JPH0636037A - Holonic visual recognizing device - Google Patents

Abstract

PURPOSE:To provide the holonic visual recognizing device which uses N holovisions for two-azimuth detection and can set 2N detection azimuths at equal intervals. CONSTITUTION:The N holovisions for two-azimuth detection are used and the angle between corresponding detection azimuths of adjacent holovisions is set to (pi/2N); and meaning storage parts 123 and 132 for the respective holovisions and a meaning integration part 140 have 4N relative positions, a weight function used for information interchanging between information generation parts 122 and 132 for the respective holovisions and the meaning storage parts 123 and 132 is regarded as the gradients of ridges of a regular inverted 4N-angled pyramid having its apexes arranged in 2N detection azimuths, and all pieces of information in the meaning storage parts are integrated by the meaning integration part.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a holonic system between semantic information groups created from input graphic information such as characters / images (still images, moving images) and stored semantic information groups of graphics. The present invention relates to a visual recognition device that performs pattern recognition by associating (that is, associating an information loop including top-down and bottom-up of semantic information).

[0002]

2. Description of the Related Art A model that captures the human visual system in three layers, the retina (and the lateral geniculate body), the primary visual cortex, and the memory function, and uses the non-linear oscillator pull-in phenomenon to process information is known. Holovision (Reference: Hiroshi Shimizu, Yoko Yamaguchi: "Cerebrum Information Principle and Its Application to Bio-Computers", Biological Science, Vol.37, No.1, p.
p. 26-40, 1986), which is a visual recognition model. Figure 3 shows the basic model of Holovision. As shown in FIG. 3 (a), the structure of holographic vision is information input unit 301 (R plane) / information generation unit 302 (V layer).
-It consists of three layers of the meaning memory unit 303 (S layer), which are the retina (and lateral geniculate body) of the human visual system, respectively.
Corresponds to the primary visual cortex and memory function. Information input section 3
Reference numeral 01 is an image receiving surface (in the figure, an example of 4 × 4 pixels) in which image receiving pixels are arranged in a simple square lattice, and binary pixel information (also called raw information, black circles ● in the figure represent signals) ) 30
Graphic information composed of 0s is input to the pixel. The information generation unit 302 detects the direction of the line (the direction indicated by \, |, /,-in the figure) from the array of the input elementary information S
Units 312 and line ends (in the figure,

[0003]

[Outer 1]

It is composed of an assembly of T unit 322 groups for detecting the line end displayed by. An aggregate 3 of units of the information generation unit 302 corresponding to one pixel of the information input unit 301
32 (hatched portion in the figure) is called a hyper column 332, and the basic column forming the hyper column 332 is shown in FIG.
S unit 3 that generates line direction information as shown in (b)
12 and a TL unit 34 for generating line end information in the arrow direction
1 and TR unit 342, which is a T unit 322
Consists of. Each unit forming the basic column is a non-linear oscillator. The hyper column 332 is a stack of basic columns having different orientations. Meaning storage unit 30
3 indicates the meaning of the memory figure in the direction of the line end (in the figure,

[0005]

[Outside 2]

A collection of M units 313 to be stored as a relative position (direction indicated by) and a relative position of line ends (relative positions indicated by L1, L2, L3, L4, L5, L6, L7, L8 in the figure). Consists of. The M unit is also a nonlinear oscillator. The meaning storage unit 303 can also correspond to the rotation of the figure by performing the conversion Li → Lj (i, j = 1 to 8; i ≠ j). The unit assembly 333 in different line end directions corresponding to one relative position of the meaning storage unit 303 (the hatched portion of L6 in the figure) is referred to as a memory column 333.

