JPH0614353B2 - Spatial filtering device - Google Patents

Description

Description: TECHNICAL FIELD OF THE INVENTION The present invention belongs to the technical field of spatial filters.

[Technical Background of the Invention and Problems Thereof] Spatial filtering is used as a basic method of image processing. The types of filters are smoothing filters (low-pass filters) for removing noise components and low filters. Examples thereof include a high-pass or edge enhancement filter for enhancing high-frequency components by pressing frequency components to make it easier to see a fine structure of an image, and a band-pass filter for enhancing a frequency band having important information.

Here, the high pass filter and the band pass filter can be realized by using the smoothing filter.
For example, as shown in FIG. 1A, if a smoothing image B is drawn from the original image line A through the smoothing filter 1, a high-pass filter 3 having the characteristics shown in FIG. As shown in FIG. 2 (a), the band pass filter 4 having the characteristics shown in FIG. 2 (b) can be realized by passing through the high pass filter 3 after smoothing filtering.

Next, for the smoothing filter, the FFT method, the convolution method using the general coefficient impulse response, and 3
There are a method of applying a × 3 filter a plurality of times and a method of using a uniform weight filter algorithm.

This uniform weight filter algorithm method is referred to in the literature (Toriwaki, Yokoi, Fukumura: High-speed algorithm in image processing by computer, information processing, Vol.17, No.3, P.215.
221 (1976)), it has the highest speed among the various smoothing filter methods mentioned above.

Next, the uniform weight filter algorithm in the above document
II will be explained.

The impulse response (rectangular shape) of the filter is And is a constant value.

First, in the first step, a working image X ' _{(i, j)} as shown in FIG. 3 (a ₎ is created. Where X ' _{(i, j)} is Is represented. This is because the working image X ′ _{(i, j)} has four vertices, that is, (1,1), as shown in FIG.
This means that it is an added value of the original image densities in a quadrangle (indicated by diagonal lines) having (i, 1), (1, j), and (i, j). The number of calculations required to create this work image X ' _{(i, j)} is two additions per pixel, and the addition method is as shown in FIG. That is, the addition order for creating the working image X ′ _{(i, j)} is, for example, from the left end to the right end of the first line (j = 1), and then from the left end to the right end of the second line (j = 2). And so on, and W (contents of the work register) is X _{(1, j)} , X _{(2, j)} , ..., X in the raster scan direction.
To calculate _{(i-1, j)} , W ← W + X _{(i, j)} … (3) X ′ _{(i, j)} ← W + X ′ _{i, j-1)} … (4) If done, a working image X ' _{(i, j)} is obtained.

Next, in a second step, a smoothing image Y _{(i, j)} is obtained based on the work image obtained in the step 1. This smoothing image Y _{(i, j)} is The calculation method is as shown in FIG. That is, As a result, the added value in the shaded area, that is, the smoothed image value is obtained, so that the operation speed becomes extremely high regardless of the size of the support size (the number of operations is 4 times per pixel).

However, the above filter algorithm cannot perform the filtering process using the impulse response in the shape of a diagonal quadrangle.

Further, a general smoothing filter including a filter based on the above filter algorithm cannot obtain a sufficiently satisfactory characteristic in a situation where the spatial frequency characteristic is required to be strict. The reason is that, as is clear from the spatial frequency characteristic diagram (support size = 35) shown in FIG. 4, for example, there is a considerably large ripple (a gain oscillates) LP near the cutoff frequency. . This ripple causes a pseudo image, and particularly when a pseudo image is generated in a medical image diagnostic apparatus or the like,
The doctor may make a misdiagnosis.

Although it is possible to realize a filter with suppressed ripples by using the direct execution algorithm, there is a drawback that the calculation speed is drastically reduced.

An important factor that determines the filter performance is the directionality of the spatial frequency characteristics. This shows, for example, how the frequency characteristics change in the ω ₁ direction, the ω ₂ direction, and the oblique direction. The directionality of the frequency characteristic in the uniform weight filter is as shown in FIG. 5A (support size = 19 × 19), and the frequency characteristic in the ω ₁ direction and the ω ₂ direction is not good (the characteristic varies depending on the direction). different).

