JPS5983277A - Two-dimensional information transmitting system - Google Patents

Abstract

PURPOSE:To transmit information in a middle direction other than signal line directions, by using alternately parallel operating devices, each of which consists of operators arranged at two-dimensional lattice points and signal lines connecting adjacent lattice points, to transmit information. CONSTITUTION:In two-system P<1>-side circuit and P<0>-side circuit having the same function which consist of switching circuits 1 and 2, a maximum value detecting circuits 3 and 4, delay circuits 5 and 6, a register 7 for P<0>, and a register 8 for P<1>, a signal is inputted from a nearby lattice operator through the circuit 1. In the circuit 1, two signal lines out of signal lines C00-C71 of peripheral lattice points are selected in accordance with the direction of transmission. If this direction is the middle direction between 0 and 1, the signal line C01 is connected to the P<0>-side circuit, and the signal line C10 is connected to the P<1>-side circuit. In respective circuits, maximum values of the input and registers 7 and 8 are calculated with circuits 3, 4 and they are delayed in circuits 5 and 6 by a one-unit time and are returned to registers 7 and 8 and are sent to the circuit 2 for output. In the circuit 2, the P<0>-side is connected to the signal line C41, and the P<1>-side is connected to the signal line C50.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an information propagation method in a two-dimensional arithmetic unit.

Information propagation on a two-dimensional plane forms the basis of pattern recognition technology, and conventionally the simplest one is the horizontal,
In the vehicle plane, only information propagation in two diagonal directions (in the case of a square lattice quantization space) has been put to practical use. However, this is insufficient; for example, if the line segment between the starting point S and the ending point E is horizontal, as shown in Figure 1(a), there is no problem, but as shown in Figure 1(b), If there is even a slight nod to the horizon, this simple information propagation will end up being a complete mess. However, if there is only one line on a white background as shown in (b) & C of the same figure, information can be propagated by using one outline of the line, but as shown in (c) of the same figure. If the lines are intersected as shown in the figure, it becomes difficult to propagate along the desired line.These problems will be explained in detail below.

First, to explain the wiring problem between each arithmetic unit, the second
The figure is a wiring diagram of the arithmetic units arranged in a square lattice with adjacent grids (nearby areas). ) is a wiring diagram with the Keima direction within the 5x5 area added to it. Figures (a) and (b)
) The number of wires connected to one point is 16 trees, but if you expand the neighborhood, you will see that the wiring becomes 9 more complex. Therefore, normally, calculations are performed independently (in parallel) at each point within a limited range (locally) in the vicinity, and information about a wider range (globally) is gradually gathered by repeating the calculations. is taken. This is called iterative local parallel operation.

However, when the directionality of wiring with the neighborhood (direction of calculation) is limited in this way (eight directions in the case of Figure 2 (2)), information is propagated in the intermediate direction (direction of information). This becomes a very serious K/41G problem.

This will be explained using an example different from the line extraction in Figure 1, in which how a point on a white background is surrounded by surrounding black background (this is called a "closed state"). do.

Now, the adjacent point of one point Z on two dimensions is expressed as shown in FIG. 4 using the vector r1 to the adjacent point based on the definition of the direction shown in FIG.

1) (Z) is used as the expression element for information at point 2. In addition, this expression is expressed as p (7; t ) when it is necessary to specify the repetition time t, and as p <z>, k=o, 1, etc. when there are multiple elements.

Further, all calculations regarding directions are performed modulo 8. That is, 7 + 1 (mod 8) = 0. 2-3
(mail 8) = 7.

Here, the initial value at the repetition time t = 0 is given according to the density of point 2 of the two-dimensional figure, and the following repetition performance: i (t: 1.2.3... ...) P<Zr t )=max(P(Z; t I L
When P (Z +rh t I ))-・al is Mlj, the value "1" in the i direction is propagated by the second term on the right side, and from the first term on the right side, once the value "1" is propagated, it does not change. , This operation propagates whether there is a black point (1″) in direction 1, that is, whether it is “closed” or not. In general, by finding the above numbers for the surrounding 1 (i = 0 to 7), it is possible to approximately estimate 9 closed states, but here, in order to simplify the discussion, we will limit it to only the i direction. ing.

