JPH0944655A - Method and device for processing image - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a satisfactorily processed image by emphasizing only the component of a required frequency band in an image without emphasizing unwanted components such as noise. SOLUTION: At a multi-resolution decomposition processing means 2, an image signal S inputted from an image input means 1 is decomposed into the images of multiple resolution by a method such as Laplacian pyramid. The emphasis degree of the desired frequency band in the decomposed image is decided based on the signal value of image in the lower frequency band, and emphasizing processing is performed to the image by an emphasizing processing means 3. The emphasis-processed image and the other images are restored by a restoring processing means 4 and a processed image signal S' is provided. The processed image signal S' is reproduced as a visible image at an image output means 5.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image processing method and apparatus for performing image processing on a predetermined frequency band of an original image.

[0002]

2. Description of the Related Art In various fields, an image signal representing an image is obtained, the image signal is subjected to appropriate image processing, and then the image is reproduced and displayed. For example, in order to improve the diagnostic performance of a radiation image, a method of performing frequency enhancement processing such as blurring mask processing on an image signal has been proposed by the applicant (Japanese Patent Laid-Open No. 55-163772). This frequency processing is a process of adding a product obtained by subtracting the blur mask signal from the image signal representing the original image and multiplying the product by a degree of emphasis, and by doing so, a predetermined spatial frequency component is emphasized in the image. is there.

As another method for performing frequency processing on an image signal, Fourier transform, wavelet transform,
By converting the image into a multi-resolution image by subband conversion, etc., the image signal representing the image is decomposed into signals in multiple frequency bands, and among these decomposed signals, signals in the desired frequency band are emphasized. There has been proposed a method of performing a predetermined image processing such as.

In recent years, in the field of image processing, a method called Laplacian pyramid has been proposed as a new method for converting an image into multiple resolutions (for example, Japanese Unexamined Patent Publication No. H6-6 / 1994).
No. 301766). In this Laplacian pyramid, the original image is masked with a mask similar to a Gaussian function, and then the image is sub-sampled to reduce the number of pixels by half to halve the original image. A blurred image of size is obtained, pixels having a value of 0 are interpolated into the sampled pixels of this blurred image to return to an image of the original size, and this image is further masked by the mask described above. A blurred image is obtained, and the blurred image is subtracted from the original image to obtain a detailed image in a predetermined frequency band of the original image. By repeating this process for the blurred image obtained, N blurred images each having a size of 1/2 ^2N of the original image are created. Here, since an image masked by a mask similar to a Gaussian function is sampled, a Gaussian filter is actually used, but the same processing as when a Laplacian filter is applied is performed. An image is obtained. Since an image in the low frequency band having a size of 1/2 ^2N is sequentially obtained from the image of the original image size in this manner, the image obtained as a result of this processing is called a Laplacian pyramid.

Regarding this Laplacian pyramid, Burt PJ, "Fast Filter Transforms for Image"
Processing ”, Computer Graphics and Image Proces
Sing16, 20-51, 1981; Crowley JL, Stern R.
M., “Fast Computation of the Difference of Low ・ Pass
Transform "IEEETrans.on Pattern Analysis and Mac
hine Intelligence, Volume 6, Issue 2, March 1984, Mallat
SG, “A Theory for Multiresolution Signal Decompo
sition ； The Wavelet Representation ”IEEE Trans.on
Pattern Analysis and Machine Intelligence, 11
Volume 7, Issue 1989; Ebrahimi T., Kunt M., "Image
compression by Gabor Expansion ”, Optical Engineer
ing, Vol. 30, No. 7, pp. 873-880, July 1991, and Pi.
eter Vuylsteke, Emile Schoeters, “Multiscale Image
Contrast Amplification ”SPIE Vol.2167 Image Proce
Details are described in ssing (1994), pp551-560.