Incidentally, FIG. 7B shows a cross section taken along the line I-I of FIG.

[Object of the Invention] The present invention has been made in view of the above circumstances, and has an oblique quadrangular spatial filter capable of performing a filtering process by an oblique quadrangular impulse response at high speed, and has excellent spatial frequency characteristics, Moreover, it is an object of the present invention to provide a spatial filtering device capable of performing a filtering process with an octagonal impulse response at high speed.

[Outline of the Invention] An outline of the present invention for achieving the above-mentioned object is to have a uniform weight filter for calculating an added value of a rectangular region in the vicinity of a target image, and an inclination of 45 ° with respect to a vertical direction. , And a group of working registers that holds the added value of the original images in the two straight line directions that are opposed to each other with the vertical direction as the axis, and the value of a specific pixel of the original image in the region surrounded by the two straight lines starting from that position. A working image memory to be an added value, an adding unit that adds pixels of an original image to the register group, a working image obtained immediately before in the working image memory, and an added value of the working register group. An oblique quadrilateral space having storage means for sequentially obtaining work images and storing the work images in the work image memory, and computing means for performing addition and subtraction of pixel values at four vertices of a quadrangle having a 45 ° inclination on the work image memory. Vertical with filter It is characterized by continuous connection.

[Embodiment of the Invention] An embodiment of the present invention will be described below with reference to the drawings.

Here, first, the principle of the present invention will be described.

Image size is n × n, and two sets of work register group W
(W ₁ ~W _2n-1) and W is prepared _{'(W' 1 ~W '2n} -1). As shown in FIGS. 6A and 6B, the work registers W and W'retain the added values of the one-dimensional pixels in the diagonal direction, respectively.

First, a work image X ' _{(i, j)} is created (step 1). This working image X ′ _{(i, j)} is the original image X _{(i, j).}
Is a value obtained by adding in the shaded area as shown in FIG. 7, and can be obtained at high speed by the sequential calculation method described later.

FIGS. 8A and 8B are explanatory diagrams for explaining the creation of the working image X ′ _{(i, j)} by the sequential calculation method. The sum X ′ of the work images in the area indicated by the broken line in FIG.
_{(i, j)} has already been obtained, and the work register W
_{i-j + n} is the one before X _{(i, j)} , that is, the original line X
The sum up to _{(i-1, j-1)} is held, and the register W '
_{i + j} is one before X _{(i + 1, j)} , that is, X _{(i + 1, j-1)}
It is assumed that the sum up to
_{A case where (i + 1, j)} (total sum of original images in the area indicated by the dashed line) is obtained will be described. Registers W _{i-j + n} , W ′
_{i + j} is _{expressed as} W _{i-j + n} ← W _{i-j + n} + X _{(i, j)} (7) W ′ _{i + j} ← W ′ _{i + j} + X _{(i + 1, j)} … (8 ), X ′ _{(i + 1, j)} is _replaced by X ′ _{(i + 1, j)} ← X ′ _{(i, j)} −W _{i-j + n} + W ′ _{i + J}
… Obtained by (9). Hereinafter, sequential calculation is performed while increasing i in the horizontal direction. When the sequential calculation of one line is completed, the added values up to the vertical j-th line are held in the respective work registers W and W '.

Next, the calculation of X ' _{(1, j + 1) on} the next vertical line, that is, the j + 1 line will be described with reference to FIG. The work register W ′ _{j + 1} holds the added value up to the j-th line. Here, the work register W ′ _{j + 1} is updated by W ′ _{j + 1} ← W ′ _{j + 1} + X ′ _{(1, j + 1)} (10) to obtain X ′ _{(1, j + 1)} Is obtained by X ′ _{(1, j + 1)} ← X ′ _{(1, j)} + W ′ _{j + 1} (11).

In this way, the working image X ' _{(i, j)} having the added value of the triangular area can be obtained.