Next, the VS relay 304 in the figure has a function of converting the absolute position of the line end detected by the information generator 302 into a relative position. If the information input unit 301 is made of a 4x4 pixel horovision as an example, the VS relay operates by multiplying the signal of the information generating unit by a weighting function as shown in FIG. Is the process of setting the maximum value of the signal in the above as the signal at the array position. In this case, the weighting function corresponds to the slope of the ridge of an inverted regular octagonal pyramid whose apex coincides with the direction of line detection. Similarly, the SV relay 305 in the figure has a function of converting the relative position of the line ends stored in the meaning storage unit 303 into an absolute position. If the information input unit 301 is a 4 × 4 pixel holovision as an example, S
The operation of the V-relay is a process of multiplying the signal of the meaning storage unit by a weighting function as shown in FIG. 3 (d) and setting the maximum value of the signal at the same array position as the signal at that array position. The weighting function in this case also corresponds to the gradient of the ridgeline of an inverted regular octagonal pyramid whose apex coincides with the azimuth of line detection, but an interpolated gradient is used for the array position corresponding to the ridgeline. The position conversion function enables graphic recognition independent of position and size. As both of these relays operate at the same time, the information generation unit 302 and the meaning storage unit 303 are synchronized with respect to the line end information (the mutual pulling action of the collection of nonlinear oscillators) and desynchronized (of the nonlinear oscillators). Suppression of mutual aggregation) is performed, and the line end information group that represents the input figure (that is, the set of line end information that represents the relationship between the line ends) and the line edge information group that is stored as the figure meaning (That is, a set of line end information representing the relationship between the line ends) is associated with each other, and graphic recognition is established.

As described above, the set of line end information generated by the information generation unit from the line information of the graphic input to the information input unit is bottom-up to the semantic storage unit as the semantic information of the graphic (this is performed by the VS relay). ), And top-down the semantic information of the graphic stored in the meaning storage unit to the information generation unit (this is S
In the information loop formed by the V relay), the non-linear oscillator groups of the semantic storage unit and the information generation unit, which carry the semantic information, are in the pull-in state, that is, at the semantic information level of the input figure and the stored figure. With the configuration and principle of matching, it is possible to recognize figures regardless of position, size, or rotation.

Next, how the holographic vision recognizes a quadrangle will be described with reference to FIG. (1) First, a quadrangle having 3 dots on each side is input to the image receiving surface of the information input unit 301.

(2) In the information generator 302, the S unit 312, which detects the direction of the line from the arrangement of the dots of the input information by the pulling action of the nonlinear oscillator, as shown in FIG.
Line information of the azimuths of and | is generated.

(3) Next, from the line information generated by the S unit 312, the T unit 322 (specifically, the TR unit 342 and the TL unit 341 corresponding to the direction of the line end).
Generates line end information in the directions of ←, →, ↓, and ↑ by the pulling action of the nonlinear oscillator.

FIG. 3 (f) shows a mechanism for generating line end information for horizontal line information. In the figure, the solid line represents mutual excitatory coupling, and the dotted line represents mutual inhibitory coupling. When line information indicated by diagonal lines is present on the image receiving surface 301, the shaded S unit 312 is excited, and the shaded TL unit 351 is excited to generate line end information.

(4) On the other hand, the quadrilateral meaning information stored in the M unit 313 of the meaning storage unit 303 is such that the arrows representing the relative position and direction of the line ends are set to the relative positions of L2, L4, L6 and L8. It is expressed by placing two of each. For example, L6 on the circumference of the figure is the memory column 3 of FIG. 3 (a).
33 represents that the arrow has information in its direction, that is, the nonlinear oscillator is excited. The strength of information (that is, the amplitude strength of the nonlinear oscillator) is represented by the length of the arrow. Since this information set (that is, the semantic information represented by the eight arrows) is represented in the retracted state of the nonlinear oscillator, all the oscillator groups related to memory vibrate with the same phase.

(5) Semantic information group (that is, bottom-up semantic information group) generated from the line end defined by the absolute position generated in the above step (3), and generated in the above step (4) The semantic information group (that is, the top-down semantic information group) generated from the line end defined by the relative position is the SV relay and VS.
An information loop is formed by the relays, which are associated with each other by pulling in the non-linear oscillator, and figure recognition is established.

[0015]

However, in the above-mentioned horovision, since the direction of the line is detected from the arrangement of the pixel information input to the image receiving surface of the simple square lattice array of the information input unit 301, it belongs to the detection direction. Only the elevation angle from the X axis determined by the pixel coordinates x and y (both natural numbers) can be selected as the angle θ of the detection azimuth = tan (y / x). For this reason, the detection direction is increased (specifically, it is increased to 5 or more directions).
It is difficult to equalize the angles between the detection directions when doing, and the multi-direction detection function such as the hypercolumn of the human primary cerebral cortex in which the detection directions are set at intervals of about 10 degrees is realized by holography. I can't. If the detection orientation intervals are not the same during multi-directional detection, a complicated figure may be misrecognized as different from the original figure when it is rotated. It is also difficult to obtain the high precision.