Then, the images in all the frequency bands of the Laplacian pyramid thus obtained are subjected to processing for enhancing the image values, and the images in the respective frequency bands subjected to this enhancement processing are reversed. A method for obtaining a processed image by conversion is described in JP-A-6-301766. Since the image processed in this way is emphasized in each frequency band, the image is substantially the same as the image subjected to the blur mask processing with the masks of a plurality of sizes in the blur mask processing described above.

[0007]

However, in the method described in Japanese Patent Laid-Open No. 6-301766, when an image in a certain frequency band is subjected to an emphasis process, an object requiring an emphasis process in the image is required. Since unnecessary components such as noise are emphasized in addition to the components such as the contours of the image, the image obtained as a result of processing is also emphasized in addition to the components such as the contours of the subject. For this reason, the image obtained as a result of the image processing is conspicuous with noise and is difficult to see.

In view of the above circumstances, it is an object of the present invention to provide an image processing method and apparatus which can obtain a high quality processed image by emphasizing only necessary components in the image.

[0009]

An image processing method and apparatus according to the present invention converts an image into a multi-resolution space, decomposes the image into images for each of a plurality of frequency bands, and Among the images of the predetermined frequency band, the portion corresponding to the portion where the absolute value of the signal value in the image of the lower frequency band than the predetermined frequency band is relatively large is multiplied by the enhancement coefficient of the enhancement degree, and the enhancement coefficient is multiplied. It is characterized in that a processed image is obtained by performing inverse multi-resolution conversion on an image of a frequency band multiplied by a coefficient and an image of another frequency band.

Converting into a multi-resolution space means that an image signal is decomposed into images for each of a plurality of frequency bands by a predetermined filter such as Laplacian pyramid, wavelet transform, and sub-band transform.

[0011]

EFFECTS OF THE INVENTION In an image for each of a plurality of frequency bands obtained when the image is converted by multi-resolution conversion, a component such as a contour of a subject included in the original image is included in the image even in the low frequency band. It is included in the ingredients. However, components such as noise are included in the image in the high frequency band, but are not included in the image in the relatively low frequency band. Therefore, in an image in a lower frequency band than a predetermined frequency band for enhancing the image,
A portion having a relatively low absolute value of a signal value has noise even if a portion corresponding to that portion of an image in a predetermined frequency band has a signal value, but a component of the image represented by the signal of that portion is noise. Very likely. Therefore, if the entire image in a predetermined frequency band is emphasized, not only a necessary component such as a subject but also an unnecessary component such as noise will be emphasized. The present invention has been made in view of this point.

That is, the image processing method and apparatus according to the present invention is capable of processing a signal value in an image in a frequency band lower than a predetermined frequency band to be emphasized among images in a plurality of frequency bands converted into a multi-resolution space. A portion having a relatively large absolute value has a higher degree of enhancement of the portion of the image in the predetermined frequency band corresponding to this portion. As a result, the portion of the image in the low frequency band, which can be regarded as noise in the predetermined frequency band, where the absolute value of the signal value is relatively small has a lower degree of emphasis than other portions. By performing the enhancement processing on the image of the predetermined frequency band in this way, unnecessary components such as noise in the image of the predetermined frequency band are less emphasized than necessary components of other subjects, so that they are conspicuous. Disappear. Therefore, the processed image obtained by inversely transforming the image of the frequency band and the image of the other frequency band on which the enhancement process is performed, the unnecessary components such as noise in the components of the predetermined frequency band become inconspicuous, It is possible to obtain a high-quality image in which necessary components such as the contour of the subject are emphasized.

[0013]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing an outline of an apparatus for carrying out an image processing method according to the present invention. As shown in FIG. 1, an apparatus for carrying out an image processing method according to the present invention comprises an image input means 1 for inputting an image to the apparatus,
Of the images decomposed into a plurality of frequency bands by the multiresolution decomposition processing means 2 for performing the multiresolution decomposition processing on the input image, the image of a predetermined frequency band will be described later. Enhancement processing means 3 for performing such enhancement processing, and restoration processing means 4 for restoring the image in the frequency band subjected to the enhancement processing by the enhancement processing means 3 and the image in another frequency band to obtain a processed image.
And an image output means 5 for reproducing the processed image restored by the restoration processing means 4 as a visible image.

