JP2013016168A - Method and device for positioning text area in image - Google Patents

Abstract

The present invention discloses a method and apparatus for positioning a text region in an image.
A method for positioning a text region in an image according to the present invention includes estimating a size of text that can exist around each pixel in an input image, and based on the size of the text and the region difference. Extracting a candidate stroke area from the input image; identifying a true stroke area from the candidate stroke area; and merging the true stroke area to form a text area.
[Selection] Figure 2

Description

  The present invention relates generally to image processing, and in particular to a method and apparatus for positioning a text region in an image.

  In various applications that perform indexing, search, classification, and the like on an image, it is necessary to extract information about the image content from the image. An image usually has text information. This text information has a relatively high correlation with the content of the image. Therefore, acquisition of such text information is important in image applications. Usually, first, a text area in an image is positioned, and then text information is obtained by performing extraction and optical character recognition (OCR) processing on the image block where the text area is located. The images can be divided into images based on natural scenes and images based on artificially added text. Since the artificially added text image has human involvement, it is relatively easy to position the text area from within it. Since an image based on a natural scene has an image as a unit, it is difficult to distinguish an image in a text area from an image in a non-text area. For this reason, it is difficult to position the text area from within. The present invention can process relatively complex images including images from natural scenes, focusing on positioning text regions in the image.

  The following presents a simplified summary of the invention in order to provide a basic understanding of some aspects of the invention. It should be understood that this overview is not an exhaustive overview regarding the present invention. It is not intended to identify key portions of the invention, nor is it intended to limit the scope of the invention. Its purpose is simply to present some concepts in order to simplify the story and to precede the detailed description that follows.

  An object of the present invention is to provide a method and an apparatus capable of positioning a text region in an image with respect to the above-mentioned problems of the prior art. This technical measure can locate text regions quickly and accurately from an image and applies to any kind of image.

  In order to achieve the above object, according to one aspect of the present invention, a method for positioning a text region in an image is provided. The method includes estimating a size of text that can exist around each pixel in the input image, and extracting candidate stroke regions from the input image based on the size of the text and the degree of difference between the regions. And specifying a true stroke area from the candidate stroke areas and merging the true stroke areas to form a text area.

  According to another aspect of the present invention, a text region positioning device in an image is provided. The apparatus is configured to estimate a candidate stroke area from the input image based on an estimation unit for estimating a size of text that can exist around each image in the input image and a difference between the size of the text and the area. An extraction unit for extraction, a specific unit for specifying a true stroke area from the candidate stroke areas, and a merge unit for merging true stroke units to form a text area are included.

  According to another aspect of the present invention, a storage medium is further provided. The storage medium has program code readable by a machine. When the program code is executed in the information processing apparatus, the program code causes the information processing apparatus to execute the method according to the present invention.

  According to another aspect of the present invention, a program is further provided. The program has a computer executable command. When the command is executed in the information processing apparatus, the command causes the information processing apparatus to execute the method according to the present invention.

It is a figure which shows the example of the image by the natural scene which should be processed. 4 is a flowchart illustrating a method for positioning a text region in an image according to an embodiment of the present invention. It is a schematic diagram which shows the structure of an image pyramid. It is a flowchart which shows the detail of step S201 of FIG. It is a flowchart which shows the detail of step S202 of FIG. It is a flowchart which shows the detail of step S203 of FIG. It is a flowchart which shows the detail of step S204 of FIG. It is a flowchart which shows the detail of step S702 of FIG. 7A. It is a schematic diagram which shows the chain structure which connects all the stroke areas. It is a schematic diagram which shows the chain structure after line division. It is a schematic diagram which shows the chain structure after character division. It is a schematic diagram which shows the processing result of the positioning method of the text area in the image which concerns on one Example by this invention. It is a block diagram which shows the structure of the positioning device of the text area in the image which concerns on one Example by this invention. 1 is a block diagram schematically showing a computer for realizing a method and apparatus according to an embodiment of the present invention.

  Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the drawings. In the interest of clarity and brevity, not all features of a practical embodiment are described herein. However, it should be understood that in the process of developing any such practical example, the rules limited by the embodiments must be increased. For example, a constraint condition such as a constraint condition related to a system and a job is satisfied, and the above-described constraint condition may vary depending on the embodiment. It should be noted that although development work is very complex and time consuming, those skilled in the art who benefit from the present disclosure will also understand that such development work is merely a routine task.

  Here, what is further to be described is only the device structure and / or processing steps closely related to the technical measures of the present invention in the drawings, so as not to obscure the present invention by unnecessary details. However, other details not so much related to the present invention are omitted. It should also be pointed out that elements and features described in one drawing or embodiment of the invention may be combined with elements and features shown in one or more other drawings or examples. It is a good point.

  Hereinafter, a flow of a method for positioning a text area in an image according to an embodiment of the present invention will be described with reference to FIG.

  FIG. 1 shows one specific example of an image that can be processed by the present invention. In the image of the natural scene shown in FIG. 1, there is a background as a house and a foreground as a traffic sign. However, as described above, since the image has a pixel as a unit and there is no tag artificially added to the text area in the image, it is difficult to extract the text area accurately and quickly.

  As shown in FIG. 2, a method for positioning a text area in an image according to an embodiment of the present invention includes estimating a size of text that can exist around each pixel in an input image (S201), Extracting a candidate stroke area from the input image based on the size of the text and the difference between the areas (S202); identifying a true stroke area from the candidate stroke area (S203); Merging the true stroke area to form a text area (S204).

  Hereinafter, step S201 of FIG. 2 will be described in detail with reference to FIGS.

FIG. 3 shows the structure of the image pyramid. One pyramid image exists for each layer Ln (n ≧ 1). The pyramid image corresponding to the first hierarchy L1 is the original input image. FIG. 1 shows a specific example of an input image. Each layer Ln has a scaling factor sc _n for the first layer L1. For each layer Ln (n> 1), the input image is reduced at an equal ratio by the scaling coefficient sc _n to obtain a pyramid image of this layer. For example, the input image is 8 × 8 pixels in size. When the step width is 1/2, the scaling factor for the first layer L1 of the second layer is sc ₂ = 1/2, and when the nearest neighbor interpolation method is employed, the second of 4 × 4 pixel size A hierarchical pyramid image is obtained. By analogy sequentially, a plurality of scaled image pyramids can be obtained from the input image as shown in FIG. For the nth layer (n ≧ 1), if the step width is step, it is clear that the scaling factor of the nth layer is sc _n = step ⁿ⁻¹ . The scaling factor of each pixel in the input image is calculated by the certainty factor of the corresponding pixel (including the pixel itself) in the pyramid image of each layer and the scaling factor of the pyramid image of each layer. This will be described later with reference.

