KR100954137B1 - Edge-based text localization and segmentation algorithms for automatic slab information recognition - Google Patents

Abstract

PURPOSE: An edge-based text localization algorithm and a segmentation algorithm for automatic slab information recognition are provided to recognize slab management numbers rapidly by performing character string extraction and character division in high success rate based on border images. CONSTITUTION: A border image is extracted from a color image(S110). Based on the extracted border image, a candidate string is extracted (S120). In the extracted candidate string, up and down locations of the character string are decided through width to height ratio of the character string and a character point(S130). Right and left locations of the character string is decided through the slope and thickness of characters(S140). In the image, the character string is divided into the characters(S150).

Description

Edge-based Text Localization and Segmentation Algorithms for Automatic Slab Information Recognition}

The present invention relates to a technique for extracting a management number for slab information recognition, and more particularly, to a method for extracting a management number for slab information recognition for recognizing a management number in a slab image.

In recent years, a high level of control technology is required to unmann the steelworks crane to accurately pick up the slab, and advanced equipment is required.

As a result, a greater amount of slabs are being produced, which requires a thorough management of them.

In order to manage such a large amount of slabs, each slab uses a unique management number, and administrators directly check whether the management number of the slab sent from the host computer matches the management number listed on the current slab. It is done.

Automating these steps will save labor costs for businesses, bring them closer to the steel plant's unmanned systems, and deliver significant improvements in work efficiency.

However, the management number recognition of the slab is not suitable for using a barcode or Radio Frequency IDentification (RFID).

In general, the work performed inside the steel mill is forced to undergo a high temperature treatment process.

In such an environment, the use of RFID or bar code that can withstand high temperatures is expensive, and the cost of production increases, resulting in economic disadvantages.

For this reason, management number recognition is most appropriate through image processing.

In general, there are two main ways to recognize characters.

The first is to extract the area containing the characters first, and then extract the characters from the area.

The second method is to extract the area containing the characters and the characters at the same time by using the contrast value.

While the first method has a high accuracy, it requires a lot of memory and processing time. The second method has a disadvantage that the processing speed can be improved, but the accuracy is poor.

Given these advantages and disadvantages, it is advantageous to choose the first method for the extraction and recognition of slab control numbers in steel mills.

Because, once a slab is moved by a crane, it takes some time to move it back.

Of course, it is not a long time, but because it is enough time to recognize the characters in the video can be said to be somewhat free from the constraints of time.

In addition, since a large amount of slabs have to be managed and discriminated, accuracy is essential.

So, even if it takes a long time because the computation is slow, it is advantageous to choose a more accurate method.

Therefore, in the first method, the projection is performed on the X and Y axes using a filter that extracts only vertical components to extract the area containing characters.

In this way it is possible to roughly identify the area where the characters are located. Next, the separated characters are recognized using the similarity method.

Similarity method is a method of recognizing a character by selecting whether the input image is the most similar among existing models using a kind of model.

However, the steel shop's work site is very different from the general environment, making it unsuitable for applying general algorithms.

For example, in the case of license plate recognition, a relatively clean image can be obtained due to less noise and damage.

On the other hand, in the case of the slab image, there was a problem that it was difficult to obtain a clean image due to a weak environmental factor such as a lot of dust in the steel mill, uneven lighting, and damage of characters due to heat.

The present invention has been made to solve the problems of the prior art as described above, an object of the present invention, in particular, to provide a management number extraction method for recognizing slab information for extracting the management number of the slab even in a harsh environment. .

To this end, a management number extraction method for recognizing slab information according to the present invention includes: a boundary image extraction step for extracting a boundary image; A candidate string extracting step for extracting a candidate string from the extracted boundary image; An up and down position determining step for determining an up and down position of the character string in the extracted candidate character string; A left and right positioning step for determining a left and right position of a character string in the character string in which the vertical position is determined; And a method for extracting a management number for recognizing slab information, the method comprising: a string dividing step for dividing a character string into characters in the character string in which the left and right positions are determined.

According to the present invention, there is an effect that can generate a boundary image even if the management number is lost a lot by heat in the complex environment of the lighting change is severe.

Therefore, based on this boundary image, it is possible to extract and recognize the slab management number more accurately and quickly with a high success rate up to not only string extraction but also character division.

Hereinafter, a method for extracting a management number for recognizing slab information according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

In describing the present invention, when it is determined that the detailed description of the related known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted.
In addition, the control number has a series of information about the manufacturing location and size of the slab, it is first to clarify that the extraction of the control number must be preceded in order for the program to recognize the slab information through the image.