The number of calculations for obtaining this work image X ' _{(i, j)} is
About 4 times of addition and subtraction per pixel (4 in steady state)
Times, 0 times in the first line, and twice when going in the vertical direction at i = 0).

Next, as shown in FIG. 9, the added value Y of the diagonal rectangular area
_{Find (i, j)} (step 2).

This Y _{(i, j)} is Y _{(i + j)} ← X ′ _{(i, j + L)} + X ′ _{(i, jL)} based on the work image X ′ _{(i, j)} obtained previously. -X ' _{(iL, j)} -X' _{(i + L, j)} (12). Here, L in the equation (12) indicates the support size.

Therefore, the filter based on the diagonal rectangular impulse response is
This can be realized by seven additions and subtractions (including the number of calculations in the first step).

An octagonal impulse response filter can be realized by using the oblique rectangular impulse response filter thus obtained and the rectangular impulse response filter already described with reference to FIG. That is, when the original image is first filtered by the quadrangular impulse response, and the result is filtered by the diagonal quadrangular impulse response, the result is equal to the result by the octagonal impulse response. This is, for example, the first
As shown in FIG. 0, the convolution integral of a uniform quadrangular impulse response A having a side length 2a and a weight of 1 and an oblique quadrangular impulse response B inscribed therein is
It is due to being equal to the octagonal impulse response C. Here, the weight of the octagonal impulse response C is not uniform, drawing the i direction (vertical, lateral direction) than the the k direction (diagonal direction), as indicated by h ₁ to h ₅ respectively, secondary It is a combination of curves or linear curves.

In addition, h _{1 to} h ₅ are respectively h ₁ = 2a ² −k ² (13) h is a ^{_{4 = 2a 2 -i 2 ... (}} 16) h 5 = (2a-i) 2 ... (17).

When filtering processing is performed using such an octagonal impulse response, the directionality of the spatial frequency characteristic can be improved and the ripple component can be reduced. Further, the filtering processing speed becomes extremely high by the sequential calculation method.

Next, the first embodiment of the present invention based on the above principle will be described.
The description will be made with reference to FIGS. 1 to 14 as well.

FIG. 11 is a block diagram showing the configuration of the spatial filtering device according to the present invention. FIG. 5A shows an original image memory for storing an original image, and 5b is a first filter unit (that is, uniform weight) for inputting an original image read from the original image memory 5a and performing a filtering process by a rectangular impulse response. (Filter) 5c is a first filtered image memory for storing the output of the first filter unit 5b, and 5d is an image read out from the first filtered image memory 5c, and a diagonal rectangular impulse response is used. A second filter unit (that is, a diagonal quadrilateral spatial filter) 5e for performing the filtering process is a second filtering image memory for storing the filtering process result by the second filtering unit 5d, and the second filtering image memory 5d. The memory contents of
As described above, this is the result of filtering the original image with the octagonal impulse response.

Next, regarding the configuration of the second filter unit 5d,
It will be described with reference to FIG.

FIG. 12 is a block diagram showing the configuration of the second filter unit 5d in FIG. FIG. 7 shows a work image creating means for creating a work image X ' _{(i, j)} based on the image X _{(i, j)} read from the first filtered image memory 6 (details will be described later). ). This work image creating means 7
The working image X ′ _{(i, j)} created by
It is stored in the second filtered image memory 5e via (details will be described later).

Incidentally, 10 is the first filtered image memory 5c,
Work image memory 1 described later in work image creating means 7
7 (FIG. 13) and the second filtered image memory 5
It is an address control means for controlling the write / read address of e.

