TW319856B - Graph-text segmentation method of geographical image data - Google Patents

Abstract

A graph-text segmentation method, which is used to segment text and graph on geographical image after computer scanning process, comprises of the steps: (1) one preprocessing step includes: (a) binary processing, which selects one threshold value according to histogram of gray scale and pixel number, and by comparing gray value with threshold value to set black point or white point; (b) thinning processing, which performs thinning process to binarized geographic image; (c) patching distortion or redundant point after thinning process; (2) fetching feature point of graph image, in which the geographic image feature point is defined as one black point, among each feature points according to image appearing way is divided to multiple types with predetermined number, and feature point of the geographic image is fetched by one 3x3 mask to check the feature point type, and classify it;(3) extracting text block in image, which uses coordinate value of feature point to calculate clustering of each feature point, then classifies feature point cluster with larger clustering into one same group, if feature point number in the same group is larger than clustering density threshold, then the cluster is classified as text cluster, otherwise as graph cluster; the clustering calculation is based on maximum-minimum clustering algorithm, in which the selection of parameter initial cluster center is with respect to two point with the most far distance in all feature points, and parameter cluster radius size is average height of the smallest text; (4) after segmenting text cluster from geographic image, connecting left feature points of graph cluster.

Description

S856 A7 B7 Printed by the Employee Consumer Cooperative of the Central Bureau of Standards of the Ministry of Economy V. Description of the invention (1) The present invention relates to a method for separating images and texts, in particular to a map image that can be computerized and processed with map image data Separation of graphics and text in the data 'This method can effectively analyze image data, especially the part where graphics and text are glued or overlapped, so that the error rate of the analyzed image is minimized. Background of the invention: Information is not developed In the era of Geographic Information Management Technology (Geographic Information Management Technology) to help people solve the problem of a large number of information explosions, it provides a set of methods for effectively managing large amounts of geographic data, but users of geographic information systems (GIS) need to invest a large amount. The cost of obtaining important map information from the original map, because the traditional manual digitization method is inaccurate and expensive. Since the early 1980s, people have used scanners to provide an economical way to scan existing original maps into network maps and convert them into vector data. However, there is no set of effective and reliable The method can provide a high degree of accuracy and automatic conversion, especially for most products, without the ability to recognize land and pictures. Related papers on graphic vectorization have been successively proposed in recent years, but due to the characteristics of various countries on the map, the existing map automation processing system does not have a discriminatory method to deal with Chinese plots perfectly. Therefore, a set of Chinese map automation has been developed The processing system is necessary. However, in the development of this system, we found that the separation of graphics and text is a major bottleneck in map automation. Only after separating graphics and text, and then separately performing text recognition and graphics interpretation, can the entire automated system be established . _40083.DQC / SLH-4-This paper scale is applicable to China National Standard Rate (CNS) Λ4 specification (210X297mm) (please read the precautions on the back before filling this page)

3.δδ56 Α7 Β7_ V. Description of invention (2) (please read the notes on the back before filling out this page) There have been many papers in the past that discussed the topic of graphic separation, such as Fletcher using Hough Transform (Hough Transform) with the word Character Blocking (Character Blocking), to separate the text and lines on an electronic document map, the method is roughly as follows: (1) First calculate the connected block (Connected Component), and find out all the connected blocks on the map. (2) Find the density of dots in ghosts in each connected area, and simply distinguish text and line graphics according to the density. (3) For the text block, according to the central coordinates, applying the principle of Huo's transformation, find the letters on the same straight line, and then find the word (Word) according to the letter spacing. The advantage of this method is that it can clearly distinguish between strings, characters and characters. The maximum and minimum ratio of the text can be up to five times, and the direction of the text is not limited. However, not all text on the map is arranged in a straight line In addition, the graphics and text on the map are glued together, and the overlapping part of the graphics and text is often mistaken as a block, so the rate of miss is very high, so it cannot be effectively applied to the separation of graphics and text on the map. (See L. A. Fletcher and R. Kasturi, "A Robust Algorithm Printed by the Consumer Labor Cooperative of the Central Standards Bureau of the Ministry of Economy for Text String Separation from Mixed Text / Graphics

