JPS63313283A - Image normalizing method - Google Patents

Abstract

PURPOSE:To obtain a normalized image of feature quantity reduced in noise and to improve a recognition rate by directly extracting the feature from an original pattern and matching a difference between linear density and the coordinates of feature quantity calculation. CONSTITUTION:The linear density hx', hy' of an inputted binary image f(i, j) are found out from equations 1, 2. A constant alpha is added to the inside linear density of a circumscribed rectangle in a graphic to loosen nonlinearity. Then, feature extraction is executed by 3X3 mask operation to obtain a feature image. Then, the discrete deviation of the coordinates between the measured linear density and the feature image is matched. Then, accumulated density functions Hx(i), Hy(j) are found out from equations 3, 4, the sums are found out from equations 5, 6, an added map is found out from equations 7, 8, and then a normalized feature image buffer is found out by equations 9, 10.

Description

[Detailed description of the invention] [Industrial application field] The present invention is useful for character recognition 9, character recognition in drawing reading, etc.

The present invention relates to an image normalization method as a preprocessing when recognizing linear patterns such as symbols.

[Conventional technology]

The correlation method (superposition method) for -m is well known as a method for recognizing character 9 figures. This is unknown pattern 2
The correlation Rk between the dimensional image P(i, j) and the dictionary (standard) pattern image Qk(i, j) is defined as the following formula, and if the pattern belongs to k that maximizes this Rk, then This is the way to do it.

Rk=(Qk*P)/1Qkl-IPI 1By the way, since this correlation method is sensitive to changes in pattern position and size, in order to compensate for this, prior to correlation calculation, the spread of the image of the unknown pattern and , alignment of misalignment is routinely performed. This work is called normalization. In addition,
From equation (1) above, the amount of product-sum calculations to find the correlation increases in proportion to the image size (N x N) - the square of the side length, that is, N2, and the amount of product-sum calculations to find the correlation increases in proportion to the size of the image (N x N) - the square of the side length, that is, N2, and When there are many types of images (for example, over 2000 types of characters), the amount of calculation becomes an obstacle to realizing a practical device, so normalization is usually performed along with image reduction.

When performing such normalization, it has conventionally been common to perform linear normalization using an affine transformation formula such as the following formula.

Here, (x, y) are old coordinates, (x', y') are new coordinates, a, b, c, d, Xo! yo is a constant.

However, such a linear normalization method is insufficient for images that undergo irregular (partial) deformation, such as handwritten characters. For this reason, for example, "Journal of Electronics Engineers of Japan
'84/11. VOL, J67-D, No. 11
, pp. 1379-1383, a nonlinear normalization method called "fine density equalization" is proposed. In this method, a characteristic value called linear density is defined for sample values of a pattern sampled at a sampling interval δ, and the characteristic value is converted so that the product of the sampling interval and the characteristic value becomes constant. As a result of this conversion, a non-linear transformation is performed in which the spacing between dense lines is lengthened and spacing between sparse lines is shortened, thereby homogenizing the line spacing in the space and making effective use of the space.

On the other hand, with improvements in superimposition methods, it has become possible to separate images into separate feature planes based on the characteristics of each point (pixel) and superimpose them. Therefore, as shown in the flowchart of FIG. 9, a method combining these two methods can be considered.

[Problem that the invention seeks to solve]

However, such a method causes the following problems. In other words, the image that has been subjected to line density equalization is the original image shown in Figure 10 (a), where dense lines are copied many times and coarse lines are omitted, so when viewed microscopically, This results in a rough image as shown in FIG. 10 (b), which may act like noise during feature extraction.

In particular, in image processing, 3
Since local operations using ×3 masks are often used, it is a big problem that local features become coarse.

Therefore, an object of the present invention is to provide a normalization method that can reduce noise and improve recognition rate.

[Means for solving problems]

Line density measurement and feature extraction are performed from the original image, and after matching the discrepancy in the discrete coordinate system due to the difference between the linear density measurement method and feature extraction method, the feature image is nonlinearly normalized using the linear density obtained from the original image. do.

[Effect]

Nonlinear normalization method using linear density equalization,
By combining this method more effectively with the feature plane superimposition method, we aim to reduce quantization errors and improve recognition performance.

[Embodiments of the invention]

FIG. 1 is a schematic flowchart illustrating an embodiment of the invention. As is clear from the figure, the normalization according to the present invention is divided into four stages: ■ Cotton density measurement ■ Feature extraction from the original image ■ Coordinate stand ■ Reduction normalization calculation of the feature image by adding pixels will be carried out. This will be explained in detail below.

