A line segmentation method for Tibetan ancient book documents based on local baseline

Abstract

The present invention discloses a line segmentation method for Tibetan ancient book documents based on local baselines, which is to firstly perform tilt correction and horizontal line projection on the binarized Tibetan ancient book document image, perform smoothing operation on the line projection map and estimate the text line; secondly, process the image in blocks according to the peak value, select the baseline expansion area, rotate and line project each area and obtain the maximum peak value, so as to obtain the baseline position; then, according to the distance constraint between the baseline and the peak, locally shift the baseline upward, according to the connected domain analysis constraint, locally shift the baseline downward according to the adjacent baseline constraint, complete the local baseline optimization, and finally, according to the baseline, splice the upper vowel with the corresponding line to complete the line segmentation. The present invention can solve the problems of line tilt, line distortion, line overlap and the like commonly existing in Tibetan document images, so as to be more conducive to the subsequent Tibetan character recognition, and finally realize the digital protection of precious Tibetan ancient books.

Description

Technical Field

[0001] The invention belongs to the field of image processing, and in particular is a Tibetan ancient book document line segmentation method based on local baseline. Background Art

[0002] Tibetan ancient books and documents are the literature preserved by the Tibetan people, and their mining and application are of great value. The mathematical application of Tibetan ancient books and documents plays an important role in the dissemination of culture, and the recognition of Tibetan ancient books and documents is an important step in digitization. Among them, line segmentation is an important step in Tibetan character recognition, and the quality of segmentation directly affects the effect of subsequent Tibetan character recognition. Chen Yuanyuan et al. (2021) proposed a segmentation method based on an adaptive run-length smoothing algorithm. By analyzing the connected domain of the results, the connected domain is divided into four categories using the height and width of the connected domain, and then the text line is segmented; this method is aimed at printed Tibetan, and cannot effectively deal with distortion and adhesion. Li Jincheng et al. (2021) proposed a line segmentation method that combines the text core and the expansion growth algorithm; first, the text core area is obtained through the Tibetan sound node and the baseline; then the Tibetan characters of each text line are connected through the baseline to obtain the pseudo-text connected area; finally, the text line is obtained using the expansion growth algorithm. This method has a significant effect on slightly tilted text lines, but it is not effective for problems such as stroke breaks and line adhesion. Hu et al. (2020) used line detection, horizontal projection and other operations to roughly estimate the range of the text line baseline, and then divided the image into eight blocks in the horizontal direction, and projected each block in turn to determine the local baseline. Then, the adhesion area was separated according to the analysis of the connected components, and the adhesion area was assigned according to the centroid coordinates to complete the line segmentation. This method works well when dealing with line distortion and line adhesion, but the baseline it determines still has some deviations. Hu et al. (2021) proposed a text line segmentation method that combines local baselines and connected components. First, the text line areas are grouped according to Tibetan features, and the text lines are segmented in the horizontal direction. Each segment is projected to obtain a local baseline. Then, the position relationship and neural network are used to detect the adhesion between characters between adjacent text lines or the adhesion between characters and the upper vowels of the next line, and the distance-based watershed algorithm is used to complete the segmentation of the adhesion area. This method can effectively deal with line distortion, but it cannot handle broken strokes and isolated areas well. Summary of the invention

[0003] The present invention aims to solve the deficiencies of the above-mentioned prior art and proposes a Tibetan ancient book document line segmentation method based on local baseline, in order to solve the common problems of line tilt, line distortion, line overlap, etc. in Tibetan ancient book document images, thereby being more conducive to subsequent Tibetan character recognition and ultimately realizing the digital protection of precious Tibetan ancient books.

[0004] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0005] The present invention provides a method for segmenting ancient Tibetan document lines based on a local baseline, which includes the following steps:

[0006] Step 1: Tilt correction and text line estimation:

[0007] Step 1.1, obtain the binary image R of the Tibetan ancient book document and use affine transformation to perform tilt correction to obtain the corrected image R′;

[0008] Step 1.2: Perform horizontal line projection on the corrected image R′ to obtain the line projection image Line projection After smoothing, we get the peak position of each row, let P m Indicates the peak position of the mth row; according to the peak position P of the mth row m Get the rough location of the mth text line; m belongs to [1,M]; M represents the total number of text lines in the binary image R;

[0009] Step 2: Local block processing:

[0010] Step 2.1: Local baseline positioning:

