j decomposition, and is obtained by the following equation.
1st stage f ₁ = f−Gσ ₀ ² * f (Equation 13)
Stage j f _j = f _j-1 −LOGσ _j-1 ² * f _j-1 (1 ≦ j ≦ J) (Equation 14)

When the restoration is expressed by equations from equations 7, 8 and 11, K ^-t f = Gσ _2-t * f + Σ ^J _{j = 1} LOGσ _2-t * f _j + K ^-t f _{J + 1} (Equation 15)
However, in order to stabilize the equation, σ _J ² > t, σ _{J + 1} ² ≤ t (Condition 1)
Must.

At this time, K ^−t f _{J + 1} of Equation 15 is unstable, but since this term is the highest frequency component of the deteriorated image, most of it is noise. Therefore, this term is ignored. Therefore, the restored image I'(x, y) to be obtained is given by the following equation.
I'(x, y) = Gσ _2-t * f + Σ ^J _{j = 1} LOGσ _2-t * f _j (Equation 16)

J is the number of decompositions of the deteriorated image, and t is the restoration depth.
σ _j is the j-th diffusion (decomposition frequency), and is a geometric progression with σ _{j + 1} = kσ _j (k <1) (Equation 17).

In this embodiment, the restored image I'(x, y) is converted into the following equation in order to calculate the equation 16 with a finite number of pixels.
I '(x, y) = Σ y' Σ x '(Gσ 0 2 -t (x', y ') * f (x · x', y · y ')) +
Σ ^J _{j = 1} (Σ y'Σ x'(LOG σ _j ² _-t (x', y') * f _j (x ・ x', y ・ y'))) (Equation 18)

Equations 12 and 18 are equations that can be implemented by a discrete program, but it is necessary to determine six parameters for the degraded image f (x, y). That is, the number of decompositions J, the restoration depth t, the Gaussian / LOG kernel size n, the initial values σ ₀ and σ _{1 of} diffusion, and the ratio k.

The kernel is a matrix (x', y') that performs convolution calculations of finite pixels, and is a square matrix of n × n (n is an odd number) with a center. These have a specific relationship from the above condition 1.
σ ₁ ≧ 1.44 σ ₀
σ _{j + 1} ≧ 0.52 σ _j
The smaller the number of decompositions, the more the processing ends.
σ ₁ = 1.44 σ ₀
k = 0.52
Will be. However, if the ratio k is coarse (small), the decomposition accuracy deteriorates and sufficient restoration cannot be obtained.

To calculate the kernel size (the size of the square matrix) automatically, do as follows.
The kernel of the embodiment is a Gaussian / LOG two-dimensional matrix. Both Gaussian / LOG have the property that the center of the matrix is the origin and the value approaches zero as the distance from the center increases.
Therefore, let r be the distance from the center, and σ, which is less than 1/256 on the circumference of radius r, can be easily approximated as σ = 0.3 · r + 0.8.
To find the kernel size from σ, use the above formula
It can be approximated as r = (σ −0.8) /0.3 + 1.
(+1) at the right end of the above equation is necessary to make a square matrix larger by one element so that there is no rounding error on each outer side when making a manufacturing method matrix.
From this, the kernel size can be calculated as (r × 2) +1.

The fidelity expressed in dB with respect to the ratio k is as follows.
k = 0.7: 45 dB, k = 0.8: 49 dB, k = 0.9: 55 dB, k = 0.95: 59 dB. To ensure fidelity of 50 dB, k = 0.8 or higher is required.

The necessary conditions are as described above, but in order to quickly determine the parameters for practical use, it is necessary to automatically calculate the parameters.
In this proposal, when two parameters, the restoration amount (restoration depth) t and the accuracy (ratio) k, are determined, the other parameters are automatically calculated using an algorithm that is determined arithmetically.
1. Initial value of diffusion σ ₀ = √ (t + 0.5) Numerical value obtained empirically 2. Initial value of diffusion σ ₁ = 1.44σ _{0 From} Non-Patent Document 1 3. Gaussian / LOG kernel size n = 2 ((() σ ₁ -0.8) /0.3+1.0)+1
Approximately odd number calculated back from σ = 0.3 (n / 2-1) +0.8 from Non-Patent Document 2. 4. Number of decompositions J = σ _j ≤ 0.7 The center of j LOG is within (0 ± 0.5, 0 ± 0.5) If σ = 0.7 and the accuracy k is close to 1.0, the number of decompositions J may become very large.

In this case, the number of decompositions is reduced by replacing σ _{j + 1} = kσ _j (k <1) in Equation 17 with an arithmetic progression as in the following equation.
σ _j ＝ σ ₁・ (σ ₁・ kσ ₁ ) ・ (j−1) (Equation 19)

The restoration processing algorithm is mainly divided into three types: preparation processing, decomposition processing, and reconstruction processing.
First, in the preparatory process, the following process is required. F _j is required to process Equation 18 representing the restored image I'(x, y). f _j is calculated by Equation 12. Equations 12 and 18 require a Gaussian kernel with a diffusion value of σ ₀ ² and J LOG kernels with a diffusion value of σ ₁ ^{2 to} σ _J ² . These kernels are calculated by Equations 1 and 10 from the initial values σ ₀ and σ _{1 of} diffusion automatically calculated from the input parameters restoration amount (restoration depth) t and accuracy (ratio) k.