It is an object of the present invention to provide a holonic visual recognition device in which the information input unit 301 can set the angles between the detection directions even on the image receiving surface of a simple square lattice array, paying attention to such problems. is there.

[0017]

In order to achieve the above-mentioned object, the present invention has an information input section which constitutes an image receiving surface with pixels arranged in a simple square lattice, and two [π / 2] different detection directions. And an information generation unit that generates line end information corresponding to line information from the image input information of the information input unit, and a meaning storage unit that has a meaning of a graphic as a set of line end information. (N is an integer greater than or equal to 2) number of holographic devices, the corresponding detection direction of N adjacent holographic devices forms an angle of [π / 2N], and the relative position of the meaning storage unit is 2N. The weighting function, which is arranged in the azimuth direction and is used in information exchange between the information generation unit and the meaning storage unit of each holographic device, is 2
Provided is a holonic visual recognition device having a meaning integration unit that integrates the N meaning storage units, with the gradient of the ridgeline of an inverted regular 4N pyramid having vertices arranged in N orientations.

[0018]

With the above construction, the present invention makes it possible to set N detection azimuths at equidistant angles by using N holographic devices, and the N holographic devices can be independently set. Since it operates, it can be configured as N parallel processing systems, and it has a great practical value as a high-accuracy and high-speed processing pattern recognition apparatus.

[0019]

FIG. 1 shows an embodiment of the present invention. Hereinafter, description will be given with reference to the drawings.

FIG. 1 shows three types of simple square lattices for sampling an image. Based on 101 in the figure, 1
The lattice of 02 is inclined by [π / 4], and the lattice of 103 is further inclined by [π / 8]. If there are a plurality of such grids, the pixels may be almost the same in three kinds of grids such as 104 and 105, or in some cases, the grid 10
In some cases, one pixel corresponds to two pixels of the grid 102 or the grid 103 to the same extent. Generally, when N kinds of grids (however, all the pixels forming the grid have the same size) are tilted at a constant angle (for example, [π / 2N]) and exist at the same time, a pixel 1 The pixels of the other grids corresponding to the number of pixels are the four pixels in the closest vicinity.
When N = 4, the relationship is as shown in FIG. Therefore, even if pixels of different grids do not have a one-to-one correspondence, there is no problem that line edge information that should appear in the same place is greatly separated by two or more pixels.