1 is a flowchart illustrating a management number extraction method 100 for recognizing slab information according to an exemplary embodiment of the present invention.

Referring to FIG. 1, in the method for extracting a management number for recognizing slab information according to the present invention, a boundary image extraction step (S110) for finding a boundary image of a color image and a candidate string based on the extracted boundary image are extracted. String extraction step (S120), and the vertical position of the string to determine the vertical position of the string using the ratio of the height and width of the string and the character score in the extracted candidate string (S130), and the vertical position is Left and right positioning of the character string to determine the left and right position of the character string using the slope and thickness of the character in the determined image (S140), and the step of dividing the character string to the character in the image in which the left and right position of the character string is determined (S150) It is configured to include).

Referring to the management number extraction method 100 for slab information recognition according to an embodiment of the present invention configured as shown in Figure 1 as follows.

Here, a method of extracting a management number for recognizing slab information according to an exemplary embodiment of the present invention configured as shown in FIG. 1 will be described sequentially with reference to FIGS. 1 to 6.

Referring to FIG. 1, first, a boundary image extraction step (S110) of searching for a boundary image in a color image is performed.

Preferably, the color image is an original color image including a slab management number.

The boundary image extraction step S110 is divided into a process of applying a weighted median filter based Anisotropic Diffusion (WMFAD) (S111) and a boundary detection process (S112).

An example of extracting the boundary image will be described with reference to FIG. 2.

2 is a photograph showing an example of the boundary image extraction step of the management number extraction method for slab information recognition according to an embodiment of the present invention.

 Referring to FIG. 2, the original color image a becomes the same image as the processed image b after the anisotropic diffusion filter is applied.

Preferably, the anisotropic diffusion filter is a filter based on anisotropic diffusion to reduce noise of the image and to increase homogeneity for each region.

Finally, a boundary image c is extracted from the processed image b.

After performing the boundary image extraction step (S110) as described above, a candidate character string is extracted based on the extracted boundary image c (S120).

At this time, an image in which only a closed outline remains is extracted from the boundary image (c).

Preferably, the closed contour is divided into a nearly closed connected contour and a perfectly closed connected contour.

For example, when the boundary of a circle-shaped image is extracted, a circle-shaped image is obtained. The case where the outline is perfectly connected is called a perfectly closed outline, and the case where the outline is almost completely is called a nearly closed outline.

That is, in the present invention, an image in which the boundary is completely connected or almost connected is referred to as a perfectly closed outline or a nearly closed outline.

The closed outline refers to a closed outline connected to a boundary when referring to the boundary image (c).

In the boundary image c from which the closed contour is extracted, closed contours whose height (length) is smaller than the minimum height and larger than the maximum height are removed. At this time, the minimum height and the maximum height are experimentally determined the maximum height and the minimum height of the characters in the thousands of images.

In other words, the letters written on the slab are to measure the size of the letters written on the thousands of images to remove the image of the height determined to be higher or lower than the average size of the letters.

Thereafter, the closed contour is subjected to a process of giving a character score in the extracted image.

3 is an exemplary view showing an example of an image showing a character score and a closed outline in the management number extraction method for slab information recognition according to an embodiment of the present invention.

As shown in (d) of FIG. 3, if two points 310 and 320 meet each other by drawing a horizontal line A in the candidate character, a score of one point is obtained. However, if one or three or more horizontal lines meet, you won't score. In this way, all the horizontal lines of the candidate characters are run and the scores are summed.
Similarly, the same process is repeated for the vertical. Unlike the horizontal line (A), which runs from left to right, the points are summed using points that meet by drawing vertical lines from the top to the bottom. This score is combined with the score obtained using the horizontal line to create a textual score.
Then, by comparing the scores obtained by each candidate character is selected as a closed contour that is above the reference character score, and the rest is deleted to extract the image remaining only the closed contour.

delete

delete

That is, since the letters are densely arranged in a certain area, they are extracted on the premise that if the character score is high, they are likely to be letters.

FIG. 3 (e) is an image of the boundary image (c) processed by the method of (d) to be deleted leaving a component having a high score over a predetermined score.

The criteria for low and high scores are as follows. In (d) of FIG. 3, 20% of the height H and 20% of the width W are both used as the reference character scores to determine the closed contour.

It is considered that the components having a score higher than a certain score are likely to be letters.

Therefore, referring to FIG. 1, a closed outline set composed of left to right directions is found by using only one closed outline set as those with overlapping positions of closed outlines in the extracted image, leaving only closed outlines having a score higher than a certain score. .

Thereafter, a vertical position that may include all of the closed outline sets is determined (S130).

