JPH031712B2 - - Google Patents

Description

[Detailed Description of the Invention] [Technical Field of the Invention] The present invention provides a highly practical character reading device that is capable of stably and efficiently extracting features for broad classification identification by reading a large number of characters including kanji. Regarding.

[Technical background of the invention]

The features of an input character pattern are extracted and compared with the features of dictionary patterns registered in advance in a dictionary, and the input character pattern is read and recognized according to the result of the comparison. However, when a large number of characters including Chinese characters are to be read in this character recognition, the amount of processing required to match the above-mentioned features becomes extremely large, and the processing efficiency deteriorates significantly.
Therefore, conventionally, prior to the above recognition process, characters to be recognized are roughly classified based on a plurality of general characteristics.
This reduces the number of characters to be compared in the identification stage, thereby improving processing efficiency and speeding up the processing speed. At this time, it is necessary to use the same features as the features used for the above-mentioned major classification and the features used for candidate character identification, but this feature is necessary to prevent deformation of character patterns caused by handwritten characters and various noises. It needs to be sufficiently stable.

However, conventionally, the features used to broadly classify character patterns have focused on the complexity of the character lines, the shape of the peripheral part of the character, and even the rough shape of the entire character. There are things that have been done. In addition, these are introduced in detail in, for example, the following literature.

Sakai, Watanabe “Current status of printed kanji recognition” Information Processing, Vol.22 No.4 PP.274-279 (April 1981) Figure 1 shows an example, focusing on the shape around the original character pattern. Then, the character lines existing in the rectangular scanning areas U, D, L, and R on the top, bottom, left, and right of the surrounding area are detected, and the amount is set to "0", "1",
It is quantized into three levels of "2" and used as feature data. Also, what is shown in Fig. 2 is the circumscribed edges u, d, which are in contact with the original character pattern.
Focusing on l, r, these circumscribed edges u, d, l, r
The area of the character background that is in contact with is determined and used as a multidimensional feature vector. The difference is that the former (Figure 1) focuses on the character line area, while the latter (Figure 2) focuses on the area of the background area, but both are characterized by the area quantity around the character pattern. It can be said that this is true.

[Problems with background technology]

However, the feature extraction described above is very effective in cases where the pattern is standardized, such as in printed characters, but it is very effective in cases where the pattern is standardized, such as in printed characters, but the pattern is significantly deformed even though it has characteristics as a character, such as in handwritten characters. In such cases, it is extremely unstable. Furthermore, even if the characters are printed, if the character pattern is chipped or faded, feature extraction becomes extremely unstable. In other words, there is a problem in that feature extraction cannot always be performed stably against deformation of character patterns and various noises.

[Purpose of the invention]

The present invention has been developed in consideration of these circumstances, and its purpose is to read a large number of characters, including handwritten kanji, and to identify the major classifications of input original character patterns. It is an object of the present invention to provide a highly practical character reading device that can extract features stably and efficiently.

[Summary of the invention]

The present invention focuses on the fact that much of the classification information (features) necessary for character classification identification exists in the periphery of the character pattern, and calculates the length of the boundary line between the character line part and the background part in the periphery of the character pattern. The input original character pattern is calculated for each circumscribing side, and the average amount of unevenness of the character pattern is calculated from the relative value of the length of these border lines with respect to the length of the circumscribing frame, and this is classified into major categories. The present invention relates to a character reading device that extracts characteristics of characters.

〔Effect of the invention〕

According to the present invention, a novel concept of a boundary line length between a character line part and a background part corresponding to each circumscribed side of the character pattern surrounded by circumscribed sides touching the input character pattern, that is, the length of the boundary line around the character This method introduces the degree of unevenness of the part, in other words, the complexity of the shape, and is characterized by the relative size of the length of these border lines with respect to the circumscribed frame, so it is difficult to prevent changes in the size of the character pattern or misalignment of the character pattern. Characteristic information can be obtained in a sufficiently stable manner against various noises, and major classifications of input patterns can be stably and efficiently identified, so character recognition processing can be significantly improved. Furthermore, even when dealing with complex character patterns such as kanji, handwritten characters, and even printed characters, the above features can be achieved using the simple method described below. Since it can be easily obtained, it can have a highly practical and remarkable effect in character recognition processing.

[Embodiments of the invention]

Hereinafter, one embodiment of the present invention will be described with reference to the drawings.