  FIG. 4 is a flowchart showing details of step S201 in FIG.

  First, as described above, a plurality of layers of pyramid images, that is, image pyramids (step S401) are generated based on the input image.

  After that, the probability that text exists around each pixel is estimated for the pyramid image of each layer (step S402). Specifically, the pyramid image of each layer is scanned with a scan window of a certain size, and the probability that text exists around each pixel of the pyramid image of each layer is acquired. In this embodiment, the size of the scan window of the pyramid image in each layer is fixed as the size of the original input image. For each pixel, a local texture in the scan window centered on the pixel, for example, a gradient direction histogram HO (Histograms of Oriented Gradient) G feature is calculated. The calculated HOG features can be input into a trained classifier, which can return the certainty (probability) that the text exists around the pixel.

  When training the classifier, it is already known whether there is a text region around one pixel. That is, it is already known whether one pixel is one of the pixels constituting the text area. For example, 1 indicates that a text area exists around the pixel, and 0 indicates that no text area exists around the pixel. HOG feature is calculated for each pixel in one image with a scan window of a certain size, and the certainty of whether there is text around the calculated HOG feature and the pixel represented by 1 or 0 To the classifier. A classifier subjected to a large amount of training can be applied to step S402.

  It should be noted that the local texture is not limited to the HOG feature described above. For example, wavelet features may be calculated. Wavelet transform is performed on the pixels in the scan window centered on one pixel, and the statistic of the wavelet coefficient obtained by the transform, for example, an average value and / or variance is used as the wavelet feature of the pixel. For the classifier, a classifier such as Waldboost or AdaBoost can be selected.

In step S403, first, for each pixel of the input image, the coordinates of the corresponding pixel in the pyramid image of each layer are calculated. For example, if the scaling factor sc ₂ is 2 to the pyramid image of the second hierarchy corresponding coordinates in the input image is (a, b) pixel coordinates in the pyramid image of the second hierarchy (2a, 2b) Corresponds to a certain pixel. That is, when the input image is multiplied by the scaling factor of the pyramid hierarchy corresponding to the abscissa and ordinate of the pixel, the abscissa and ordinate of the corresponding pixel in the pyramid image corresponding to the scaling factor are obtained. If the scaling factor is a non-integer value, the calculation result is rounded to an integer to obtain the abscissa and ordinate of the corresponding pixel. It is clear that for one pixel in the input image, the first level pyramid, ie its corresponding pixel in the input image, is itself. Accordingly, for each pixel in the input image, there is one corresponding pixel in each layer of the pyramid image.

Let P _i be a set composed of pixels in each layer of the pyramid image (including the first layer pyramid image, that is, the input image itself) to which the i-th pixel s _i in the input image corresponds. It is assumed that j belongs to P _i and is a pixel corresponding to s _i in a pyramid image of a certain hierarchy, and the corresponding scaling coefficient of the pyramid image is sc _j . Assume that the certainty that text exists around j calculated in step S402 is w _j . Probability that text exists around pixel s _i in the input image according to equations 1 and 2 below

及びスケーリング係数ｓｃ_ｉを算出する。

And a scaling coefficient sc _i is calculated.

ただし、

However,

は、それぞれ、確信度及びスケーリング係数の正規化因子を示す。

Denote the normalization factor of the certainty factor and the scaling factor, respectively.

In the above formulas 1 and 2, w _j is used as a weighting coefficient, and the information of the corresponding pixel in the pyramid of each layer is projected onto the original input image as the pyramid image of the first layer.

Based on the scaling factor sc _i and the size of the scan window, the size of text that can exist around the pixel s _i in the input image can be calculated. For example, indicating the size of the scanning window length length or width width of the scan window, length / sc _i or width / sc _i a circle and radius or length / sc _i and width / sc _i the length and, A rectangular area having a width can represent the size of text that can exist around the pixel s _i in the input image.

As a variation of the formula 2, replacing the sc _j in length / sc _j, the left side of the equation becomes length / sc _i. As can be seen from the above description, the physical significance of Expression 2 is that the certainty factor w _j of the pixel j corresponding to each pixel s _i in the input image in the pyramid of each layer and the size of the text corresponding to the pyramid layer. On the basis of the length length / sc _j , the text size length / sc _i that can exist around each pixel s _i in the input image is calculated.

As a candidate of the formula 1 and 2, among the pixels corresponding to the pixel s _i in the input image, and a confidence factor w _j and scaling factor sc _j of the highest pixel j confidence w _j, around the pixel s _i Probability that the text exists

とスケーリング係数ｓｃ_ｉとする。

And scaling factor sc _i .

  According to steps S401 to S403, it is possible to acquire the probability that text exists around each pixel and the size of text that can exist around each pixel in the original input image.

  Here, by adopting an image pyramid and performing scaling conversion, it is possible to detect characters of any size. Relatively, if the image is scanned with a constant size window without scaling changes, i.e. only the first hierarchy of the above example exists, there will still be text around each pixel. Probability can be obtained, and the size of text that can exist around each pixel is the size of the scan window. In this case, the present invention can still be realized, but since there is no scaling conversion, only characters with a constant size can be detected.

  It should be noted that the scaling conversion method is not limited to the above-described zooming of each layer of pyramid image scaling without changing the scan window. The input image may be scanned a plurality of times by using a scan window (that is, scaling conversion) whose size changes without zooming the input image. In this case, similarly, the probability that text exists around each pixel in the input image and the size of text that can exist around each pixel can be calculated based on the result of each scan.

  Hereinafter, step S202 in FIG. 2 will be described in detail with reference to FIG.

  FIG. 5 is a flowchart showing details of step S202 in FIG.

  As described above, in step S202, a candidate stroke area is extracted from the input image based on the size of the text and the area difference that can exist around each pixel calculated in step S201.

  Specifically, the similar regions in the input image are continuously merged by repeating the merge of regions (initially pixels) until a predetermined condition is satisfied. Thereby, a plurality of areas merged as candidate stroke areas are obtained. The standard for merging mainly considers the degree of difference within an area, the degree of difference between areas, and the size of text that can exist around the area.