Next, the operation of the image processing method according to the present invention will be described. FIG. 2 is a block diagram for explaining the processing performed by the multi-resolution image decomposition processing means 2 in FIG. In this embodiment, it is assumed that the image signal S is decomposed into a multi-resolution image by the Laplacian pyramid method. As shown in FIG. 2, the digital image signal S representing the original image is processed by the multi-resolution decomposition processing means 2
When input to, the filtering means 10 filters by a low-pass filter. This low-pass filter substantially corresponds to a two-dimensional Gaussian distribution on a 5 × 5 grid as shown in FIG. 3, for example. This low-pass filter is applied to images of all resolutions as described later.

The image signal S filtered by such a low-pass filter is sampled every other pixel by the filtering means 10 to obtain a low resolution approximate image g ₁ . This low-resolution approximate image g ₁ has a size of ¼ of the original image. Next, the interpolation means 11 interpolates pixels having a value of 0 in the sampled intervals of the low resolution approximate image g ₁ . This interpolation is performed by inserting rows and columns having a value of 0 for each column and each row of the low-resolution approximate image g ₁ . In this way, the low-resolution approximate image g ₁ in which the pixel of which the value is 0 is interpolated is blurred, but the pixel of which the value is 0 is inserted every other pixel, so that the change in the signal value is not smooth. Has become.

After the interpolation is performed in this way,
Further, the low-resolution approximate image g ₁ thus interpolated is filtered again by the low-pass filter shown in FIG. 3 to obtain a low-resolution approximate image g ₁ ′. The low-resolution approximate image g ₁ ′ has a smoother change in signal value as compared with the low-resolution approximate image g ₁ which is interpolated as described above. In addition, it is an image in which high frequencies higher than half of the original image are disappeared in terms of frequency band.
This makes the size of the image 1/4 and the value is 0 every other pixel.
3 is interpolated and further filtered by the low-pass filter shown in FIG. 3, so that an image in a frequency band whose spatial frequency is higher than half is blurred by the Gaussian function.

Next, the subtractor 12 subtracts the low-resolution approximate image g ₁ ′ from the original image to obtain a detailed image b ₀ . This subtraction is performed between the signals of the pixels corresponding to the original image and the low-resolution approximate image g ₁ ′. Here, since the low-resolution approximate image g ₁ ′ is an image in a frequency band higher than half of the spatial frequency of the original image as described above, the detailed image b ₀ is less than half of the original image. The image shows only the upper frequency band.
That is, as shown in FIG. 4, the detailed image b ₀ represents an image in the frequency band of N / 2 to N of the Nyquist frequency N of the original image.

Next, the low resolution approximate image g ₁ is input to the filtering means 10 and subjected to filtering processing by the low pass filter shown in FIG. Then, the low-resolution approximate image g _{1 that} has been subjected to the filtering process is sampled every other pixel by the filtering means 10, and a low-resolution approximate image g ₂ is obtained. The low-resolution approximate image g ₂ has a size of ¼ of the low-resolution approximate image g ₁ , that is, 1/16 of the original image. Then the interpolation means
At 11, the pixels having a value of 0 are interpolated in the sampled intervals of the low resolution approximate image g ₂ . This interpolation is performed by inserting rows and columns having a value of 0 for each column and each row of the low-resolution approximate image g ₂ . In this way, the value is 0
Although the low-resolution approximate image g ₂ in which the pixels are interpolated is blurred, pixels having a value of 0 are inserted at every other pixel, so that the change in the signal value is not smooth.

After the interpolation is performed in this way,
Further, the low-resolution approximate image g ₂ thus interpolated is filtered again by the low-pass filter shown in FIG. 3 to obtain a low-resolution approximate image g ₂ ′. The low-resolution approximate image g ₂ ′ has a smoother change in signal value than the low-resolution approximate image g _{2 on} which the above-described interpolation is performed. Further, as compared with the low-resolution approximate image g ₁ , the image in the frequency band higher than half is disappeared.

