KR101048399B1 - Character detection method and apparatus - Google Patents

Abstract

In order to achieve the above object, the character detection method according to the present invention is a method of detecting a region including a character in an image, the method comprising: converting an input image into a gray image, and a first character candidate in the gray image Checking a region, applying a tensor voting to the gray image to identify a second character candidate region, and identifying a region where the identified first and second character candidate regions coincide with each other. Detecting a region to extract a character.

Image, character, area, recognition, tensor, voting, boundary, detection, object

Description

Character detection method and apparatus {METHOD FOR TEXT DETECTION AND APPARATUS FOR THE SAME}

The present invention relates to an image processing method and apparatus, and more particularly, to a method and apparatus for detecting a character contained in an image.

Recently, a huge amount of video data is generated through a digital camera or a mobile phone camera. Understanding characters in natural images can be used as a very important purpose between assistive devices for the visually impaired and interpreter systems for foreigners. Detection of the text area included in the image is the first step for text recognition. Since natural images can be photographed under various conditions, various studies have been conducted to detect only desired character regions. Backgrounds in a natural image can be very complicated, and the character region existing therein also has various types of formats (eg, fonts), sizes, and orientations. The image contains various elements of uneven lighting, reflections, shadows, and highlight marks.

On the other hand, the character detection method can be largely divided into a texture-based method and a region-based method. Texture-based methods may require high computational complexity by derivatives or frequency-based approaches, and are relatively good at detecting small characters. However, the frequency characteristics may appear differently depending on the font type of the character, and there is a problem that is difficult to adapt to such a change. Region-based methods use spatial cohesion such as edge information or characterization of connected elements in character strokes. The edge-based method detects edges from character outlines, merges them, and then filters some non-literal regions using some heuristic elements. Such a region-based method has a problem in that the rate of misrecognition of characters is high when there are many parts having similar characteristics of characters in an image.

SUMMARY OF THE INVENTION The present invention has been made in view of the above-described point, and an object thereof is to propose a character detection method and apparatus capable of increasing the precision of character detection and increasing the detection rate of characters under conditions in which the size and direction of the characters are not constant. .

According to an aspect of the present invention, there is provided a method of detecting a character including a character in an image, the method comprising: converting an input image into a gray image, and converting the input image into a gray image. A process of identifying a text candidate area, a process of identifying a second text candidate area by applying tensor voting to the gray image, and identifying a region where the identified first and second text candidate areas coincide with each other, And extracting a character by detecting an area in which it is included.

According to another aspect of the present invention, there is provided a character detecting apparatus. An apparatus for detecting an area including a character in an image includes: an image converting unit converting an input image into a gray image, and a gray image output from the image converting unit; A first character candidate region checking unit for checking a first character candidate region by applying an edge area detection algorithm to the second character candidate, and a second character candidate by applying tensor voting to the gray image output from the image converting unit A second character candidate area checking unit for checking an area, a character detection unit detecting the first and second character candidate areas, detecting an area containing a character, and detecting a character included in the area including the character; It includes.

According to the character detection method and apparatus of the present invention, it is possible to reduce the false recognition rate of character detection and to ensure the detection of robust characters under conditions in which the size and direction of the characters are not constant.

In addition, by applying tensor voting to character detection, information necessary for position search, noise reduction, and detection rate of the character may be generated, and by using the same, the accuracy of character detection may be increased and the detection rate of the character may be increased.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the following description, specific matters such as specific elements are shown, which are provided to help a more general understanding of the present invention. It is self-evident to those of ordinary knowledge in Esau.

The following terms are defined in consideration of the functions of the present invention, and may be changed according to the intentions or customs of the user, the operator, and the like. Therefore, the definition should be based on the contents throughout this specification.

1 is a block diagram showing a schematic configuration of an apparatus for detecting a text area according to an embodiment of the present invention. Referring to FIG. 1, an apparatus for detecting a text area according to an embodiment of the present invention includes an image converter 110 for converting a color image expressed in RGB into a gray level using an RGB average value, and the gray image. A first character candidate area checking unit 120 for identifying a first character candidate area from objects within the second character candidate area checking unit 130 for identifying a second character candidate area from objects in the gray image; Equipped. In addition, the character area detecting apparatus includes a character detecting unit 140 which checks the first and second character candidate areas output from the first and second character candidate area checking units 120 and 130, respectively, and finally detects the text area. .

