JPH0139156B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a method for extracting character features used in character recognition, and is intended to enable feature extraction of portions that are hidden by conventional methods.

In character recognition, characteristics of characters viewed from the left and from above are determined and used as data for character recognition. For example, taking the number 3 shown in Figure 1 as an example, the handwritten number 3 is scanned with a line scanner to obtain a binary video signal, temporarily stored in memory, and the video signal of number 3 is cut out from the memory ( (generally scanned and stored in the memory along with other handwritten characters, etc.), each horizontal scan from left to right is a collection of points that first convert from 0 (background) to 1 (character) on the scan. , and use it as the outline feature of the number 3 seen from the left. According to this logic, line segments l ₁ , l ₂ , and l ₃ are the features of the character viewed from the left (herein referred to as left features). Similarly, if we find the first set of points that change from 0 to 1 on each vertical scanning line scanning from top to bottom, the second
Line segments l ₄ and l ₅ in the diagram are obtained, and these become the features of the character viewed from above (upper features). These left features, upper features, and right features obtained in the same way,
The downward features are easy to extract, and the numbers 0, 1, 2
This provides good feature data for characters with simple patterns such as ..., and can be effectively used for character recognition. For example, there is a line segment l ₂ that intersects between the above line segments l ₁ and l ₃ , but there is no other number than 3 in the range 0 to 9, so the target is either 0 to 9. It is possible to recognize the number 3 using only one characteristic of cigarettes. However, there are many features that are hidden in the background and are not extracted. For example, in the left feature, large concave parts P ₁ and P ₂ are not extracted,
In the case of upper features, only the part that forms a roof or eaves is extracted, and everything below it is hidden and not extracted. Therefore, the feature extraction is insufficient for characters with slightly complex patterns such as kanji and kana, and some characters cannot be separated or identified from other characters.

Therefore, the present invention aims to enable accurate and quick character recognition by making it possible to extract features from hidden parts. That is, the present invention scans a character pattern in the horizontal or vertical direction, extracts line segments that make up the outline of the character pattern as a collection of black-and-white transition points between the background and characters, and uses these as character features for character recognition. In the feature extraction method, a line forming the outline of a character pattern part hidden behind the character pattern part that first appears in the scanning direction is a collection of odd-numbered black and white transition points other than the first in each horizontal and vertical scanning line direction. The present invention is characterized in that the characters are extracted and used as the character features, which will be described in detail below with reference to the drawings.

The line segments l ₁ , l ₂ . . . are recognized as follows. That is, since the scanning lines l ₀ , l ₁ , and l ₂ have different Y coordinates, finding the coordinates (X coordinates) of the first point where 0 → inversion occurs for each Y coordinate, (Y ₀ , 0), (Y ₁ ,
0), (Y ₂ , 0), (Y ₃ , X ₃ ), (Y ₄ , X ₄ )... are obtained. Here, 0 indicates "there is no point that causes 0 → 1 inversion", but if such a point is also simply expressed as Xi (i = 0, 1, 2...), the above data becomes (Xi, Yi ). These data can be easily processed by recording the Xi data at the Yi address in the memory. That is, the address counter is sequentially incremented by 1 and Y ₀ , Y ₁ , Y ₂ , etc.
… Read address data X ₀ , X ₁ , X ₂ …,
When determining ΔX=X _i+1 −Xi, the difference ΔX is small for line segments l ₁ , l ₂ , and l _{3 ,} so it can be determined from this that they are continuous line segments. Also, if ΔX<0, it can be said to be a line segment going down the left shoulder, and if ΔX>0, it can be said to be a line segment going down the right shoulder. At the boundary between line segments l ₁ and l ₂ and l ₂ and l ₃ , ΔX suddenly becomes large. In this case, the line segment is broken,
It can be said that they are connected by line segments parallel to at least the horizontal scanning line. It is this discontinuity point and the point that takes on a certain value from Xi-0 that determines the end of the line segment. The starting end of the line segment _l1 is the latter, and the ending is the former, the both ends of the line segment _l2 are the former, and the starting end of the line segment _l3 is the latter. The fact that both ends are discontinuous points, such as the line segment l ₂ , means that when numbers 0 to 9 are connected, it can be said that there are character pattern portions at both ends, and both ends are closed. On the other hand, both ends of line segments l ₁ and l ₃ can be said to be open. As mentioned above, having a line segment with both ends closed is a major feature of the number 3.

