JPH0256688A - Character segmenting device - Google Patents

Abstract

PURPOSE:To obtain respective characters, which are included in concentration picture data, as a binary picture by repeatedly execute re-binarization so as to unify character lines, which go to be a different label, regarding the label not to be unified to the character line as a noise and removing such a label. CONSTITUTION:The concentration picture data are divided into several small areas and an optimum binarizing threshold for character area detection is computed in each small area by using a binarizing means 1200 by divided areas. Then, the concentration picture data are binarized. Next, by using a re- binarizing means 1800 by character areas, the propriety of the character picture data in each character area to be outputted from a binarizing means 1700 by character areas is decided and the concentration picture data are re-binarized so as to satisfy a condition for the binary picture data of satisfactory quality. Then, the satisfactory binary picture data of the respective characters are outputted. Thus, the respective character areas can be exactly detected from the concentration picture data, which are outputted from an image-pick up means 1000, and further, the detected character picture data go to be the satisfactory binary picture data of suitable character line width not to accompany the noise.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a character cutting device, and more particularly to an improvement in a character cutting device that extracts and outputs areas where individual characters exist from grayscale image data containing characters.

[Prior art] Engraved characters are resistant to aging and dirt, and are widely used in various production processes to indicate production instructions and important information for management.Especially in the automobile production process, stamped characters are It is widely used as a unique number for parts, engines, etc.

To recognize such stamped characters, a combination of a character cutting device and a character identification device is usually used.

Conventional character cutting devices image the engraved characters using an ITV camera or the like to obtain a grayscale image, and convert each pixel data of this grayscale image into binary image data.
Each character was detected and extracted from the value image data.

Then, the character recognition device performs pattern matching in which the individual character image data extracted in this way is sequentially overlaid and compared with standard character image data registered in advance, and the degree of coincidence or similarity between the two images is determined. Characters were identified based on.

However, with conventional character extraction devices, characters are extracted by binarizing the entire image data using a fixed threshold value, so there may be uneven shading or noise in the background of the image data. When the contrast between a character part and a background part changes for each character, there is a problem in that it is difficult to accurately detect and cut out a desired character.

In other words, with the currently widely used stamping machines, it is difficult to control the stamping conditions so that they are always the same, and the depth of the stamp often varies for each stamped character, and the steel plate being stamped is If it is thin, unevenness may occur around the characters due to uneven stamping pressure. Therefore, in the grayscale image data of stamped characters, the contrast in density between the character part and the background part changes for each individual character, and furthermore, the density of the background part varies partially.

Also, when the number of digits of characters to be recognized is large.

For example, when there are many digits like a vehicle number,
The angle and orientation of each character relative to the light source becomes uneven.

For this reason, in images of engraved characters, even if the characters are the same, the density of the character portions often differs from character line to character line.

Further, if anti-rust treatment or painting is applied after the engraved characters are engraved, uneven shading occurs in the background of the image due to unevenness in the anti-corrosion treatment or painting, or changes in lighting conditions. Furthermore, small scratches and dirt easily adhere to the area around the engraved characters,
Therefore, a lot of noise appears in the background of the image.

However, excessive shading noise occurs in the background part of the image, and the contrast between the text part and the background part changes for each character.
Furthermore, if we use conventional technology that binarizes the entire image data with a fixed threshold value and cuts out the characters for grayscale image data in which the density differs for each part of the character line even within one character, as shown in Figure 7. As shown in Figure 2, there are problems in that an image that is supposed to be a single character is partially cut off, and adjacent characters sometimes come into contact with each other and cannot be separated.

Also, even if it is extracted as a single character, part of the character may be crushed, resulting in a partially thick character image, or conversely, part of the character line may be too thin, or noise may remain in the character area. There is a problem in that variously deformed character images occur.

In order to solve such problems, Japanese Patent Application Laid-Open No. 60-144
A technique disclosed in Japanese Patent No. 884 is known.

This conventional technology suppresses variations in image density caused by the angle and orientation of each character with respect to the light source, and when imaging the engraved characters, oblique light is irradiated onto the engraved surface from at least two different directions. It is characterized by inputting the grayscale image data of the engraved characters corresponding to each irradiation, converting it into a binary image, and forming the final binary image by calculating the logical sum of these binary images.

However, this conventional technology also binarizes the entire input grayscale image data using a fixed threshold, so even if the binarization threshold is successively reset, individual character parts cannot be detected accurately. The problem was that it was difficult to

Furthermore, as another technique for detecting characters from grayscale image data, there is known a technique disclosed in Japanese Patent Application Laid-Open No. 60-211583, which focuses on changes in density between a character portion and a background portion.

This conventional technology calculates the minimum value among the density levels of the pixel of interest and surrounding pixels for each pixel of gray-scale image data captured by a camera, and sets a threshold value using that value. It is characterized by evaluating the difference value of the density level between the pixel of interest and surrounding pixels to detect whether it is an edge part of a character or not.

