CN108564078B - A method of extracting the axis of Manchu word image - Google Patents

Abstract

The method for extracting the axis of Manchu word images belongs to the field of text segmentation, and is used to solve the problem of improving the accuracy of Manchu segmentation. It is for the extraction of the central axis of the Manchu word image, which directly affects the accuracy of segmentation. In order to provide the accuracy of segmentation, it is necessary to improve the accuracy of the central axis extraction. Therefore, to locate the central axis and detect its width, you can The central axis is accurately extracted.

Description

A method of extracting the axis of Manchu word image

technical field

The invention belongs to the field of text segmentation, and relates to a method for extracting the central axis of a Manchu word image.

Background technique

Manchu is the language used by the Manchu, Xibo and other ethnic minorities in my country. It was popularized and used as a legal script in the Qing Dynasty, forming a large number of precious Manchu documents. As the Manchu language is on the verge of disappearing, the Manchu language and cultural heritage needs to be rescued and protected urgently, which has been recognized and valued by the state and all sectors of society. The study of the optical character recognition technology of Manchu is particularly important for the protection and inheritance of the cultural heritage of the Qing Dynasty. Manchu is a phonetic script with a total of 38 letters, including 6 vowels, 22 consonants, and 10 specific letters for spelling Chinese loanwords. Manchu writing adopts the rules of word order from top to bottom and lines from left to right. For Manchu character recognition, it is often necessary to first divide Manchu characters into basic units (such as letters, etc.), and then identify them. Therefore, improving the accuracy of Manchu character recognition can start from improving its segmentation accuracy.

SUMMARY OF THE INVENTION

In order to solve the problem of improving Manchu segmentation accuracy, the present invention proposes the following technical solutions:

A method for extracting the central axis of a Manchu word image, comprising the following steps:

S1. locate the central axis of the Manchu word image;

S2. Detect the width of the central axis of the Manchu word image.

As a supplement to the technical solution: the step S1 specifically includes:

S1.1. Invert the image of Manchu words, let the pixel value of the text part take 1, and the pixel value of the background part take 0;

S1.2. Use the morphological refinement function of the MATLAB image processing toolbox to realize the morphological refinement of Manchu word images;

S1.3. For the morphologically refined Manchu word image, use Hough transform to determine the column coordinates corresponding to the thinned central axis, and the column coordinates are used as the position of the central axis of the Manchu word image, wherein the Hough transform is defined. The angle of the search line is θ=90, only the vertical line is searched, and the line connecting the same vertical position, the spacing is less than the height of the Manchu word image and the length is greater than 1 pixel is a straight line, and the central axis is obtained. the center position.

As a supplement to the technical solution: the step S2 specifically includes:

S2.1. Determine the search area for the maximum run-length proportional method;

S2.2. Determine the width of the central axis of the Manchu word image by applying the maximum run-length ratio method to the Manchu word image in the search area;

S2.3. Calculate the left and right boundaries of the central axis from the central position of the central axis of the Manchu word image and the width of the central axis.

As a supplement to the technical solution: the step S2.1 is specifically:

The search area for the maximum run ratio method is determined by the range specified by the following formula:

Among them, sl is the left boundary of the limited search range, sr is the right boundary of the limited search range, baseline is the center position of the central axis, round means rounding to the nearest integer, and W is the width of the Manchu word image.

As a supplement to the technical solution: the steps of the maximum run-length ratio method in the step S2.2: scan each line of the word image in the Manchu word image search area, and count the run length of continuous black pixels and the number of times the length appears, then there is The run length of the maximum number of occurrences is the width of the axis in the Manchu word image.

As a supplement to the technical solution: calculate the left and right boundaries of the central axis of the step S2.3 by the following formula;

Where: bl is the left border of the central axis, br is the right border of the central axis, baseline is the center position of the axis of the Manchu word image, baseline_width is the width of the axis of the Manchu word image, and round means rounding to the nearest integer.

Beneficial effects: For the extraction of the central axis of the Manchu word image, the accuracy of segmentation is directly affected. In order to provide the accuracy of segmentation, the accuracy of the extraction of the central axis must be improved. Therefore, the central axis is positioned and its width is detected. The central axis can be accurately extracted.