Next, regarding the configuration of the work image creating means 7,
It will be described based on FIG. FIG. 11 shows the original pixel X _{i, j)} read from the first filtered image memory 5c and the value W _{i-j + n} (i-) read from a work register group (w) 13 described later. j + n is an adder unit for adding the register numbers, 13 is a working register group (W), which is composed of a plurality of registers and holds the output of the adder unit 11,
Is the sum of the original pixel X _{(i + 1, j)} read from the first filtered image memory 5c and the value _{W'i + j} read from the work register group (W ') 13 described later. An adder unit 14 for operation, which comprises a plurality of registers, and which holds the output of the adder unit 12 (W '),
Reference numeral 15 is an adder that adds the output of the adder 12 and a work image X ' _{(i, j)} read from a work image memory 17 described later, and 16 is the output of the adder 11 and the adder. A subtraction unit for calculating the difference from the output of 15, and a work image memory 17 for storing the output of the subtraction unit 16.

Incidentally, the work register group (W) 13, the work register group (W ′) 14 and the work image memory 17
The address read / write is performed by the address control unit 10. The stored contents of the work image memory 17, that is, the work image X ' _{(i, j),} is used for the calculation of the calculating means 8 (FIG. 12), as described later.

Next, the structure of the calculating means 8 will be described with reference to FIG. This computing means 8 is, as shown in FIG.
The calculation is performed on the diagonal rectangular area B inscribed in the rectangular area A processed by the first filter unit 5b. FIG. 18 is an addition unit for adding the work images X ′ _{(i, j + L)} and X ′ _{(1, jL)} read from the work image memory 17, and 19 is the work image memory 17 Pixel X ' _{(iL, j)} read out and work image X' _{(i + L, j)}
And an adder unit 20 for adding and
9, which is a subtraction unit for calculating the difference with 9, and the output Y _{(i, j)} of the subtraction unit 20 is stored in the second filtered image memory 5e.

Since the filtering process by the rectangular impulse response can be easily realized by executing the algorithm II described in the above-mentioned document, the detailed configuration of the first filter unit 5b in FIG. 11 is omitted. To do.

Next, the operation of the device configured as described above will be described.

First, the creation of a work image by the work image creating means 7 (step 1) will be described. Address control means 1
With the address control of 0, the original pixel X sequentially scanned and read from the first filtered image memory 5c.
_{(i, j)} and the value W _{i-j + n} read from the work register group (W) 13 are added by the adder 11, and the result is held in the same register (W _{i-j + n} ). To be done. Ie W
_{i-j + n} is updated (formula (7) above). Further, the original image X read from the first filtered image memory 5c
_{(i + 1, j)} and the value _{W'i + j} read from the work register group (W ') 14 are added by the adder 12, and the result is stored in the same register ( _{W'i + j} ). Retained. That is, _{W'i + j} is updated (formula (8)). Further, the output of the adder 12 and the work image X ′ _{(i, j)} read from the work image memory 17 are added by the adder 15, and the result is input to the subtractor 16 and It is used for subtraction from the output of the adder 11. Therefore, this subtraction unit 1
The output X ' _{(i + 1, j)} of ₁ becomes as shown in the previous equation (9),
This X ' _{(i + 1, J)} is stored at the address (i + 1, j) of the work image memory 17. The above operation is performed in the lateral direction (i
In the increasing direction).

Although the function for the vertical scanning is not shown, it can be similarly performed. That is, as already described with reference to FIG. 8 (b), when the sequential operation for one line is completed, the work register group (W) 13 and the work register (W ') 14 have vertical j lines. The added value up to this point will be retained. Therefore, the next vertical line (j +
(1 line) updates the value _{W'j + 1} read from the work register group (W ') 14 as in the expression (10), and can be easily obtained by the expression (11).

In this way, the work image X _{(i, j)} is created in the work image memory 17.

Further, the number of calculations for creating this work image X ′ _{(i, j)} (step 1) is 4 times per pixel in the steady state (once in each of the adders 11, 12, 15 and the subtractor 16). Will be).

Next, the work image X'created in step 1 above.
_{Based on (i, j)} , the calculating means 8 calculates the added value Y _{(i, j)} of the diagonal quadrangular area (see FIG. 9) (step 2). Work images X ′ _{(i, j + 1)} and X ′ read from the work image memory 17 at different addresses.
_{(i, jL)} is added by the adder 18, and the work images X ′ _{(iL, j)} and X ′ _{(i + L, j)} are added by the adder 19. Then, the difference between the output of the adder 19 and the output of the adder 18 is calculated by the subtractor 20. This is the execution of the equation (12), and the output Y of the subtraction unit 20
_{(i, j)} is temporarily stored in the second filtered image memory 5e as a filtering result by the diagonal rectangular impulse response.