Note that for propagation on the line shown in the wc1 diagram, a new quantity 1) (z r t ) representing distance may be introduced and the following procedure may be performed. That is, as seen in this example, the discussion of closed state propagation is easier because the distance does not include 5 scalars and iA. However, the discussion below is directly applicable to propagation on a line, with the caveats noted above.

In the above example, the directionality of the neighborhood and the directionality of information (1=0.1
．． 2.・・・・・・・・・7) is a match, but if this does not match, for example, Keima (jan” L
Let us consider the propagation of information from the direction (=266°).

Now, in order to propagate the value "I n from the intermediate direction between direction 1 and i + 1, we transform equation (2) and write ,
t'LP CZ+rl+□;t-1)) River... Old Mark... (5) As shown in Figure 5, the i direction and i+L direction (direction +01 in the figure) and 11
If there is "l" somewhere between '), I)(Z)=1, and the information is not necessarily propagated from the intermediate direction between i and i + 1.

Similarly, equation (2) can be expressed as P(Z; t)=max, (P(Z; t 1)4m1n
(P(ZrrI;t −1), )'(Z→−r++r
;t-t ))-(Gl, then i and i
VC, 1゛ in all directions.

Without it, 1' (Z) will not become 1'.

As shown in this example, if the direction of calculation deviates from the direction of interest, it is a very difficult problem to prevent ``information dissemination and information disappearance'', so it is necessary to make the propagation more precise.'' I got ai+ with +yti while I had to solve it.

This invention was made in view of the above points, and uses a method in which a plurality of convenient elements are prepared in the direction of information propagation, and the wiring is alternately shifted to use them. The purpose is to provide a method for propagating information in different directions.

Hereinafter, a pre-propagation step 11 of an embodiment of the present invention will be explained, and an embodiment of the embodiment will be explained based on the drawings.

Let us consider the unidirectional propagation of the above-mentioned value "1 ++" in the Keima direction jan-' - direction by performing calculations in the 3x3 neighborhood. Here, the unidirectional propagation is defined as the equation (2). Like propagation, it means that only information on a straight line in that direction is propagated.

First, in order to propagate information in two directions, in this invention, two elements P0(Z) and P'(Z) are used at each point.
have.

The initial value at the repetition time t=o is (k=
0.1) is given.

Next, the local parallel operation for the repetition time t=L, 2.3, . . . is expressed by the next generation.

P'(Z;t)=max(P'(Z;t1)+P
'(Z0r4.t 1))P'(Z; t )=m
ax(P'(Z; t-1), PO(Z+r,
Q,, t −1)) (8) The first term on the right side of equation (8) is the value of the own point one time before. The second term is 1 (for 20, the neighborhood r1 is 1 (=
The value of 1 is propagated by crossing the value of =() in the neighborhood "i+'-s" for k=1.

Finally, after completing the repeated calculation, P(Z)=max(P'(Z;oo))...
, , , , , , , , , (g) Find the value P (Z) at point 2 using k-0,1. Note that the repetition is performed for a sufficiently large value of t or for all ZKs,
J"(Z;t)=1"(Z; I-1)
) It stops at kudzu (Z; t) = culm (Z; 1-1). Here, for convenience, the time t at which it stopped is expressed as ■, but in reality it does not repeat λ%a infinite times.

FIG. 6 shows f=0 in equation (8). i.e. direction '0
It is a diagram showing how information is propagated in the intermediate direction between ' and 1''. In the diagram, first, only point A is a black point and all others are white points. Then, how this information is transmitted in the Y + 4th line. , is propagated only to point G. As can be seen from Fig. 6 and equation (8), the information starting from the point in the y-th row is Line point B
It is propagated only to January and Jll of point C. Furthermore, at the next time, it is propagated only to the dots Po and p+ in the 12th row of y. In other words, the information that was divided into two points, B and C in the y-th, 1st line, has been propagated to Keima's place 1a in the y+2nd line without being interfered with by other information. As a result of this repetition K, the signal is propagated only to the point G in the y-1-4th row, and one-way propagation along the shore in the Keima direction is realized using a 3×3 neighborhood model. At this time, odd numbered rows (y+i, y+3...
...), the information is scattered over two points, but in the final evaluation, it is thought that information of A is conveyed for both points from the (9th) second pregnancy, and information is missing in the odd numbered rows. It's not a translation.