Images ”, IEEE Trans. Pattern Anal. Machine Intell, pp 910-918, Vol. 10, 1988) In order to solve the problem of image and text bonding, the method proposed by Kasturi in 1990 also aimed at the problem of image and text bonding. The block method is as follows: ⑴ First complete the method proposed by Fletcher, and do a preliminary separation of text. 40083.D0C / SLH · 5 ~ This paper scale is applicable to the Chinese National Standard (CNS) A4 specification (210X297 mm)- -A7 printed by the Staff Consumer Cooperative of the Central Bureau of Standards of the Ministry of Economic Affairs 5. Description of the invention (3 ⑵ Find the suspected image in the link block, and then restore its eight neighbors to find suspicious characters in the surrounding area until All points have been found. (3) Check whether the suspicious character can be inserted into a string. However, although this method can solve the simple problem of image and text bonding, it still cannot solve the problem of overlapping images. (See R . Kasturi and ST B〇w ,,, A System for Interpretation of Line Drawings11 JEEE Trans, Pattern Anal., Vol '· 12, No. 10, pp 978 · 991, 199〇), in addition, Wahl also proposed for document analysis Separate text The method of graphics and images, the steps of this method include: (1) According to the principle of run-length code, the black point of each column is coded as 1 and the white point is 0. (2) One of services] If the number of white dots between black dots is less than a certain critical value, change 〇 to 1. ⑶Similarly, repeat the steps (丨) and (2) for each row (γ _direction). ⑷ AND results. ⑸ Calculate the point density of the block obtained in (4). ⑹ Distinguish between text, graphics and images according to the size of the point density. Although this method can effectively distinguish text, graphics and images , But applied to the map, still can not solve the problem of image and text adhesion and image overlap. (See FM Wahl, Μ.K. Wang, and R, G. Casey, MBlock Segmentation and Text Extraction in Mixed

Text / Image Documents ", Comput. Vision, Graphics, and Image Processing, V〇l. 20, pp. 375-390, 1982)

Boatto et al. Also proposed a method to detect the side and separate the "picture and text bonding" and the factory picture and text # "but it is applied to the Chinese map, for the complex 40083.D0C / SLH _ β _ Μ's Zhang scale fit Compensation Standard (CNS) Α4 specifications (21Qx 297 public (please read the precautions on the back before filling in this f)) Printed A7 B7 printed by the Central Standard Falcon Bureau of the Ministry of Economic Affairs Consumer Cooperative V. Invention description (4) in Chinese The problem of overlapping characters and graphics is still not effectively dealt with (see L. Boatto et al. '' Detection and Separation of Symbols

Connected to Graphics in Line Drawings 丨 ', International

Conference on Pattern Recognition, Vol. 2, pp. 545-548, 1992) As for other related theories and technologies, some patents have been published, such as the Republic of China Patent No. 3 0 9 8 7, "H-Brush Frame Word Processing" has been revealed Graphic processing method of printed fonts, but this method is limited to framed printed fonts. For handwritten fonts on the map, even connected with graphics, it is helpless. 3 Given the current graphic processing methods still cannot be effectively solved The problem of image and text sticking and image and text overlapping, the present invention provides a new method of image and text separation. This method extracts feature points, and then judges text and lines according to the density of feature points. This method can separate different sizes. The text or the text is rotated, JL can effectively deal with the problems of "graphic and text bonding" and "graphic and text overlapping". According to the present invention, before the image and text are separated, some pre-processing must be done, such as binarization, thinning, and feature point extraction. The purpose of binarization is to remove unnecessary noise and improve the quality of the map. Thinning is decreasing, the storage space of the image, all graphics become a pixel (1_pixel width), so the intersection of any line can be determined to be a pixel, so that the characteristics of the text symbols on the map are all It can be clearly highlighted, because the feature point clustering of text symbols is higher than that of graphics, and this principle can be used to separate graphics and text. Regardless of the size of the text, the different direction of the text, or the location of the text ^, it can be processed. Yes, the most common problems in the analysis of documents such as “text” and “text” overlap J can be easily solved.