[0021] Then, in the subtracter 12, subtracting from the low-resolution approximate image g ₁ low-resolution approximate image g ₂ 'is carried out, the detail image b ₁ is obtained. This subtraction is a low-resolution approximate image g
_This is performed between signals for pixels corresponding to ₁ and the low-resolution approximate image g ₂ ′. Here, since the low-resolution approximate image g ₂ ′ is a blurred image in a frequency band higher than half of the spatial frequencies of the low-resolution approximate image g ₁ as described above, the detail image b ₁ has a low resolution. Approximate image g
_It is an image showing only the frequency band above half of ₁ . That is, as shown in FIG. 4, the detail image b ₁ is only in the frequency band above half of the low-resolution approximate image g ₁ , that is, N / 4 of the Nyquist frequency N of the original image.
It represents an image in a frequency band of up to N / 2.
As described above, a low-pass filter having a Gaussian distribution is used to perform a filtering process to obtain a detailed image. However, since the filtered image is subtracted from the low-resolution approximate image, the Laplacian filter is substantially used. The result is the same as when the filtering process is performed.

Then, the above-mentioned processing is performed by filtering means.
The low-resolution approximate image g _k (k = 1 to N) filtered and sampled by 10 is sequentially repeated to obtain n detailed images b _{k (} k = 1) as shown in FIG.
~ N) and the residual image g _L of the low resolution approximate image.
Here, the resolution of the detailed image b _k decreases in order from b ₀ , that is, the frequency band of the image decreases.
Against the Nyquist frequency N of the original image, detail images b _k represents the frequency band of ^{N / 2 k + 1 ~N /} 2 k, the size of the image is a 1/2 ^2k times the original image. That is, the highest resolution detail image b ₀ has the same size as the original image, which is a quarter of the size of the next high resolution detail image b ₁ original image detail image b _0. As described above, since the detail image becomes smaller from the one having the same size as the original image, and the detail image is substantially the same as the one to which the Laplacian filter is applied, the multi-resolution according to the present embodiment is used. The transformation is called the Laplacian Pyramid. Further, the residual image g _L can be regarded as an approximate image having a very low resolution of the original image, and in an extreme case, the residual image g _L is composed of only one image representing the average value of the original image. Becomes The detailed image b _k and the residual image g _L thus obtained are stored in a memory (not shown).

Next, the detailed image b thus obtained
_The emphasis processing means 3 performs emphasis processing on _k . The details of the emphasis process will be described below.

A detailed image b for each of a plurality of frequency bands obtained by performing multi-resolution conversion on the image as described above.
_{At k} , a component such as the contour of the subject included in the original image has a signal value of a certain magnitude even in the image in the low frequency band. However, although a component such as noise is included in the image in the high frequency band, it disappears in the image in the low frequency band and has a signal value close to 0. For example, as shown in FIG. 5, a detailed image b
Details of _k and a low frequency band than the detail images b _k image b _{k + 1}
Comparing with, at points A, B and C, both detail images have signal values. However point D
Has a signal value in the detail image b _k ,
The signal value is 0 in the detailed image b _{k + 1} . Therefore, it can be considered that the point D of the detailed image b _k is an unnecessary portion like noise, and the other points A, B and C are necessary portions like the contour of the subject.

Therefore, in an image in a lower frequency band than a predetermined frequency band to which the image is to be emphasized, a portion where the absolute value of the signal value is relatively low is a signal value corresponding to the portion of the image in the predetermined frequency band. Even if it has
The component represented by the signal in that part is very likely to be noise. Therefore, the signal value of the detail image in the low frequency band (the detail image b _{k + 1} in which the one-level frequency band is low in the present embodiment) in the lower frequency band than the detail image b _{k in the} frequency band to be emphasized is detected. As a result of signal value detection,
For the portion of the detailed image b _k corresponding to the portion of the detailed image b _{k + 1} where the absolute value of the signal value is relatively small, the enhancement coefficient is made smaller than that of the other portions, and the enhancement is performed. That is, as shown in FIG. 6, the enhancement degree f is increased as the absolute value of the signal value of the detailed image b _{k + 1} is increased. Then, the detail image b _k is multiplied by the emphasis degree f (b _{k + 1} ) based on the detail image b _{k + 1} thus determined to obtain the emphasized image b _kp .