The above is the propagation operation of the two-dimensional information propagation method of this invention.
This will be explained based on Figures 7 and 8.

FIG. 7 is a block diagram showing the circuit configuration of the arithmetic unit at each grid point, and FIG. 8 is a wiring diagram between the grid points.

In FIG. 7, 1 and 2 are switching circuits, 3 and 4 are maximum value detection circuits, 5 and 6 are delay circuits, T is a register for po, and 8 is a register for pl.

As shown in the figure, it is composed of two circuits 5 & C1 with the same function. That is, maximum value detection circuit 3, delay circuit 5.
The system configured by PO register 1 is P in equation (8).
The maximum value detection circuit 4. Delay circuit 6. P
A system composed of registers 8 for 1 calculates Pl in formula 8). Here, 1μ dog value detection circuit 3° delay circuit 5.
Systematic P00 circuit consisting of PO register γ, maximum value detection circuit 4. Delay circuit 6. )・The system composed of the 1 register 8 is called the 131 side circuit. A signal from a nearby lattice calculator is input to the P00 circuit and the Pl side circuit via the switching circuit 1. In the switching circuit 1 for human power, there are 16 signal lines C60""'C connected to the surrounding grid points.
Two signal lines are selected depending on the direction of propagation of 7+5.

For example, the % ratio in the intermediate direction between 0'' and °°l'' is (Connect F1 line eo+ to the po side circuit, C4° to the P11 circuit, Fi,
#ru.

The two input signals of the swinging circuit 1 are respectively expressed as P' (Z-1-++, f, -
1), I)'(Z+1., t-1). No. 8
), the maximum value of this input and the register γ for IJ0 is calculated in the maximum value detection circuit 3, and then the delay circuit 5 calculates the maximum value for one unit time. After the delay, it is returned to the PO register and simultaneously sent to the output switching circuit 2.

The P11 circuit that calculates P'(Z;
send to

Similarly to the input switching circuit 1, the output switching circuit 2 also connects two signal lines to 16 output signal lines C6.
Selectively connect to O-C7+ depending on the propagation direction. +0
In the case of intermediate direction between 1 and l', Pollll wo, C4
+, connect the PI side with C, OK. The signal output from C41 is connected to the signal line C8 at the lower grid point as shown in Figure 8.
The input signal to C1 and the signal output from C50 become a human input signal at the lower left signal line C1°.

In addition, from the description of one embodiment of the above invention, the following (11-
(5) Changes such as those described above and combination thereof are easily possible.

(1) In the above embodiment, propagation in the jan ” 2 direction was described, but by increasing the number of elements P,
It can be extended to unidirectional propagation other than jan-.1. That is, the general angle tact −′′! For unidirectional propagation in one direction, m Pk (k=o, 1.2
．． . . .m-1) and expand equation (8) as follows. However, it is assumed that m>n and m and n are positive numbers having no common divisor. It is clear that the generality of the problem is not lost.

P (l; t) = where k + n (moct l+n) is k + n
It is the number modulo 1n of , that is, the remainder of (k ten n) 7m. Furthermore, equations (7) and (9) are expressed as k=0.
1.2...blasphemy - IK then 'C'.

Because of the busyness of doing this, one line of (m i'H; will be shifted by the number element, and the propagation of the In line will be shifted by nXm elements. Also, one point has m elements. Hold it flat, and when it protrudes, the 5th & Cth
Since ri:ri propagates from ri1t, the movement in the column direction is (nXm
)/m=n, and propagation of n1 rows causes movement of exactly n columns. The propagation of this general angle jan-'' is m-4゜n
== 3K is illustrated in FIG. 9. Further, at this time, the initial value of equation (7) may be set to only one element as follows.

・・・・・・・・・・・・・・・・・・・・・α0 Conversely, by adding 1″ to all elements of the sunspot in the force formula,
Ultimately, it is also possible to make a determination based on only one element.

(2) Although the above embodiments described unidirectional propagation,
Wider propagation is possible by combining two or more unidirectional propagations. Broad propagation means that information is propagated from a certain direction range as shown in FIG. 10(C). FIG. 10(b) shows the propagation of equation (8) as shown in FIG. 6, and FIG. 10(a) shows the unidirectional propagation in the vertical direction. The propagation in Figure 1(a) uses only one element P.
- Can be subtracted by setting i = O in equation (2). By combining the propagations in Figure 10 (a) and (b) (propagation in equations (2) and (8)), wide propagation as shown in Figure 10 (C) is created by two elements 1) It can be realized by the following formula using 20.1 for k.