^ 0085.D0C / SLH

This paper scale is applicable to the Chinese National Falcon (CNS (please read the notes on the back before filling this page)

3 i & S56 A7 ～ ._____ B7____ Friends, invention description (5) 1 * ~~ If the principle and advantages of the present invention are explained, a preferred embodiment will be described in conjunction with the following drawings, among which: _ 一 系 本Invention flow block diagram Figure 1 is the detailed flow chart of the pre-processing in Figure 1 Figure 3A is the original image of the example of the two-valued processing in Figure 2 Figure 3B is the result of the two-valued conversion in Figure 3A Figure 4 is the series One thin line, the link code representation used for the conversion results. Figure 5 is the detailed flowchart of the graphic separation in Figure 1. Figure 6 is the definition of the feature point types in Figure 5. Figure 7 is used to explain that the text is a feature point group. Figure 8 shows that the clustering classification in Figure 5 will cause the line segment ab to disappear due to the classification error. Figure 9 shows that the line segment repair in Figure 5 uses the branch and chase method. Figure 10 shows that the object of line segment repair in Figure 5 is mining. The one with the smallest angle (d 1 > d2 > d3) FIG. 11A is an image of the input wall image according to an embodiment of the present invention. FIG. 11B is a thin line image of A. FIG. 11 is a C line diagram ^-B The result of thinning and repairing Figure 11D is the feature points extracted in Figure 11C. Printed by Voice Cooperative (Please read the ‘Notes on the back before filling this page

• IT Figure 11E is the image of the line left after drawing 11C from the text, Figure 11F is the result of the repair of the line segment of Figure 11E. Detailed description of the invention: The purpose of the present invention is directed to the aforementioned text The lack of separation technology, and provides a method for separating text and graphics when processing maps, different treatments are performed for parts of different nature, and text symbols are recognized. Graphics are vectorized to complete the map input Automation, the present invention proposes a new type of graphic separation technology that can solve all of the above _ ^ 〇〇85.D〇C / SLH_-8-t @ ® {CNS J A4 ^ (TIo'x 297 ^^.) ------__ The Central Standards Bureau of the Ministry of Economic Affairs, the Consumer Labor Cooperative Printed A 7 B7 V. Difficulties in the description of invention (6). The system architecture of the graphic separation of the present invention is shown in FIG. 1. First, a scanner is used to input the map original 101 into a personal computer, and a two-dimensional array (Array) is stored. Map pre-processing l〇2 (preprocessing), the quality of the image is processed better, in order to separate the graphics and text to carry out β This pre-processing method can be processed by the program written in visuaic ++ language, but not Limited to this. The steps of the pre-processing 102 described above can be achieved by methods currently available on the market. A preferred embodiment according to the present invention is shown in FIG. 2, which includes a binary processing 20, a thinning 2 〇2, and a thinning repair 2 0 3 other steps. Among them, the processing principle of the binary value 201 is to select a critical value based on the statistical graphs of gray scale and number of points. If the gray scale value is greater than this critical value, the point will be changed to a black point. Otherwise, it will be changed to a white point. Three illustrate the experimental results of the two-valued 001, Figure 3A is the original image, and Figure 3B is the two-valued result. Then, the thinned data is thinned to 202. According to this preferred embodiment, the thinning can use the method of Zhang & S ue η, and the result of thinning is obtained by a bonding code (C haincode) Means that this method is based on the coding method shown in Figure 4. All lines are stored not only their starting and ending coordinates, but also their key codes, so that when calculating the area

Get a more accurate value. (See T.Y, Zhang and C.Y. Suen ,,, A

Fast Thinning Algorithm for Thinning Digtal Patterns ", Comm. ACM, Vol. 27, pp. 236-239, 1984) However, the above thinning will have some problems, such as quadrants becoming two three-points and redundant points, etc., Therefore, a thin line repair step 2 0 3 needs to be performed. The purpose of the thin line repair 2 0 3 is to enable all feature points to be represented correctly, so that its coordinates can be used to accurately calculate the text 40083.DOC / SLH * 9 ~ The size of this paper adopts the Chinese National Standard (CNS) A4 specification (2 丨 OX 297mm) (Please read the precautions on the back before filling this page)

3iSS56 A7 ~ -----------___________ B7 V. Description of the invention. ) ^ — The clustering of words. The next step to enter the invention is the graphic separation step 103. As shown in Figure 5, the graphic separation step is first to extract the feature points of the image data (5〇1) and then measure the clustering of each feature point Character (step 5〇2), according to the clustering property of the arm measurement to do the cluster classification (step 5 0 3), judge the part of the text and graphics by the clustering of the classification, and separate the text (step aggregation 504), and finally repair the broken part of the map caused by the extraction of text. According to the present invention, the map image data is first defined as a feature point, and the feature point is defined as a black dot, and each feature The types of points are divided into three types, such as degree called, degree = 3 * degree == 4, as shown in π. When the feature points of the data are to be extracted, the way to judge them is to use a 3 X 3 mask ( mask) to determine the type of feature points where the number of degrees' is defined as the number of changes in the clockwise direction -O.

