JPH0614374B2 - Binarization method for grayscale images - Google Patents

Description

The present invention relates to a binarization method for a grayscale image that is widely used in the field of image processing such as character recognition.

<Prior Art> Conventionally, in a character recognition device (OCR) or the like, since a character placed under uniform illumination conditions of known brightness is targeted,
Binarization was possible with a constant binarization threshold over the entire image. Further, in an apparatus for recognizing an object such as a machine part, since binarization is performed with a constant threshold value, there are restrictions on the color of the object and its background and illumination conditions. On the other hand, in image processing other than recognition, OC that originally handles objects that are close to binary
Different from R etc., the grayscale image is once differentiated and then binarized by a constant threshold (differential binarization method), or the threshold value is dynamically changed for each part of the image (dynamic 2 Valuation method) has been devised. For example, JP-A-59-1
Japanese Patent No. 146887 divides an image into partial regions, and determines a local threshold value of a point at the center of the partial region by an average value of brightness (or density) of the points in the partial region. The threshold for any point in the image is determined by interpolation of four local thresholds around that point.

<Problems to be Solved by the Invention> However, regarding the above-mentioned conventional differential binarization method, since the entire image is binarized with a constant threshold value, the threshold value is set so as not to damage an object such as a character or an object. When set to, there was a problem that a lot of noise was generated in the background. On the other hand, when the conventional dynamic binarization method as described above is used as the preceding stage of character recognition processing, 1) a low contrast area such as a character line portion thicker than the size of the partial area or a background portion not including characters , An improper threshold is given due to background noise rather than the contrast between the characters and the background.

2) There is a problem that the size of the partial area must be dynamically determined according to the size of the characters. Therefore, the conventional method has drawbacks such as noise in the background portion of the binary pattern, which makes it difficult to cut out and recognize characters, and it is necessary to reset the number of partial areas depending on the nature of the grayscale image to be handled. was there. Further, in the conventional dynamic binarization method, the noise generated in the background can be suppressed to some extent by increasing the size of the partial area, but on the contrary, the binarization processing for the illumination and shadow of the scene is stable. There was a problem that impaired sex. Therefore, it is an object of the present invention to provide a binarization method for a grayscale image that overcomes the drawbacks of the conventional method.

<Means for Solving Problems> In order to solve the above problems, the present invention provides a pre-processing as a previous stage for detecting and recognizing characters and figures in a three-dimensional scene particularly under non-uniform illumination conditions. In addition, as a dynamic binarization method of a differential image, characters and figures are surely separated from the background regardless of their size, and background noise is reduced as much as possible.

<Operation> In this way, an appropriate threshold value is corrected by correcting an inappropriate threshold value obtained for a low contrast partial area such as a character background with a threshold value and a weight value obtained for the surrounding partial areas. Can be determined.

<Embodiment> FIG. 2 shows an embodiment of a recognition processing procedure of characters in a scene by the binarization method of the grayscale image of the present invention, 1 is a binarization processing, 2 is a character cutting processing, and 3 is a character. This is recognition processing.
First, in the binarization processing 1, a grayscale image pattern is input, and a monochrome binary binary image pattern is output. In the character cut-out process 2, the binary image pattern is scanned in the vertical and horizontal directions, and the character is cut out on a character-by-character basis by the circumscribed quadrangle. Next, in character recognition processing 3, the cut-out characters are recognized one by one and the result is output.

FIG. 3 shows a detailed processing procedure of the binarization processing 1.
Is a region division process, 12 is a threshold value determination process, 13 is a weight value determination process, 14 is a first interpolation process, 15 is a second interpolation process, and 16 is a binary determination process.

First, the grayscale image pattern G (i, j) is subjected to the region division processing 1
1 is a partial area of a predetermined size (for example, 16 × 16 pixels) (partial area number is (m, n))
Is divided into This division processing does not necessarily have to be equal division as shown in FIG. 4 (a), and the areas may overlap or be separated from each other as shown in FIGS. 4 (b) and 4 (c). .

In the threshold value determination process 12, a threshold value is determined for each of the divided partial areas from the appearance frequency of the gray value of the pixel in the partial area. Various methods are conceivable for determining the threshold value from the appearance frequency of the density value. For example, a simple method using the average value of the density values, the median value, the average value of the maximum value and the minimum value as the threshold value, or the threshold value Considering the processing as a two-class problem that separates two classes into black and white, and using the density value with the highest degree of separation (maximum interclass variance) as the threshold (discriminant analysis method)
and so on.

