CN104992176A - Inscription oriented Chinese character extracting method - Google Patents

Abstract

The present invention provides an inscription oriented Chinese character extracting method, mainly comprising two parts: a Chinese character region positioning method based on Area Voronoi and a single character segmentation method based on a communication field and a bounding box; and the method comprises the steps of: firstly, for a distribution rule and feature of Chinese characters in an inscription image, acquiring communication fields in the image as generators, and representing the contour of each communication field by using a Freeman encoding chain; constructing an Area Voronoi diagram of the image by using a contour boundary extension method; then according to a standard deviation between the communication fields, merging adjacent Area Voronoi regions so as to complete the positioning operation of a Chinese character region; after acquiring an image of the Chinese character region, performing pre-processing on the image and removing a certain noise influence; and by concluding structure features of the Chinese characters, combining a communication field analysis method with a bounding box overlapping judgment method, and performing single character extraction on the Chinese characters. According to the inscription oriented Chinese character extracting method provided by the present invention, for an arrangement rule of ancient inscriptions, the extraction work for a single Chinese character can be better completed, and the method has a certain noise immunity.

Description

A kind of Chinese character extracting method towards an inscription on a tablet

Technical field

The invention belongs to computer vision and computer graphic image process field, the specifically a kind of method of image Chinese version location and character extraction, the method conveniently for the research of Chinese character in ancient times, can be of great significance probing into Chinese character development law tool at all times.

Background technology

An inscription on a tablet, as a kind of traditional art form of calligraphy, have recorded Chinese character development from ancient times to the present and the course of differentiation.But because an inscription on a tablet generally all lives through the very long history age, also to suffer the impact of spontaneous corrosion etc. simultaneously, make it be unfavorable for permanent preservation and propagation.Along with the development of the universal of computing machine and related scans equipment, an inscription on a tablet can be scanned into digital picture and store in a computer, effectively extracts Chinese character in an inscription on a tablet and is of great significance follow-up structure analysis method and the development law tool of probing into Chinese character at all times.

The method main following three kinds: one of at present carrying out String localization to image is the method based on connected domain, as the people such as Zhang propose a framework based on high-order markov random file, Chinese character region can be determined whether according to the posterior probability of maximum-likelihood criterion and node state.Such algorithm has good locating effect for the image that background is comparatively single, but its performance depends on initial connected domain to be generated, and the image comparatively complicated for background texture then faces test; Two is the methods based on edge, and these class methods lay particular emphasis on the high-contrast between word and background, and time efficiency is higher, but can not remove complicated grain background, often combines with other algorithms, as initial coarse positioning; Three is the methods based on texture, the people such as Wu utilize the Gaussian filter of three kinds of different scales to image carry out filtering go forward side by side line linearity conversion, propose a kind of multi-dimension texture segmentation text positioning method.These class methods have higher robustness usually, but the calculated amount of carrying out texture analysis is large, complexity is high owing to needing, comparatively consuming time.

Present stage carries out the research of Chinese character extraction for an inscription on a tablet image and seldom, due to the Chinese character arrangement disunity in an inscription on a tablet, the structure of Chinese character is comparatively complicated, makes general Chinese character extracting method cannot be directly applied for this field thus obtain good effect.

Summary of the invention

For overcoming above-mentioned shortcoming, the object of the invention is to the Chinese character Regional Distributing Regularity probed in an inscription on a tablet image, the design feature of research Chinese character, analyzes existing String localization and individual character segmentation technique simultaneously, and the Chinese character for an inscription on a tablet image obtains and proposes simply a kind of and effective solution.

In order to achieve the above object, the present invention proposes a kind of Chinese character extracting method towards an inscription on a tablet, it is characterized in that comprising the following steps:

Step (1), the Area Voronoi adopting the method for extended boundary to build an inscription on a tablet image scheme, and each Voronoi area represents a connected domain, and each connected domain can be divided into Chinese character connected domain and non-Chinese character connected domain according to wherein whether comprising Chinese character; Due to the discreteness of Chinese character, the separation between individual Chinese character stroke can be caused in the process building Area Voronoi figure, thus belong in different connected domains, therefore for Chinese character connected domain, only comprise part Chinese character, be called the assembly of Chinese character;

Step (2), according to the otherness between connected domain element width standard deviation, remove non-Chinese character connected domain, and merge all Chinese character connected domains, thus the whole Chinese character region in an inscription on a tablet image can be oriented;

Step (3), TV smooth operation is carried out to Chinese character area image remove certain noise spot, carry out binaryzation afterwards, make image only comprise black and white two kinds of pixels and facilitate subsequent operation; Calculate the row projection histogram in bianry image, according to the intermittent rule of an inscription on a tablet calligraphy column distribution, be partitioned into single-row Chinese character region;

Step (4), the connected domain extracted in Chinese character image row, this connected domain may comprise single complete Chinese character, also may be the combination of the different piece of different Chinese character, therefore needs to recombinate to connected domain; Analyze nine kinds of common structural characteristics of Chinese character, the relation between connected domain boundary rectangle is summarized as three kinds of situations: upper and lower relation, left-right relation and overlapping relation, effectively extract Chinese character according to these three kinds of relations.