However, this conventional technology cannot accurately display characters when the difference in density between the character part and the background part changes for each individual character, such as in an image of engraved characters, and the density differs within one character. There was a problem in that it was difficult to detect and extract.

[Problems to be Solved by the Invention] The present invention has been made in view of such conventional problems, and its purpose is to solve the above-mentioned conventional problems,
From shading image data in which the background part contains noise, etc., the density contrast between the character part and the background part changes for each character, and the density differs in each part of the character line of one character. However, it is an object of the present invention to provide a character cutting device that can accurately detect desired characters and cut out each detected character so that it becomes a binary image containing no noise and less character deformation.

[Means for Solving the Problems] In order to achieve the above object, as shown in FIG. , in a character cutting device that detects and cuts out individual characters, divides the grayscale image data into several small areas, calculates the optimal binarization threshold for character area detection for each small area, and divides the grayscale image data into several small areas. Based on binary image data outputted from divided area binarization means 1200 for character area detection, which binarizes data for each small area, and binary image data outputted from the divided area binarization means 1200, it is possible to determine whether a character string exists. Character string area detection means 1400 for detecting a small area, and character string area detection means 14
Character area detection means 1500 that sequentially detects small areas where individual characters exist in a small area where the character string detected as 00 exists; and character area detection means 1500 that sequentially detects small areas where individual characters exist; It is characterized in that it includes: binarization means for each character region 1700 that calculates a valorization threshold value and binarizes grayscale image data for each small region, and cuts out and outputs individual character image data.

Here, the dividing area 2111 conversion means 1200
divides the entire grayscale image data into a predetermined number of divisions corresponding to the size of the characters, creates a density histogram for each divided small area, and calculates the is threshold value from the histogram. It is preferable to form the grayscale image data so as to obtain it and binarize it for each small area.

Further, the character string area detection means 1400 scans the binarized image data in parallel with the arrangement of character strings to create a projection distribution, and detects an area where the character string exists based on the projection distribution and the size of the characters. The character area detection means 1500 scans the binary image data of a small area in which a character string exists in a direction perpendicular to the arrangement of character strings to create a projection distribution, and combines the projection distribution with the character It is preferable to sequentially detect small areas in which individual characters exist based on the size.

Further, the character area-by-character binarization means 1700 creates a density histogram of the grayscale image data for each small area in which each character exists, and calculates an optimal threshold value for each small area from each histogram to obtain the grayscale image data. It is preferable to form the image so as to binarize it.

Further, at the subsequent stage of the character area-specific binarization means 1700, the quality of the binary image data for each character area outputted from the character area-specific binarization means 1700 is determined to satisfy the conditions for a high-quality character image. Rewrite the grayscale image data for each character area.
Re-binarization means 180 for character areas to obtain character images
It is preferable to provide 0.

Further, the character area-by-character re-binarization means 1800 performs labeling processing on the binary image data for each character area, and determines the degree of clustering of character lines and the degree of character line width. It is preferable to increase or decrease the threshold value to re-binarize the grayscale image data in the character area, and to repeat this process to obtain a high-quality character image.

Furthermore, between the character area detection means 1500 and the character area-specific degree binarization means 1700, breaks in character lines within one character and contact between two or more characters are detected, and the character area is corrected. It is preferable to provide a character area correction detection means 1600 for detecting the correction.

Further, an imaging means 1000 and a divided area binarization means 1200
It is preferable to provide a character candidate density extracting means 1100 between and Q, which alleviates the unevenness of shading in the background part of the shading image data output from the imaging means 1000 and extracts a density that becomes a candidate for the character part.

Furthermore, the character string area detection means 1100 performs a process of replacing an area where the density changes greatly in a narrow range of the grayscale image data with the same density as the surrounding background, and replaces the processed grayscale image data with the original grayscale. It is preferable to extract the density of a region that is a character candidate by alleviating unevenness in the background portion of the grayscale image data by obtaining differential image data from the image data.

Further, between the binarization means 1200 for each divided region and the character string region detection means 1400, the binarization means 1200 for each divided region
It is preferable to provide a noise removing means 1300 for removing noise from the binary image data output from.

[Point of View of the Present Invention] As described above, in the conventional character cutting technique, the engraved characters are photographed using the imaging means 1000.
The shading image data output from 00 is converted into binary image data, and the shading image data is subjected to differential processing or difference means and then binarized and converted into image data representing the edge portion of the character. Characters were extracted from value image data.

However, all of these conventional techniques do not adequately address the fact that the shading image data of engraved characters is different from the shading image data of printed characters, and the density of both the character portion and the background portion tends to change partially. was not taken into consideration.