As for the type of the feature point, the coordinate value of the feature point is used as the basis for calculating the clustering. For the feature points with larger clustering, we will summarize them to the same cluster (c 1 uster). After all the feature points have been assigned, we can simply use the cluster density to determine whether a cluster is a text cluster or a graphic cluster. Q The cluster detection step 5 of this invention can Printed by the Employee Consumer Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs. Please read the precautions on the back and then fill out this page.) Group Minimum Rule (Maximum Minimum Distance Clustering

Algorithm), because the text has the following characteristics on the wall: (1) All Chinese characters exist at a distance from each other on the map. (2) After extracting the feature points, the Chinese characters are found to be a collection of feature points that are " combined with each other ’, as shown in Figure 7. Therefore, we can simply apply the maximum-minimum distance clustering rule to 40083. DQC / SLH-10 _ this standard 4 financial " transfer (CNS) A4 specifications (21QX Biechuchu) A7 B7 Ministry of Economic Affairs Central Standards Bureau staff consumption Printed by the cooperative. 5. Description of the invention (8) Calculate the distance between any two feature points, and classify the feature points into clusters when allowed / succeeded (see JT, Tou and R rr * Gonzalez, Pattern Recognition

Principles, Addison-Wesley Publishing Company. 1981) ○ The steps of the child grouping rule are briefly described as follows: (1) Choose a special feature arbitrarily as the grouping center (center) Z0. (2) In all but ZG〇 卜The feature point towel is the one farthest away from ^ as the ⑶. For all feature points, calculate the distance from the center of the finance. ⑷In the distance obtained by ⑶, find a minimum value MiNS. ⑸ Then among all minimum values Find—the maximum value. ⑹If the maximum value is MAXS> Z〇Zi / cluster radius, a new cluster center is generated, and repeat ⑶. ⑺Classify all points to the cluster to which they belong. At all feature points After the classification is completed, several clusters can be obtained, and then cluster classification 5 0 3, if the number of feature points in a cluster is greater than one "Cluster Density Threshold" (Cluster Density Threshold) ", then the cluster Clustering is classified as text clustering, otherwise it is a cluster of line graphics. Since the present invention considers the clustering of text from the perspective of feature points, regardless of the size or rotation of the text, the location of the placement, and the stickiness of the text Or text and text Common problems can be solved easily.

In the clustering step, there are 3 parameters that directly affect the result of the classification. The first is to select the initial cluster (Initial ClusteOZoiZ). The second is the size of each cluster radius. The last is each cluster. Feature point 40083.00C / SLH This paper scale is applicable to China National Standard (CNS) A4 specification (2I〇X297 public daughter) (please read the precautions on the back before filling this page)