Since σ _{0 to} σ _J are reconstructed, they are stored in an array or the like by Equation 17. However, when the number of decompositions J is larger than a certain level (specified by the memory capacity of the mounting), Equation 19 is used.

Next, in the decomposition process, the following process is performed. Once the kernel is ready, according to the first half of Equation 12, Gσ ₀ ² * f _{0 obtained} by applying the Gaussian kernel Gσ ₀ ² with a diffusion value of σ ₀ ² to the deteriorated image f (x, y) and the deteriorated image f (x, y). Find the difference with and let this be f ₁ . Then, according to the second half of the equation 12, and f _1, we obtain the difference between LOGσ ₁ ² * f ₁ to which the LOG kernel LOGσ ₁ ² diffusion values sigma ₁ ² to f _1, which is referred to as f _2. Repeat this process until the following f _J, obtained J Like the separation image f ₁ ~f _J degraded image f (x, y).

Next, in the reconstruction process, the following process is performed. According to Equation 18, the restored LOG kernel having the diffusion values σ ₁ ² _{-t to} σ _J ² _-t is applied to the decomposed images f _{1 to} f _J , and the sum of all the images is calculated. Next, a restored Gaussian kernel having a diffusion value of σ ₀ ² _-t is applied to the deteriorated image f (x, y), and all the images to which the above restored LOG kernel is applied are added to the image to restore the image I'(. Let x, y).

Here, the algorithm is based on grayscale. Therefore, the input image (deteriorated image) is first converted into the YUV color space, and the restoration process is applied only to the Y image (Y plane).

Next, the image restoration method according to the present invention will be described with reference to the flowchart shown in FIG. First, the deteriorated image (input image) f (x, y) is converted into the YUV color space (step SP1). Next, the Y image (Y plane) f ₀ in the YUV color space is converted into 32 bits (step SP2).

In step SP3, a diffusion value array of σ ₀ to _J is created from the input parameters (restoration depth t, ratio k) (Equation 17 or Equation 19). In step SP4, a Gaussian kernel G σ ₀ ^{2 having} a diffusion value σ ₀ ² is created (Equation 1). In step SP5, to create a LOG kernel LOGσ ₁ ² ~LOGσ _J ² diffusion value σ ₁ ² ~σ _J ² (Equation 10).

In step SP6, the restoration depth array is created with σ as −t, simply by using the array contents of the diffusion value array as −t. In step SP7, the same work as in steps SP4 and SP5 is performed with the abdominal depth array in step SP6 to create a restored kernel. Up to this point, steps SP1 to SP7 are preparatory processing steps.

Then, in step SP8, Y image diffusion value of f ₀ in step SP4 sigma ₀ ² of Gaussian kernel Jishiguma ₀ image ² was applied Gσ ₀ ² * f ₀ the separation image after one exploded subtracted from the Y image f ₀ to generate the _{f 1 (f 1 = f 0} -Gσ 0 2 * f 0). In step SP9, the image LOGσ _j-1 ² * f _{j in which} the LOG kernel LOGσ _j-1 ² of the diffusion value σ _j-1 ² of step SP5 is applied to the decomposed image f _j-1 after (j-1) decomposition. _-1 is subtracted from the decomposed image f _j-1 after (j-1) decomposition to generate the decomposed image f _j after j decomposition (f _j = f _j-1 −LOGσ ₁ ² * f _j-1 ).

In step SP10, it is determined whether or not j (1 ≦ j ≦ J) is equal to J. If j <J, the process returns to step SP9 to generate the decomposed image f _j , and if j = J, the decomposed image f _j is generated. Proceed to step SP11. Up to this point, steps SP8 to SP10 are decomposition processing steps, and J images of decomposition are obtained.

Next, in step SP11, the restored LOG kernels LOGσ ₁ ² _{-t to} LOGσ _J ² _-t created in step SP7 are applied to the decomposed images f _{1 to} f _J , respectively, and the sum of all the images is obtained to create a total image. To do. In step SP12, the restored Gaussian kernel Gσ ₀ ² _-t created in step SP7 is applied to the Y image, and the summed image created in step SP11 is added.

In step SP13, the Y image is returned to the bit depth of the deteriorated image (input image) f (x, y). In step SP14, the image obtained in step SP13 is combined with the UV plane of the deteriorated image (input image) f (x, y) as a Y plane, and output as a restored image I'(x, y). Steps SP11 to SP14 are reconstruction processing steps.

FIG. 3 shows a line buffer operation method in the decomposition processing / restoration processing, and is the case of J = 3 in the decomposition processing step (step SP8 to step SP10) and the reconstruction processing step (step SP11).