■ Linear density measurement input binary image f(i, j) (t=1.2...M,
j=1, 2.・N), Xi in each direction of the x and y axes,
Projection of the number of lines to the X and y axes in yi (or
=Xi, 3'=3'J, the number of intersections between the vertical line and the figure f) is the line density hx'(i)s hy'(j
) is calculated as follows.

hy' (J) = Σf(i,j)・f(i-1,j
)...(3) However, T is the inverse of 1, 0 of f, and is specially set as f(i, 0)=Of(0, j)"0.

FIG. 2 is an explanatory diagram for explaining the concept of cotton density.

In the figure, the circle rOJ indicates the place where the number of lines is counted. For example, the projection hx'(i) of the number of lines onto the X axis at x=(=i) is "3", and the same < y
Projection of the number of lines to the y-axis at J (=j) hy'(j
) becomes 1". Perform the above operation for Xi = l, 2...M 7j = 1.2...N, and find the linear density. Here, , character patterns "i", "ni", etc., which have a part where the line density is "0" in the center of the image, if the line density is left as is, the center will be crushed (the white part will be completely clogged). )
Therefore, as shown in FIG. 3, a constant α is added to h x', h,' inside the circumscribed rectangle of the figure to alleviate nonlinearity.

In addition, Figure 3 (a) shows the linear density hx' before relaxation, and (b)
respectively indicate the linear density hX after relaxation. That is, the linear density hx, h after relaxation is expressed as follows.

However, i, = min (i I hx' (i
)≠0)i,=max(i l hx'
(i)≠0) However, js = lll1n (j
l h,' (j)≠OL = wax (j
l hy' (j)≠0 (4) In this way, the cotton densities hx(i) and hy(J) are measured.

(2) Feature Extraction In feature extraction, the original image is processed using a 3×3 mask operation to obtain a feature extraction image (or feature image). Specifically, the mask calculation is, for example, a Laplacian, and if nine pixels in the mask are "O", the pixel value output as the central pixel value is also generally "0".

■ Coordinate matching attempts to match the measured linear density and the discrete coordinate deviation with the feature image. In other words, feature extraction is 3×3
Since the linear density measurement is performed using a 1×2 mask, a coordinate shift as shown in FIG. 4 occurs.

In addition, the same figure (a) shows the feature extraction range, 11 is the original image, 12 is the feature image, 13 is the feature extraction mask, 14 is the movement range of the mask, and the same figure (b) shows the cotton density measurement range. 16 is an original image, 17 is a linear density measurement range, and 18 is a linear density definition range.

That is, as is clear from the definition of linear density, the range where the result of feature extraction is not "0" is always larger by one pixel on each side than the range where the linear density is not "0". Then, the pixels of the feature images that are not “O” on the outermost sides of both sides are those 1
It is thought that this is caused by a pixel in the original image inside the pixel, so for the outermost pixels on both sides, the line density is defined to be equal to the line density one step inside, thereby matching (correcting) the coordinate shift. . FIG. 5 is for explaining this, and FIG. 5(a) shows the cotton density before expansion, and FIG. 5(b) shows the linear density after expansion. That is, 19A and 19B are expanded portions, which are formed by copying the value of the line density one pixel inside.

■ Pixel addition In this step, a cumulative linear density function is defined, a summation map is obtained that divides the feature image, and a normalized feature image is obtained according to the map. Therefore, the cumulative linear density function is defined over integers.

......(5) In addition, h, (0), h, (0), h, (M+1), h,
(N+1) is the result of coordinate alignment.

Next, the sum of linear densities GX and Gy is expressed as GX = HX (M +1), Gy = Hy (N
``1). This is explained in Figure 6, where hX in Figure (A) is sequentially accumulated to obtain Figure (A).
(b) The total sum GX is determined as shown in (b). Here, HX
, Hy on continuous coordinates, the interval (i, i+
l), HX(x) is HX(x) = (
x −1)Hx(i + 1)+ (i + l −X
)HX(1). Similarly, for y in the interval (j, j +1), H, (y) = (y-j) Hy (j + 1) + (j +
1-y)Hy(j).

Next, as shown in FIG. 7, a coordinate sequence Xm,)'n of the addition map is created from the cumulative linear densities Hx, Hy, and the original image f
(i, j) (i-1...M, j=1...N) The characteristic image P(i, j) (i=0・
...M+1.3=0...N+1), the normalized feature image Q(k, f) (k=1...
K, f=1...).

