JPH0883346A - Method and device for detecting size of feature pattern of signal - Google Patents

Abstract

PURPOSE: To easily detect the size of a feature pattern of a signal. CONSTITUTION: This signal feature pattern size detecting device is provided with a storage means 2 for storing an input signal, a scale setting means 4 for setting up plural scales for the input signal, a measure detecting means 6 for detecting the measure of the input signal in each scale by the use of plural set scales, and an inherent size determining means 8 for determining the inherent size of the input signal based upon the detected measures.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a signal characteristic pattern size detecting method and apparatus mainly used for recognizing a characteristic pattern in a signal.

[0002]

2. Description of the Related Art Conventionally, character recognition technology for printing, handwritten characters and the like has been important, and there are many applicable fields such as sorting of mails and input of handwritten characters, and they have been put to practical use. However, with the current technology, it is difficult to cut out as a pattern on a character-by-character basis, and it is common to supplement the cut-out using the meaning of the character (knowledge information such as radical). If the character size can be determined in the process before using the knowledge information, the subsequent recognition can be easily performed. Therefore, the development of the character size detection technology is required.

Further, in the preprocessing of recognition of an image in which a plurality of textures (a state in which fine pattern patterns are uniformly distributed) are mixed, it is desired to detect the size of each texture or to detect the pitch period of voice. However, there is also a demand for facilitating subsequent processing such as recognition.

It is known to use the fractal dimension to detect the size of the characteristic patterns of these signals. The fractal dimension means that the complexity of a signal (image signal) is represented by a unique value (see Yamaguchi, Hata, Kigami, Iwanami, Applied Mathematics I'Mathematical Fractals', April 1993) There are several. Here, the blanket-covering method, which is one of the methods for obtaining the fractal dimension, is explained (T.
Peli, V.Tom, B.Lee, “Multi-Scale Fractal and Correl
ation Signatures for Image Screening and natural C
lutter Suppression, “SPIE vol.1199 Visual Commum.
and Image Processing IV, 1989). An image curved surface (three-dimensional) consisting of a set of points whose length is the magnitude of the brightness value (integer value) of the pixel in the direction perpendicular to the screen for each pixel in the two-dimensional screen, and a scale. _Consider a sequence {ε _k } (k = 0, 1, ...) Of ε _k (> 0). For a certain scale ε _k , a blanket with a width of ε _k is placed above and below the image curved surface. The upper surface of the blanket covered on the pixel (i, j) is u _{i, j} (ε _k ), the lower surface of the blanket covered on the bottom is b _{i, j} (ε _k ), and the luminance value of the pixel (i, j) is If (integer value) is g _{i, j} , u _{i, j} (ε _o ) = b
_{i, j} (ε _o ) = g _{i, j} recursively,

[0005]

[Equation 1] Ask for. Here, ε _o = 0. From now on, the change B (ε _k ) of the brightness surface of the blanket is

[0006]

[Equation 2] Therefore, this B (ε _k ) is called a measure for the scale ε _k here. That is, the scale corresponds to a unit for measuring the measure. Here, as shown in FIG. 13, when the logarithm of the measure is plotted on the vertical axis and the logarithm of ε is plotted on the horizontal axis, a straight line having a slope a may be obtained, and D = 2- which is obtained from the slope of the straight line.
a is called fractal dimension.

In general, a measure is an amount corresponding to a volume or area that is uniquely determined for a certain scale, and the linear relationship that appears when the logarithm of the measure and the scale is taken is characterized by the fractal dimension.

Conventionally, a method of extracting a characteristic region of an image using such a fractal dimension has been studied. For example, in Japanese Patent Laid-Open No. 4-170685, the target image is divided into smaller blocks, and the block division is continued until the fractal dimension does not change depending on the block size. However, in the conventional method using such a fractal dimension, the region cannot be extracted unless the block division is performed.

Further, as in Japanese Patent Laid-Open No. 3-269782, there is known a method for extracting a character area from a character image using a fractal dimension. In this method, the fractal dimension in the vertical direction and the fractal dimension in the horizontal direction are obtained with respect to all the pixels of the image, and it is determined pixel by pixel whether or not to enter the area indicating the character area from these values. This kind of pixel-based area division requires time and effort, and it has not been possible to roughly identify a character and easily specify its size.

As described above, in the conventional method, there is a problem that the size detection of the signal characteristic pattern requires a considerably troublesome procedure and the size cannot be easily detected.

The present invention has been made in consideration of the above circumstances, and an object of the present invention is to provide a method and apparatus for detecting the size of a signal characteristic pattern, which can easily detect the size of the signal characteristic pattern. To do.