Work to generate decomposed image f ₁ , decomposed image f ₂ , decomposed image f ₃ from Y image f ₀ and restore to decomposed images f _{1 to} f ₃ Apply LOG kernel LOGσ ₁ ² _{-t to} LOGσ ₃ ² _-t . The work of obtaining the sum of all the images and creating the summed image is processed in parallel in each line buffer (LBLV1 to LBLV4) at each processing time (1 to 8).
Here, the line buffer is a temporary storage memory having a size of the number of horizontal pixels of the input image × the size of the kernel vertical size.

Specifically, at the first time (top of FIG. 3), the Y image f ₀ is read from the line buffer (LV1), and the decomposed image f ₁ is written in the line buffer (LV2).
At the next time (second stage from the top of FIG. 3), the read modify (f ₀ − = f ₀ * G−) is written in the line buffer (LV1), and the decomposed image f ₂ is written in the line buffer (LV3).
At the next time (third stage from the top of FIG. 3), the read modify (f ₁ − = f ₁ − * L1-) is written in the line buffer (LV2), and the decomposed image f ₃ is written in the line buffer (LV4). ..
At the next time (fourth stage from the top of FIG. 3), read modify (f _2- = f _2- * L2-) is written to the line buffer (LV3), and read modify (f ₃ −) is written to the line buffer (LV4). = F _3- * L3-) is written.
At the next time (fifth stage from the top of FIG. 3), (f ₃ −) is read from the line buffer (LV4), and (f ₂ −) + (f ₃ −) is written in the line buffer (LV3).
At the next time (6th step from the top of FIG. 3), (f ₂ −) + (f ₃ −) is read from the line buffer (LV 3), and (f ₁ −) + (f ₂ ) is added to the line buffer (LV 2). -) + (F ₃ −) is written.
At the next time (7th stage from the top of FIG. 3), (f ₁ −) + (f ₂ −) + (f ₃ −) is read from the line buffer (LV2), and (f ₀ ) is set in the line buffer (LV1). Write-) + (f ₁ −) + (f ₂ −) + (f ₃ −).
By performing parallel processing in each line buffer at each processing time in this way, it is possible to speed up the decomposition processing and the restoration processing.

In FIG. 4, (a) is an image before restoration and (b) is an image after restoration of the image deteriorated due to lens aberration.
Further, FIG. 5 shows a generally used test chart, in which (a) is an original image, (b) is a blurred image of the original image, and (c) is an image obtained by restoring the blurred image by the method of the present invention. ..
From these images, it can be seen that the deteriorated image due to the Gauss-shaped blur due to the aberration of the lens is restored by the method of the present invention.

According to the present invention, there is provided an image restoration method capable of restoring an observed image (deteriorated image) in which a function representing an image deterioration process is a Gaussian function to obtain a restored image as close as possible to the original image without blurring. be able to.

1 ... Image restoration device, 2 ... Preparation processing means, 3 ... Disassembly processing means, 4 ... Reconstruction processing means.

Claims (2)

This is an image restoration method for restoring a deteriorated image to obtain a restored image as close as possible to the original image without blurring. A preparatory process for converting a deteriorated image into a YUV color space and a YUV color obtained by this preparatory process. A decomposition processing step of decomposing a Y image (Y plane) in space into a plurality of decomposed images, and a decomposition processing step of reconstructing the decomposed image obtained by this decomposition processing step to obtain a Y image (Y plane) and the preparatory processing step. It is equipped with a reconstruction processing process that combines with the UV image (UV plane) converted in step 1 to create a restored image.
The decomposition treatment step comprises the following steps 1 to 3.
Process 1. Y in the image f ₀ of the diffusion values sigma ₀ ² Gaussian kernel Jishiguma ₀ image ² was applied Gf ₀ * _{σ 0} ²⁾ is subtracted from the Y image f ₀ to generate the decomposed image f ₁ of the decomposed once (f ₁ = F−f ₀ * Gσ ₀ ² ).
Pull the process 2.1 times degradation after separation image f ₁ image f ₁ * LOGshiguma ₁ ² according to the LOG kernel LOGshiguma ₁ ² diffusion values sigma ₁ ² from separation image f ₁ after one exploded twice degradation The later decomposed image f ₂ is generated (f ₂ = f ₁ −f ₁ * LOGσ ₁ ² ).
Process 3. The same operation as in step 2 is repeated to generate decomposed images f _J , and J decomposed images are obtained.
The image restoration method, wherein the reconstruction processing step comprises the following steps 4 to 7.
Process 4. The restored LOG kernels LOGσ ₁ ² _{-t to} LOGσ _J ² _-t created in the preparatory processing step are applied to the decomposed images f _{1 to} f _J , respectively, and the sum of all the images is obtained to create a total image.
Process 5. The restored Gaussian kernel Gσ ₀ ² _-t created in the preparatory process step is applied to the Y image, and the sum image created in step 4 is added.
Process 6. The Y image is returned to the bit depth of the input image f (x, y).
Process 7. The image obtained in step 6 is combined with the UV plane of the input image f (x, y) as a Y plane and output as a restored image I'(x, y). The image restoration method according to claim 1, wherein the decomposition processing step and the reconstruction processing step are simultaneously processed in parallel by a plurality of line buffers.