In the following, a holovision for two-direction detection will be referred to as N (N
The case where N = 2 will be described as an example of the holonic visual recognition device that uses 2 or more integers. Figure 2
1 shows a configuration of a holonic visual recognition device including two holographic vision detecting two azimuths. The first holographic vision is an image receiving surface in which pixels are arranged in a simple square lattice (in the figure, an example of 4 × 4 pixels). Information input section 121 consisting of
And a group of S units 126 for detecting the direction of the line (two directions indicated by | and-in the figure) from the array of pixel information of the input figure and the end of the line (indicated by ↑ ↓, → ← in the figure) Information generation unit 122 composed of an aggregate of T unit 127 groups for detecting the line ends), and the meaning of the memory figure in the direction of the line ends (directions indicated by ↑, →, ↓, ← in the figure). Relative position of line ends (L1, L2, L3, L4, L5, L6, L7, L8 in the figure)
4N at the relative position displayed by 4)). The VS relay 124 has a meaning storage unit 123 made up of an aggregate of M units 128 groups, and the VS relay 124 has a vertex for the line end information generated by the information generation unit 122. The meaning storage unit 123 multiplies the gradient of the ridgeline of the inverted regular octagonal pyramid 150 that matches the direction detected by the above (that is, the 4N direction) as a weighting function.
The SV relay 125 has a function of sending to the, and the slope of the ridge of the inverted regular octagonal pyramid 150 whose vertex matches the direction detected by the entire system (that is, 4N direction) with respect to the line end information stored in the meaning storage unit 123. Is multiplied as a weighting function and sent to the information generating unit 122. Similarly, the second holography corresponds to the information input unit 131 (however, the information input unit 121 corresponds to the image receiving surface (4 × 4 pixel is an example in the figure) in which pixels are arranged in a simple square lattice pattern. Direction is rotated by an angle of [π / 2N = π / 4]) and the direction of the line (two directions indicated by | and-in the figure) from the arrangement of the pixel information of the input figure. The information generating unit 132 including an aggregate of the T unit 137 group that detects the S unit 136 group and the line end (the line edge displayed by ↑ ↓, → ← in the figure) to be detected, and the meaning of the stored figure is the line end. Direction (the direction indicated by ↑, →, ↓, ← in the figure) and the relative position of the line end (L1, L2, L3, L4, L5 in the figure,
There are 4N relative positions displayed by L6, L7, and L8, but L1, L2, L3, L4, L5, L6, and L of the meaning storage unit 123.
7, L8 is a relative position rotated by an angle of [π / 2N = π / 4])
And a semantic storage unit 133 composed of an aggregate of groups.
The relay 134 multiplies the line edge information generated by the information generation unit 132 by the gradient of the ridgeline of the inverted regular octagonal pyramid 150 whose vertex coincides with the direction detected by the entire system (that is, 4N direction) as a weighting function, and stores the meaning. The SV relay 135 has a function of sending to the unit 133, and the SV relay 135
A function of multiplying the line edge information stored in 3 by the slope of the ridge of the inverted regular octagonal pyramid 150 whose vertex coincides with the direction detected by the entire system (that is, the 4N direction) as a weighting function, and sending it to the information generating unit 132. Have. Further, the semantic information in the semantic storage units 123 and 133 is integrated by the semantic integration unit 140 (in the case of the present embodiment, the semantic information is superimposed).
To be done. The operation of the meaning integration unit 140 is as shown in FIG.
The lattice 102 (corresponding to the first horovision) is [π / 2N = π] for the lattice 102 (corresponding to the second horovision).
/ 4], the relative position of the line end information is shown in FIG.
As shown in, the L1 of the lattice 101 and the L1 of the lattice 102 appear at a position rotated by [π / 4N] = π / 8. After that, when the meaning information of the meaning storage units 123 and 133 is integrated (superposed in this embodiment) by the meaning integration unit 140 of FIG. 2, L1 of the lattice 101 and L8 of the lattice 102 are displayed at the position D1 of the meaning integration unit 140. Therefore, even in a system in which there are two two-direction detecting holovisions, the same information as when one four-direction detecting holovision is used can be expressed. Specifically, when a right-angled triangle as shown in FIG. 5 is input as image information, the line end information as the semantic information generated in the first horovision is shown in FIG. 6 and is generated in the second horovision. FIG. 7 shows the line end information as the meaningful information. When these pieces of line end information are integrated (superposed) by the meaning integration unit 140, the result is as shown in FIG. This is exactly the same as the semantic information obtained in the conventional four-direction detection horovision.

FIG. 9 shows an embodiment of the holonic visual recognition device when four holographic visions with two detection directions (that is, N = 4) are used. The points are the four azimuths detected by the two azimuth detection horovision system (in this case, 2
Corresponding to N = 8 and 8 directions), the relative position has 4N = 16 meaning storage units 223 to 226, and the meaning integration unit 240 integrates (superimposes) the semantic information of the meaning storage unit.
4N = 16 relative positions corresponding to the detected direction of 2N = 8, and the weighting function used in the information exchange between the information generation unit and the meaning storage unit coincides with the detected direction. This is to use the slope of the ridge of the inverted regular 16-sided pyramid 250.

Although N = 2 and 4 have been described as the embodiments of the present invention, the same can be extended to N> 4 or more.

[0024]

As described above, according to the present invention, N pieces of two azimuth detecting holovisions are prepared and the corresponding detection azimuths of the adjacent holovisions are set to be inclined by [π / 2N]. It is possible to detect 2N azimuths at evenly spaced angles,
Moreover, since N holographic visions can be processed independently, N parallel configurations are possible, and as a result, high-accuracy direction detection capable of interpolation and high-speed arithmetic processing are achieved at the same time.
The practical effect of the present invention is extremely large.

[Brief description of drawings]

FIG. 1 is a pixel correspondence diagram showing a principle of the present invention when a plurality of simple square lattices overlap each other.

FIG. 2 is a diagram showing a holonic visual recognition device including a first two-direction detecting horovision, a second two-direction detecting horovision, and a meaning integrating unit in the first embodiment of the present invention.