[0012]

According to the present invention, there is provided a signal characteristic pattern size detection device, which stores an input signal, scale setting means for setting a plurality of scales for the input signal, and the setting means. A measure detecting means for detecting a measure of the input signal with respect to each scale using a plurality of scales, and a unique size determining means for determining a unique size of the input signal based on the detected measure. It is characterized by being

Further, in the method for detecting the size of a characteristic pattern of a signal according to the present invention, an input signal is stored, a plurality of scales are set for the input signal, a measure of the input signal for each scale is obtained, and this determination is made. The characteristic size of the input signal is determined based on the measured measure.

[0014]

According to the present invention having such a configuration, the input signal is stored, a plurality of scales are set for the stored input signal, and a measure of the input signal is set for each of the set plurality of scales. Is obtained, and the characteristic size of the input signal is obtained based on the obtained measure. As a result, unlike the conventional case, it is not necessary to perform block division or area division, so that the size of the characteristic pattern of the signal can be easily detected.

[0015]

1 shows the configuration of a first embodiment of a signal characteristic pattern size detection apparatus according to the present invention. The size detecting device of this embodiment includes a storage unit 2 and a scale setting unit 4.
And a measure detecting means 6 and a characteristic size determining means 8. The configuration and operation of the size detection device of this embodiment will be described with reference to the flowchart shown in FIG. First, the input signal is stored in the storage means 2. A plurality of scales are set by the scale setting means 4 with respect to the stored input signal, and the measure B used to obtain the fractal dimension of the input signal is measured by the measure detecting means 6 using the set scales. Detected (Fig. 2
Step F1). This measure B is, for example, as shown in FIG. 4, a signal curved surface g (image curved surface described in the related art).
Can be obtained by covering a signal curved surface g with a cube whose one side is a scale ε and counting the number of cubes that can cover the signal curved surface g. This measure detection method is called the box-counting method, for example, by BB Chaudhuri, N. Sarkar, P. Kunde,
“Improved fractal geometry based texture sementat
ion technique ”IEE Proceedings-E, vol.140, No.5, S
ep, 1993.

In such a measure detection, several kinds of scales ε _i (i = 0, _i) set by the scale setting means 4 are set.
...), and the change amount (Δlog B) of these measures (log B) with respect to the scale ε is detected by the characteristic size determining unit 8 (see step F2 in FIG. 2). At this time, the peculiar size determination means 8 determines whether or not the change amount of the measure satisfies a predetermined condition. When this condition is satisfied, for example, the scale ε _T when the amount of change in the measure becomes equal to or larger than a predetermined value is output from the unique size determination unit 8 to the outside as the unique size (see step F3 in FIG. 2). Then, when the measure is measured on all the prepared scales, the detection of the characteristic size is ended, and when the measure is not measured on the all scales, the scale is switched and the above-mentioned steps are repeated (step F4 in FIG. 2). (See F5).

In this embodiment, a cube is used to detect the measure, but this embodiment is not limited to this. For example, in the case of a one-dimensional signal such as a voice signal, a line segment of length ε may be used as a scale, and if the two-dimensional scale of voice strength and time is considered, the measure is measured with a square area on all sides of the scale. Is also good. This can be determined according to the signal of interest and the desired feature pattern.

A two-dimensional signal such as an image signal may be regarded as a three-dimensional area including a square area and a luminance direction, and the volume of a cube, the dynamic range of a signal with respect to a scale, or its logarithm may be used. As a simplified method, the number of signal changes in squares on all sides of the scale may be used. Further, the blanket-covering method described in the related art or another method for obtaining the fractal dimension may be used. The measure algorithm can also use a graphic with the same dimensions as the commonly used dimensions of the signal for the measurement.

When measuring the measure of a signal on the set scale, the cutting method of the scale may be overlapped or the measurement may be carried out while overlapping little by little. Measures in the case of overlapping are averaged by the number of overlapping. At this time, you may measure with all possible cutting methods, or you may perform it at random,
It may be determined and measured using another rule.

The change of the measure with respect to the scale may be determined by whether the amount of change of the measure is positive or negative, or by using the difference or variance of the absolute values of the amount of change.

In the above-mentioned embodiment, the measures are obtained for a plurality of scales, and the peculiar size of the signal (the size of the characteristic pattern of the signal) is detected depending on whether or not the changes in these measures satisfy a predetermined condition. This detection principle will be described.

It is assumed that the measure detecting means 6 is given a signal pattern as shown in FIG. 6 (a). When the scale is small, the change is obtained with a blanket expressing this change as shown in FIGS. 6 (b) and 6 (c), but when the scale is ε _T or more, the upper blanket surface and the lower blanket surface are measured. Both become flat, and the uneven pattern of the signal is buried in the blanket. This is shown in Figure 1.
A graph similar to that of No. 3 is shown in FIG. In the conventional fractal dimension, this change was linearly approximated, but in reality, it is observed as a straight line that bends at ε _T. That is, it is observed as a graph that bends on a scale having a change amount of the measure as shown in FIG.
By detecting _T , it is possible to detect the pattern size of the characteristic pattern of the signal.