The first text candidate area checker 120 receives a gray image from the image converter 110 and applies the gray image to an edge detection algorithm. In general, the characters included in the image exist within a range of a predetermined size due to the characteristics of the shape. Accordingly, the first text candidate area checking unit 120 sets an area expected to be a text area among the boundary areas detected through the boundary area detection algorithm, as the first text candidate area in consideration of general character characteristics. For example, the first character candidate region checking unit 120 may include a boundary region detection unit 121 that processes a boundary region detection algorithm for the gray image, and a first character candidate region for the boundary regions. One character candidate area detection unit 123 may be provided.

In addition, the boundary region detector 121 may extract a boundary region by applying a Canny Edge Algorithm to the gray image.

Meanwhile, the gray image 15 output from the image converter 110 is also provided to the second character candidate area checker 130. Accordingly, the second character candidate region checking unit 130 confirms the second character candidate region by applying tensor voting to the gray image 15. For example, the second character candidate region checking unit 13 may include a candidate object detector 131 for detecting a candidate object to be used as an input token of tensor voting from a gray image, and a tensor voting on the output values of the candidate object detector 131. A tensor bot processing unit 133 for processing the framework and a second character candidate area detection unit 135 for detecting the second character candidate area in consideration of the characteristics of the characters included in the image and the result of the tensor voting. Can be.

The point assumed to be the center point of the character area is preferably used as an input token of the tensor voting. In addition, the characters included in the image may be generally arranged on a straight line having a horizontal component, or may be arranged in the form of a smooth curve in the horizontal direction. Accordingly, the candidate object detector 131 performs a vertical component detection Sobel operation so that the characters included in the gray image can be divided into vertical components in consideration of the arrangement characteristics of the characters. It is preferable to detect and set the object of the boundary area as a candidate object. That is, when f (x, y) is defined as the brightness of the pixel at the position represented by (x, y), the horizontal or vertical component may be defined by Equation 1 below. Correspondingly, the candidate object detector 131 may detect a vertical component by using a Sobel edge filter and an operation of Equation 3 as shown in Equation 2 below, and detect the candidate object.

The candidate object detector 131 may remove candidate objects outside a predetermined size range among the detected candidate objects in consideration of characteristics of shapes and / or sizes of characters included in an image.

In addition, the candidate object detector 131 may detect a center point of the candidate object detected as described above and generate a list for each center point in order to facilitate tensor voting of the candidate object.

Although, in one embodiment of the present invention, the characters included in the image is generally arranged on a straight line having a horizontal component, or may be arranged in a gentle curved form in the horizontal direction, the present invention is limited to this It is not. For example, the characters included in the image may be arranged on a straight line having a vertical component, or may be arranged in the form of a gentle curve in the vertical direction. In this case, the candidate object detector 131 may be arranged on the gray image 15. It is also possible to perform a horizontal component detection Sobel operation so that the included characters can be divided into horizontal components.

The tensor bot processing unit 133 applies center points of the candidate object as an input token of tensor voting, and processes a tensor voting framework for the input token. Tensorboating framework is an integrated framework developed by the Computer Vision group of the University of Southern California (USC), and its detailed description is omitted.

The tensor voting processor 133 processes the tensor voting framework for the input token, and outputs a saliency of the center points of the input candidate object.

Furthermore, among the center points of candidate objects to which tensor voting is applied, the curve protrusion degree of the tensors corresponding to the center points of candidate objects corresponding to the character area has a relatively high value compared to other tokens. Therefore, in consideration of the degree of curve protrusion of the candidate objects corresponding to the character area, the second character candidate area detection unit 135 detects the second character candidate area from the candidate objects. That is, the second character candidate area detector 135 sets the candidate object as the second character candidate area when the curve protrusion degree of the tensors corresponding to the candidate objects has a value that is relatively larger than a predetermined threshold value. When the curve protrusion of the tensors corresponding to the candidate objects has a value equal to or less than a predetermined threshold value, the second character candidate area is detected by removing the candidate object.

Further, the character detection unit 140 determines a region where the first character candidate region and the second character candidate region coincide as a character region, and finally detects the character by applying the character region to a character recognition algorithm.

Preferably, the character detecting unit 140 performs an expansion operation on the center points of the candidate objects arranged in the horizontal direction using the center points of the candidate object to remove the non-character area included in the second character candidate area. By doing so, neighboring candidate objects can be connected. Accordingly, the character detector 140 may recognize only candidate objects to which candidate objects are connected to each other as the second character candidate area, and remove other second character candidate areas.