By the way, the number 3 has depressions P ₁ and P ₂ , so
If this is also detected, the feature of number 3 will have been extracted even better. This concave part is line segment l ₁ ,
l It could not be extracted because it is in the shadow of ₃ , but the extraction logic is not "the first 0 → 1 change point" but "the third (odd numbered if generalized) 0 →
1 change point", the shaded part can be extracted. That is, as shown in Fig. 3, if we look at the 0-1 inversion of the scanning line _l a passing through the concave part P ₁ , it is
There are four points Q ₁ , Q ₂ , Q ₃ , and Q ₄ , and the inversion point Q ₃ that defines the contour of the shaded portion l ₆ is the third. Line segments l ₆ and l ₇ can be extracted using the logic of this "third reversal point", and by combining these with the line segment l ₂ obtained using the method shown in Figure 1, that is, the logic of the "first reversal point", a concave portion is formed. A line segment l ₈ that penetrates to the deepest part of is obtained. line segment
l Recognizing the number ``3'' becomes more reliable and easier when using ₈ . That is, when we calculate ΔX of this line segment, it changes from positive to negative to positive to negative, which clearly represents the characteristics of the number 3. Combining the line segment l ₈ with the line segments l _{1 and} l ₃ , that is, in the vertical direction, the line segments l ₁ , l ₈ , and l ₃ exist in this order, and in the horizontal direction, the line segments l ₁ and l Based on the logic that there is a line segment l ₈ to the right of ₃ whose ends overlap with the line segments l ₁ and l ₃ , the numeral 3 can be separated and identified from others even if it is handwritten in a very rough manner.

FIG. 4 shows an example of a handwritten "chi", and FIG. 5 shows an example of a handwritten "te". It is not uncommon for these differences to be determined by the presence or absence of the protrusion R. However, since this is hidden in the upper character pattern portion B, it cannot be extracted using the logic of the first change point in FIG. On the other hand, if we use the logic of the third change point, especially the third change point in the vertical scanning line, line segments l ₉ and l ₁₀ can be extracted (both ends are the ones extracted at the "first change point"). ). If the line segments l ₉ and l ₁₀ can be extracted, the presence or absence of the protrusion R can be checked based on whether the change in the Y coordinate is uniform (with the vertical direction being X), and it becomes possible to identify the tip and the tip. There are many confusing characters, such as katakana ``u'' and ``ka'', ``ku'' and ``u'', ``mi'' and ``shi'', ``riton'', and ``niton'', which are difficult to distinguish using simple left and top features when written in a rough manner. It is something. These can also be expected to be identified by extracting the shadowed parts using the "odd-numbered" logic, or by combining it with a simple left-hand feature, etc.

In the recognition of handwritten numbers and kana, a large number of features, such as 1000 types, are used, and these constitute a tree circuit of about 50 stages, and character recognition is performed by walking through the tree circuit. The features extracted by the present invention are also used in addition to one of them. Note that if the features are appropriate, results can be obtained by passing through a relatively small number of stages, and character recognition speed can be increased. In this respect, the method of the present invention is extremely effective.

[Brief explanation of drawings]

1 and 2 are explanatory diagrams of the conventional method, and FIGS. 3 to 5 are explanatory diagrams of the method of the present invention. In the drawing, l ₁ to _{l 10} are line segments, and l ₆ , l ₇ , l ₉ , and l ₁₀ are line segments of hidden character pattern parts.

Claims (1)

[Claims] 1. Scanning a character pattern horizontally or vertically,
In a feature extraction method that extracts line segments constituting the outline of a character pattern as a collection of black-and-white transition points between the background and characters, and uses them as character features for character recognition, 1 in each horizontal and vertical scanning line direction is used. A line segment forming the outline of a character pattern part hidden behind a character pattern part that appears first in the scanning direction is extracted as a collection of odd-numbered black-and-white change points excluding the character pattern part, and this is used as the character feature. A character feature extraction method used for character recognition.