[0011] Step 2.2.1: Divide the rectified image R′ into blocks to obtain H local image blocks. Let the i-th local image block be Z i ; i∈[1,H];

[0012] Step 2.2.2: Use the i-th local image block Z i and the peak position P of the mth row m , get the i-th local image Z i The peak position P′ of the mth row in m,i ; According to the peak position P′ m,i Get the i-th local image block Z i The mth baseline extension area S′ in m,i ; Expand the area S′ for the mth baseline m,i Perform the jth rotation direction γ j The horizontal row projection under the baseline extension area is obtained by m,i,j , j represents different rotation directions; according to the baseline expansion area row projection map H′ m,i,j Get the projection value V′ of the maximum row m,i,j and its coordinate position D′ m,i,j , thus obtaining the projection value of the maximum row under all rotation directions, and selecting the rotation direction and coordinate position corresponding to the maximum projection value, respectively denoted as γ m,i,max and D′ m,i,max, finally according to γ m,i,max , S′ m,i and D′ m,i,max Get the i-th local image block Z i The mth local baseline B′ in m,i ;

[0013] Step 2.3: Local baseline optimization:

[0014] Step 2.3.1, calculate the peak position P′ of the mth row m,i The row coordinates of the mth local baseline B′ m,i The difference between the row coordinates of the center point, if the difference is greater than the set threshold V, then P′ m,i Assign to B′ m,i , to complete the correction process of the local baseline shifting upward; otherwise, keep the mth local baseline B′ m,i constant;

[0015] Step 2.3.2: According to the peak position P of the mth row m Draw its horizontal line L m , and calculate the horizontal line L m The number of connected domains N that pass through the Tibetan region in the rectified image R′ m , set the quantity threshold T 1 ; if N m >T 1 , then for the mth baseline extension area S′ m,i Expand up and down to obtain the mth baseline extension area S″ after expansion m,i ; To complete the correction of the local baseline tilt; otherwise, keep the mth baseline extension area S′ m,i unchanged, and S′ m,i Assign to S″ m,i ;

[0016] Step 2.3.3: Follow the process from step 2.2.2 to step 2.3.2 to complete the update of all baseline extension areas and local baselines;

[0017] Step 2.3.3: Calculate the i-th local image block Z i The mth local baseline B′ in m,i , the i-1th local image block Z i-1 The mth local baseline B′ in m,i-1 , the i+1th local image block Z i+1 The mth local baseline B′ in m,i+1 , and determine whether they all satisfy B′ m,i -B′ m,i-1 >T 2 , B′ m,i -B′m,i+1 >T 2 , B′ m,i-1 ≈B′ m,i+1 If so, then B′ m,i+1 Assign to B′ m,i , to complete the correction process of the downward shift of the local baseline; otherwise, keep the mth local baseline B′ m,i constant;

[0018] Step 3: Text line segmentation:

[0019] Step 3.1: Based on the mth local baseline B′ m,i , merge the mth local baselines of all local image blocks and obtain the mth text line baseline B m According to B m At the coordinate position in the corrected image R′, find the K pixel area expanded above it and get the mth expanded area A m ; in A m The text area where the upper vowel is located is retained in the text area, thereby obtaining the mth upper vowel area U m ; Then subtract U from the corrected image R′ m After that, we get the mth image E after the upper vowel separation. m ;

[0020] Step 3.2: Using the mth image E m The text line baseline B m Perform line segmentation on the rectified image R′ to obtain the preliminary segmentation result W of the mth text line m ;

[0021] Step 3.3: Merge W m and U m And get the segmentation result F of the mth text line m , and then the segmentation result F m The reverse tilt correction is performed to obtain the segmentation result of the mth text line in the binary image R, and then the segmentation result of the binary image R is obtained.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. The present invention overcomes the influence of adhesion, overlap and broken strokes commonly found in ancient Tibetan books on the line segmentation of ancient Tibetan books by adopting the local baseline method, thereby effectively realizing the line segmentation of ancient Tibetan books.

[0024] 2. The present invention adopts a method of smoothing the horizontal line projection map, takes 30 lines backward for the text line whose line projection map is greater than a threshold, and assigns the average value of these 30 lines to the text line, thereby overcoming the problem of double peaks appearing in the text line in the horizontal projection map, so that each text line contains only one peak, thereby improving the accuracy of peak locating text lines.