In response, the present inventors have investigated how to extract all individual characters as high-quality binary images from grayscale image data in which the densities of both text and background parts change little by little. As a result of his research, he came to focus on each point of the garment.

The first point of focus of the present invention is that when detecting character areas where individual characters exist from grayscale image data, the grayscale image data is divided into several small areas, and an optimal threshold value is determined for each small area. The purpose is to calculate the optimum threshold value and to binarize the gray scale image data for each small area using the optimum threshold value.

That is, when considering binary image data for detecting characters, it is important that the character lines of individual characters are not interrupted and that the characters are not in contact with each other.

Therefore, the entire image is divided into several small areas by a predetermined number of divisions, and the 21ia threshold value for character area detection is calculated for each small area, and the image data is
Value.

Then, when performing character string and character detection, the 2#! obtained in this way! By adding information about the size and spacing of characters to the image data, and also simply checking for breaks and contacts, it is possible to accurately detect individual character areas.

The second point of focus of the present invention is that when cutting out the image data of each character based on the detected individual character areas, the optimal binarization for character extraction is performed using the density histogram in the small area of each detected character. The purpose of this method is to obtain a threshold value by calculation, and use the thus obtained binarization threshold value for cutting out each character to binarize the grayscale image data for each small region of each character.

By doing so, each character can be extracted as a binary image from the grayscale image data.

As described above, according to the present invention, in accordance with the first and second points of view, by dividing grayscale image data into small areas and creating binary image data used for character string and character area detection, Individual characters can be extracted as high-quality binary image data even from grayscale image data in which the density of both text and background parts changes little by little.

By the way, in the binary image data of each character extracted in this way, there are cases where character lines are broken, noise occurs in the background, or character line widths are changed due to local variations in the character and background parts. Some characters are significantly different from other characters.

The third point of focus of the present invention is to cut out each character as a good binary image by re-binarizing the binary image data of such characters.

That is, in order to check for each detected character area how the character lines are clustered, to distinguish between character lines and noise, and to check whether the character line width is appropriate, the present invention uses binary character image data. Labeling is performed on the label, and re-binarization is repeated to integrate the character lines with different labels.Then, labels that are not integrated into the character lines even after repeated re-binarization are treated as noise and removed. do.

By repeating this re-binarization process,
Each character included in the grayscale image data can be cut out and output as a high-quality binary image.

[Effect] Next, the present invention will be explained in detail.

First, when characters are read by an imaging means 1000 such as an ITV camera, the imaging means 1000 outputs grayscale image data including the characters.

L12 - FIG. 2(a) shows a part of the character string included in such grayscale image data.

The first feature of the present invention is that the divided area binarization means 1200
The purpose of this method is to divide the grayscale image data into several small regions, calculate the optimal binarization threshold for character area detection for each small region, and convert the grayscale image data into 21 using the following method.

FIG. 2(b) shows one small region obtained by dividing the grayscale image data of FIG. 2(a) into predetermined sizes (dashed lines indicate dividing lines).

In the present invention, in order to calculate the binarization threshold for text area detection for each small area, a density histogram of image data is created for each small area, and the threshold value is determined from this density histogram. .

Figure (c) is a diagram illustrating the frequency of each density level by dividing the individual elements (called pixels) of the image data within the small area of figure (b) into their respective density levels. (referred to as a density histogram), as shown in Figure (c).
represents a case in which the imaging means 1000 outputs an image in which the text portion is dark and the background portion is bright; from this density histogram, it is found that there are two types of images that have a high frequency near the density levels indicating the text portion and the background portion. It will be understood that mountains are formed.

Therefore, as shown in (c) in the same figure, when a small area includes characters, the valley between the peaks of the text area and the background area is found and binarized for character area detection. Determined as threshold value 4 Furthermore, if the small area does not include any characters, the density level is concentrated almost only in the background area, so there is only one peak. Therefore, in this case, the level darker than the mountain in the background is
It is determined as a binarization threshold for character area detection.

When the grayscale image data is binarized using the optimal binarization threshold for character area detection determined for each small area in this way, the binary image data as shown in FIG. 2(d) is obtained. can get.

The segmented region binarization means 1200 of the present invention repeatedly performs such binarization for each small region and converts the entire grayscale image data into binary image data.

S -- In the grayscale image data, the density unevenness in the background part is significant, and the character image data binarized by the divided area binarization means 1200 has collapsed character lines (particularly in the character lines forming a closed curve). If there is a lot of noise in the background or contact between characters, the binarization means for each divided area 120 is used.
It is preferable to provide the character candidate density extraction means 1100 before the character 0.

This character candidate extracting means, when the imaging means tooo outputs shading image data with significant shading unevenness in the background as shown in FIG. The process of replacing each pixel with the brightest density level (background level) of the surrounding pixels is repeated, and after replacing all the dark levels of the text area with the bright levels of the background area, each pixel At each time, the difference from the original grayscale image data is calculated, and grayscale image data with this difference as a new density level is output.