  First, in step S501, the difference between regions and the difference within the region are calculated for adjacent regions in the input image. The degree of difference within a region is, for example, the largest color difference within the region. For the gradation image, the largest color difference in the area is the absolute value of the difference in gradation value between the pixel having the highest gradation value and the pixel having the lowest gradation value among all the pixels in the area. For example, for each color image indicated by (Y, Cr, Cb),

の値を算出することができ、領域内の各画素間の最大の

Value can be calculated and the maximum between each pixel in the region

の差の絶対値を領域内の最大色差とする。領域間の差異度は、階調画像について二つの領域の諧調の平均値の差の絶対値をとることができ、カラー画像について２つの領域の

The absolute value of the difference between is the maximum color difference in the region. The degree of difference between the areas can be an absolute value of the difference between the average values of the gradations of the two areas in the gradation image, and the difference between the two areas in the color image can be obtained.

の平均値の差の絶対値をとることができる。まず、隣接した領域は隣接した画素であり、隣接した画素の領域内の差異度は０であり、隣接した画素の領域間の差異度は、例えば隣接した画素の階調値の差の絶対値（階調画像について）、又は隣接した画素の

The absolute value of the difference between the average values can be taken. First, the adjacent area is an adjacent pixel, the difference degree between the adjacent pixel areas is 0, and the difference degree between the adjacent pixel areas is, for example, the absolute value of the difference between the gradation values of the adjacent pixels. (For gradation image) or of adjacent pixels

の差の絶対値又は隣接した画素の輝度の差の絶対値（カラー画像について）である。

Or the absolute value of the difference in luminance between adjacent pixels (for a color image).

In step S501, if the adjacent regions and _C 1, _{C 2,} a set of pixels adjacent to the pixel region _{C 2} of the pixel region _{C 1} and _{C 1,} region of the pixel region _{C 2} C the set consisting of adjacent pixels as _one pixel and C _2. As described above, except for the embodiment in which the difference in the region and the difference between the regions in the range of C ₁ and C ₂ are calculated, the above-described method is adopted in the range of only C ₁ and C ₂ , and C _1, the difference of the C ₂ in the area and to calculate the difference degree between the region may represent the difference of the differences of and regions within the region of C _1, C _2.

  Next, in step S502, based on the size of text that can exist around each pixel in the input image estimated in step S201, the size of text that can exist around the region is estimated and estimated. The degree of difference within each region is adjusted based on the size of text that can exist around the region.

The step states that it is not entirely accurate to determine whether to merge adjacent regions based solely on the following discoveries by the inventor, i.e., differences within and between regions. It is based. If it is known that the size of the text that can exist around the adjacent regions C ₁ and C ₂ is small, it is likely that C ₁ and C ₂ should not be merged. If it is known that the size of the text that can exist around the adjacent regions C ₁ and C ₂ is large, it is highly likely that C ₁ and C ₂ should be merged. Therefore, it is possible to more accurately determine whether or not the adjacent regions C ₁ and C ₂ should be merged by considering the size of text that can exist around the region and adjusting the degree of difference in the region. Can do.

If the region C ₁ contains only one pixel, the text that can exist around the region C ₁ in the input image using the scaling factor sc _i of the only pixel in the region C ₁ and the size of the scan window Can be calculated. For example, indicating the size of the scanning window length length or width width of the scan window, length / sc _i or width / sc _i a circle and radius or length / sc _i and width / sc _i the length and, rectangular area that the width can represent the magnitude of which can be present around the area C ₁ of the input image text. When the area C ₁ includes more than one pixel, the size of text that can exist around the area C ₁ in the input image is determined by the average value of the scaling factor sc _i in the area C ₁ and the size of the scan window. Can be calculated. For example, when the length of the scan window is represented by the length or width width of the scan window and the average value of sc _i is a_sc _i , a circle having a radius of length / a_sc _i or width / a_sc _i , or length / a_sc rectangular area of the _i and width / a_sc _i the length and width may represent the magnitude of which can be present around the area C ₁ of the input image text.

Next, in step S503, it is determined whether or not the difference between the adjacent areas is smaller than the minimum value of the difference in the adjusted area of the adjacent areas. If the determination result is affirmative, the process proceeds to step S504, and after the adjacent areas C ₁ and C ₂ are merged, the process proceeds to step S505. If the determination result is negative, the process directly proceeds to step S505.

  It should be noted that steps S501 to S503 are performed for all current adjacent areas.

  In step S505, it is determined whether or not all of the current adjacent areas satisfy the merge condition in the determination in step S503. If the determination result is negative, that is, if there is at least one newly merged area, the process returns to step S501. If the determination result is affirmative, it means that none of the current adjacent areas can be merged, that is, all candidate stroke areas have already been extracted.

  Hereinafter, exemplary expressions 3 and 4 for realizing the above steps S501 to S503 are shown.

Ｄｉｆ（Ｃ_１，Ｃ_２）は領域間の差異度を示す。Ｄ（Ｃ_１，Ｃ_２）がｔｒｕｅである場合には領域Ｃ_１、Ｃ_２をマージし、Ｄ（Ｃ_１，Ｃ_２）がｆａｌｓｅである場合には領域Ｃ_１、Ｃ_２をマージしない。ｍｉｎ（）は最小値をとることを示し、Ｉｎｔ（Ｃ_１）、Ｉｎｔ（Ｃ_２）はそれぞれ、領域Ｃ_１、Ｃ_２の領域内の差異度を表す。τ（Ｃ_１）、τ（Ｃ_２）はそれぞれ、領域Ｃ_１、Ｃ_２のスケーリング正則化項であり、領域Ｃ_１、Ｃ_２の周りに存在可能なテキストの大きさを表す。上述のように、ステップＳ２０１により、１つの画素の周りに存在可能なテキストの大きさが得られるとともに、その中の各画素の周りに存在可能なテキストの大きさに基づいてＣ_１、Ｃ_２の周りに存在可能なテキストの大きさ

Dif (C ₁ , C ₂ ) indicates the degree of difference between regions. When D (C ₁ , C ₂ ) is true, the regions C ₁ and C ₂ are merged, and when D (C ₁ , C ₂ ) is false, the regions C ₁ and C ₂ are not merged. min () indicates a minimum value, and Int (C ₁ ) and Int (C ₂ ) indicate the degrees of difference in the regions C ₁ and C ₂ , respectively. tau _(C 1), a tau _{(C 2),} respectively, the scaling regularization term of area _C 1, _{C 2,} represents the magnitude of the present text that can be around the area _C 1, _{C 2.} As described above, the size of the text that can exist around one pixel is obtained by step S201, and C ₁ and C ₂ are based on the size of the text that can exist around each pixel therein. Size of text that can exist around

を算出することができる。

Can be calculated.