2A to 2D, 3A to 3D, 4A to 4B, and 5A to 5C are views illustrating an image generated by the character area detection apparatus according to an exemplary embodiment of the present invention. Hereinafter, the operation of the text area detection apparatus will be described with reference to an exemplary view of an image generated by the text area detection apparatus.

First, FIG. 2A illustrates the original image 11, FIG. 2B illustrates the gray image 15, FIG. 2C illustrates the boundary region image 21 including the boundary region, and FIG. 2D illustrates the first character. A first text candidate image 25 including a candidate region is illustrated.

The image converter 110 receives the original image 11, performs gray conversion on the original image 11, and outputs a gray image 15. The gray image 15 is provided to the boundary region detector 121, and the boundary region detector 121 applies a boundary region detection algorithm such as Canny Edge Detection Algorithm to the gray image 15. The boundary area is extracted and the boundary area image 21 is output.

3A illustrates a gray image 15 output from the image converter 110, and FIG. 3B illustrates an image of the boundary region 31 of the candidate object detected by the candidate object detector 131. 3c illustrates a candidate object image 33 that filters an object that is not expected to be a text area from the candidate object boundary region image 31 in consideration of characteristics of characters included in the image, and FIG. 3d illustrates a candidate object image. An image (ie, candidate object center point image) 35 from which the center point is extracted is illustrated. 4A illustrates a tensor voting image 41 in which the candidate object center point image 35 is tensored, and FIG. 4B shows a second character candidate region in which a second character candidate region is detected from the tensor voting image 41. An image 45 is illustrated.

The candidate object detector 131 receives the gray image 15 illustrated in FIG. 3A, performs a vertical component detection Sobel operation to generate the boundary region image 31, and then determines the size of the boundary region in advance. By removing the area out of the range, the candidate object image 33 filtering the object that is not expected to be the text area is formed. The candidate object detector 131 checks the center point of each candidate object included in the candidate object image 33 to generate the candidate object center point image 35, and labels the center points of each candidate object. Next, a center point list of each candidate object is generated. Next, the center point list is provided to the tensorboating processor 133, and the tensorboating processor 133 performs the tensorboating process by applying the center points included in the center point list as an input token. As a result of the tensor voting process, the tensor bot processing unit 133 generates a curve saliency of the tensors corresponding to the center points and a normal vector of the text area by the curved normal vectors of the center points. The curve saliency and the normal vector of the character region are reflected in the candidate object image 33 to output the tensor boating image 41 illustrated in FIG. 4A. In the tensor boating image 41, the center point of each candidate object is represented by a red point, the curve saliency is represented by the length of the green line, and the normal vector of the character area is represented by the arrow of the green line. As indicated in the tensor voting image 41, the curve saliency is represented by a large value in the region where the character is located, and their normal vector appears similar to the normal vector of the character region. Accordingly, in consideration of the curve saliency of the region where the character is located and the characteristics of the normal vector, the second character candidate region detector 135 is a candidate having a curve protrusion of a relatively smaller value than the predetermined threshold value. The second character candidate area image 45 is generated by removing the object and removing the candidate object having the normal vector formed in the horizontal direction.

FIG. 5A illustrates an expanded image 51 including candidate objects of a second text candidate area connected through an expansion operation, and FIG. 5B illustrates a text area extraction image in which only the first and second text candidate areas are extracted. 53), FIG. 5C illustrates a character image 55 including a character detected from the character region extracted image 53.

In order to remove the non-character area from the second character candidate area image 45, the character detection unit 140 performs expansion operation on the center points of the candidate objects arranged in the horizontal direction using the center points of the candidate object. Connect the objects. Accordingly, the character detector 140 generates an expanded image 51 including candidate objects of the second character candidate region. The character detecting unit 140 recognizes only candidate objects to which candidate objects are connected to each other as the second character candidate area, and removes other second character candidate areas. In addition, the character detection unit 140 identifies a region where the first character candidate region and the second character candidate region coincide with each other, and extracts only the region where the first and second character candidate regions overlap with each other. After generating the, the character area is finally detected by applying to the character area character recognition algorithm included in the extracted text area image 53, and generates a character image (55).

6 is a flowchart illustrating a sequence of a method for detecting a text area according to an embodiment of the present invention.

Referring to FIG. 6, in the method for detecting a text area according to an embodiment of the present invention, an original image (eg, 11 of FIG. 2A) is input (210), and the input original image 11 is a gray image (eg, 2B, step 15).