By doing this, from the imaging means 1000 to the sixth
Even when grayscale image data as shown in Figure 6(a) is output, parts where the density level does not change sharply are softened, and the original grayscale is restored as shown in Figure 6(b). It is possible to obtain grayscale image data in which character parts are emphasized compared to image data.

Therefore, such character candidate density extraction means 1100
By using this, even if there is significant unevenness in shading in the background part of the shading image data output from the imaging means 1000, the divided area binarization means 1200 can be used to convert one shading image into a single shading image without being affected by this. The data can be converted into binary image data for detecting the entire character area.

Noise and Column By the way, the binary image data for character area detection outputted from the segmented area binarization means 1200 as described above often contains fine noise.

Therefore, the 2 bits output from the divided area binarized image means
Preferably, fine noise is removed from the value image data using a noise removal means 1300.

FIG. 3 shows an example of binary image data obtained in this manner.

The apparatus of the present invention uses a character string area detection means 1400 to detect a small area where a character string exists from this binary image data.

That is, this character string area detection means 1400 creates a distribution map by accumulating black pixels (pixels representing character parts) included in the binary image data shown in FIG. 3 in a direction parallel to the arrangement of character strings, From this distribution map, a small area representing the range of character strings is detected. This situation is shown in FIG. 3(a).

Then, in the cumulative black pixel distribution map, the part with large changes is used as the boundary line representing the range of the character string, and the character size information (in this diagram, the height of the character) is used to represent the range of the character string. Define the small area.

Sentence 3'' L1 amount M stages Next, the apparatus of the present invention uses the character area detection means 1500 to sequentially detect small areas where individual characters exist from the small areas where character strings exist. Specifically, within the range of the detected character string, the black pixels forming the character part are accumulated in the direction perpendicular to the arrangement of the character strings, and a distribution map as shown in FIG. 3<b) is created.

In this distribution map, the parts with large changes are used as the boundaries representing the range of characters, and the information on the size of the characters (in this diagram, the width of the characters) and the spacing between characters are used to create small areas representing the range of characters. Confirm.

At this time, the character area detected by the character area detection means 1500 may not always be accurate due to breaks in characters or contact between characters.

Therefore, in the present invention, the character area detection means 150
It is preferable to input the character area outputted from 0 to a character area correction detection means to detect and correct any breaks in characters or contact between characters.

-12. When the small area for each individual character is detected in this way, the device of the present invention uses the character area specific binarization means 1700 to determine the optimum value for each small area where each character exists. A binarization threshold for character segmentation is calculated. Then, using the determined threshold value, the grayscale image data output from the character image data extraction means 1100 is extracted for each small region (for each character region).
Character image data is output by binarizing it.

That is, the character area-specific binarization means 1700 creates a histogram of image data for each small area in which individual characters exist, and calculates an optimal threshold value for each image area from each histogram. The grayscale image data is then binarized for each character area and output as character image data.

In this way, according to the character cutting device of the present invention, not only can individual character regions be accurately detected from the gray scale image data output from the imaging means 1000, but also character images can be extracted from the detected character regions. It can be output as high quality binary image data.

The second feature of the present invention is that character areas are detected in this way, and the binarized image data of each character is further removed by removing character breaks, noise, and checking the character line width. , to obtain better quality character image data.

This lick, the character cutting device of the present invention,
Using the digitization means 1800, the binarization means 170 for each character area
Determine the quality of the character image data for each individual character area output from 0, re-binarize the grayscale image data so that it satisfies the conditions for a high-quality character image, and output high-quality binary image data for each character. are doing.

That is, the character area-based re-binarization means 1800 labels the detected and binarized character image data for each character area. Labeling processing is a process of attaching the same label to adjacent pixels in order to indicate whether or not they are a group of consecutive pixels. For example, rl shown in FIG.
Taking J as an example, this character image data is classified into three labels: label 1, label 2, and label 3.

If the number of labels for black pixels in the character area (pixels in the character part) is 2 or more, as in this "L", either the character line is broken, noise exists, or both appear at the same time. That's what I'm doing.

Therefore, in order to determine whether or not the character line is broken, the character area of the grayscale image data is re-binarized by changing the character cutting threshold value to obtain binary image data of each character.

Figure 4(b) shows the case where the threshold value is changed so that the number of black pixels increases; by doing this, label 1
The character lines of label 2 and are connected, and these two labels are 1
will be integrated into However, in this case, the label 3 representing noise still remains, and a new noise is generated to become label 4. Here, a collection of pixels of each label will be called an "island".

After performing the labeling process, it is determined whether or not to re-binarize the character area included in the grayscale image data.This determination is made as follows.