In other words, the characteristics of the character string are taken into consideration that they are configured in a direction from left to right with a similar height.

The positions above and below the determined string consist of the real string and the candidate string.

As a result of summing up the scores of the closed contours included in each of the upper and lower positions, it is determined that the number of strings included in the image is the real strings in the order of having the highest score.

4 is a picture showing a real string and a candidate string according to an embodiment of the present invention.

Referring to FIG. 4, first, candidate strings are generated through a process of tying closed contours at similar heights with respect to the extracted image, leaving only the closed contours having a score higher than the predetermined score.

The average X coordinate (horizontal axis) between the closed contours in the string is obtained and the closed outline that is farthest from the X coordinate is erased.

Then, create a string again with the remaining closed outlines, and calculate the ratio of the height and width (width / height) of the recreated string to determine the candidate string as a value between 5 and 8.

This is because the characters displayed on the slab have a ratio of height and width of 5-8.

1, the left and right positions of the character string are determined using the width and thickness of the character obtained through the slope of the character boundary in the image (S140).

Here, the step (S140) of determining the left and right positions of the character string may include first determining an approximate left and right position (S141), selecting a width occupied by each character (S142), and selecting a thickness of each character ( S143). In the process of determining the approximate left and right positions, Equation 1 below is used.

AR _max : ratio of height and width of string (width / height)

H: height of string, b: coordinates indicating position

The process of selecting a width occupied by the character from the selected candidate string through the above process uses a slope and a direction. As illustrated in FIG. 3E, the boundary image of the closed contour may be synthesized as illustrated in FIG. 5F.

5 is an exemplary diagram illustrating a process of obtaining a width and a thickness of a character using a tilt and a direction in a boundary image according to an exemplary embodiment of the present invention.

If the number 511 of FIG. 5 is one representative image, the inclination and the direction of the representative image are obtained.

Equation 2 below is an equation for calculating the slope.

The slope is calculated using Equation 2 above.

5 (g) shows a boundary image of FIG. 511.

As shown in (g) of FIG. 5, when the coordinates indicate the position of the surrounding white color based on the central white color, l ₁ = -1, k ₁ = -1, l ₂ = 0, k ₂ = 1 And then the slope can be calculated. The direction according to the slope selects a direction from the light region to the dark region using brightness information in the contrast image with respect to the regions shown in R1 520 and R2 530 of FIG. 5 in the contrast image.

When the inclination and the direction obtained through the above process are applied to the closed contour as shown in (h) of FIG. 5, a straight line obtained from the part shown in 514 of FIG. 5 to obtain the inclination and the direction is shown in 515 of FIG. 5. You will meet at the part.

On the contrary, as shown in FIG. 5 (i), the unnecessary portion 516 of the closed outline is eliminated to determine the width of the character.

In the process of selecting the thickness of the character, as shown in (k) of FIG. 6, the slope and the direction 520 are obtained at an arbitrary point as shown in FIG. The slope and direction 521 are again obtained with respect to the meeting point 518.

When the magnitude of the angle formed by the two slopes is smaller than 90 °, the distance between the two points is calculated and obtained.

At this time, the average of the distance between two points for all closed contours in the candidate string is the thickness of the character.

Then, using this thickness, as shown in 519 of FIG. 5, an area is secured to be the same thickness for all characters in the candidate character string, and the space shown in (a) of FIG. Enter the data value from the color image.

Finally, referring to FIG. 1, the character string is divided into characters in the image (S150).

6 is an exemplary view illustrating a process of finding a real string from a candidate string according to an embodiment of the present invention.

Referring to FIG. 6, in the step S150 of splitting the character string, it is assumed that the spacing between the characters is constant as shown in FIG. 6 (k). As shown in FIG. After drawing the dividing line at the same interval for the character string, the sum of the averages is calculated using the pixel values between the dividing lines, and the sum of the variances is calculated using the pixel values on the dividing line, and the sum of the averages and the The candidate string m with the largest divided value of the sum of variances is selected.

Also, for more accurate division of characters, as shown in FIG. 6 (n), each red line is moved within a certain range to move to the position where the deviation is the smallest, and finally the real character strings 611 and 612 are determined. .

Finally, as shown in 613 of FIG. 6, the position of the empty space is determined, and the position and the position of S ₇ shown in FIG. 6 (k) are compared to determine whether the string is flipped or not.

That is, referring to FIG. 6 (k), (k) shows an example of a management number of a certain form printed on the slab.

S ₇ of the control number is a blank space.