Description of drawings

Figure 1 is a flowchart of the construction of Manchu component set;

Fig. 2 Manchu component segmentation flowchart;

Fig. 3 traditional method Manchu word image central axis extraction error example diagram;

Fig. 4 adopts the maximum run-length ratio method limited by the region to determine the width of the Manchu central axis, wherein: (1) the maximum run-length ratio method error example chart, (2) the search range chart limited by the present invention, (3) the result chart of the method of the present invention ;

5 is an effect diagram of axis extraction in the method of the present invention;

Figure 6 is a flowchart of Manchu component segmentation;

Fig. 7 Manchu component segmentation result diagram, wherein: (1) Weak segmentation phenomenon diagram, (2) Weak segmentation region through fine segmentation diagram, (3) Over segmentation phenomenon diagram, (4) Over segmentation region through merge diagram, (5) Partial segmentation result graph.

Detailed ways

，同时是字母a、e和n的字中形，在识别中可根据相邻字母的拼读规则加以区别。(5)部分满文字母具有相同的组成部分。例如字符

(字母o的字头形)，可以看做是由字符

(字母e的字头形)和字符

(字母o的字中形)两部分组合而成。因此以满文字母为基本分割单元则容易出现过分割和弱分割的现象。(6)某些字母组合不具备可分性。例如

(bo)，切分成

(字母b)和

(字母o)非常困难。From the perspective of optical character recognition technology, Manchu has the following characteristics: (1) According to the different positions in the word, the same letter in Manchu generally has four different forms: independent, prefix, middle and suffix. . There are a total of 114 letters in Manchu with different glyphs. (2) The words in the same column of Manchu documents are located near the same central axis, and the words between the two columns of printed Manchu are basically not crossed, which is conducive to column extraction. There is a certain interval between Manchu words in the same column of Manchu text, which is beneficial to word extraction. (3) Manchu words are composed of one or more Manchu letters connected by the vertical axis, and there is no gap between letters in the same word. However, the place where the letters and letters are joined is located on the axis of the Manchu word image, and it can be considered to segment the Manchu letters by using the pixel characteristics at the axis. (4) Some Manchu letters have the phenomenon of "one shape and many characters". e.g. characters

, which is the middle form of letters a, e and n at the same time, can be distinguished according to the spelling rules of adjacent letters in recognition. (5) Some Manchu letters have the same components. e.g. characters

(the initial shape of the letter o), which can be seen as a combination of characters

(the initial form of the letter e) and the character

(the middle form of the letter o) is a combination of two parts. Therefore, taking the Manchu alphabet as the basic segmentation unit is prone to over-segmentation and weak segmentation. (6) Certain letter combinations are not divisible. E.g

(bo), cut into

(letter b) and

(letter o) is very difficult.

Based on the above-mentioned characteristics of Manchu, this embodiment proposes an idea of re-deconstructing Manchu words with components, using Manchu components (hereinafter referred to as components) as the basic unit of segmentation and recognition, which can solve the problem of using Manchu letters as the basic segmentation unit The problem of over-segmentation and weak segmentation caused by the Manchu parts set includes Manchu letters, part of letters or letter combinations, and letter combinations. This is because if letters are used as the basic segmentation unit, as in the previous analysis, the problems of over-segmentation and weak segmentation are prone to occur, and the subsequent classifiers used to identify letters are bound to produce recognition errors for the over-segmented and weakly segmented parts, or even fail to recognize them. The Manchu component set proposed by the present invention (method) is designed and constructed based on the results of the segmentation method, that is, the common over-segmentation (part of a letter or a letter combination) and weak segmentation (a letter combination) It is no longer considered a "wrong", but a "correct" segmentation, so the classifiers designed later can recognize these parts, thereby reducing the recognition errors caused by segmentation errors. For the understanding of Manchu parts, you can refer to the analogy with the recognition of English words. Take the English word study as an example, you can directly identify the whole word study; you can also divide the whole word into letters such as s, t, u, d, y, identify the letters respectively, and then combine them into the word study; if it is difficult to divide into letters, It is relatively easy to divide into parts. For example, it is easy to divide into: st, u, dy, (where st, u, and dy are all parts), then identify the parts, and then combine the parts into words. However, Manchu has the above characteristics because of the above characteristics. , the segmentation of its components is not as easy as the English example. As shown in Figure 1, the construction process of the Manchu component set is: refer to the Manchu alphabet, the national standard of the People's Republic of China "General Multi-octet Coded Character Set for Information Technology" The Mongolian component set in Xibo, Manchu, and documents [1-2], proposes an initial set of Manchu components (hereinafter referred to as "initial set") containing 99 initial components, and juxtaposes each Manchu Component's Flag=0. Use Manchu segmentation to segment Manchu word images, and count and analyze the segmentation results: if the segmented component does not belong to the initial set, add the component to the initial set, and set the component's Flag=1; If the component belongs to the initial component set, set Flag=1 of the corresponding component. Check whether there is a component in the initial set with Flag=0, determine whether there is a component that has never appeared in the segmentation result, and delete the component from the initial set if there is such a component. Organize and output Manchu parts sets. There are 106 components in the Manchu component set, see Appendix 1 for details. The literature mentioned in it [1-2]:

[1]Hongxi Wei,Guanglai Gao.A keyword retrieval system for historicalMongolian document images[J].International journal on document analysis andrecognition,2014,17(1),33-45.

[2] Liangrui Peng, Changsong Liu, Xiaoqing Ding, Jianming Jin, Youshou Wu, Hua Wang, Yanhua Bao. Multi-font printed Mongolian document recognition system[J]. International journal on document analysis and recognition, 2010, 13(2) :93-106.

As shown in Figure 2, the segmentation steps of the full text are as follows:

S1. Manchu paper documents are converted into digital image documents that can be stored and processed by a computer through photoelectric conversion equipment, and image preprocessing (smoothing, binarization) is performed on the digital images of Manchu documents;

S2. Layout analysis (tilt correction, column segmentation and word segmentation);

S3. Extract Manchu word images;

S4. Position normalization;

S5. Central axis extraction;

S6. According to the relationship between the Manchu parts and the position of the central axis, the Manchu parts are divided.

Among them, the tilt correction uses the Hough transform method to determine the tilt angle of the layout, and then rotates the image and corrects it back to the vertical text state; the Manchu document that has been tilt-corrected uses the vertical projection method for column segmentation, and the horizontal projection method is used to segment words and extract full text. Manchu words in the text image, and then perform position normalization on the Manchu word images. The preprocessing of the Manchu word image is completed by the above steps, and the height of the Manchu word image is recorded as H and the width as W. It should be noted that the position normalization of the Manchu word image is to cut off the redundant white background edge of the Manchu word image. The process shown in Figure 2 is to perform image flipping for the convenience of programming, and it is displayed that the Manchu word is removed. The black edge of the word. In Figure 2, the picture with white characters on a black background is the picture of the so-called image flipping. The original image should be black characters on a white background, but for the convenience of programming, it is more convenient to flip it to white characters on a black background to remove the edges of the four corners, so the image with the black edges removed after flipping is directly given.

In this embodiment, the extraction of the central axis of the Manchu word image directly affects the accuracy of segmentation, and the specific scheme thereof will be described in detail below.

For the extraction of the central axis of the Manchu word image, that is, step S5, the vertical projection method and the maximum cumulative vertical projection method are generally used in the prior art. However, the above two methods have the central axis positioning offset and the central axis width estimation error, etc., such as shown in Figure 3. This embodiment provides a method for extracting the central axis of a Manchu word image, including the following steps:

S5.1. Central axis positioning of Manchu word images:

First, the Manchu word image is inverted, that is, the pixel value of the text part is 1 and the pixel value of the background is 0. Using the morphological refinement function of the MATLAB image processing toolbox, a 3×3 structural element template is used, each template contains 9 pixels, and each pixel can only take 0 or 1, so the template has 512 different forms, the template is divided into 8 morphological refinement of Manchu word images in each direction. For the refined Manchu word image, the Hough transform is used to determine the column coordinates corresponding to the refined central axis, which is the position of the central axis of the Manchu word image. In the extraction of the central axis of the Manchu word image, the angle of the Hough transform search line is limited to θ=90, that is, only the vertical line is searched and connected to the same vertical position, the spacing is less than the height H of the word image and its length is greater than A straight line of 1 pixel is a straight line, that is, the center position of the central axis is obtained, which is recorded as baseline. The central axis of the Manchu word image refers to the column coordinate position of the central axis of the Manchu word in the image in a Manchu word image, rather than the center line of an image.