319356 A7 B7 '-------- Employee consumption cooperation of the Central Bureau of Economics of the Ministry of Economic Affairs, printed by the company V. Description of invention (the number of 9 (that is, the critical value of cluster density) Usually, at the beginning, the cluster center is selected at any point As ^, then select the point farthest from 20 among the remaining points as 21. However, in practice, such z〇, ^ selection method does not necessarily guarantee that% 2 丨 is the farthest distance, there may be two points X iX j is greater than z 〇zi, so that the classification will be wrong, that is, some unnecessary cluster centers will be generated, because the value of the cluster radius will change due to the value of z〇Zi, so the present invention: Ensure that the minimum number of clusters can be generated. The selection of ^ and Z 1 is based on the two clusters with the farthest distance among all feature points as the initial cluster center. The second parameter is the selection of the cluster radius, if too large, many The feature 4 of the line will be known as a text cluster. If there are exactly two ends of a line segment that are classified as a text cluster, the line segment will be left behind, making the image and text separation incorrect. Figure 8 illustrates this Case β Xi Conversely, if the cluster radius is too small, a Chinese character may be subdivided into Small clusters, and cluster gatherings with too few feature points are mistaken for line graphics, and the actual measurement results, this value is the most appropriate for the average height of the smallest font. S The third parameter is in the cluster How many feature points can be summed up as text? In principle, the number of feature points in the text cluster should be greater than 丨, otherwise, the graphic cluster should be equal to 1, but to avoid artificial negligence, or small and long For the noise generated by the line segment, the preferred embodiment of the present invention sets the critical value of the cluster density to 3, which means that if the number of feature points in a cluster is greater than 3, the cluster will be text, otherwise It is a line graphic. Renzhang because the selected cluster radius is set to the average height of the smallest font _i °° e3.D0C / SLH _ 12-National National Rate (CNS) A4 specification (2) 0 > < 29 因 嫠) ----- (please read the back page first: > i matters, then fill out this page)

A 7 ______B7_ printed by the employee consumer cooperative of the Central Bureau of Standards and Economics of the Ministry of Economic Affairs 5. Invention description (10), so it is possible that several small characters are combined into a large collection, so the above method cannot guarantee that each picture and text is separated. The text clusters just represent a word, so it is necessary to further cut the text. The method used in the present invention is to use the neighbor-growing method to find 8 neighbors while updating the block size of the text. Finally, use a rectangular frame to frame this text. However, this method is only an illustrative example and is not intended to limit the present invention. After completing the text extraction 504, all text clusters will be removed, leaving the line and graphic clusters. Now the feature points in these clusters are connected, and the graphic and text separation is completed successfully. The method used is branching. Trace to ensure the existence of each line segment. The branch tracking of the present invention uses a two-way tracking method. As shown in Figure 9, line segment ab will be traced once by A-B, and reverse B-A will be traced once again. The purpose of this is to ensure that When any point of A and B is classified as text, line segment AB will not disappear. However, when "picture and text stick" or "picture and text overlap" occurs, the wire segment is bound to be interrupted after the picture and text are separated, so the present invention still needs to do the work of line segment repair 5 0 5, traditional repair The line segment method is to connect two line segments that meet all the following 3 conditions: (1) the two lines have the same slope; (2) the two broken lines have an endpoint degree = l; (3) the endpoint of two degree = l, which The distance is less than a certain critical value. However, in this way, when the overlapping of pictures and text occurs on a long and narrow line, an error will occur, as shown in Figure 10, when AB is truncated by a larger character and CD is by a smaller character Truncated, so as far as a £ is concerned, CG 'FB and HD all have the same slope and have one end point. 40083.DOC / SLH-13-This paper scale is applicable to the Chinese National Standard (CNS) A4 said grid (2 丨0X297mm) '' (Please read the notes on the back before filling this page)

* 1T 1 A7 B7 Fifth, the description of the invention (11) 8 I, but because d 3 is the smallest, EG will be connected, in order to avoid this error, the repair method adopted by the present invention is as follows: (please read the notes on the back first Please fill in this page again) (1) Take the AE vector as the benchmark; (2) Find the inner product of AE and EF, M, EG, and find its cool angle; (3) The angle with the most +, the vector is the repair Object. Through the above-mentioned feature point extraction, cluster detection, cluster classification, and text extraction and repair steps, the factory map image data can be accurately parsed out. Because the invention first defines image data in a "feature point" manner, and according to each The clustering of feature points is classified as text or graphics, so it can effectively separate the text from the map image, and the part where the graphics and text are glued or overlapped can also be effectively analyzed. The present invention can thus provide an accurate and effective method for image and text separation. The descriptions of the above embodiments are used to explain the principle of the present invention, not to limit the scope of the present invention. Therefore, the present invention can be modified by other methods. As long as the basic concepts are the same, they will not depart from the spirit of the present invention. The patent scope of this invention shall be as defined in the patent application scope listed below. The experiment results printed by the Employee Consumer Cooperative of the Central Bureau of Standards of the Ministry of Economic Affairs. The entire system of image and text separation was completed in a PC 3 8 6 compatible computer environment. All experimental images are 1024x768 Pixels. This is to facilitate the use of video memory ( V ide ο M em ο ny) as a buffer (Buffer). This method tests more than 20 images. The results are quite successful, here we give an example to introduce β The first experiment tests a cadastral map, as shown in Figure 11 (a), Figure 10 A0083.DOC / SLH ___ U _ This paper scale is applicable to Chinese national standards (CNS) A4 specification (210X297mm) A7 A7 Printed by the employee consumer cooperative of the Central Standardization Bureau of the Ministry of Economic Affairs V. Invention description (U)-⑻ is the result of thinning the line "Figure 11 ⑷ is the repairing of the thin line" Eleven ⑷ is for all feature points, Figure bu ⑷ is the result of Figure ^, Figure eleven (f) is the repair result of broken line, in which the picture and text have been successfully separated, and the text is repeated-the text in the overlapping part is also Divided away, the line is: the ground is repaired. The advantages of the present invention can be briefly described as follows: (1) Simple calculation can be easily made into hardware and added to the scanner without adding a lot of extra costs. (2) Using the characteristics of the clustering of character feature points, graphics and text can be separated. Even if the text is too small and different, the text is rotated, and the text is glued or overlapped, which can be effectively solved. (3) After the text with overlapping text is separated, the part where the graphic is removed can use the characteristics of directivity and angle to make up the line graphic β