[0025] 3. The method of the present invention divides the document image into 40 blocks in the horizontal direction by adopting an image segmentation method, thereby solving the problem of inaccurate baseline detection of long texts in the prior art and improving the accuracy of baseline detection.

[0026] 4. The method of the present invention adopts the method of baseline extension area. According to the position of each text line peak in each image, the upper 20 lines and the lower 7 lines of this position are intercepted as the baseline extension area, thereby overcoming the influence of the core area of ​​​​the text on the local baseline detection, and narrowing the scope of baseline detection, thereby improving the efficiency and accuracy of baseline detection.

[0027] 5. The method of the present invention adopts a method of rotating the baseline extension area within a fixed range and horizontally projecting the baseline extension area. The baseline extension area is rotated between (-5°, +5°) with a step size of 0.5 and horizontally projected. The maximum value and corresponding angle of each projection image are counted. After the rotation is completed, the position of the maximum projection value in the complete image is rotated around the midpoint by a corresponding angle as the local baseline, which overcomes the problems of local adhesion and tilt in the text line, thereby reducing the influence of line adhesion, line tilt and line overlap on baseline detection.

[0028] 6. The method of the present invention adopts a local baseline optimization method. First, it is determined whether the local baseline is shifted upward according to the horizontal distance between the text line peak and the local baseline. If so, the local baseline is replaced by the current segment of the peak; then, it is determined whether the text line has a large inclination according to the number of connected domains of the text line crossed by the peak. If the number of crossings exceeds a threshold, the baseline extension area is expanded; finally, it is determined whether the local baseline is shifted downward according to the horizontal distance between the local baseline and the adjacent baseline. If the horizontal distance exceeds the threshold, it is replaced by the next segment of the local baseline, thereby solving the problem of local baseline detection errors caused by excessive projection values ​​at upper vowels, large inclination of the text line, and excessive projection at the core area of ​​the text line, and improving the accuracy of baseline detection.

[0029] 7. The method of the present invention performs text line segmentation through baselines. First, the local baselines of each text line are merged to obtain the text line baseline. Each text line is divided into a text core area and an upper vowel area according to the text line baseline. Finally, the text line is segmented according to adjacent text line baselines to complete line segmentation, thereby solving the problems of broken strokes, text line tilt and text line adhesion, and improving the line segmentation accuracy of Tibetan ancient book document images. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a structural schematic diagram of a method for segmenting lines of ancient Tibetan books based on a local baseline according to the present invention;

[0031] Figure 2 It is a flow chart of a method for segmenting lines of ancient Tibetan books based on a local baseline according to the present invention;

[0032] Figure 3 A schematic diagram of a binarized Tibetan ancient book image of a Tibetan ancient book document line segmentation method based on a local baseline according to the present invention;

[0033] Figure 4 A schematic diagram of a tilt-corrected Tibetan ancient book image of a Tibetan ancient book document line segmentation method based on a local baseline according to the present invention;

[0034] Figure 5 It is a schematic diagram of horizontal line projection of an ancient Tibetan book document image according to a method for segmenting ancient Tibetan book documents based on a local baseline of the present invention;

[0035] Figure 6 It is a schematic diagram of horizontal line projection of a smoothed Tibetan ancient book document image of a Tibetan ancient book document line segmentation method based on a local baseline of the present invention;

[0036] Figure 7 A schematic diagram of a baseline expansion area of ​​a Tibetan ancient book document line segmentation method based on a local baseline according to the present invention;

[0037] Figure 8 A schematic diagram of a text core area of ​​a Tibetan ancient book document line segmentation method based on a local baseline according to the present invention;

[0038] Fig. 9 A schematic diagram of the baselines of each text line of a Tibetan document image in a Tibetan ancient book document line segmentation method based on a local baseline of the present invention;

[0039] Fig.10 It is a schematic diagram of each text line of a Tibetan document image with upper vowels removed in a Tibetan ancient book document line segmentation method based on a local baseline of the present invention;

[0040] Fig.11 A schematic diagram of vowels on each text line of a Tibetan document image according to a local baseline-based Tibetan ancient book document line segmentation method of the present invention. DETAILED DESCRIPTION

[0041] In this embodiment, a Tibetan ancient book document line segmentation method based on local baseline is provided. Figure 1 and Figure 2 As shown, the following steps are included:

[0042] Step 1: Tilt correction and text line estimation:

[0043] Step 1.1, obtain the binary image R of the Tibetan ancient book document and use affine transformation to perform tilt correction to obtain the corrected image R′, as shown in Figure 3 and Figure 4 As shown, by performing tilt correction on the binary image R, the influence of the overall image tilt on the row segmentation can be solved;

[0044] Step 1.2: Perform horizontal line projection on the corrected image R′ to obtain the line projection image Line projection After smoothing, we get the peak position of each row, let P m Indicates the peak position of the mth row, such as Figure 5 and Figure 6 As shown, through the smoothing operation, the peak of each text line can appear only near the text line baseline, so that the text line position can be better located; according to the peak position P of the mth line m Get the rough location of the mth text line; m belongs to [1,M]; M represents the total number of text lines in the binary image R;

[0045] Step 1.2.1: Perform smoothing, the specific steps are:

[0046] Step 1.2.2, let Height be the height of the image, then the smoothed row projection graph is shown in formula (1). By replacing the projection value exceeding the threshold with the average value of the projection values ​​of the next 30 rows, a better smoothing effect can be achieved.

[0047]

[0048] Step 1.2.3: Use the smoothed line projection image By Filter The peak P with a projection value greater than 500 and a horizontal distance between adjacent peaks greater than 50 m , determine P m The number of text lines is m, and the text lines are roughly located using the peak positions.

[0049] Step 2: Local block processing:

[0050] Step 2.1: Local baseline positioning:

[0051] Step 2.2.1: Divide the rectified image R′ into blocks to obtain H local image blocks. Let the i-th local image block be Z i ; i∈[1,H], by dividing the image into blocks, the influence of text line distortion and line adhesion on baseline detection can be avoided, thus obtaining better detection effect;

[0052] Step 2.2.2: Use the i-th local image block Z i and the peak position P of the mth row m , get the i-th local image Z i The peak position P′ of the mth row in m,i ; According to the peak position P′ m,i Get the i-th local image block Z i The mth baseline extension area S′ in m,i ; Expand the area S′ for the mth baseline m,i Perform the jth rotation direction γ j The horizontal row projection under the baseline extension area is obtained by m,i,j , j represents different rotation directions; according to the baseline expansion area row projection map H′ m,i,j Get the projection value V′ of the maximum row m,i,j and its coordinate position D′ m,i,j , thus obtaining the projection value of the maximum row under all rotation directions, and selecting the rotation direction and coordinate position corresponding to the maximum projection value, respectively denoted as γ m,i,max and D′ m,i,max , finally according to γ m,i,max , S′ m,i and D′ m,i,max Get the i-th local image block Z i The mth local baseline B′ in m,i ,By rotating line projection of the baseline extension area, the ,baseline of text lines with different tilt angles can be detected, which solves the ,effect of line tilt on the baseline of text lines;

[0053] Step 2.2.2.1: According to the peak position P′ m,i Get the i-th local image block Z i The mth baseline extension area S′ in m,i In the figure, the baseline extension area is a rectangular area consisting of 20 pixels extending upward from the peak and 7 pixels extending downward, such as Figure 7 As shown, the baseline extension area selected by the peak can include the baseline of each text line, thereby enabling faster and more accurate detection of the baseline in the baseline extension area.

[0054] Step 2.2.2.2, according to γ m,i,max , S′ m,i and D′ m,i,max Get the i-th local image block Z i The mth local baseline B′ in m,i The specific steps are to convert D′ m,i,max The corresponding global position B m,i In S′ m,iRotate around its midpoint, as shown in formula (2):

[0055] B′ m,i,x =B m,i,x ×cosβ-B m,i,y ×sinβ+B m,i,n ×cosβ+B m,i,h ×cosβ+B m,i,n (2)

[0056] In formula (2), B m,i,x and B m,i,y is the coordinate of the starting point or end point before rotation, B′ m,i,x and B m,i,y is the coordinate of the starting point or end point after rotation, B m,i,n and B m,i,h is the coordinate of the midpoint of the baseline, where B m,i,x D′ m,i,max The corresponding global position, β is γ m,i,max The corresponding radians are defined as shown in equations (3) and (4):

[0057] β=γ m,i,max ×π÷180 (3)

[0058] B m,i,x =P′ m -(L m -D′ m,i,max,x ) (4)

[0059] In formula (3) and formula (4), P′ m is the row where the local peak is located in the original image, L m is the range selected from the peak upward, D′ m,i,max,x is the row where the region baseline is located, i is the number of the image block, and m is the number of each row.