S5.2. Axis width detection of Manchu word images

S5.2.1. Use the maximum run-length ratio method of the width of the central axis: first scan each line of the Manchu word image, and count the run-length of consecutive black pixels and the number of occurrences of this length; scan all the lines in turn, then the maximum number of occurrences The run length is the width of the axis of the Manchu word image, denoted as w ₀ . The maximum run-length ratio method is effective to detect the width of the axis in the Manchu word image, but there is still an error situation as shown in Figure 4(1). The reason for this error is that the maximum run-length ratio method is to perform continuous black pixel run-length statistics on the entire Manchu word image, and the deformation of different Manchu fonts seriously interferes with the global statistical results of the maximum run-length ratio method. Statistics on Manchu writing show that the width of the axis in Manchu generally does not exceed 1/2 of the word width W, so the search area of the maximum run-length ratio method is limited, and the search area of the algorithm is limited to the range specified by formula (1), It is called the area-limited maximum run-scale method.

In formula (1), sl is the left boundary of the limited search range, sr is the right boundary of the limited search range, baseline is the center position of the central axis, and round means rounding to the nearest integer. The limited search area weakens the statistical influence of Manchu free and branch strokes on the width of the central axis, and then the maximum run-length ratio method is used to complete the detection of the width of the central axis in the Manchu word image after the limited search range. The results are shown in Figure 4 ( 3) shown.

S5.2.2. From the width baseline_width of the central axis and the center position baseline of the central axis, calculate the left boundary bl and the right boundary br of the central axis according to formula (2).

A total of 400 Manchu images with different font sizes were randomly selected, and the central axis was extracted by the area-limited maximum run-length ratio method and the vertical projection method in this embodiment. The results are shown in Table 1. A partial example of correctly extracting the central axis using the method of the present invention is shown in FIG. 5 . The experimental results show that the morphological refinement and Hough transform can accurately locate the axis position of Manchu word images, and the maximum run-length probability method limited by the region can correctly determine the width of the axis of Manchu word images.

Table 1 Statistical table of the results of axis extraction in Manchu word images

method of the invention vertical projection correct number of samples 397 210 number of wrong samples 3 190 Correct rate 99.25% 52.50%

In this embodiment, the accuracy of Manchu character segmentation is a bottleneck problem to improve the Manchu recognition accuracy, and the specific scheme thereof will be described in detail below.

For Manchu component segmentation, that is, step S6, as shown in Figure 6, it includes:

S6.1. Manchu parts are roughly cut;

S6.2. Weak segmentation decision and fine segmentation of candidate segmentation regions;

S6.3. Over-segmentation decision and merging of candidate segmentation regions.

The above steps are described in detail below:

S6.1. Rough cutting of Manchu parts

Since the Manchu parts are connected by the central axis, the Manchu words are firstly divided into left, middle and right parts with the central axis as the center. Among them, the range of the left part is from column 1 to column bl-1 of Manchu words, and the range of the right part is from column br+1 to column W of Manchu words. The left part and the right part are projected horizontally, denoted as pl and pr. The segmentation cost function of the i-th row is defined as:

Cost(i)=pl(i)+pr(i), i=1,2,...,H(3)

只有满足条件：Ideally, the value of the cost function for dividing a row should be 0, that is, the left and right parts of the row have no strokes other than the central axis. However, in practice, due to the influence of noise caused by preprocessing such as scanning, tilt correction, and binarization, too strict constraints on segmented rows will lead to serious weak segmentation problems. Let T1 be the rough segmentation threshold of Manchu parts, and determine the value of T1 through a large number of experiments.

Only if the conditions are met:

Cost(i)≤T1 (4)

The row is the candidate segmentation row, and the sequence composed of all the candidate segmentation rows that satisfy the condition (4) is recorded as Can_seg. Among them, the determination experiment of the value of T1 is to select different multiples of baseline_width as T1, and these multiples are all fractions of <=1, execute the Manchu component segmentation method, and select the segmentation method for the image comparison after segmentation. The T1 corresponding to the Manchu word image with better effect is finally selected as the above T1 value.