にそれぞれ、経験によって得られる係数値Ｋを乗算すると、領域Ｃ_１、Ｃ_２のスケーリング正則化項τ（Ｃ_１）、τ（Ｃ_２）が得られる。ＭＩｎｔ（Ｃ_１，Ｃ_２）は、中間算出結果となり、隣接した領域の調整後の領域内の差異度の最小値である。

Are multiplied by a coefficient value K obtained by experience, respectively, to obtain scaling regularization terms τ (C ₁ ) and τ (C ₂ ) of regions C ₁ and C ₂ . MInt (C ₁ , C ₂ ) is an intermediate calculation result, and is the minimum value of the degree of difference in the adjusted area between adjacent areas.

  Hereinafter, step S203 of FIG. 2 will be described in detail with reference to FIG.

  FIG. 6 is a flowchart showing details of step S203 in FIG.

  As described above, a plurality of candidate stroke areas are acquired in step S202. Thereafter, in step S203, the candidate stroke area erroneously extracted from the background is filtered and removed to identify the true stroke area.

  In step S202, the conditional random field CRF model is adopted, and at the same time, the relationship between the individual stroke characteristics and the adjacent strokes is considered, and whether or not the candidate stroke area is a true stroke area is specified accurately. .

First, in step S601, candidate stroke areas that correlate with each other are specified from the candidate stroke areas. Specifically, based on the size information of the candidate stroke area and the distance between the candidate stroke areas, it is determined whether or not the candidate stroke areas are correlated. One candidate stroke area is one communication area, and the width and height of the outer cut rectangle of the communication area are the width w and height h of the candidate stroke area, and between the center of mass of the two candidate stroke areas i and j Is represented by dist (r _i , r _j ), and min [] represents the minimum value, and it is determined whether or not two candidate stroke areas are correlated according to the following equation (5).

式５を満たした候補筆画領域は、相関すると考えられる。全ての候補筆画領域に対して以上の判断を行うと、候補筆画領域近隣図が得られる。なお、候補筆画領域をノードとし、相関する候補筆画領域の対応するノードが互いに接続する。

Candidate stroke regions that satisfy Equation 5 are considered to be correlated. When the above determination is made for all candidate stroke areas, a candidate stroke area neighborhood diagram is obtained. Note that the candidate stroke area is a node, and the corresponding nodes of the correlated candidate stroke areas are connected to each other.

  In step S602, a true stroke area is specified according to the following equation 6.

Ｅは確信度を示し、訓練された分類器によって与えられるものである。Ｘは特徴の観測値を示す。Ｙは真の筆画領域であるか否かを示す。Ｇは候補筆画領域近隣図を示す。Λはパラメータを示す。ｘ_ｉは、例えば候補筆画領域のサイズ、確信度などの特徴である。上記ステップＳ２０１では、一つの画素の周りにテキスト領域が存在する確信度を算出することで、候補筆画領域の周りにテキスト領域の存在する確信度を算出して（例えば、候補筆画領域における各画素の周りにテキスト領域の存在する確信度の平均値をとる）候補筆画領域ｉの特徴ｘ_ｉとすることができる。ｙ_ｉは、候補筆画領域ｉが真の筆画領域であるか否かを示す。ｙ_ｉが１である場合、ｉは真の筆画領域である。ｙ_ｉが０である場合、ｉは真の筆画領域ではない。λ_uni、λb_iは、分類器を訓練することによって得られたパラメータである。Ｎ_ｉはｉと相関する全ての候補筆画領域の集合である。ｊはＮ_ｉにおける一つの候補筆画領域である。ｘ_ｊは候補筆画領域ｊの特徴を示し、例えば、候補筆画領域ｊにおける各画素周りにテキスト領域の存在する確信度である。ｙ_ｊは候補筆画領域ｊが真の筆画領域であるか否かを示し、ｙ_ｊが１である場合、ｊは真の筆画領域である。ｙ_ｊが０である場合、ときに、ｊは真の筆画領域ではない。λ_ijは、重み係数であり、ｊとｉの相関度を反映している。次の式７、８を採用して重み係数λ_ijを算出する。

E indicates confidence and is given by a trained classifier. X represents the observed value of the feature. Y indicates whether or not it is a true stroke area. G shows a neighborhood drawing of a candidate stroke area. Λ indicates a parameter. x _i is a feature such as the size of a candidate stroke area and the certainty factor, for example. In step S201, the certainty factor that the text region exists around one pixel is calculated by calculating the certainty factor that the text region exists around one pixel (for example, each pixel in the candidate stroke region). Can be used as a feature x _i of the candidate stroke area i). y _i indicates whether or not the candidate stroke area i is a true stroke area. If y _i is 1, i is the true stroke area. If y _i is 0, i is not a true stroke area. λ _uni and λb _i are parameters obtained by training the classifier. N _i is a set of all candidate stroke areas correlated with i. j is one of the candidate stroke region in N _i. x _j indicates the characteristics of the candidate stroke area j, and is, for example, the certainty that a text area exists around each pixel in the candidate stroke area j. y _j indicates whether or not the candidate stroke area j is a true stroke area. If y _j is 1, j is a true stroke area. If y _j is 0, sometimes j is not a true stroke area. λ _ij is a weighting factor and reflects the degree of correlation between j and i. The following formulas 7 and 8 are employed to calculate the weighting factor λ _ij .

各候補筆画領域ｉに対して、それと相関する全ての候補筆画領域（すなわち、Ｎ_ｉ）及びｉ自身を利用して、ｉの所在しているテキスト行ｌ_ｉをフィットする。具体的には、特徴区間において一つの点は一つの候補筆画領域の特徴を示す。Ｎ_ｉ及びｉの対応する点をフィットし、同一のフィッティング曲線に属する点の対応する候補筆画領域を、ｉの所在のテキスト行ｌ_ｉに属すると特定する。ｊは、Ｎ_ｉに属する、ｉと相関する候補筆画領域の一つである。dist（j,l_i）はｊの質量中心からｌ_ｉまでの距離である。

For each candidate stroke area i, all candidate stroke areas (that is, N _i ) and i that correlate with the candidate stroke area i are used and the text line l _{i where i} is located is fitted. Specifically, one point in the feature section indicates the feature of one candidate stroke area. The corresponding points of N _i and i are fitted, and the corresponding candidate stroke area of the points belonging to the same fitting curve is specified as belonging to the text line l _i where i is located. j belongs to N _i, is one of the candidate stroke region correlated to i. dist (j, l _i ) is the distance from the center of mass of j to l _i .