B _kp = b _k × f (b _{k + 1} ) ... (1) In this way, the detail image b _k in the predetermined frequency band is subjected to the emphasizing process, so that the detail image b _{k in the} predetermined frequency band is _obtained. Unnecessary components such as noise in 3 have less emphasis than other necessary components, and thus become inconspicuous.

In this enhancement processing, the detailed image b
There are four points in the detailed image b _k that correspond to one pixel of _{k + 1} . Therefore, in determining the coefficients for enhancement processing interpolates a pixel value of the detail image b _{k + 1} corresponding to the four pixels of the detail images b _k, respectively corresponding to four pixels A value is obtained and the emphasis coefficient is determined based on this value. Further, as a representative of the detail images b _{k + 1} 1 single pixel values corresponding to the four pixels of the detail image b _k, the 1
You may make it determine the emphasis degree of four pixels based on one pixel value.

Then, the detail image b _{k in} the frequency band and the detail image in the other frequency band which have been subjected to the emphasis process are inversely transformed. The inverse conversion process is performed in the restoration processing means 4 as follows.

FIG. 7 is a diagram showing details of the inverse transformation of the detailed image. First, the residual image g _L is interpolated between the pixels by the interpolating means 14 to obtain an image g _L ′ having a size four times the original size.
It is said. Next, in the adder 15, the interpolated image g
_L 'and perform the addition at the lowest resolution detail image b _n-1 of the corresponding pixels comrades addition image (g _L' get + b _n-1). Then, this added image (g _L ′ + b _n−1 ) is interpolated.
It is input to 14, and the interpolation means 14 interpolates between each pixel to obtain an image b _n-1 ′ having a size four times the original size.

Next, this image b _n-1 ′ is subjected to the addition of the pixels corresponding to the one-step high resolution image b _n-2 of the detail image b _{n-1 in} the adder 15, and the added addition is performed. The signal (b _n-1 ′ + b _n-2 ) is interpolated by the interpolator 14 at each pixel interval, and the image b having a size four times as large as the detail image b _n-2 is obtained.
_n-2 .

The above processing is repeated and the same processing is applied to the emphasized image b _kp . That is, addition of the emphasized image b _kp and the one-step low resolution image b _k-1 ′ subjected to the above-described processing is performed in the adder 15, and the addition signal (b
_kp + b _k-1 ′) is interpolated between the pixels in the interpolating means 14 to obtain an interpolated signal b _kp ′. This processing is sequentially performed on the higher frequency detail image, and finally, in the adder 15, the interpolation image b ₁ ′ and the detail image b of the highest resolution are obtained.
Addition with ₀ is performed to obtain the processed image signal S '.

The processed image signal S'obtained in this way is input to the image output means 5 and displayed as a visible image. The image output means 5 may be a display means such as a CRT, a recording device for performing optical scanning recording on a photosensitive film, or for that purpose, an image signal is temporarily stored in an image file such as an optical disk or a magnetic disk. It may be a device.

With respect to the detail image thus converted into the multi-resolution by the Laplacian pyramid, the enhancement coefficient for the detail image in the desired frequency band is the signal of the detail image in the lower frequency band than the desired frequency band. By setting based on the above, unnecessary components such as noise are not emphasized so much in the image of the desired and frequency bands, and only necessary components such as the outline of the subject are emphasized. Therefore, in the processed image obtained by inversely transforming the detail image subjected to the enhancement process and the detail image other than that, the image in the desired frequency band is enhanced, but noises and the like in this frequency band are generated. Unnecessary portions are not emphasized, so that it becomes a good one suitable for observation / interpretation in which noise is not noticeable.