In operation 230, a boundary character algorithm is applied to the gray image (15 of FIG. 2B) to identify a first character candidate region, and a first character candidate image (25 of FIG. 2D) is generated.

In operation 240, tensor voting is applied to the gray image (15 of FIG. 2B) to identify a second character candidate region, and a second character candidate image (45 of FIG. 4B) is generated. Furthermore, in order to apply tensor voting to the gray image (15 of FIG. 2B), detection of an input token to be input to tensor voting is required. Accordingly, in step 240, a candidate object to be used as an input token is detected, each center point of the candidate objects is detected, and the center points are used as the input token.

Finally, in step 250, a region where the first character candidate region and the second character candidate region coincide is determined as a character region, and the character region is finally applied to a character recognition algorithm to detect a character. Furthermore, in operation 250, the neighboring candidates are expanded by performing expansion operations on the center points of the candidate objects arranged in the horizontal direction by using the center points of the candidate objects to remove the non-character areas included in the second text candidate area. You can connect objects. Accordingly, only candidate objects to which candidate objects are connected to each other may be recognized as the second character candidate area, and other second character candidate areas may be removed.

In detail, in operation 250, expansion operations are performed on the center points of the candidate objects arranged in the horizontal direction by using the center points of the candidate objects included in the second text candidate area image 45. Accordingly, candidate objects included in the second text candidate area image 45 are connected, and an expanded image (51 of FIG. 5A) including the connected candidate objects is generated. After checking the region where the first character candidate region and the second character candidate region coincide with each other, a character region extracted image (53 of FIG. 5B) extracted only the region where the first and second character candidate regions overlap is generated. Then, the character is finally detected by applying the character region character recognition algorithm included in the character region extracted image 53 to generate a character image (55 of FIG. 5C).

7 is a flowchart illustrating an operation sequence of step 230 of FIG. 6 in detail.

First, in step 231, the gray image (15 of FIG. 2B) is input, and the gray image 15 is applied to an edge detection algorithm. Accordingly, in step 231, the boundary area of the objects included in the gray image 15 is checked, and a boundary area image (21 of FIG. 2C) including the boundary area is generated. For example, the edge detection algorithm may be a Canny Edge Algorithm.

In operation 233, each region formed by the boundary regions detected by the boundary region detection algorithm is set as the first character candidate region.

On the other hand, generally, the characters included in the image exist within a range of a predetermined size due to the characteristics of the shape. Accordingly, in step 235, the non-character area (ie, non-character area) that is not expected to be non-letter is removed from the first character candidate areas in consideration of general characteristics of the character, and the final character candidate area is removed. And generate a first character candidate image (25 in FIG. 2D).

FIG. 8 is a flowchart illustrating an operation sequence of step 240 of FIG. 6 in detail.

First, in step 241, a candidate object for detecting the second character candidate region is detected from the gray image 15. Characters included in the image are generally arranged on a straight line having a horizontal component, or may be arranged in the form of a smooth curve in the horizontal direction. Accordingly, in consideration of the arrangement characteristics of the characters, a vertical component detection Sobel operation is performed to detect a boundary region so that the characters included in the gray image 15 can be divided into vertical components. It is preferable to set the object of the region as a candidate object. That is, when f (x, y) is defined as the brightness of the pixel at the position represented by (x, y), the horizontal or vertical component may be defined by Equation 1 above. Correspondingly, a vertical component is detected and a candidate object is detected by using a Sobel edge filter such as Equation 2 and an operation of Equation 3. That is, after performing a vertical component detection Sobel operation on the gray image 15 illustrated in FIG. 3A to generate a boundary region image (31 in FIG. 3B), an area in which the size of the boundary region is out of a predetermined range is obtained. The candidate object is detected by filtering an object that is not expected to be a text area, and a candidate object image (33 of FIG. 3C) including the detected candidate object is formed.

Further, in step 242, in order to facilitate the application of tensor voting, the center point of each candidate object included in the candidate object image 33 is checked, and the candidate object center point image 35 is generated. In operation 243, the center points of the candidate objects are generated by labeling the center points of the candidate objects.

In order to perform tensor voting on the candidate object, in step 244, the labeled candidate objects are sequentially selected according to a predetermined rule. For example, the predetermined rule may be to select the center point of the candidate object in the order of the number assigned to the center point of the candidate object.