[0060] Step 2.3: Local baseline optimization:

[0061] Step 2.3.1. Calculate the peak position P′ of the mth row m,i The row coordinates of the mth local baseline B′ m,i The difference between the row coordinates of the center point, if the difference is greater than the set threshold V, then P′ m,i Assign to B′ m,i , to complete the correction process of the local baseline shifting upward; otherwise, keep the mth local baseline B′ m,i This operation can correct the incorrect baseline detection result when the projection value of the upper vowel in the baseline expansion area is large, thereby improving the accuracy of baseline detection.

[0062] In a specific implementation, the threshold V is set to 30 in the first and second quarters of the text line and to 15 in the middle of the text line.

[0063] Step 2.3.2: According to the peak position P of the mth row m Draw its horizontal line L m , and calculate the horizontal line L m The number of connected domains N that pass through the Tibetan region in the rectified image R′ m , set the quantity threshold T 1 ; if N m >T 1 , then for the mth baseline extension area S′ m,i Expand up and down to obtain the mth baseline extension area S″ after expansion m,i ; To complete the correction of the local baseline tilt; otherwise, keep the mth baseline extension area S′ m,i unchanged, and S′ m,i Assign to S″ m,i This operation can correct the incorrect baseline detection results caused by the large inclination of the text line and the absence of the true baseline position in the area when projecting the baseline extension area, thereby improving the accuracy of baseline detection.

[0064] If N m >T 1 , expand the baseline up and down to expand the area S′ m,i Get S″ m,i The processing process is to use S′ m,i Then expand 40 pixels upward and 20 pixels downward to get S″ m,i .

[0065] Step 2.3.3: Follow the process from step 2.2.2 to step 2.3.2 to complete the update of all baseline extension areas and local baselines;

[0066] Step 2.3.4: Calculate the i-th local image block Z i The mth local baseline B′ in m,i , the i-1th local image block Z i-1 The mth local baseline B′ in m,i-1 , the i+1th local image block Z i+1 The mth local baseline B′ in m,i+1 , and determine whether they all satisfy B′ m,i -B′ m,i-1 >T 2 , B′ m,i -B′ m,i+1 >T 2 , B′ m,i-1 ≈B′ m,i+1If so, then B′ m,i+1 Assign to B′ m,i , to complete the correction process of the downward shift of the local baseline; otherwise, keep the mth local baseline B′ m,i unchanged; this operation can correct the incorrect baseline detection result caused by the large projection value in the core area of ​​the text in the area when the baseline expansion area is projected, thereby improving the accuracy of baseline detection, such as Figure 8 In this embodiment, the threshold value T 2 =10.

[0067] Step 3: Text line segmentation:

[0068] Step 3.1: Based on the mth local baseline B′ m,i , merge the mth local baselines of all local image blocks and obtain the mth text line baseline B m ,like Fig. 9 As shown; According to B m At the coordinate position in the corrected image R′, find the K pixel area expanded above it and get the mth expanded area A m ; in A m The text area where the upper vowel is located is retained in the text area, thereby obtaining the mth upper vowel area U m ; Then subtract U from the corrected image R′ m After that, we get the mth image E after the upper vowel separation. m ,like Fig.10 and Fig.11 As shown;

[0069] In this embodiment, in A m Remove non-upper vowels from the region and get the upper vowel region U m The processing process is to m The contour is detected in the area and the contour C is obtained. q , q represents the number of the contour, and the threshold T is set 2 =40, T 3 =500, T 4 =20, T 5 = 8. Remove contour C q The area C q,areas <T 2 ∪C q,areas >T 3 、Contour centroid C q,x -B′ m >T 4 and outline height and width C q,h <T 5 ∩C q,w <T 5 The contour of the upper vowel area U mThis operation can easily handle the strokes of text line adhesion, overlap and break.

[0070] Step 3.2: Using the mth image E m The text line baseline B m Perform line segmentation on the rectified image R′ to obtain the preliminary segmentation result W of the mth text line m ;

[0071] Step 3.3: Merge W m and U m And get the segmentation result F of the mth text line m , and then the segmentation result F m The reverse tilt correction is performed to obtain the segmentation result of the mth text line in the binary image R, and then the segmentation result of the binary image R is obtained.