40083.0OC / SLH This paper scale is applicable to China National Standard (CNS) A4 (2 丨 Ο297297) (Please read the precautions on the back before filling this page)

Claims (1)

Patent Application No. 84107775 3 i 9¾¾¾ Patent Amendment (April 86) A8 B8 C8 DS AW;], u The Ministry of Economic Affairs, Central Standard Falcon Bureau Beigong Consumer Cooperative System 6. Patent application scope = 3 1 • A method for separating graphics and texts 'to separate text and graphics on map images after computer scanning'. The method includes the following Steps: (1) A pre-processing step, including ... (a) Two-value processing 'This step is to select a critical value based on the statistical graph of gray levels and points. The comparison between the gray value and the critical value is set to Black dots or white dots; (b) Thin line processing: thinning the map image after binarization processing; and (〇 correction of distortions and redundant points after thin line processing. (2) Image The feature point of the map image is defined as a black dot. The feature points are divided into a preset number of various types according to the way the image appears. The feature points of the map image are captured by a 3> ; < 3 mask (mask) check the type of the feature point, and give the category; (3) extract the text block in the image, this step is to use the coordinate value of the feature point to calculate the group of each feature point Clustering, and then clustering Larger feature point clusters are grouped into the same cluster. If the number of feature points in the same cluster is greater than the "cluster density critical value", the cluster is classified as a text cluster. Otherwise, it is regarded as a line. The clustering of the graph; the step of calculating the clustering is based on the maximum-minimum clustering rule, the parameters used in the initial selection of the clustering center are for the two furthest points of all the feature points, and the parameter clustering The size of the radius is set to the average height of the smallest font; and (4) After the text is separated from the map image, the remaining line drawing paper scale is applicable to the Chinese standard Falcon (CNS) A4 specification (210 X 297mm) (Please read the precautions on the back before filling in this page) Connect the feature points of the '1Τ 泉 A8 B8 C8 D8' patent application type cluster. 2 · According to the method of item 丨 in the scope of patent application , Where, • The critical value of cluster density n is 3. 3. According to the method of item 丨 in the patent application park, where the connection method of step (4) is a two-way tracking method, to avoid any feature point being classified Shoulders An error β is generated according to the method of the first item in the scope of the patent application, in which step (4) still includes a graphic repairing step, which includes: (1) The vector specifying the endpoints of the two line segments as a reference; (2) The inner product of this vector datum and other adjacent line segment vectors, and find its cool poetry; and (3) The angle with the smallest angle is repaired. 5. According to the method of item 1 of the patent application garden, where the feature point type is A 3x3 mask that meets the point of degree = l, 3, or 4. (Please read the precautions on the back before filling out this page) The staff of the Central Standards Bureau of the Ministry of Economic Affairs of the Ministry of Economic Affairs and Jewelry Cooperative Society are jealous of this paper. Falcon (CNS) A4 see grid (2I0X297 public)