は経験によって得られる正規化因子である。ｅｘｐ［］は、自然対数ｅを底とする指数関数である。

Is a normalization factor obtained through experience. exp [] is an exponential function with the natural logarithm e as the base.

は回帰誤差である。以上から分かるように、ｊとｌ_ｉの距離が遠いほど、

Is the regression error. As can be seen from the above, the farther the distance between j and l _i is,

が小さくなる。

Becomes smaller.

の使用により、ｉと相関する候補筆画領域ｊの重み係数と異なるようにし、ｉと同一のテキスト行に属するｊがｉに対する影響がより大きくなる。よって、ｉから離れた特徴の類似する候補筆画領域がｉに大きい影響を与えることが回避される。

Is different from the weight coefficient of the candidate stroke area j correlated with i, and j belonging to the same text line as i has a greater influence on i. Therefore, it is avoided that a candidate stroke area having a similar feature away from i has a large influence on i.

In the above equation, E (x _ij , y _i , y _j , λ _bi ) can be used instead of E (x _i , x _j , y _i , y _j , λ _bi ), and x _ij is a candidate The absolute value of the difference between the average values of the certainty factors in which the text area exists around each pixel in the stroke areas i and j can be taken. x _ij can also take the distance between the center of mass of the candidate stroke areas i and j, and can better reflect the relationship between the areas. E (x _i , y _i , λ _uni )
Is the certainty as to whether or not the single candidate stroke area is a true stroke area (this is due to the case where the value presented is y _i ), and E (x _i , x _j , y _i , y _j , λ _bi ) indicates the certainty factor when the values to be presented are y _i , y _j ,

は、相関する候補筆画領域間の関係を示す。

Indicates the relationship between the correlated candidate stroke areas.

For each candidate stroke area i, it is assumed whether i and all j for i are true stroke areas, ie, the values of y _i , y _j are assumed. All pricing possibilities and corresponding features (ie, x _i , x _j and / or x _ij ) are input into the trained classifier and from the classifier E (x _i , y _i , λ _uni ) Instead, the value of E (x _i , x _j , y _i , y _j , λ _bi ) is returned, and E (X, Y, G, Λ) is calculated. When E (X, Y, G, Λ) is maximized, the corresponding value of y _i and y _j is taken as the result of specifying the true stroke area.

  That is, in step S602, various assumption combinations regarding whether or not all candidate stroke areas in the input image are true stroke areas are generated. For each combination of various assumptions, the first certainty factor of each candidate stroke area is calculated based on the combination of the above assumptions and the pixel characteristics in each candidate stroke area, and the candidate strokes correlated with each other. Based on the characteristics of the pixels in the region, a second certainty factor corresponding to the first certainty factor is calculated. Next, based on the first certainty factor and the second certainty factor, a certainty factor representing the case of the assumed combination is calculated. A hypothetical combination corresponding to the case where the certainty level representing the hypothetical combination is the highest is taken as the result of specifying the true stroke area. In addition, when the candidate stroke areas correlated with each other belong to the same text line, a relatively large weighting factor is applied to the corresponding second certainty factor in the calculation of the certainty factor representing the case of the assumed combination. Give.

  In step S203, the true stroke area is acquired. In step S204, the true stroke area is merged to form a text area.

  Hereinafter, step S204 of FIG. 2 will be described with reference to FIGS. 7A-7B and 8A-8C.

  FIG. 7A is a flowchart showing details of step S204 in FIG. FIG. 8A is a schematic diagram of a chain structure that connects all true stroke areas. FIG. 8B is a schematic diagram of the chain structure after line separation. FIG. 8C is a schematic view of the chain structure after character division.

  In step S203, the true stroke area has already been specified. In step S204, these stroke areas are merged to form a text area.

  First, in step S701, the connection relation between the stroke areas is specified based on the distance between the stroke areas. The distance between the stroke areas can be represented by the Euclidean distance between the centers of mass of the stroke areas. As shown in FIG. 8A, based on the distance between the stroke areas, a minimum spanning tree algorithm can be employed to connect all stroke areas according to a chain structure. Since the minimum spanning tree algorithm is a known algorithm in this technical field, it will not be described here.

  In FIG. 8A, when the relationship between the stroke areas is determined based only on the distance between the stroke areas, the stroke areas belonging to different characters in the same line and the stroke areas belonging to different lines are connected because the distance is short. It is clear that there is a possibility. Therefore, attention is paid to removing such erroneous connection in steps S702 and S703 described later.

  In step S702, the connection relationship between the stroke areas belonging to different text lines is removed. FIG. 7B shows a detailed flowchart of step S702 in FIG. 7A.

In step S7021, it is determined whether or not the Euclidean distance between the two stroke regions connected by one connection side in the chain structure is greater than the threshold value th _ed . If the determination result is negative, the process directly proceeds to step S7023. If the determination result is affirmative, the connection side is disconnected (step S7022), and the process proceeds to step S7023.

  Similar to the above situation, if there is only a distance, there is still a possibility that an erroneous connection is left. Accordingly, in steps S7023 to S7025, the erroneous connection is further detected and disconnected.

  When steps S7021 and S7022 are performed, there is a possibility that one chain structure generated by the original minimum spanning tree algorithm has already been divided into a plurality of chain structures. Steps S7023 to S7025 described later are executed for each chain structure.

  In step S7023, the stroke areas belonging to the same chain structure are fitted to one center line l. For example, the center of mass of the stroke regions belonging to the same chain structure is fitted to one center line l using the least square method.

It is determined whether or not the distance from each stroke area belonging to the chain structure to the center line l is greater than a predetermined threshold th _le (step S7024).

If the determination result is affirmative, it indicates that there is at least one text line on each side of the center line l. Therefore, the connection side straddling the center line 1 in the chain structure is cut. (Step S7025)
In step S7025, one chain structure becomes two new chain structures, so the process returns to S7023 again to continue the determination.

  If the determination result in step S7024 is negative, it indicates that there is only one text line in the current chain structure. Therefore, there is no connection side between the text lines, and step S702 is completed, and the process proceeds to step S703, where the erroneous connection between the characters belonging to the same text line is disconnected. The processing result of step S702 is shown in FIG. 8B.