In the above-described embodiment, the Laplacian pyramid method is used to convert an image into a multi-resolution image, but the method is not limited to this. For example, wavelet transform or subband transform is used. Other methods such as the above may be used for conversion into a multi-resolution image.

The wavelet transform has been recently developed as a frequency analysis method and is applied to stereo pattern matching, data compression, etc. (OLIVIER RIOUL and MARTIN VETTERLI; Wa).
velets and Signal Processing, IEEE SP MAGAZINE, P.14
-38, OCTOBER 1991, Stephane Mallat; Zero-Crossingsof
a Wavelet Transform, IEEE TRANSACTIONS ON INFORM
ATION THEORY, VOL.37, NO.4, P.1019-1033, JULY 1991
).

This wavelet transform is

[0037]

[Equation 1]

In the equation, the signal is converted into a frequency signal for each of a plurality of frequency bands. That is, if the filtering process is performed by changing the period and reduction ratio of the function h and moving the original signal, it is possible to create a frequency signal adapted to a desired frequency from a fine frequency to a coarse frequency.

On the other hand, the sub-band transform not only obtains images of two frequency bands by one kind of filter like wavelet transform, but also obtains images of plural frequency bands at once by using plural kinds of filters. It is a conversion method that also includes.

Then, as in the case of the Laplacian pyramid described above for the image for each of a plurality of frequency bands obtained by the wavelet transform or the sub-band transform as described above, this frequency is applied to the image of the desired frequency band. The image in the desired frequency band is emphasized by setting the emphasis coefficient based on the signal value of the image in the frequency band lower than the band, and the image in the desired frequency band is emphasized. Is not emphasized, it is possible to obtain a good image suitable for observation and interpretation in which noise is not noticeable.

Further, in the above-described embodiment, the enhancement coefficient of the image in the desired frequency band to be subjected to the enhancement process is determined based on the signal value of the image in the frequency band one step lower than that frequency band. However, the determination may be made based on not only the one-step low frequency band but also the two-step, three-step or more low-frequency band images.

[Brief description of drawings]

1 is a block diagram of an apparatus for implementing an image processing method according to the present invention.

FIG. 2 is a diagram showing details of multi-resolution decomposition processing means.

FIG. 3 is a diagram showing a low-pass filter.

FIG. 4 is a diagram showing a detailed image for each of a plurality of frequency bands on which a Laplacian pyramid is applied.

FIG. 5 is a comparative diagram of the detailed image b _k and the detailed image b _{k + 1} .

FIG. 6 is a graph showing the degree of emphasis

FIG. 7 is a diagram showing details of restoration processing means.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Image input means 2 Multi-resolution decomposition processing means 3 Enhancement processing means 4 Restoration processing means 5 Image output means 10 Filtering means 11 Interpolation means 12 Subtractor 14 Interpolation means 15 Adder

Claims (2)

[Claims] 1. An image is converted into an image in each of a plurality of frequency bands by converting the image into a multi-resolution space, and the image of a predetermined frequency band among the plurality of frequency bands is subjected to the predetermined image. A part corresponding to a part where the absolute value of the signal value in the image in the lower frequency band than the frequency band is relatively large is multiplied by an emphasis coefficient with a larger emphasis degree, and the image in the frequency band and the other frequency bands multiplied by the emphasis coefficient are multiplied. An image processing method characterized in that a processed image is obtained by performing inverse multi-resolution conversion on the image. 2. An image decomposing means for decomposing the image into an image for each of a plurality of frequency bands by converting the image into a multi-resolution space, and an image of a predetermined frequency band among the plurality of frequency bands. , An emphasis coefficient multiplication means for multiplying an emphasis coefficient having a higher emphasis degree to a portion corresponding to a portion where an absolute value of a signal value in an image in a lower frequency band than the predetermined frequency band is relatively large, and the emphasis coefficient is multiplied. An image processing apparatus comprising: an inverse conversion unit that obtains a processed image by performing inverse multiresolution conversion on an image in a frequency band and an image in another frequency band.