Claims

1. A method for segmenting lines of ancient Tibetan documents based on local baselines, which is characterized by the following steps: Step 1: Tilt correction and text line estimation: Step 1.1, obtain the binary image R of the Tibetan ancient book document and use affine transformation to perform tilt correction to obtain the corrected image R′; Step 1.2: Perform horizontal line projection on the corrected image R′ to obtain the line projection image Line projection After smoothing, we get the peak position of each row, let P m Indicates the peak position of the mth row; according to the peak position P of the mth row m Get the rough location of the mth text line; m belongs to [1,M]; M represents the total number of text lines in the binary image R; Step 2: Local block processing: Step 2.1: Local baseline positioning: Step 2.2.1: Divide the rectified image R′ into blocks to obtain H local image blocks. Let the i-th local image block be Z i ; i∈[1,H]; Step 2.2.2: Use the i-th local image block Z i and the peak position P of the mth row m , get the i-th local image Z i The peak position P′ of the mth row in m,i ; According to the peak position P′ m,i Get the i-th local image block Z i The mth baseline extension area S′ in m,i ; Expand the area S′ for the mth baseline m,i Perform the jth rotation direction γ j The horizontal row projection under the baseline extension area is obtained by m,i,j , j represents different rotation directions; according to the baseline expansion area row projection map H′ m,i,j Get the projection value V′ of the maximum row m,i,j and its coordinate position D′ m,i,j , thus obtaining the projection value of the maximum row under all rotation directions, and selecting the rotation direction and coordinate position corresponding to the maximum projection value, respectively denoted as γ m,i,max and D′ m,i,max , finally according to γ m,i,max , S′ m,i and D′ m,i,max Get the i-th local image block Z i The mth local baseline B′ in m,i ; Step 2.3: Local baseline optimization: Step 2.3.1, calculate the peak position P′ of the mth row m,i The row coordinates of the mth local baseline B′ m,i The difference between the row coordinates of the center point, if the difference is greater than the set threshold V, then P′ m,i Assign to B′ m,i , to complete the correction process of the local baseline shifting upward; otherwise, keep the mth local baseline B′ m,i constant; Step 2.3.2: According to the peak position P of the mth row m Draw its horizontal line L m , and calculate the horizontal line L m The number of connected domains N that pass through the Tibetan region in the rectified image R′ m , set the quantity threshold T1; if N m >T1, then the mth baseline expansion area S′ m,i Expand up and down to obtain the mth baseline extension area S″ after expansion m,i ; To complete the correction of the local baseline tilt; otherwise, keep the mth baseline extension area S′ m,i unchanged, and S′ m,i Assign to S″ m,i ; Step 2.3.3: Follow the process from step 2.2.2 to step 2.3.2 to complete the update of all baseline extension areas and local baselines; Step 2.3.3: Calculate the i-th local image block Z i The mth local baseline B′ in m,i , the i-1th local image block Z i-1 The mth local baseline B′ in m,i-1 , the i+1th local image block Z i+1 The mth local baseline B′ in m,i+1 , and determine whether they all satisfy B′ m,i -B′ m,i-1 >T2, B′ m,i -B′ m,i+1 >T2, B′ m,i-1 ≈B′ m,i+1 If so, then B′ m,i+1 Assign to B′ m,i , to complete the correction process of the downward shift of the local baseline; otherwise, keep the mth local baseline B′ m,i constant; Step 3: Text line segmentation: Step 3.1: Based on the mth local baseline B′ m,i , merge the mth local baselines of all local image blocks and obtain the mth text line baseline B m According to B m At the coordinate position in the corrected image R′, find the K pixel area expanded above it and get the mth expanded area A m ; in A m The text area where the upper vowel is located is retained in the text area, thereby obtaining the mth upper vowel area U m ; Then subtract U from the corrected image R′ m After that, we get the mth image E after the upper vowel separation. m ; Step 3.2: Using the mth image E m The text line baseline B m Perform line segmentation on the rectified image R′ to obtain the preliminary segmentation result W of the mth text line m ; Step 3.3: Merge W m and U m And get the segmentation result F of the mth text line m , and then the segmentation result F m The reverse tilt correction is performed to obtain the segmentation result of the mth text line in the binary image R, and then the segmentation result of the binary image R is obtained.

Images