Note that the number of calculations when calculating the added value Y _{(i, j)} of the diagonal quadrangular region (step 2) is three (one in each of the adders 18 and 19 and the subtractor 20). Therefore, the subtraction circuit for the filtering process using the diagonal quadrangular impulse response becomes 7 times (4 times in step 1 and 3 times in step 2) regardless of the support size, and the processing speed becomes high.

By using the oblique quadrangle filter described above, the filtering process using the octagonal impulse response can be performed at high speed, as already described with reference to FIG. That is, the first filter unit 5b may perform the filtering process using the rectangular impulsive response, and then the second filter unit 5d may perform the filtering process using the diagonal rectangular impulse response (or vice versa).

Here, the first filter unit 5b and the second filter unit 5
The total number of arithmetic circuits in d is 11 times per pixel (4
+7), which can be performed at high speed.

Although one embodiment of the present invention has been described above, it is needless to say that the present invention is not limited to the above embodiment and can be appropriately modified within the scope of the gist of the present invention. .

[Effects of the Invention] According to the present invention described above, an octagonal impulse response having an excellent spatial frequency characteristic is provided by using an oblique quadrilateral spatial filter capable of performing filtering processing by the diagonal quadrangular impulse response at high speed. It is possible to provide a spatial filtering device capable of performing the spatial filtering processing by the method at high speed.

[Brief description of drawings]

1 (a) and 1 (b) are explanatory diagrams for explaining the realization of a high-pass filter, and FIGS. 2 (a) and 2 (b) are explanatory diagrams for explaining the realization of a band-pass filter, and FIG. (A), (b), (c) are explanatory diagrams for explaining the uniform weight filter algorithm, and FIG. 4 is a characteristic diagram showing a spatial frequency characteristic of the uniform weight filter (rectangular impulse response). FIGS. 6 (a) and 6 (b) are characteristic diagrams showing the directionality of the spatial frequency characteristic in the uniform weight filter, FIGS. 6 (a), 6 (b) and 7 and 8 (a), 8 (b).
Is an explanatory view for explaining the principle (first step) of the filtering processing by the diagonal rectangular impulse response, and FIG. 9 is an explanatory view for explaining the principle (second step) of the filtering processing by the diagonal rectangular impulse response. FIG. 10 is an explanatory diagram for explaining the realization of an octagonal impulse response, and FIG. 11 is a block diagram showing the configuration of a spatial filtering device according to the present invention.
FIG. 12 is a block diagram showing the configuration of the diagonal quadrangular spatial filter, and FIGS. 13 and 14 are block diagrams showing the detailed configuration of the main parts of FIG. 12, respectively. 5b ... first filter section (uniform weighting filter), 5d ... second filter section (oblique quadrilateral spatial filter), 7 ... work image creating means, 8 ... computing means, 11, 12 ... Adder, 13 ... Working register group (W), 14 ... Working register group (W ').

Claims (1)

[Claims] 1. A uniform weight filter for calculating an added value of a rectangular region in the vicinity of a target image and 45 in the vertical direction.
A working register group that has an inclination of ° and holds the added value of the original images in two straight line directions that are opposite to each other with the vertical direction as the axis
A working image memory in which the value of a specific pixel is the added value of the original image in the area surrounded by the two straight lines starting from that position, an adder for adding the pixels of the original image to the register group, the working Storage means for sequentially obtaining a work image by the work image obtained immediately before in the image memory and the added value of the work register group and storing the work image in the work image memory; And a diagonal quadrangular spatial filter having arithmetic means for performing addition and subtraction of pixel values of four vertices in a quadrangle having an inclination of 45 ° are cascade-connected.