PO(Z; t)=max(P'(z; t-1)
I PI (Z-1-ro; t-1)) P'(Z; t
) = rnax (P' (Zy t 1 ) + PO (Z
+ro+ t 1'LP' (Z+rl, t 1))
・・・・・・・・・・・・・・・・・・・・・(
12 In general, 1 an −+−shi・tan −′−!
! -L("-<y" r IIJ and m11112
m1m2 nI and ml and n2 have no common divisor, +11<
11111 n2 < ml) unidirectional propagation is n11 each.

Although m2 elements pk are required, the propagation with a width within their directional range is the least common multiple of ml and In2.

If you have 1 element, you can actually eliminate it.

(3) In the above embodiment, the propagation of the "closed 9" state was described as well as the propagation of the value "1", but this invention is applicable to the length of the line, the number of lines, and the distance of the line other than the closed state.

It can also be applied to propagation of edge directionality, etc.

(4) In the above embodiment, 1 in a predetermined direction and an intermediate direction between i + 1, which is referred to as i + 2.
ノ; As I mentioned about dissemination, the element Tl for each direction
' (z) + + "It can also be applied to locally parallel front needles in which 0 to 7 have = 0.1 and each 1+4 element performs calculations by interaction with other elements in its 3x3 neighborhood. As mentioned above, this allows a two-dimensional 6-closed Pa state to be expressed.

(5) In the above embodiment, calculations using a 3×3 a neighborhood on a square lattice were described, but operations other than 8 neighborhoods, such as 16
It is also possible to use a square lattice (for example, a hexagonal lattice).

As explained above, the two-dimensional information propagation method according to the present invention consists of a number of conductors arranged at two-dimensional grid points and a line No. 11 that interconnects these performers at adjacent grid points. When performing directional information propagation only from multiple signal lines, multiple information elements are transmitted at each grid point. , and used them alternately to propagate information, so information propagation force t can be achieved in intermediate directions other than the direction of the signal line, and the wiring of the signal line in the hardware is simple, making it easy to apply the features to be extracted. It has an extremely excellent effect of widening the range.

[Brief explanation of the drawing]

Figure 1 is a two-dimensional square grid representation, where (a) is a horizontal line segment, (b) is a slightly inclined line segment, and (C1 is a diagram showing a line segment with an inclination). , Figure 2 shows the wiring diagram for the adjacent grids of the device. Figure (a) is the wiring diagram for the 3x3 range, Figure (b) is the wiring diagram for the 5x5 range, and Figure 3 is the wiring diagram for the 3x3 range. Direction definition diagram, Figure 4 is a definition diagram of vector representation of neighboring points, Figure 5 is an explanatory diagram of one black point 1 information propagation, Figure 6 is 3 × 3
An explanatory diagram of information propagation from the intermediate direction (Keima direction) between 0'' and 11゛ by the neighborhood, Figure 7 is a block circuit diagram of the ↓9 operator at each grid point for propagation in the Keima direction, and Figure 8 is Performance 1
Wiring diagram with the grid near the 9th device, Figure 9 is the general angle +n
-+ "Explanatory diagram of directional propagation. Figure 10 is an explanatory diagram for configuring wide propagation from two unidirectional propagations. Figure (1)
) is a diagram showing the propagation in FIG. 6, and FIG. 6 <c> is a diagram showing the propagation by combining the propagation shown in FIG. (8 is the PL register. 515 Fig. 1a) Fig. 2(a) Fig. 3 Fig. 4 Fig. 6 Fig. 7 Coo Co+ ---C41Coo -C70C
71Figure 8

Claims (1)

[Claims] Using a parallel computing device consisting of computing units arranged in two-dimensional lattice points and signal lines adjacent to the computing units and interconnecting the grid points, or a computing device that simulates the same, the directionality When performing information propagation with , information can be propagated in an intermediate direction other than the direction of the signal line by preparing multiple information elements at each of the grid points and using them alternately to propagate the information. A two-dimensional information propagation method characterized by the fact that it is pregnant.