As described above, according to this embodiment,
Unlike the conventional case, it is not necessary to perform block division or area division, so that the size of the signal characteristic pattern can be easily detected.

When the detecting device of the above embodiment is used for the image signal, the distance from the edge of the image to the desired area can be detected as one of the peculiar sizes. This is because the measurement is performed in the same state as when there is a large change in the image signal at the edge of the image as shown in FIG. On the other hand, if the edge of the image is folded back and the measure is obtained by matching the size of the given image with the scale, there will be no change in the image signal at the edge of the image.
The distance from the edge of the image is not measured, but another distance such as the edge-to-edge distance of the region in the image is detected. In FIG.
The horizontal axis represents the image position, and the vertical axis represents the brightness of the image signal at that position. In FIG. 8A, since the brightness is zero and the unique size d is detected in the range of image position 0 <x <x1, x =
The measure is detected as having a region change at the positions of x1, x2, x3, and x4. As a result, d2 and d3 are detected as unique sizes. On the other hand, in FIG. 8B, x = x
Since the luminance is folded back at the positions of 2 and x4 to perform the measure detection, the distance from the edge of the image is not detected, but only the distance d1 of the specific region in the image is detected.

Next, an image in which textures are mixed (see FIG.
An example in which the present embodiment is applied when dividing (see (a)) will be described with reference to FIG. This example is an example of texture division. For all possible scales, consider a cube with one side equal to the scale ε, as shown in FIG. The brightness curved surface of the stored texture image is covered with a minimum required cube, and the number thereof is taken as a measure B (ε) for this scale. Observe the rate of change (gradient) of the measure that decreases in proportion to the scale, and obtain the unique scale that is the feature quantity. In this example, the eigenscales εT ₁ and εT ₂ represent the distances d ₁ and d ₂ from the edge of the image to the texture transition, and the texture can be easily divided at this distance. As shown in FIG. 9B, it is also possible to know the property of the complexity of the switched texture from the slope of the measure change up to εT ₁ (fractal dimension), the slope from εT ₁ to εT ₂ , and the slope after εT _2. it can.

An example in which the present embodiment is applied to knowing the size and position of a desired area having a characteristic pattern will be described with reference to FIG. This example is an example of obtaining the size and position of a region. For example, if the measure is obtained in the same manner as the texture division example described above, the unique scales ε ₁ , ε ₂ , ε ₃ , and ε ₄ of the size of N × N pixels are obtained (FIG. 10).
(See (b)). When the edges of the image are not folded, the natural scale corresponds to d ₁ , d ₂ , d ₃ and d ₄ . There are several possible ways of knowing which can correspond to which. For example, as shown in FIG. 11, the obtained ε ₁ , ε ₂ ,
Assuming a region (wavy line in FIG. 11) of all sizes determined by ε ₃ and ε ₄ , the image is divided by ε ₁ and the eigenscale of the remaining region is detected again. The correct position is determined based on whether or not the inherent scale of the re-detected result is consistent with the assumed case. In this example, FIG.
Since ε ₂ , ε ₃ −ε ₁ and ε ₄ are detected in (b), (c) and (d), it is determined that ε ₁ = d ₁ is correct. Similarly, the diagrams of FIGS. 11 (a), (b), (c), and (d) are represented by ε.
Judgment is made in the area divided by ₂ , and by repeating this, the unique size and the actual distance can be associated.

Further, in the above example of obtaining the position of the area, if the method of folding the edge of the screen is used together, the area size of r ₁ and r ₂ can be directly detected, and the location of the area can be easily specified. For example, there is also a method in which one side is fixed to one of two types and the other side is made to correspond to a scale, the measure is measured again in units of rectangles, and the vertical and horizontal characteristic sizes are separately determined.

As another method for detecting the size of a desired area, a method of measuring a measure with a rectangle having vertical and horizontal g ₁ and g ₂ scales can be used. In a rectangle having the same size as the desired area, the amount of change in the measure changes abruptly, and the set of two sides of the rectangle at this time becomes the eigenscale. Besides rectangles, triangles and circles,
It is possible to meet demands such as three-dimensional figures.

When various patterns are mixed in the signal as described above, if there is one size that appears frequently,
This size can be detected from the amount of change in the measure as an intrinsic scale. For example, it is effective when detecting the character size from the character area. As preprocessing for character recognition, the position of a rough character region is detected by projection that scans vertical and horizontal pixel values, and detailed character size detection is performed by using the present invention.
Characters can be cut out easily and accurately. It is also possible to detect the position of a character area by projection using an example of area division. As still another application, the thickness of the character can be detected from the character image.