In the weight value determination process 13, the divided partial regions (m,
For each n), the average value A _mn of the brightness of the pixels in the partial area and its variance value σ ² _mn are obtained, and the weight value W _mn indicating the contrast is determined based on this. The weight value W _mn has, for example, the following four types.

[Method a]: W _mn = σ ² _mn [Method b]: W _mn = σ 2 _mn [Method c]: W _mn = σ ² _mn / A _mn [Method d]: W _mn = σ _mn / A _mn Weight value When using variance or standard deviation as (method a,
In b), it is determined that the higher the contrast of the partial area is, the higher the possibility that the important information is included in the image processing is included, and a higher weight value is given to the threshold value given in the partial area. This is effective when the lighting conditions are relatively good. In addition, when a value obtained by dividing (normalized) the variance or standard deviation by the average value of brightness is used as the weight value (methods c and d), for example, in a lighting condition in which character strings having the same contrast are not uniform. If the original contrast of the character string is the same, even if the lighting intensity is different, the weight is set to the same value. This is effective when the values are not uniform.

The first interpolation processing 14 includes the threshold value θ _mn obtained in the threshold value determination processing 12 and the weight value W obtained in the weight value determination processing 13.
_This is a process of correcting the threshold value θ _mn based on _mn , where the threshold value after the first interpolation processing is θ ′ _mn , and the threshold value before the first interpolation processing is θ
_{When mn} , there are two methods given by the following equation.

[Method 1] [Method 2] Here, ω _mn = 0: W _mn <φ1 (constant) W _mn : otherwise However, if the weight value is W _mn ≧ φ2 (φ2 is a constant), this processing is not performed.

By repeating this processing, the threshold value of the partial area to which a large weight value is given is propagated to the surrounding partial areas, and the threshold value of the small area that does not include characters is corrected, so that the occurrence of background noise can be suppressed.

FIG. 5 is an explanatory diagram of the second interpolation processing 15, showing an arrangement of arbitrary four adjacent partial areas, and the threshold value θ corrected by the first interpolation processing is provided at the center of each partial area. ′
It is assumed that ₁ , θ ′ _2n , θ ′ ₃ and θ ′ ₄ are given. At this time, a quadrangle having the centers of the four partial regions as vertices is considered, and the threshold T (i, j) corresponding to an arbitrary pixel P (i, j) in the quadrangle is set to four thresholds θ ′ ₁ ,
It is obtained by interpolation of θ ′ ₂ , θ ′ ₃ and θ ′ ₄ . That is, the threshold value T (i, j) in the pixel P (i, j) is the horizontal and vertical lengths of the quadrangle L ₁ , L ₂ , from the upper left apex of the quadrangle to the pixel P (i, j). When the horizontal and vertical distances are α and β, they are given by the following equation.

T (i, j) = θ ′ ₁ · (L ₁ −α) (L ₂ −β) / L
₁ · L ₂ + θ ′ ₂ · (L ₁ −α) · β / L ₁ · L ₂ + θ ′ ₃ · α · (L ₂ −β) / L ₁ · L ₂ + θ ′ ₄ · α · β / L ₁ · L ₂ surface shape represented by the formula is not a clearly plane since there to go through any four points, it can be shown that a straight line is on four sides. For example, considering the value of T (i, j) on the upper side of FIG. 5, substituting j = 0 and β = 0 into the above equation, the second and fourth terms on the right side become zero. For, Becomes It can be proved by the same calculation for the other edges, and it can be shown that linear interpolation is meant on the edges.

In the binary determination process 16, a binary pattern B (i, j) is obtained from the magnitude comparison between the grayscale image G (i, j) and the threshold value T (i, j) obtained in the second interpolation process 15.