First, as a result of the labeling process, if the number of labels in the binarized character image data is 2 or more, the number of pixels is calculated in order of label numbers, and the largest "island", the second [island]
9 Next, for the largest "island", check whether the height and width of the "island" are suitable for the font size, and calculate the pixels between the largest "island" and the second "island". Check whether the numbers are sufficiently different.

In the case of "L" shown in Figure 4 (a), the number of labels is 2 or more, and there is not much difference between the largest "island" and the second [island], so the threshold is set to increase the number of pixels. By changing the values and re-binarizing, character image data as shown in FIG. 2(b) is obtained.

In the same figure (b), the largest "island" (label 1) satisfies the font size, and the second "island" (label 3)
Since the difference between the two is sufficient, there is no need to re-binarize any further. Therefore, the "islands" of labels 3 and 4 are judged to be noise and removed, and the "islands" of labels 3 and 4 are removed, as shown in the same figure (c). Obtain character image data.

The number "5" shown in FIG. 3(a) is also re-binarized in the same manner.

In addition, other characters shown in the same figure (a>) have 1 label.
Therefore, instead of re-binarizing immediately, first calculate the ratio of the total number of pixels of the "island" to the length of the outline (this is known to be a value corresponding to the average character line width) Calculate and determine whether this value is appropriate. If this value is not appropriate, for example, if the character line width is thin as shown in 0 shown in FIG.

In this manner, according to the present invention, individual character areas can be accurately detected from the grayscale image data output from the imaging means 1000, and the detected character image data has an appropriate character line width. This results in high-quality binary image data without noise.

Then, the character area cutting means cuts out each character detected in this way from the image data and outputs it.

[Effects of the Invention] As explained above, according to the present invention, the imaging means 100
The grayscale image data output from 0 is divided into several small areas, and for each divided small area, fi13i! is applied to each grayscale image data. automatically finds the threshold value and binarizes it,
Further, the process of converting the binary image data obtained in this manner into 21ii again is repeated so that the binary image data obtained in this way becomes better quality binary image data.

Therefore, each character included in the shading image data can be accurately detected without being affected by uneven shading or noise in the background, changes in the contrast between the text and background, or differences in the density of each character line. It has the effect of being able to be cut out.

Furthermore, according to the present invention, by appropriately determining the division size of the grayscale image data at the beginning, each threshold value can be set each time even when there are fIi image conditions or changes in illuminance of the light source. It can be calculated automatically without modification.

This method 9 has the effect that when cutting out various characters from grayscale image data at various production sites, it is possible to significantly save labor compared to the conventional technology.

[Example] Next, preferred embodiments of the present invention will be described based on the drawings.

FIG. 5 shows a preferred embodiment of the character recognition device 10 to which the present invention is applied. The letter A is engraved. This stamped character A is, for example, a 19-digit character string consisting of alphanumeric characters and symbols.

The character recognition device 10 of the embodiment is for automatically reading the engraved character A, and includes an imaging means 100 for photographing the engraved character A, a character cutting device 200 according to the present invention, and a character identification device 300. It consists of

The imaging device 100 illuminates the engraved character A with a light source 16 and images it with a television camera 14. In the engraved character A, the clear character part is in shadow and dark, and the surface part of the vehicle body panel 12 is displayed as a bright and dark gray image.

The grayscale image data output from the television camera 14 is converted into a digital signal by the A/D conversion circuit 18 and then output to the character cutting device 200.

文ヱ”JLL漠1 The character extraction device 200 of this embodiment inputs the grayscale image data output from the A/D conversion circuit 18 to the character candidate density extraction circuit 20.

(a) This character candidate density extraction circuit 20 is formed to reduce the unevenness of density in the background part of the grayscale image data and extract and output the grayscale image data in which the character part is emphasized. image memory 22, density replacement circuit 24, second
image memory 26, a difference circuit 28, and a third image memory 30.

After the grayscale image data outputted from the A/D conversion circuit 18 is stored in the first image memory 22, the density of the character part is replaced by the density of the background part by the density replacement circuit 24. The image is stored in the image memory 26 of.

Then, the difference circuit 28 converts the grayscale image data stored in the first image memory 22 into the second image memory 26.
A difference calculation is performed on the grayscale image data stored in the gradation image data, and this is stored in the third image memory 30.

By doing this, for example, the A/D conversion circuit 18
Therefore, as shown in FIG. 6(a), even when gray image data with uneven shading in the background is output,
In the image memory 30, as shown in FIG. 6(,b), it is possible to obtain grayscale image data in which the unevenness of grayscale in the background area has been alleviated and the density of the area serving as a character candidate has been extracted.

Then, the grayscale image data stored in the third image memory 30 is outputted to the binarization circuit 32 for each divided area.