FIG. 3 shows the feature extraction process of the input character pattern according to the present invention, and here, a handwritten input character pattern "Ten" is shown. The feature extraction process is performed by first detecting a circumscribing frame surrounding a given character pattern. This circumscribing frame has a left circumscribing edge L that touches the leftmost edge of the character pattern, a right circumscribing edge R that also touches the rightmost edge, an upper circumscribing edge U that touches the uppermost edge of the character pattern, and a lower circumscribing edge D that touches the lowermost edge. It is determined by detecting. In addition, each of the above circumscribed sides U, D, L, R scans the character pattern,
It can be easily obtained by calculating the limit value of the position coordinates of the character line.

After that, each of the obtained circumscribed edges U, D, L,
With R as the reference position (reference line), each circumscribed side U,
Scanning is performed from D, L, and R in the direction toward the center of the character pattern, and the length of the scanning line until reaching the character line is determined. However, from the viewpoint of signal processing, it is preferable that the maximum length of the scanning be a constant ratio depending on the size of the character pattern. Then, the sum of the differences between these scanning line lengths is calculated for each circumscribed side U, D, L, and R, respectively, and this value is detected as the boundary line length between the background part and character line part of the character pattern. . For example, the length l L of the boundary line on the circumscribed side _L is defined as the maximum length h _H (h>1) as shown in Figure 3, and scanning is performed in the direction toward the center of the character pattern with L as the reference position. ,
Find each length. However, if the scanning line does not reach the character line, the scanning is terminated when the length h _H is reached. Then, for the scanning lines obtained in this manner, the difference in length (absolute value) between adjacent scanning lines is added as shown in FIG. 4, and then the vertical width V of the character is added. The length of the contour line tangent to the circumscribed side L _L (the length of the thick line in Figure 4) obtained in this way is stabilized against the width of the character line of the character pattern, the inclination and positional shift of the character, and covers a wide area. It is a feature quantity. Then, by similar processing, the contour line lengths of the background portions that are in contact with the other circumscribed sides R, U, and D, respectively, are determined.

After that, these values are normalized for the size of the circumscribing frame of the character pattern, that is, the height and width, and the border line length is calculated as C _L = l _L /H + V C _R = l _R /H + V C _U = l _U /H + V Calculate as C _D = l _D /H+V. However, H and V in the above formula indicate the horizontal width and vertical width of the character pattern circumscribing frame. As a result, information indicating the characteristics of the surrounding area of the input character pattern can be expressed as a set of boundary line lengths {C _L , C _R , C _U , C _D }
It is required as. This information corresponds to the average degree of unevenness of the outer shape seen from the circumscribed sides U, D, L, and R of the character pattern, in other words, the complexity of the surrounding area of the character, and therefore, the character pattern can be roughly classified. This sufficiently reflects the characteristics of character patterns for identification. In other words, the characteristics of the surrounding area of the character pattern obtained in this way {C _L ,
By using C _R , C _U , C _D }, it becomes possible to effectively, stably and reliably classify and identify the character patterns.

For example, when the handwritten characters "Ko" in various shapes are input as shown in _{FIG .} , l ₃ , these are all l ₄ (=H ₄
+hV ₄ ). Therefore, if we normalize these feature information by font size,
As long as the ratio of H and V is the same, the feature values are the same. In other words, {C _L , C _R , C _U , C _D } can be said to be a feature that is stabilized against character deformation.

FIG. 6 is a schematic configuration diagram showing an embodiment of an apparatus of the present invention which performs such processing to extract features for broad classification identification of character patterns.

A character pattern 1 to be read is read by a scanning photoelectric conversion device 2 consisting of, for example, a television camera.
The data is photoelectrically converted and input by scanning the character surface. The image signal of the character pattern 1 inputted via the photoelectric conversion device 2 is inputted to the binary quantization device 3, where it is discriminated at a discrimination level determined based on, for example, the background density and is binary quantized. . The quantized character pattern signal is then stored in a pattern storage device 4 such as a one-frame memory as a binary pixel signal corresponding to the scanning position. The character pattern pixel signals stored in the storage device 4 are scanned in a predetermined direction and read out, and subjected to the feature extraction process described above.