In step 245, a tensor voting process is applied to the center point of the selected candidate object by applying the center point of the candidate object selected in step 244, and as a result, curve saliency of the tensors corresponding to the center points. ), And the normal vector of the text area by the curved normal vectors of the center points. In operation 245, the curve saliency and the normal vector of the character region are reflected in the candidate object image (33 of FIG. 3A) to output the tensorboating image 41 illustrated in FIG. 4A.

In the tensor boating image 41, the center point of each candidate object is represented by a red point, the curve saliency is represented by the length of the green line, and the normal vector of the character area is represented by the arrow of the green line. As indicated in the tensor voting image 41, the curve saliency is represented by a large value in the region where the character is located, and their normal vector appears similar to the normal vector of the character region. Therefore, in consideration of the curve saliency of the region where the character is located and the characteristics of the normal vector, in step 246, it is checked whether the protrusion of the candidate object has a value that is relatively greater than a predetermined threshold. . If the protrusion of the candidate object has a value that is relatively larger than a predetermined threshold value, the candidate object is likely to be a text area, and therefore, in step 247, the candidate object is set as the second text candidate area. do. On the other hand, if the protrusion of the candidate object has a value equal to or smaller than a predetermined threshold value, since the candidate object is unlikely to be a character area, the process proceeds to step 248 to form the candidate object as the second character candidate area. Remove from As a result, a second character candidate area image (FIG. 5C 45) is generated from the tensor voting image 41.

Further, in step 249, it is checked whether all the tensor voting for the labeled candidate objects have been processed, and if there is a candidate object that has not been tensored in the list of the candidate objects labeled, steps 244 to 248 are performed. Tensor voting of the candidate object is repeatedly performed. On the other hand, if all tensor voting for the candidate object has been processed, the tensor voting process of the candidate object is terminated.

Further, in order to evaluate the performance of the above-described character detection method, the performance of detecting the characters included in the image was tested. In the experimental example according to the exemplary embodiment of the present invention, twelve label images having a resolution of 640 × 480 shown in FIG. 9 were used.

And, Comparative Example 1 according to the edge-based method and Comparative Example 2 according to the connection element-based method were compared with the embodiment according to the present invention. In the experimental example of the present invention, the number (S) of characters included in an image, the number (T) of character regions successfully detected, and the number (F) of misdetected regions were measured. Recall Rate was calculated to compare the character detection performance. The comparison results are shown in the table below.

Comparative Example 1 Comparative Example 2 Example S 1161 1161 1161 T 1016 908 880 F 77 391 450 Precision 92.96% 69.9% 66.17% Detection rate 87.51% 66.17% 75.80%

As disclosed in Table 1, the method according to the embodiment of the present invention shows relatively good precision and recall rate compared to the method according to Comparative Examples 1 and 2.

The character detection method according to the present invention can be embodied as computer readable codes on a computer readable recording medium. Computer-readable recording media include all kinds of recording devices that store data that can be read by a computer system. Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical disk, and the like, and may also include those implemented in the form of carrier waves (eg, transmission over the Internet). do. The computer-readable recording medium may also be distributed over a networked computer system and stored and executed in computer readable code in a distributed manner.

Although the present invention has been described above by means of limited embodiments and drawings, the present invention is not limited thereto, and various modifications and changes may be made by those skilled in the art to which the present invention pertains.

1 is a block diagram showing a schematic configuration of a character area detection apparatus according to an embodiment of the present invention;

2A is an exemplary diagram of an original image illustrated in an apparatus and method for detecting a text area according to an embodiment of the present invention;

2B is an exemplary view of a gray image illustrated in an apparatus and method for detecting a character area according to an embodiment of the present invention;

2C is an exemplary diagram of a boundary region image illustrated in an apparatus and method for detecting a character region according to an embodiment of the present invention;

2D is an exemplary diagram of a first character candidate image illustrated in an apparatus and method for detecting a character region according to an embodiment of the present invention;

3A is an exemplary diagram of a gray image illustrated in an apparatus and method for detecting a text area according to an embodiment of the present invention;

3B is an exemplary diagram of a boundary region image of a candidate object illustrated in an apparatus and method for detecting a character region according to an embodiment of the present invention;

3C is an exemplary view of a candidate object image illustrated in an apparatus and method for detecting a character region according to an embodiment of the present invention;

3D is an exemplary diagram of a candidate object center point image illustrated in an apparatus and method for detecting a character region according to an embodiment of the present invention;

4A is an exemplary diagram of a tensor boating image illustrated in an apparatus and method for detecting a text area according to an embodiment of the present invention;

4B is an exemplary view of a second character candidate region image illustrated in an apparatus and method for detecting a character region according to an embodiment of the present invention;

5A is an exemplary diagram of an expanded image illustrated in an apparatus and method for detecting a text area according to an embodiment of the present invention;

5B is an exemplary view of a text region extracted image illustrated in an apparatus and a method for detecting a text region according to an embodiment of the present invention;

5C is an exemplary diagram of a character image illustrated in an apparatus and method for detecting a character region according to an embodiment of the present invention;

6 is a flowchart illustrating a sequence of a method for detecting a text area according to an embodiment of the present invention.