The set of candidate segmentation rows obtained by the rough segmentation of Manchu parts will have the following three situations:

1) Take the first row of the image as a candidate segmentation row, which is obviously an unreasonable candidate row, so it should be deleted from the candidate segmentation row set;

2) The continuous adjacent lines of the image starting from the first line/the continuous adjacent lines of the image ending with the last line (the Hth line) are all unreasonable candidate line subsections, so they should be selected from the candidate segmentation set. delete these subsections;

3) Except for the sub-segment composed of consecutive adjacent rows in 2), only one candidate segmentation row in the middle position is needed, and the rest are not required; therefore, the candidate row in the middle position should be used instead of the entire continuous adjacent row. subsection.

From the above, there are often redundant candidate segmentation lines in Can_seg. For this reason, the following strategies are adopted to further delete the redundant candidate segmentation lines in Can_seg:

(1) If there is only one candidate segmentation row in Can_seg, and it is the first row, delete the row; otherwise, go to step (2);

(2) Find the sub-segment conti_subseg composed of continuous candidate segmentation lines, if the starting line of the sub-segment is the 1st line, or the end line of the sub-segment is the H line, then delete all the lines of the sub-segment; otherwise, go to step (3);

(3) In the continuous candidate sub-segment conti_subseg, in the order from small to large, replace all rows of the sub-segment with the median (for an even number of candidate rows, take the average of the two middle values and then round up);

(4) Output a new segmented row sequence Can_seg_new for deleting redundant candidate segmented rows.

S7.2. Weak segmentation decision and fine segmentation of candidate segmentation regions

其中，对T2的值的确定实验，是为选择不同倍数的baseline_width作为T2，这些倍数都是<＝1的分数，执行满文部件切分方法，对于切分后的图像比较，选择出切分效果更好的满文单词图像所对应的T2，最终选择为上述T2值。Roughly cut Manchu parts may have weak cuts. Statistical results show that the height of Manchu parts generally does not exceed 5 times baseline_width, so the weak segmentation judgment threshold T_less=5 is set. The height hl of each segmented region in Can_seg_new is calculated, and the segmented region with height hl>(T_less×baseline_width) is determined as a weakly segmented region. For weakly segmented regions, the above-mentioned coarse segmentation method and fine segmentation threshold T2 are used to perform secondary segmentation, and are stored in the Seg1 sequence. The fine segmentation threshold T2 relaxes the constraints on the segmentation candidate rows on the basis of the coarse segmentation, and is determined through a large number of experiments

Among them, the determination experiment of the value of T2 is to select different multiples of baseline_width as T2, and these multiples are all fractions of <=1, perform the Manchu component segmentation method, and select the segmentation method for the image comparison after segmentation. The T2 corresponding to the Manchu word image with better effect is finally selected as the above T2 value.

S7.3. Over-segmentation decision and merging of candidate segmentation regions

After rough segmentation and fine segmentation, there may also be over-segmented regions in the Seg1 sequence. Statistical results show that the height of Manchu parts is generally greater than baseline_width, so the over-segmentation judgment threshold T_over=1 is set. Calculate the height ho of each segmented area in Seg1, and the segmented area with height ho<(T_over×baseline_width) is judged as an over-segmented area and needs to be merged. The merge will have the following situations:

1) Counting from top to bottom, if the first segmented area is judged to be over-segmented, it can only be merged with the second area;

2) From bottom to top, the penultimate area is judged to be over-segmented, it can only be merged with the penultimate area;

3) If the over-segmented area is located in the middle, the adjacent upper and lower areas need to be considered. Calculate the height h_up of the area merged with the area above, and the height h_lw of the area merged with the area below, and select the merging scheme with the smaller height after merging;

4) If the combined heights of the upper and lower regions are equal, that is, the merging scheme cannot be determined according to 3), then calculate the number of connected domains after merging with the upper and lower regions respectively, and select the merging scheme with a small number of connected domains;

5) Output the segmented rows after the regions are merged.

To do this, merge over-segmented regions using the following rules:

(1) If the first segmented area is over-segmented, merge with the second segmented area; otherwise, go to step (2).

(2) If the penultimate segmented area is over-segmented, it will be merged with the last segmented area; otherwise, go to step (3).