FIG. 3 is a diagram in which the simple square lattices used in the first and second horovisions shown in FIG. 2 are overlapped.

FIG. 4 is a diagram in which the semantic information (the relative position of the line end expressing the line end information and its detection direction) of the meaning storage unit used by the first holographic and the second holographic shown in FIG. 2 is superimposed.

FIG. 5 is a diagram showing a graphic input in the first embodiment.

FIG. 6 is a diagram showing line end information generated by the first holography for the input figure shown in FIG.

FIG. 7 is a diagram showing line end information generated by a second holography for the input figure shown in FIG.

FIG. 8 is a diagram showing the first and second input figures shown in FIG.
Showing the line-end information of the meaning integration part that integrates (superimposes) the line-end information generated by the holographic

FIG. 9 is a configuration diagram of a holonic visual recognition device according to a second embodiment of the present invention.

FIG. 10 is a basic configuration diagram of holographic vision in a conventional example.

FIG. 11 is a diagram showing a main part of the conventional example.

FIG. 12 is a diagram showing a weighting function used in the conventional example.

FIG. 13 is a diagram showing a weighting function used in the conventional example.

FIG. 14 is a diagram showing how the conventional holographic vision recognizes a rectangle.

FIG. 15 is a diagram showing a mechanism in which the conventional holographic vision system generates line end information for horizontal line information.

[Explanation of symbols]

 121, 131 Information input section 122, 132 Information generation section 123, 133 Meaning storage section 124, 134 VS relay 125, 135 SV relay 140, 240 Meaning integration section 150, 250 Gradient of regular polygonal pyramid ridge representing weighting function

Claims (4)

[Claims] 1. An information input section comprising an image receiving surface in which pixels are two-dimensionally arranged, and information for generating line information corresponding to a detected azimuth and line edge information corresponding to the line information from image information of the information input section. N (N is an integer of 2 or more) Holovision devices composed of a generation unit and a meaning storage unit that has the meaning of a figure as a set of line end information and all the meaning storage units of the Holovision. A holonic visual recognition device, comprising: one semantic integration unit that integrates. 2. An information input section which constitutes an image receiving surface with pixels arranged in a simple square lattice, and two detection directions [[pi] / 2] different from each other, and from the image information of the information input section to the detection direction. N (N is 2 or more) which is composed of three parts: a corresponding line information and an information generation unit that generates line end information corresponding to the line information, and a meaning storage unit that has the meaning of a graphic as a set of line end information. (Integer) number of holographic devices and one semantic integration unit that integrates all the semantic storage units of said holographic vision,
A holonic visual recognition device, wherein the holographic vision is tilted so that a corresponding detection direction forms an angle of [π / 2N]. 3. An information input section forming an image receiving surface with pixels arranged in a simple square lattice, and two detection azimuths [π / 2] different from each other, and from the image information of the information input section to the detection azimuth. N (N is 2 or more) which is composed of three parts: a corresponding line information and an information generation unit that generates line end information corresponding to the line information, and a meaning storage unit that has the meaning of a graphic as a set of line end information. (Integer) number of holographic devices and one semantic integration unit that integrates all the semantic storage units of said holographic vision,
The detected azimuths corresponding to the holographic vision are tilted so as to form an angle of [π / 2N], and the meaning storage unit sets relative positions in 4N directions in which the angles between the azimuths are equal to the [π / 2N]. Holonic visual recognition, wherein the holonic visual recognition is characterized in that a weighting function used for information exchange between the information generating unit and the meaning storage unit is a ridge of an inverted regular 4N pyramid having vertices arranged in the 4N directions. apparatus. 4. An information input section that constitutes an image receiving surface with pixels arranged in a simple square lattice, and two detection directions [[pi] / 2] different from each other, and from the image information of the information input section to the detection direction. N (N is 2 or more) which is composed of three parts: a corresponding line information and an information generation unit that generates line end information corresponding to the line information, and a meaning storage unit that has the meaning of a graphic as a set of line end information. (Integer) number of holographic devices and one semantic integration unit that integrates all the semantic storage units of said holographic vision,
A holonic visual recognition device characterized in that if the detection directions corresponding to the holographic vision are different and the pixels on the image receiving surfaces of the holographic vision are not separated by two pixels or more, the information is at the same position.