7 is a flowchart illustrating an operation sequence of step 230 of FIG. 6 in detail;

8 is a flowchart illustrating an operation sequence of step 240 of FIG. 6 in detail;

9 illustrates twelve label images used in an experimental example according to an embodiment of the present invention.

Claims (17)

In the method for detecting a region containing a character in the image, Converting the input image into a gray image; Identifying a first text candidate area in the gray image; Identifying a second character candidate region by applying tensor voting to the gray image; Identifying a region in which the identified first and second character candidate regions coincide, detecting a region containing a character, and extracting a character; The checking of the second text candidate area may include: Detecting a candidate object expected to be a character in the gray image; Applying the detected candidate object to tensor voting, And extracting the second text candidate area in consideration of the result of applying the tensor voting. The process of claim 1, wherein the checking of the first text candidate area is performed. Extracting an edge region of at least one object included in the image from the gray image; And extracting at least one first text candidate area from the extracted boundary areas. The character detection method of claim 2, wherein the extraction of the boundary region uses a Canny Edge Algorithm. The method of claim 2, wherein the checking of the first text candidate area comprises: And detecting the non-character area included in the first text candidate area and removing the detected non-character area in consideration of the characteristics of the character. delete The method of claim 1, wherein the checking of the second text candidate area comprises: Detecting a center point of the candidate object; Applying a center point of the detected candidate object to tensor voting, And extracting the second text candidate area in consideration of the result of applying the tensor voting. The method of claim 6, wherein the extracting of the second text candidate area comprises: And extracting, as the second character candidate area, a corresponding candidate object having a salient greater than a predetermined value as a result of applying the tensor voting. The character detection method of claim 1, wherein the candidate object expected as the character is identified by performing a Sobel operation of vertical and / or horizontal components on the gray image. The character detection method of claim 1, wherein the candidate object expected to be the character is detected in consideration of structural characteristics of the character. The method of claim 1, wherein the detecting of the area containing the letter comprises: Detecting a region in which the first and second character candidate regions overlap; Performing a dilation operation in a horizontal and / or vertical direction on the second character candidate area; And extracting second character candidate regions connected through the expansion operation into an area including the character. In the device for detecting the area containing the characters in the image, An image converter for converting the input image into a gray image; A first text candidate area checking unit checking a first text candidate area by applying an edge area detection algorithm to the gray image output from the image converting unit; A second text candidate area checking unit which checks a second text candidate area by applying tensor voting to the gray image output from the image converting unit; A character detector configured to check the first and second character candidate regions, detect a region containing a character, and detect a character included in the region including the character; The second character candidate area checking unit, A candidate object detector for detecting candidate objects expected from the gray image; A tensor boating processor for applying the detected candidate object to tensor boating; And a second text candidate area detection unit extracting the second text candidate area in consideration of the result of applying the tensor voting. The method of claim 11, wherein the first character candidate area confirmation unit, A boundary region detector for detecting a boundary region by applying a Canny Edge Detection Algorithm to the gray image; And a first character candidate area detection unit detecting a first character candidate area from at least one detected boundary area through the boundary area detection unit. delete The method of claim 11, wherein the candidate object detector detects a center point of the candidate object, And the tensor voting processor applies the center point of the detected candidate object to tensor voting. The method of claim 11, wherein the second character candidate area detection unit, And extracting, as the second character candidate region, a candidate object having a saliency having a result value of applying the tensor voting relatively larger than a predetermined value. The character detection apparatus of claim 14, wherein the candidate object detector identifies a candidate object expected as the character by performing Sobel operations on vertical and / or horizontal components on the gray image. The method of claim 11, wherein the character detection unit, Detecting a region where the first and second character candidate areas overlap, performing horizontal and / or vertical dilation operations on the second character candidate area, and connecting the second character through the expansion operation. And extracting candidate regions into the region including the character.

Images