  In step S703, each chain structure obtained in step S702 represents one text line. In each chain structure, there are a plurality of stroke areas, and the stroke areas are connected by connecting edges. The frame distance bd between the connected stroke areas and the average frame distance a_bd of the entire text line (that is, the chain structure) are calculated. The frame distance between two stroke areas connected by a connecting edge refers to the distance between adjacent edges of a circumscribed rectangle of the two stroke areas. When the frame distance bd between the two stroke areas connected by the connecting edge is much larger than the average frame distance a_bd of the entire text (for example, bd> a_bd * ξ, ξ is a constant set in advance by experience). Represents that these two stroke areas should belong to different characters, and cuts their connecting edges. That is, in step S703, the connection relationship between the stroke areas belonging to different characters is removed. The processing result of step S703 is shown in FIG. 8C.

  In the above, a plurality of chain structures have been acquired. Each chain structure represents one character, and each chain structure includes a plurality of stroke areas connected by connecting edges. A circumscribed rectangle for each chain structure can be used as a text area of a corresponding character of the chain structure. FIG. 9 is a schematic diagram showing a processing result of a text region positioning method in an image according to an embodiment of the present invention.

  Hereinafter, the structure of an apparatus for positioning a text area in an image according to an embodiment of the present invention will be described with reference to FIG. FIG. 10 is a block diagram illustrating a structure of an apparatus for positioning a text area in an image according to an embodiment of the present invention. As shown in FIG. 10, an apparatus 100 for positioning a text region in an image according to this embodiment includes an estimation unit 101 for estimating the size of text that can exist around each pixel in an input image, An extraction unit 102 for extracting a candidate stroke area from the input image based on the size and degree of difference between the areas, a specification unit 103 for specifying a true stroke area from the candidate stroke area, and a text area And a merge unit 104 for merging the true stroke areas so as to form

  The estimation unit 101 estimates an image pyramid generation unit 1011 for generating a pyramid image of a plurality of layers based on the input image, and a probability that text exists around each pixel in the pyramid image of each layer. An estimation subunit 1012 and a calculation unit 1013 for calculating the size of text that can exist around each pixel in the input image based on the probability and the size of the text corresponding to the pyramid hierarchy are included.

  The extraction unit 102 includes a difference degree calculation unit 1021 for calculating a difference degree between regions and a difference degree within the region with respect to adjacent regions in the input image, and around each pixel in the estimated input image. Estimate the size of text that can exist around the area based on the size of text that can exist, and adjust the degree of difference within each area based on the size of text that can exist around the estimated area An adjustment unit 1022 for merging, and an adjacent region merging unit 1023 for merging the adjacent regions based on the difference between the adjacent regions and the difference in the adjusted region of the adjacent regions including.

  The specifying unit 103 includes a correlation specifying unit 1031 for specifying candidate stroke areas correlated with each other among the candidate stroke areas, and various types relating to whether each candidate stroke area in the input image is a true stroke area. An assumption combination unit 1032 for generating an assumption combination, and for each assumption combination, a first certainty factor of each candidate stroke area is calculated based on the assumption combination and the pixel characteristics in each candidate stroke area, and the assumption is calculated. Based on the combination and the features in the candidate stroke areas that are correlated with each other, a second certainty factor corresponding to the first certainty factor is calculated, and then, based on the first certainty factor and the second certainty factor, Corresponding to the certainty calculation unit 1033 for calculating the certainty factor representing the case of the assumed combination and the highest certainty factor representing the case of the assumed combination Assumptions combination that includes a specific subunit 1034 for a particular result of the true stroke region. When the candidate stroke areas correlated with each other belong to the same text line, a relatively large weighting factor is given to the corresponding second certainty factor in calculating the certainty factor representing the assumption combination. . The correlation specifying unit determines whether or not the candidate stroke areas correlate based on the size information of the candidate stroke areas and the distance between the candidate stroke areas. The certainty factor calculation unit fits candidate stroke regions in the feature space, identifies candidate stroke regions belonging to the same fitting curve as belonging to the same text line, and weights in calculating the certainty factor based on the regression error Calculate the coefficient.

  The merge unit 104 includes a connection unit 1041 for specifying the connection relationship between the stroke areas based on the distance between the stroke areas, and a line dividing unit 1042 for removing the connection relationship between the stroke areas belonging to different text lines. And a character division unit 1043 for removing the connection relationship between the stroke areas belonging to different characters.

  The processes in the estimation unit 101, the extraction unit 102, the specifying unit 103, and the merge unit 104 included in the text region positioning apparatus 100 in the image according to the present invention are steps S201 to S204 of the text region positioning method in the image described above, respectively. For the sake of brevity, a detailed description of these units is omitted.

  Similarly, the processes in the image pyramid generation unit 1011, the estimation subunit 1012, and the calculation unit 1013 included in the estimation unit 101 are similar to the processes in steps S 401 to S 403 described above, and are included in the extraction unit 102. The processing in the difference calculation unit 1021, the adjustment unit 1022, and the adjacent region merging unit 1023 is similar to the processing in steps S501 to S505 described above, and the correlation specifying unit 1031 and the hypothetical combination unit included in the specific unit 103, respectively. 1032, the certainty calculation unit 1033, and the processing in the specific subunit 1034 are similar to the processing in steps S 601 to S 602 described above, respectively, and the connection unit 1041, line segmentation unit 1042, and character segmentation included in the merge unit 104. Uni Preparative process in 1043 is similar to the processing in step S701-S703 described above, respectively, for the sake of brevity, a detailed description thereof will be omitted for these units.

  Further, it should be pointed out that each component module and unit in the apparatus may be configured by software, firmware, hardware, or a combination thereof. Specific means or schemes that can be used in the construction are already known to those skilled in the art and will not be described here. When realized by software or firmware, a program constituting the software is installed from a storage medium or a network to a computer having a dedicated hardware structure (for example, the general-purpose computer 1100 shown in FIG. 11). The computer can execute various functions when various programs are installed.

  FIG. 11 is a block diagram schematically illustrating a computer for implementing the method and apparatus according to the embodiments of the present invention.

  In FIG. 11, a central processing unit (CPU) 1101 performs various processes based on a program stored in a read-only memory (ROM) 1102 or a program loaded from a storage unit 1108 to a random access memory (RAM) 1103. Execute. In the RAM 1103, data necessary for the CPU 1101 to execute various processes and the like is also stored as necessary. The CPU 1101, ROM 1102, and RAM 1103 are connected to each other via a bus 1104. An input / output interface 1105 is also connected to the bus 1104.