(b) This binarization circuit for divided areas 32 divides the input grayscale image data into a plurality of small areas, calculates an optimal binarization threshold for character area detection for each small area, and calculates a binary value for each small area. By performing imaging processing, the entire grayscale image data output from the third image memory 30 is converted into 211i image data.

Specifically, the binarization circuit 32 for each divided region includes a region division circuit 34, a density histogram creation circuit 36, a binarization threshold calculation circuit 38, and a binarization circuit 40.

Then, the area dividing circuit 34 divides the grayscale image data inputted from the third image memory 30 into a plurality of small areas according to a predetermined number of divisions, and inputs the divided small areas to the density histogram creation circuit 36.

The density histogram creation circuit 36 creates a density histogram for each small area from the grayscale image data input for each individual small area, and outputs it to the binarization threshold calculation circuit 38.

The binarization threshold calculation circuit 38 determines whether or not the small area includes characters based on the degree of variation in the density level of the input density histogram. As shown in FIG. 2(c), the valley between the frequency peak representing the character density level and the frequency peak representing the background density level is determined as a binarization threshold for character area detection. Furthermore, in the case of a small area that does not include characters, a density level darker than the frequency peak of the background density level is determined, and this is used as the optimal binarization threshold for the small area.

Then, the optimal binarization threshold value for each small area obtained in this way is output to the binarization circuit 40, and the binarization circuit 40
The binarized image data is output to the noise removal circuit 42.

In this way, according to the present invention, the optimal binarization threshold value is determined for each small area and binary image formation is performed. Even if the contrast with the background is different, high-quality binary image data with fewer breaks in characters and fewer contacts between characters can be obtained.

Then, the noise removal circuit 42 considers a region with a small number of consecutive black pixels (the number of pixels corresponding to a character part) to be noise among the input binary image data, and erases it. The image data is output to the character string area detection circuit 44.

(c) The character string area detection circuit 44 scans the binary image data inputted in this way in a direction parallel to the arrangement of character strings, and generates a projection distribution map in which black pixels (pixels of the character part) are accumulated. make. Then, the part of this distribution map with large changes is defined as the range where the character string exists, and after checking whether this range is appropriate compared to the height of the character, this is detected as a small area representing the character string. , and outputs binary image data of the detected character string area to the character area detection circuit 46.

(d) The character area detection circuit 46 scans the binary image data of the character string area input in this way in the direction perpendicular to the arrangement of the character strings, and accumulates black pixels (pixels representing the character part). Create a projection distribution map. Then, a part of this distribution map with a large change is found as a set as the left end and right end of the character, and it is checked whether the widths of the left end and right end of this set are appropriate by comparing with the width of the character. Then, the left end and right end coordinates of the character confirmed to be in the character area and the character image data surrounded by the left end and right end are outputted to the character area correction detection circuit 48.

(E) The character area modification detection circuit 48 determines the center coordinates of the character from the left and right end coordinates of the character input in this manner. Then, the character pitch is determined from the difference in the character center coordinates of each adjacent character, and this is compared with a predetermined character pitch, and there are no characters that could not be detected successfully due to character breaks or contact between characters. Find out if.

FIG. 7(a) shows an example of a character boundary line detected by the character area detection circuit 46, and FIG. 7(b) shows an example of a character boundary line corrected and detected by the character area correction detection circuit 48. has been done.

That is, if the detected character area is smaller than the predetermined character pitch, the "v" shown in FIG. 7(a), for example.
It is determined that the character is broken, as shown in Figure 7 (b), and it is regarded as one character, and the left and right end coordinates are newly determined, and the center coordinates are determined to determine whether the character width is appropriate. , check whether the pitch between adjacent characters is appropriate.

Furthermore, if the detected character area is larger than the predetermined character pitch, as in the case of "4" shown in Figure 7(a), erroneous detection may occur due to breaks in characters or contact between characters. It is determined that there is a character, and the projection distribution map in that range is checked again, and the left and right end coordinates of the character are re-detected to match the specified pitch and character width, as shown in Figure 7(b). .

Then, the left end and right end coordinates of each character corrected and detected in this manner are outputted to the character region-specific binarization circuit 50.

(f) The character area-by-character binarization circuit 50 performs optimal character extraction binarization for each small area in which individual characters exist, based on the data inputted from the character area correction detection circuit 48 in this way. Calculate the threshold. Then, using the optimal binary threshold value for character extraction obtained in this manner, binary image data of each character area is created from the grayscale image data output from the third image memory 30.

Specifically, the character area-specific binarization circuit 50 includes a density histogram creation circuit 52, a binarization threshold calculation circuit 54.
It is composed of a binarization circuit 56.

Then, the density histogram creation circuit 52 generates grayscale image data corresponding to the coordinates indicating each character area inputted from the character area correction detection circuit 48 and the coordinates indicating each character area inputted from the third image memory. Based on this, a density histogram is created and output to the binarization threshold calculation circuit 54.