Next, FIG. 3 shows the configuration of a second embodiment of the signal size pattern size detecting apparatus according to the present invention. The size detecting apparatus of this embodiment is different from the size detecting apparatus of the first embodiment shown in FIG. 1 in that the determining means 10 is newly provided and is used when the input signal is an image signal. Based on the eigensize output from the eigensize determining means 8, the judging means 10 decides whether or not to re-divide the image block based on the number of eigensizes and whether or not the eigensize satisfies a predetermined condition. to decide.

If it is determined that the image data is to be re-divided, the division instruction of the image signal held in the storage means 2 is given to the storage means 2 and is divided. If it is determined not to subdivide,
Output the unique size. This unique size is used to identify the edge of the block segmented image. The condition to be given is determined according to a desired feature pattern, for example, whether an edge can be specified or an edge in a desired range can be specified.

FIG. 12 shows an example in which this is applied to area division of an image signal. In FIG. 12, the distance from the edge of the image is obtained as the unique size. Unique size is Δlo in Fig. 5
The top two are used in descending order of g B (measure change amount). First, as shown in FIG. 12A, the unique size is obtained for the entire image, and the blocks are divided by using the upper two scales (ε ₁ , ε ₂ ), for example, as in the case of finding the position of the region. The unique size is extracted again for the subdivided block (FIG. 12B). This is repeated, and the process ends when the edge is specified as shown in FIG. As a result, the boundaries of the areas included in the block can be specified in detail. Besides, FIG.
It is also possible to divide the area using a method of knowing the size and position of the area as shown in FIG. In addition, as in the example of FIG. 9, from the change amount of the measure before and after switching to the unique size,
It is also possible to know the characteristics of the complexity of the divided areas.

In the above example, the subdivision of the block is
A simple predetermined division method may be used. Further, if there is information on the area in advance, a desired unique size can be selected from a plurality of unique sizes.

The size detecting device of the second embodiment is also the first
It goes without saying that the same effect as that of the above embodiment can be obtained.

[0035]

As described above, according to the present invention, the size of the characteristic pattern of the signal can be easily detected.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a first embodiment of a signal characteristic pattern size detection device according to the present invention.

FIG. 2 is a flowchart illustrating the operation of the first embodiment.

FIG. 3 is a block diagram showing the configuration of a second embodiment of the signal feature pattern size detection apparatus according to the present invention.

FIG. 4 is an explanatory diagram illustrating a specific example of a measure detection method used in the present invention.

FIG. 5 is a graph illustrating a specific example of a detection method for detecting a unique size from a measure change amount.

FIG. 6 is a schematic diagram illustrating the principle of unique size detection.

FIG. 7 is a graph illustrating the principle of unique size detection.

FIG. 8 is an explanatory diagram illustrating that a distance from an edge of an image to a desired area can be detected as a unique size.

FIG. 9 is an explanatory diagram illustrating a case where the present invention is used for texture analysis.

FIG. 10 is an explanatory diagram illustrating an example of detecting a desired area on the screen.

FIG. 11 is an explanatory diagram illustrating an example of detecting a desired area on the screen.

FIG. 12 is an explanatory diagram illustrating a case where edge detection of an image signal is performed using the second embodiment.

FIG. 13 is a graph illustrating a method for obtaining a fractal dimension based on a conventional measure.

[Explanation of symbols]

 2 storage means 4 scale setting means 6 measure detecting means 8 peculiar size determining means 10 judging means

Claims (5)

[Claims] 1. A storage means for storing an input signal, a scale setting means for setting a plurality of scales for the input signal, and a measure of the input signal for each scale using the set plurality of scales. A signal characteristic pattern size detection device, comprising: a measure detecting unit for detecting a signal; and a characteristic size determining unit for determining a characteristic size of the input signal based on the detected measure. 2. The input signal is an image signal, and it is provided with a judging means for judging whether or not to divide the image signal into blocks based on the characteristic size determined by the characteristic size determining means. The size detection device according to claim 1. 3. The size detecting apparatus according to claim 1, wherein the unique size determining means determines the unique size from the amount of change in the measure with respect to the scale. 4. The input signal is an image signal, and the measure detecting means covers the image curved surface determined by the image signal with a figure designated by the scale for each scale and intersects with the image curved surface. 4. The size detection device according to claim 1, wherein the number of the figures is a measure of the input signal with respect to the scale. 5. An input signal is stored, a plurality of scales are set for the input signal, a measure of the input signal for each scale is obtained, and an inherent size of the input signal is determined based on the obtained measure. A method for detecting the size of a characteristic pattern of a signal, which is characterized by determining.