If S ₇ is filled with a certain character (for example, '3'), it means that the string is reversed, rotating the phase of the string 180 degrees to extract the original string.

As described above, although the present invention has been described with reference to the limited embodiments and the drawings, the present invention is not limited to the above embodiments, and those skilled in the art to which the present invention pertains various modifications and variations from such descriptions. This is possible.

Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined not only by the claims below but also by the equivalents of the claims.

1 is a flow chart illustrating a management number extraction method for slab information recognition according to an embodiment of the present invention.

Figure 2 is a photograph showing an example of the step of extracting the boundary image of the management number extraction method for slab information recognition according to an embodiment of the present invention.

3 is an exemplary view showing an example of an image showing a character score and a closed outline in the management number extraction method for slab information recognition according to an embodiment of the present invention.

4 is a photograph showing a real string and a candidate string according to an embodiment of the present invention.

5 is an exemplary view illustrating a process of obtaining a width and a thickness of a character using a slope and a direction in a boundary image according to an embodiment of the present invention.

6 is an exemplary view illustrating a process of finding a real string from a candidate string according to an embodiment of the present invention.

Claims (13)

In the management number extraction method for slab information recognition, A boundary image extraction step of finding a boundary image of the color image; A candidate string extracting step of extracting a candidate string based on the boundary image extracted in the boundary image extracting step; Determining an upper and lower position of a string using the closed outlines obtained from the candidate string extracted in the candidate string extracting step using not only the height of the character string but also the character scores that measure the similarity with the character; A left and right positioning step of determining a left and right position of a character string by using a width and a thickness of a character obtained through the slope at a character boundary in an image in which the up and down position of the character string is determined by the up and down position of the character string; And a character string dividing step of dividing a character string into characters in an image in which the left and right positions of the character strings having undergone the left and right positioning of the character strings are determined. The method of claim 1, wherein the boundary image is Management for recognizing slab information, wherein the color image, which is an original image including a control number, is detected by detecting a boundary of the color image by applying a weighted median filter based anisotropic diffusion (WMFAD) process. Number extraction method. The method of claim 1, wherein the boundary image extraction step An anisotropic diffusion filter applying process for applying an anisotropic diffusion filter to the color image, and And a boundary detection process for detecting a boundary from the image which has been subjected to the application process. The method of claim 2 or 3, wherein the anisotropic diffusion filter A control number extraction method for slab information recognition, characterized in that it plays a role of reducing image noise and increasing homogeneity of color of characters with filters based on anisotropic diffusion. The method of claim 1, wherein the extracting of the candidate character string comprises: And removing the closed contours having a height (length) of the candidate strings obtained from the extracted boundary image, which are lower than a predetermined minimum height and higher than a predetermined maximum height. The method of claim 1, wherein the closed contour is If there are two points that meet each other by drawing a horizontal line and a vertical line at one point of the candidate character, the candidate characters whose total sum of the character scores is greater than or equal to the reference character score are closed using the character score obtained through the method of determining one point. Control number extraction method for slab information recognition, characterized in that determined by the outline. The method of claim 6, wherein the reference character score is And 20% of the height and 20% of the width of the extracted character strings as the reference character scores to serve as a criterion for discriminating the closed contour. According to claim 1, wherein the step of determining the vertical position of the character string Determining an up-and-down position to include all of the closed contour sets by using a set of closed contours that overlaps the upper and lower positions of the closed contours in the extracted image, leaving only the closed contours connected by the boundary having a score of at least the standard textual score. Control number extraction method for recognizing slab information. The method of claim 8, wherein the set of closed contours A set of candidate character strings, the sum of the character scores of the closed contours included in each of the upper and lower positions, the management number extraction method for slab information recognition, characterized in that arranged in the order having a high character score. The method of claim 1, wherein the right and left positioning of the candidate character strings is performed. First, the process of determining the left and right positions using the ratio of the height and width of the candidate string, determining the width of each character through the inclination of each character boundary, and selecting the thickness of each character Management number extraction method for recognizing slab information, characterized in that configured. delete The method of claim 1, wherein the string division step After drawing the dividing line at the same interval for all candidate character strings, calculating the sum of the averages using the pixel values between the dividing lines, and calculating the sum of the variances using the pixel values on the dividing line, the sum of the averages. And selecting a candidate string having the largest dividing value of the sum of the variances. The method of claim 12, wherein the string division step is After moving each line within a certain range to move to the position where the deviation is the smallest, and finally determine the true character string, determine the position of the empty space and compare the area and position of the empty space with the character string to determine whether the string is upside down. Method for extracting a management number for recognizing slab information, further comprising determining.

Images