  An input unit 1106 (including a keyboard, a mouse, etc.), an output unit 1107 (for example, a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), a speaker, etc.), and a storage unit 1108 (a hard disk, etc. And a communication unit 1109 (including a network interface card such as a LAN card, a modem, etc.) are connected to the input / output interface 1105. The communication unit 1109 executes communication processing via a network (for example, the Internet). The driver 1110 may be connected to the input / output interface 1105 as needed. A removable medium 1111, for example, a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like is attached to the driver 1110 as necessary, so that a computer program read out from the driver 1110 can be stored in the storage unit 1108 as necessary. Installed.

  When the above-described series of processing is realized by software, a program constituting the software is installed from a network, for example, the Internet, or a storage medium, for example, a removable medium 1111.

  Those skilled in the art will recognize that such storage media is not limited to the removable media 1111 shown in FIG. 11 in which the program is stored and delivered remotely from the device to provide the program to the user. Should be understood. Examples of the removable medium 1111 include a magnetic disk (including a floppy (registered trademark) disk), an optical disk (including a compact disk / read only memory (CD-ROM) and a digital versatile disk (DVD)), and a magneto-optical disk. There are discs (including Mini Disc (MD) (registered trademark)) and semiconductor memories. Alternatively, the storage medium may be a hard disk included in the ROM 1102 and the storage unit 1108 in which a program is stored and distributed to the user together with a device including them.

  The present invention provides a program storing computer-readable instruction codes. When the instruction code is read and executed by a computer, the method according to the embodiment of the present invention can be executed.

  Similarly, a storage medium on which a program product in which the above-described computer-readable instruction code is stored is also included in the disclosure of the present invention. Examples of the storage medium include, but are not limited to, a floppy (registered trademark) disk, an optical disk, a magnetic optical disk, a memory card, a memory stick, and the like.

  In the foregoing description of specific embodiments of the invention, the features described and / or illustrated for one type of embodiment may be used in one or more other embodiments in the same or similar form. , Can be combined with features in other embodiments, or can be substituted for features in other embodiments.

  It should be emphasized that the term “comprising / having” as used herein means the presence of a feature, element, step or component, but one or more other features, elements, steps or It does not exclude the presence or addition of a component.

  The method of the present invention is not limited to being executed according to the time series described in the specification, and may be executed sequentially, in parallel, or individually according to other time series. Therefore, the order of execution of the methods described herein does not limit the technical scope of the present invention.

While specific embodiments of the invention have been described above, it should be understood that all of the above-described embodiments and examples are illustrative and not limiting. Those skilled in the art may contemplate various modifications, improvements, or equivalents to the present invention within the spirit and scope of the claims. These modifications, improvements or equivalents are considered to be within the protection scope of the present invention.
(Appendix 1)
A method for positioning a text area in an image comprising:
Estimating the size of text that can exist around each pixel in the input image;
Extracting a candidate stroke area from the input image based on the text size and the area difference; and
Identifying a true stroke area from the candidate stroke areas;
Merging the true stroke areas to form a text area.
(Appendix 2)
Estimating the size of text that can exist around each pixel in the input image comprises:
Generating a plurality of layers of pyramid images based on the input image;
Estimating the probability that text will exist around each pixel in the pyramid image of each hierarchy;
The method of claim 1, further comprising: calculating a size of text that can exist around each pixel in the input image based on the probability and the size of the text corresponding to a hierarchy of pyramids.
(Appendix 3)
The step of extracting the candidate stroke area includes:
Calculating a difference between regions and a difference within the region for adjacent regions in the input image; and
Estimate the size of text that can exist around the area based on the size of text that can exist around each pixel in the estimated input image, and size of the text that can exist around the estimated area Adjusting the degree of difference within each region based on the depth;
Determining whether or not to merge the adjacent regions based on the difference between the regions of the adjacent regions and the difference in the adjusted region of the adjacent regions;
Merging the adjacent areas when it is determined that the adjacent areas should be merged, and repeating the steps described above.
The method according to appendix 1, which means that if it is determined that all current adjacent areas should not be merged, a candidate stroke area is extracted.
(Appendix 4)
The method according to claim 3, wherein the adjacent areas are merged when the difference between the adjacent areas is smaller than the minimum difference in the adjusted area of the adjacent areas.
(Appendix 5)
The method according to appendix 3 or 4, wherein the adjacent region when starting the extraction step is an adjacent pixel, and the degree of difference between the regions includes a color difference.
(Appendix 6)
The step of specifying the true stroke area includes:
Identifying candidate stroke areas that correlate with each other among the candidate stroke areas;
Generating various hypothetical combinations as to whether each of all candidate stroke areas in the input image is a true stroke area;
For each hypothetical combination, the first certainty factor of each candidate stroke area is calculated based on the hypothesis combination and the characteristics of the pixels in each candidate stroke area, and the hypothetical combination and the pixels in the candidate stroke area correlated with each other. A second certainty factor corresponding to the first certainty factor is calculated based on the characteristics of the first confidence factor, and thereafter, a certainty representing the case of the hypothetical combination based on the first certainty factor and the second certainty factor. Calculating the degree,
A step of setting a hypothetical combination corresponding to the highest certainty level representing the hypothetical combination as a result of specifying a true stroke area;
A step of adding a large weighting factor to the corresponding second certainty factor in calculating the certainty factor representing the case of the hypothetical combination when the candidate stroke areas correlated with each other belong to the same text line; The method according to appendix 1, comprising:
(Appendix 7)
The method according to claim 6, further comprising the step of determining whether or not the candidate stroke areas correlate based on the size information of the candidate stroke areas and the distance between the candidate stroke areas.
(Appendix 8)
Fitting candidate stroke areas in the feature space and identifying candidate stroke areas belonging to the same fitting curve as belonging to the same text line;
The method according to appendix 6 or 7, further comprising: calculating a weighting factor in calculating the certainty factor representing the case of the assumption combination based on a regression error.
(Appendix 9)
The step of merging the true stroke area includes:
Identifying a connection relationship between the stroke areas based on the distance between the stroke areas;
Removing a connection between stroke areas belonging to different text lines;
The method according to claim 1, further comprising the step of removing a connection relationship between stroke areas belonging to different characters.
(Appendix 10)
An apparatus for positioning a text area in an image,
An estimation unit for estimating the size of text that can exist around each pixel in the input image;
An extraction unit for extracting a candidate stroke area from the input image based on the size of the text and the difference between the areas;
A specific unit for specifying a true stroke area from the candidate stroke area;
A merge unit for merging true stroke units to form a text area.
(Appendix 11)
The estimation unit is:
An image pyramid generation unit for generating a plurality of layers of pyramid images based on the input image;
An estimation subunit for estimating the probability that text will exist around each pixel in the pyramid image of each hierarchy;
The apparatus according to claim 10, further comprising: a calculation unit for calculating the size of text that can exist around each pixel in the input image based on the probability and the size of text corresponding to the pyramid hierarchy.
(Appendix 12)
The extraction unit is
A difference degree calculation unit for calculating a difference degree between areas and a difference degree within the area with respect to adjacent areas in the input image;
Estimate the size of text that can exist around the area based on the size of text that can exist around each pixel in the estimated input image, and size of the text that can exist around the estimated area And an adjustment unit for adjusting the degree of difference in each area,
The apparatus according to claim 10, further comprising: an adjacent area merging unit that merges the adjacent areas based on a difference between the adjacent areas and a difference in the adjusted area of the adjacent areas.
(Appendix 13)
The specific unit is:
A correlation identification unit for identifying candidate stroke areas correlated with each other among the candidate stroke areas;
A hypothetical combination unit for generating various hypothetical combinations as to whether each of all candidate stroke areas in the input image is a true stroke area;
For each hypothetical combination, the first certainty factor of each candidate stroke area is calculated based on the hypothesis combination and the characteristics of the pixels in each candidate stroke area, and the hypothetical combination and the pixels in the candidate stroke area correlated with each other. A second certainty factor corresponding to the first certainty factor is calculated based on the characteristics of the first confidence factor, and thereafter, a certainty representing the case in the assumed combination based on the first certainty factor and the second certainty factor. A certainty calculation unit for calculating the degree,
A hypothetical combination corresponding to the highest certainty level representing the case in the hypothetical combination includes a specific subunit that results in identifying the true stroke area;
When the candidate stroke areas correlated with each other belong to the same text line, a large weighting factor is added to the corresponding second certainty factor in the calculation of the certainty factor representing the case of the assumed combination. The device described.
(Appendix 14)
The apparatus according to appendix 13, wherein the correlation specifying unit determines whether or not the candidate stroke areas correlate based on size information of the candidate stroke areas and a distance between the candidate stroke areas.
(Appendix 15)
The certainty factor calculation unit includes:
Weights in the calculation of certainty factors that fit the candidate stroke area in the feature space, identify candidate stroke areas belonging to the same fitting curve as belonging to the same text line, and represent the case in the hypothetical combination based on the regression error The apparatus according to appendix 13 or 14, which calculates a coefficient.
(Appendix 16)
The merge unit is
A connection unit for identifying the connection relationship between the stroke areas based on the distance between the stroke areas;
A line dividing unit for removing the connection relationship between the stroke areas belonging to different text lines;
The apparatus according to claim 10, further comprising: a character dividing unit for removing a connection relationship between the stroke areas belonging to different characters.