The binarization threshold calculation circuit 54 uses this density histogram to find the valley between the peak of the character density level and the peak of the background density level, and outputs this as a binarization threshold.

In this way, the binarization threshold calculation circuit 54 calculates the optimum binarization threshold for each character and outputs it to the binarization circuit 56.

The binarization circuit 56 converts the gray scale image data corresponding to the coordinates indicating each character area into a binary image using the optimum threshold value, and converts the binary image data for each character obtained in this way into a binary image. The signal is output to the character area re-binarization circuit 58.

FIG. 6(e) shows the optimum 21iaiii image data for each character obtained in this way. As is clear from the figure, by using the circuit of this embodiment, the image data as shown in FIG. As such, high-quality binary image data for each character can be obtained.

(g) In the apparatus of this embodiment, the binary image data for each character obtained in this manner is input to the re-binarization circuit 58 for each character area.

This character area re-binarization circuit 58 repeatedly converts the image into binary image data (hereinafter referred to as "re-binarization") in order to improve the discontinuity of individual characters and the unsuitability of the character line width. It is formed to create a character image and also remove noise within the character area. Specifically, it includes a labeling circuit 60, a character line determination circuit 62, a re-binarization threshold setting circuit 64, a re-binarization circuit 66, and a noise removal circuit 68.

The labeling circuit 60 performs a process of attaching the same label to a group of black pixels (pixels corresponding to a character part) in binary image data for each character.

Then, the character line determination circuit 62 evaluates the labeled character image data and determines whether it is necessary to re-binarize it.

In other words, if the number of labels for each character image data is 1,
It is determined whether this collection of black pixels is appropriate as a character size and character line width is appropriate. If this condition is not satisfied, the re-binarization threshold setting circuit 6
4, the character image data is output to the noise removal circuit 68, and if the result is satisfied, the character image data is output to the noise removal circuit 68.

Further, when the number of labels of each character image data is two or more, the largest group of black pixels and the second group of black pixels are compared. Then, the character image data is denoised only when the largest cluster has more than twice the number of pixels as the second cluster, and the largest cluster satisfies the tolerance values for character size and character line width. It is output to the circuit 68, and the others are output to the re-binarization threshold setting circuit.

The re-binarization threshold setting circuit 64 lowers the threshold value when the number of labels of each character image data is 1 and the character line width is too thick, and raises the threshold value for all other cases, and performs re-binarization. Used as a new threshold for value conversion. Then, in the re-binarization circuit 66, the third image memory 3 corresponding to the character area is
The 0 grayscale image data is again binarized and outputted to the labeling circuit 60.

The character line determination circuit FI/i62 determines in advance the maximum number of times to repeat the above-described series of processes consisting of the labeling process and the re-binarization process, and when the specified number of times is completed, the character line determination condition is Even if the condition is not satisfied, the binarized character image data is output to the noise removal circuit 68.

The noise removal circuit 68 erases a group of black pixels having labels other than character lines in each character area, and
Output towards 0.

The character extraction circuit 70 individually cuts out the character image data together with the coordinates of the character area inputted from the noise removal circuit 68 and outputs it to the character identification device 300 .

FIG. 6(f) shows the binary image data of each character that has been re-binarized by the character area re-binarization detection circuit 58 and extracted from the character extraction circuit 70. As shown in the figure, even when character image data with discontinuous characters or uneven character line widths is output from the character area binarization circuit 50 as shown in FIG. ,
By using this character area re-binarization circuit 58,
As shown in FIG. 5F, it is possible to obtain character image data in which the discontinuity of individual characters and inappropriate character line widths are improved, and noise is removed.

文ヱILL聚1 The character identification device 300 of this embodiment includes the character cutting device 20.
The individual characters cut out by 0 are processed by a position normalization circuit 7.
2, the center of gravity of the input character is determined, the center of gravity is aligned with the reference position, and then outputted to the similarity calculation circuit 74.

The similarity calculation circuit 74 sequentially superimposes the standard pattern of each character type registered in advance in the standard pattern memory 78 and each character image data output from the position normalization circuit 72, and calculates the character types and their matches. The degree of similarity representing the degree of similarity is calculated and output to the character type determination circuit 76.

The character type determination circuit 76 compares the degree of similarity of each character with a predetermined criterion and determines whether the determined character has been correctly identified.

As explained above, the character cutting device 20 according to the present invention
By using 0, the engraved character A can be accurately cut out,
The interpretation can be performed accurately.

Note that the present invention is not limited to the embodiments described above, and various modifications can be made within the scope of the main aspects of the present invention.

For example, in the above-mentioned practical example, the explanation was given using stamped characters as an example, but the present invention is not limited to this, and can be used to recognize other types of characters as necessary.