That is, the circumscribing frame detection device 5 searches for the positional coordinates of the character line of the character pattern stored in the storage device 4, and determines its leftmost coordinate X _L and its rightmost coordinate X _R
The left circumscribed side L is determined as L=X _L , and the right circumscribed side R is determined as R=X _R. At the same time, the uppermost coordinate Y _U and the lowermost coordinate of the character line of the character pattern
Find Y _D and find the upper circumscribed side U and lower circumscribed side D. Then, the circumscribing frame detection device 5 sends information about these circumscribing sides L, R, U, and D to scanning circuits 6L, 6R,
It is given to 6U and 6D as control information, respectively. Therefore, each scanning circuit 6L, 6R, 6U, 6
In D, a reference line for starting scanning is determined from the information on the given circumscribed side, and the signal of the character pattern stored in the storage device 4 is moved in the direction from the reference line toward the center of the character pattern (from left to right). , (right → left), (top →
(lower) and (lower to upper), and the scanning line length corresponding to (contacting) each circumscribed edge is determined. The scanning line information obtained in this way is stored in the boundary line calculation section 7.
L, 7R, 7U, and 7D, and as mentioned above, calculate the total sums l _L , l _R , l _U , and l _D of the scanning line length differences, respectively, and normalize these to obtain the relative amount C _L ,
These pieces of information that require C _R , C _U , and C _D are integrated, and feature information around the input character pattern {C _L ,
C _R , C _U , C _D } are given to the feature comparison device 8.

In the feature dictionary 9, the above-mentioned feature information for each character to be recognized is registered in advance, and the feature comparison device 8 compares the similarities between these dictionary features and the obtained features of the input character pattern. (similarity) is calculated sequentially. Then, when the degree of similarity is greater than or equal to a predetermined tolerance value θ, the category of the obtained feature dictionary is output as a major classification identification result.

In this way, the embodiment device that performs feature extraction and broad classification identification based on the feature extraction according to the present invention can be realized very easily. Then, by feeding the thus obtained large classification identification results and the image signal of the input character pattern to the character recognition section at the next stage, the character recognition can be performed easily and efficiently. Moreover, as described above, it is stable against deformation of character patterns such as handwritten characters and various noises, so that the efficiency of recognition processing can be significantly improved. Therefore, it is possible to construct a highly practical character reading recognition system that recognizes many characters including Chinese characters, and has great effects.

Note that the present invention is not limited only to the above-described embodiments. For example, in the example, the boundary line length was determined as the sum of the differences in the lengths of adjacent scanning lines, but
To obtain a similar effect, the boundary line length may be determined by directly tracing the outline of the character portion. In this way, although the amount of data processing increases, it is possible to obtain more detailed and accurate border line lengths for character lines and background parts. Major classification identification may also be performed using feature information and the like. In short, the present invention can be implemented with various modifications without departing from the gist thereof.

[Brief explanation of the drawing]

Figures 1 and 2 are diagrams for explaining the concept of conventional character pattern feature extraction, Figure 3 is a diagram for explaining the concept of character pattern feature extraction according to the present invention, and Figure 4 is a diagram for explaining the concept of character pattern feature extraction according to the present invention. FIG. 5a is a diagram for explaining a specific method of character pattern feature extraction according to the present invention.
-d are diagrams for showing the effects of the present invention, and FIG. 6 is a schematic diagram of the main part of an apparatus according to an embodiment of the present invention. 1... Character pattern, 2... Scanning photoelectric conversion device, 3
...Binary quantization device, 4...Pattern storage device, 5...
Circumscribing frame detection device, 6L, 6R, 6U, 6D...scanning circuit, 7L, 7R, 7U, 7D...boundary line calculation unit,
8...Feature comparison device, 9...Feature dictionary.

Claims (1)

[Scope of Claims] 1. Means for photoelectrically converting and inputting the original character pattern, means for binary quantizing and storing the input original character pattern, and four directions of the stored original character pattern, top, bottom, left and right. means for detecting a circumscribed frame consisting of circumscribed edges; means for detecting an outline of a character background portion in a direction from each of the circumscribed edges toward the center of the original character; and an outline of a character background portion corresponding to each of these circumscribed edges; Means for detecting an average amount of unevenness around the original character pattern from the relative length of a line length with respect to the circumscribed frame, and means for broadly classifying the original character pattern from the detected average amount of unevenness. A character reading device characterized by: 2 The means for detecting the outline length of the character background area is to scan from each circumscribed edge in the direction toward the center of the original character pattern and accumulate the differences in scanning distance from the circumscribed edge to the character line. A character reading device according to claim 1.