100 apparatus 101 estimation unit 102 extraction unit 103 specific unit 104 merge unit

Claims (10)

A method for positioning a text area in an image comprising:
Estimating the size of text that can exist around each pixel in the input image;
Extracting a candidate stroke area from the input image based on the text size and the area difference; and
Identifying a true stroke area from the candidate stroke areas;
Merging the true stroke areas to form a text area. Estimating the size of text that can exist around each pixel in the input image comprises:
Generating a plurality of layers of pyramid images based on the input image;
Estimating the probability that text will exist around each pixel in the pyramid image of each hierarchy;
Calculating the size of text that can exist around each pixel in the input image based on the probability and the size of text corresponding to a hierarchy of pyramids. The step of extracting the candidate stroke area includes:
Calculating a difference between regions and a difference within the region for adjacent regions in the input image; and
Estimate the size of text that can exist around the area based on the size of text that can exist around each pixel in the estimated input image, and size of the text that can exist around the estimated area Adjusting the degree of difference within each region based on the depth;
Determining whether or not to merge the adjacent regions based on the difference between the regions of the adjacent regions and the difference in the adjusted region of the adjacent regions;
Merging the adjacent regions when it is determined that the adjacent regions should be merged, and repeating the steps described above.
The method according to claim 1, wherein if it is determined that none of all current adjacent areas should be merged, the candidate stroke area is extracted.   The method according to claim 3, wherein the adjacent areas are merged when the difference between the areas of the adjacent areas is smaller than the minimum difference in the adjusted area of the adjacent areas.   5. The method according to claim 3, wherein an adjacent region at the start of the extraction step is an adjacent pixel, and the difference between the regions includes a color difference. The step of specifying the true stroke area includes:
Identifying candidate stroke areas that correlate with each other among the candidate stroke areas;
Generating various hypothetical combinations as to whether each of all candidate stroke areas in the input image is a true stroke area;
For each hypothetical combination, the first certainty factor of each candidate stroke area is calculated based on the hypothesis combination and the characteristics of the pixels in each candidate stroke area, and the hypothetical combination and the pixels in the candidate stroke area correlated with each other. A second certainty factor corresponding to the first certainty factor is calculated based on the characteristics of the first confidence factor, and thereafter, a certainty representing the case of the hypothetical combination based on the first certainty factor and the second certainty factor. Calculating the degree,
A step of setting a hypothetical combination corresponding to the highest certainty level representing the hypothetical combination as a result of specifying a true stroke area;
Adding a large weighting factor to the corresponding second certainty factor in calculating the certainty factor representing the case of the hypothetical combination when the candidate stroke areas correlated with each other belong to the same text line. The method of claim 1.   The method according to claim 6, further comprising: determining whether the candidate stroke areas are correlated based on size information of the candidate stroke areas and a distance between the candidate stroke areas. Fitting candidate stroke areas in the feature space and identifying candidate stroke areas belonging to the same fitting curve as belonging to the same text line;
The method according to claim 6, further comprising a step of calculating a weighting factor in calculating the certainty factor representing the case of the assumed combination based on the regression error. The step of merging the true stroke area includes:
Identifying a connection relationship between the stroke areas based on the distance between the stroke areas;
Removing a connection between stroke areas belonging to different text lines;
Removing a connection relationship between stroke areas belonging to different characters. An apparatus for positioning a text area in an image,
An estimation unit for estimating the size of text that can exist around each pixel in the input image;
An extraction unit for extracting a candidate stroke area from the input image based on the size of the text and the difference between the areas;
A specific unit for specifying a true stroke area from the candidate stroke area;
A merge unit for merging true stroke units to form a text area.

Images