Further, in the above embodiment, the case where an optical Il image means is used is explained as F1, but the present invention is not limited to this, and all other types of flI image means may be used as necessary. .

Further, in the above embodiment, the case where there is one character string and is arranged side by side has been explained as an example, but the present invention is not limited to this.
Even if there are multiple character strings or characters are arranged vertically, characters can be extracted in the same way.

[Brief Description of the Drawings] Fig. 1 is a claim correspondence diagram according to the present invention, Fig. 2 is an explanatory diagram showing the binarization operation for each divided area according to the present invention, and Fig. 3 is a diagram corresponding to the claims according to the present invention. FIG. 4 is an explanatory diagram showing the character string area detection operation and character area detection operation; FIG. 4 is an explanatory diagram showing the re-binarization operation for each character area according to the present invention; FIG. FIG. 6 is an explanatory diagram showing an example of the operation of the character recognition device shown in FIG. 5; FIG. 7 is an example of the character area detection operation shown in FIG. 5. FIG. 14 ... 20 ... 32 ... 42 ... 44 ... 46 ... 48 ... 50 ... 58 ... A ... TV camera character candidate density detection circuit by division area Binarization circuit Noise removal circuit Character string area detection circuit Character area detection circuit Character area correction detection circuit Binarization circuit for each character area Re-binarization circuit for each character area... Imaging device...Character cutting device Stamped character No. Figure 4 (b) (C) Figure 4

Claims (10)

[Claims] (1) In a character cutting device that detects and cuts out individual characters from grayscale image data including characters imaged using an imaging means, the grayscale image data is divided into several small regions, and each small region is binarization means for each divided area for character area detection, which calculates an optimal binarization threshold for character area detection for each subarea, and binarizes grayscale image data for each small area; a character string area detection means for detecting a small area in which a character string exists based on binary image data output from the means; Character area detection means for sequentially detecting small areas where individual characters exist, and 2 for optimal character extraction for each small area where individual characters exist.
A character region-specific binarization means for calculating a valorization threshold and binarizing grayscale image data for each small region; and a character cutting device for cutting out and outputting individual character image data. . (2) In the apparatus according to claim (1), the dividing area-specific binarization means divides the entire grayscale image data into a predetermined number of divisions corresponding to the size of the characters, and A character cutting device characterized in that a density histogram is created for each small area, and an optimal threshold value is determined from the histogram to binarize grayscale image data for each small area. (3) In the apparatus according to any one of claims (1) and (2), the character string area detection means scans and projects the binarized image data in parallel with the arrangement of the character strings. The character area detecting means converts binary image data of a small area where a character string exists into an arrangement of character strings. A character cutting device is characterized in that a projection distribution is created by scanning in a direction perpendicular to the character, and small areas in which individual characters exist are sequentially detected based on the projection distribution and the size of the character. (4) In the apparatus according to any one of claims (1) to (3), the character area-specific binarization means generates a density histogram of the grayscale image data for each small area in which each character exists. A character cutting device that generates a histogram, determines an optimal threshold value for each small region from each histogram, and binarizes grayscale image data. (5) In the device according to any one of claims (1) to (4), after the character area-based binarization means, each character area output from the character area-based binarization means is provided. A character area-specific re-binarization means is provided for determining the quality of the binary image data, re-binarizing the grayscale image data for each character area to satisfy the conditions for a high-quality character image, and obtaining a character image. A character cutting device featuring: (6) In the apparatus according to claim (5), the character area-based re-binarization means may perform binarization for each character area.
The labeling process is performed on the value image data, and by determining the degree of clustering of character lines and the degree of character line width, the threshold value is increased or decreased to re-binarize the grayscale image data in the character area, and this process is repeated. A character cutting device characterized by obtaining high quality character images. (7) In the device according to any one of claims (1) to (6), between the character area detection means and the character area specific binarization means, one
1. A character cutting device comprising a character area correction detection means for detecting a break in a character line within one character and a contact between two or more characters, and correcting and detecting a character area. (8) In the apparatus according to any one of claims (1) to (7), between the imaging means and the divided area binarization means, a background part of the grayscale image data output from the imaging means is provided. A character cutting device characterized by comprising a character candidate density extraction means for alleviating uneven shading and extracting densities as candidates for character portions. (9) In the apparatus according to claim (8), the character candidate density extraction means performs a process of replacing an area where the density changes greatly in a narrow range of the grayscale image data with the same density as the surrounding background. By calculating the difference image data between the processed grayscale image data and the original grayscale image data, the grayscale image data is formed so that the density unevenness in the background part of the grayscale image data is alleviated and the density of the area that becomes a character candidate is extracted. A character cutting device characterized by: (10) In the device according to any one of claims (1) to (9), between the divided area binarization means and the character string area detection means,
A character cutting device characterized by comprising a noise removing means for removing noise from binary image data outputted from a binarizing means for divided regions.