(3) If the over-segmented area is neither the first nor the penultimate second, calculate the heights h_up and h_lw of the adjacent upper and lower segmentation areas respectively. If h_up<h_lw, merge with the previous segmented area; if h_up>h_lw, merge into the next segmented area; otherwise, go to step (4).

(4) If the heights of the upper and lower adjacent regions of the over-segmented region are equal, then calculate the number of connected domains num_up and num_lw merged with the upper or lower regions, respectively. If num_up<num_lw, merge with the previous segmented area; if num_up>num_lw, merge with the next segmented area.

(5) Output the segmented row sequence that merges the over-segmented regions.

From the above scheme, the result of Manchu component segmentation is obtained, as shown in Figure 7, Figure 7(1)-(2) is the result of fine segmentation of the weakly segmented area; Figure 7(3)-(4) is over-segmentation The result of the merged regions.

Further processing is performed by the above-mentioned completed Manchu parts segmentation results to identify the Manchu parts, and the recognition method further comprises the following steps in addition to the segmentation of the above-mentioned Manchu word images:

(1) Normalization of Manchu parts

Including Manchu component position normalization and size normalization.

The position normalization of Manchu parts is to take the uppermost, lowermost, leftmost and rightmost pixels of the stroke pixels as the boundaries of the Manchu part image, cut off the background part, and only keep the part with strokes. Manchu component size normalization refers to normalizing the above image after position normalization to the same size (for example: 64 pixels×64 pixels).

(2) Manchu component feature extraction

Firstly, the methods commonly used in the feature extraction of ethnic minority characters are extracted respectively, including: outline feature, grid feature, direction line feature, visual direction feature and affine invariant distance feature. These features are then fused, and principal component analysis is used to reduce the dimensionality of the fused features.

(3) Manchu component recognition

A support vector machine classifier with Gaussian kernel function is used, and a "one pair rest" multi-classifier combination rule is used to realize the recognition of a Manchu part.

(4) Post-processing of Manchu parts identification,

For the recognized Manchu parts, according to the recognition results of the upper and lower adjacent parts and the spelling rules of Manchu letters, the reorganization from parts to words is completed, so as to realize the recognition of Manchu words.

Schedule 1:

Claims (1)

1. a method for extracting a central axis of a Manchu word image is characterized by comprising the following steps:

s1, positioning a central axis of a Manchu word:

s1.1, negating the Manchu word image, and enabling the pixel value of a character part to be 1 and the pixel value of a background part to be 0;

s1.2, using a morphological thinning function of an MATLAB image processing tool box to realize morphological thinning of the Manchu word image;

s1.3, determining a column coordinate corresponding to a refined central axis by using Hough transform on the morphologically refined Manchu word image, wherein the column coordinate is used as the position of the central axis of the Manchu word, and the angle of a Hough transform search straight line is limited to beθ=90, only straight lines in the vertical direction are searched, the straight lines at the same longitudinal position are connected, the straight lines with the distance smaller than the height of the Manchu word image and the length larger than 1 pixel are taken as a straight line, and the central position of the central axis is calculated;

s2, detecting the width of the central axis of the Manchu word:

s2.1, determining a search area of a maximum run proportion method:

the search area of the maximum run length ratio method is determined by the range specified by the following formula:

（1）

wherein,slis the left boundary of the defined search range,sris the right boundary of the defined search range,baselineis the central position of the central axis,roundmeaning that the rounding to the nearest integer is performed,Wis the width of the Manchu word image:

s2.2, applying a maximum run length proportion method to the Manchu word image in the search area to determine the width of a central axis of the Manchu word image: scanning each line of the Manchu word image search area, and counting the run length of continuous black pixels and the occurrence frequency of the run length, wherein the run length with the maximum occurrence frequency is the width of a central axis of the Manchu word:

s2.3, calculating the left boundary and the right boundary of the central axis according to the central position of the central axis of the Manchu word image and the width of the central axis: calculating the left boundary and the right boundary of the central axis according to the following formula;

(2)

wherein:blis the left boundary of the central axis,bris the right boundary of the central axis,baselineis the central position of the central axis of the Manchu word image,baseline_widthis the width of the central axis of the Manchu word image,roundindicating rounding to the nearest integer.

Images