Therefore, the addition map coordinate string xk, yn (k=1
...K! ! =1...L) as shown in (6), and a normalized feature image Q is obtained according to Xh and 7n. At this time, the characteristic image P(i, D is extended from the grid point of the integer to the continuous coordinates. That is, P(x, y)=P((x)+1.(y)+1)(However, [] indicates the Gaussian symbol, so (x) is
Indicates the largest integer not exceeding . ).

Under these conditions, add together as ・・・・・・(7). The integral range of equation (7) is shown in Figure 7 (
It is R shown in c), and the result Q(k, J) is S.

FIG. 8 is a block diagram showing an example of a configuration for implementing the above-described processing using hardware.

In the figure, 1 is the original image buffer, 2A, 2B are x,
y-direction linear density measurement section, 3 is a feature extraction circuit, 4A and 4B are coordinate matching sections, 5 is a feature image buffer, 6A and 6B are map coordinate string forming sections, 7 is an integration circuit, and 8 is a normalized feature image buffer. be.

In other words, the measurement of the linear density explained in item 0 above is performed by the measurement unit 2.
This is done in A and 2B. On the other hand, the feature extraction from the original image in the buffer 1 explained in section 3 is performed by the circuit 3 and stored in the buffer 5, and the coordinate matching explained in section 0 is performed in the coordinate matching sections 4A and 4B. Map coordinate string forming section 6A
, 6B, the map coordinate string X as shown in Fig. 7 (a) and (b)
h+3'j is determined, and the integration range (R) in the characteristic image is determined by this, so the integration circuit 7 calculates the above (7).
) A normalized feature image is obtained and stored in the buffer 8. That is, in this case, the measurement of the cotton density and the extraction of features are performed in parallel, and after coordinate alignment and formation of a map coordinate string are performed, nonlinear normalization of the feature images is performed.

〔Effect of the invention〕

According to the present invention, features are extracted directly from the original pattern, and differences in linear density and coordinates for feature amount calculation are matched, so the influence of quantization errors during nonlinear transformation is reduced, and It is possible to obtain a normalized feature image with less noise, and as a result, there is an advantage that the recognition rate can be improved.

[Brief explanation of the drawing]

FIG. 1 is a schematic flowchart showing an embodiment of the present invention;
Fig. 2 is an explanatory diagram to explain the concept of cotton density, Fig. 3 is an explanatory diagram to explain the method of alleviating nonlinear distortion, and Fig. 4 is an explanatory diagram to explain the difference between the feature extraction range and the linear density measurement range. Figure 5 is an explanatory diagram to explain the coordinate matching method, Figure 6 is an explanatory diagram to explain the cumulative linear density,
FIG. 7 is an explanatory diagram for explaining a method for obtaining a normalized feature image from a summed coordinate string for normalization, FIG. 8 is a block diagram showing the hardware configuration of an embodiment of the present invention, and FIG. The figure is a flowchart showing a method of combining a normalization method using linear density equalization and a superposition method, and FIG. 10 is an explanatory diagram for explaining an original image and an image subjected to linear density equalization. Code explanation 1...Original image buffer, 2A, 2B...
・Cotton density measurement section, 3...Feature extraction circuit, 4A,
4B... Coordinate matching unit, 5... Characteristic image buffer, 6A, 6B... Map coordinate string forming unit, 7... Integrating circuit, 8... ...Normalized feature image buffer, 11.16 ...Original image, 12
... Feature image, 13 ... Feature extraction mask, 14 ... Mask movement range, 15 ...
...Width of characteristic image, 17... Line density measurement range,
18...Line density definition range, 19A, 19B...
・・・Extended part. Agent Patent Attorney Akio Namiki Agent Patent Attorney Kiyota Matsuzaki 1 Figure 112 Figure 71, i (M, N) II3! Il (I)
(mouth 214 figure g5 figure man 6 figure 1 f7 figure suitable figure

Claims (1)

[Claims] A set of feature parameters and a set of characteristic values are calculated from a set of sample values of an image pattern sampled at a predetermined sampling interval, and the calculated set of feature parameters and characteristic values are calculated. An image normalization method characterized in that after adjusting discrete coordinate deviations with a set of , the set of feature parameters is normalized so that the product of the sampling interval and the characteristic value is constant.