B (i, j) = 0: G (i, j) <T (i, j) 1: G (i, j) ≧ T (i, j) FIG. 1 shows the binarization process 1 described above. 1 is a common control unit, 102 is a frame memory (1), 103 is an address control unit, 104 is a threshold value / weight value determination unit, 105 is an interpolation processing unit (I), and 106 is a hardware configuration example. The interpolation processing unit (II), 107 is a frame memory (2), and 108 is a comparison circuit. In the figure, each unit is a common control unit 101.
It operates according to the timing signal given from the. First, the frame memory (1) 102 is a memory for one screen of a grayscale image that stores a grayscale image pattern, and the image pattern is input via a television camera or the like (not shown). The address control unit 103 receives from the frame memory (1) 102,
An address for reading the image pattern of the designated partial area is given to the threshold / weight determining unit 104. The threshold value / weight value determination unit 104 sequentially gives the partial area numbers to the address control unit 103, obtains the threshold value and the weight value for each partial area,
This is given to the interpolation processing unit (I) 105. Interpolation processing unit (I) 1
Reference numeral 05 denotes the threshold value and the weight value obtained from the threshold value / weight value determination unit 104, and the threshold value is corrected and the village interpolation processing unit (II) 10
Give to 6. The interpolation processing unit (II) 106 is an interpolation processing unit (I).
The threshold value for the entire image pattern is obtained by interpolating the threshold value for each partial area corrected by
Output to 7. The comparison circuit 108 includes a frame memory (1)
The grayscale image pattern is binarized by comparing the levels of the corresponding pixels in 102 and the frame memory (2) 107,
The result is output to the frame memory (2) 107.

<Effects of the Invention> As described above, according to the present invention, an inappropriate threshold value obtained for a low contrast partial area such as a character background is set as a threshold value and a weight value obtained for the surrounding partial areas. An appropriate threshold value is determined by correcting the threshold value.
Therefore, according to the present invention, as a pre-process of a region extraction process for recognizing a character or a figure or an object in a scene, a character, a figure or an object in a grayscale image is automatically and stably maintained regardless of its size. Can be binarized to

[Brief description of drawings]

FIG. 1 is a hardware configuration example of a binarization process showing an embodiment of a binarization method of a grayscale image of the present invention, FIG. 2 is a recognition process procedure of characters in a scene, and FIG. 3 is a binarization process. Detailed procedures of the processing, FIGS. 4 (a), (b), and (c) are examples of division of partial regions, and FIG. 5 is an explanatory diagram of the second interpolation processing. 1 ... Binarization processing, 2 ... Character cut-out processing, 3 ... Character recognition processing, 11 ... Region division processing, 12 ... Threshold value determination processing, 13 ... Weight value determination processing, 14 ... First Interpolation processing, 15 ... Second interpolation processing, 16 ... Binary determination processing,
101: common control unit, 102: frame memory
(1), 103 ... Address control unit, 104 ... Threshold / weight determination unit, 105 ... Interpolation processing unit (I), 106 ... Interpolation processing unit (II), 107 ... Frame memory (2), 108 …
… Comparison circuit.

Claims (5)

[Claims] 1. A local threshold value for each partial area of a grayscale image is determined based on the appearance frequency of the density of pixels in the partial area, and the local threshold value is determined based on the obtained local threshold value. A threshold value for the entire grayscale image is obtained, and a binary image is obtained based on the magnitude of the threshold value for the entire grayscale image and the value of the grayscale image.
In the binarization method, means for obtaining, for each of the partial areas, a weight value representing the contrast of the density for a grayscale image in the partial area, and a local threshold value and weight value given to each partial area Based on the local threshold corrected by the local threshold and the weight value given to the partial region and the local threshold corrected by the processing by the first interpolation means or the repeated processing, all the grayscale images The method for binarizing a grayscale image, comprising: a second interpolating means for obtaining a threshold value for each pixel. 2. The binarizing method for a grayscale image according to claim 1, wherein the weight value is given by a standard deviation of the brightness of the grayscale image in the partial area. 3. A weight value is an average value of the brightness of the grayscale image A _v in the partial _area, when the standard deviation and S _d, claims, characterized in that given by S d _/ A _v A method of binarizing a grayscale image according to the range (1). 4. The first interpolating means W _{mn is} a weight value for the m-th partial area (m, n) in the horizontal direction and n-th partial area in the vertical direction, before and after correction by the first interpolating means. If the thresholds of θ _mn and θ ′ _mn are Claims (1), (2), or (3), characterized in that the following process is performed once or repeatedly.
A method of binarizing a grayscale image according to the item. 5. The first interpolating means sets predetermined weighting values W _mn , φ ₁ and φ ₂ for the m-th partial area (m, n) in the horizontal direction and the n-th partial area in the vertical direction, φ ₁ ≦ φ when the ₂ becomes parameter, patents, characterized in that W is replaced with _mn ≦ phi ₁ if W _mn = 0 performs interpolation processing on the partial area, does not perform the interpolation process for phi _{2 <W mn} if the partial region Claim (1) or (2), or (3)
The method for binarizing a grayscale image according to item (4).