Further, the step building Area Voronoi figure in described step (1) is as follows:

Step (11), first connected domain division is carried out to image, the corresponding closed polygon of each connected domain, and each connected domain and the background dot that is not divided into connected domain are marked;

Step (12), adopt Freeman coding strand to represent the profile of each connected domain, and on profile and the pixel of contoured interior carry out initialization, initial value is the mark of corresponding profile;

The mode of step (13), employing spread F reeman coding strand expands the border of connected domain; Judge each point on extended contour, if this point is also positioned on another one connected domain profile simultaneously, so this pixel can be considered a Voronoi frontier point, and arranging its value is EDGE_POINT simultaneously; This point is a background dot else if, then arrange the mark that its value is corresponding profile;

Step (14), the expansion of iteration, judgement, each connected domain of mark, when pixels all in plane all no longer comprises background pixel, namely represent all pixels or belong to some Voronoi area, for the frontier point in region, this completes the structure of the Area Voronoi figure of image.

Principle of the present invention is: first for an inscription on a tablet image obtained, AreaVoronoi according to the method design of graphics picture of extended boundary schemes, each Voronoi in this figure contains an effective connected domain, this connected domain can be divided into Chinese character connected domain and non-Chinese character connected domain according to whether comprising Chinese character, by certain criterion, merge Chinese character connected domain, remove non-Chinese character connected domain, thus orient the Chinese character region in an inscription on a tablet image.Next be the extraction operation of individual Chinese character, equally first take each Chinese character in the method preliminary characterization region of connected domain analysis, but the Chinese character comprised in the region now marked off imperfect, be the result of adjacent Chinese characters combination.In order to further divided area, based on nine kinds of structures that Chinese character is common, the relation of the boundary rectangle of induction and conclusion three kinds of connected domains, be respectively: upper and lower relation, left-right relation and overlapping relation, utilize this three kinds of relations, carry out next step sub-dividing operations for the connected domain with overlap, thus extract effective Chinese character.

Chinese character at present for an inscription on a tablet image extracts research also seldom, the characteristic rule analyzing Chinese character distribution in an inscription on a tablet image deep in the present invention, and the advantage compared with the text extraction techniques of existing general pattern is:

(1) the discrete sparse characteristic distributions of Chinese character in an inscription on a tablet image, is considered, by the mode building Area Voronoi figure, Chinese character region is divided, carry out region merging technique according to the skeleton width of Chinese character simultaneously, Chinese character region can be oriented effectively, targetedly.

(2), in order to avoid the impact of noise, have employed the level and smooth method of TV first and remove noise spot tiny in an inscription on a tablet image, and then carry out binaryzation operation.Simultaneously in the process of individual character cutting, carry out induction and conclusion to common nine kinds of Hanzi structure rules, propose the segmentation algorithm based on connected domain and bounding box thereof, experimental result shows, this algorithm simple and effectively can extract individual Chinese character.

Accompanying drawing explanation

Fig. 1 illustrates in the present invention and schemes for the Area Voronoi constructed by an inscription on a tablet image;

Fig. 2 illustrates Central Plains of the present invention an inscription on a tablet image and the result figure after carrying out Chinese character zone location, and wherein, figure (a) is former an inscription on a tablet image, and figure (b) is for carrying out the result figure after Chinese character zone location;

Fig. 3 (a), (b), (c), (d), (e), (f), (g), (h), (i) illustrate in the present invention and conclude nine kinds of common Hanzi structures;

Fig. 4 illustrates the definition to connected domain bounding box overlapping relation in the present invention;

Fig. 5 illustrates the false code of individual character segmentation algorithm in the present invention;

Fig. 6 illustrates the Chinese character segmentation result figure of an inscription on a tablet image in the present invention, and wherein scheming (a) is an inscription on a tablet image, the Chinese character segmentation result figure that figure (b) is an inscription on a tablet image;

Fig. 7 illustrates the Chinese character segmentation result figure of noisy an inscription on a tablet image in the present invention, and wherein scheming (a) is noisy an inscription on a tablet image, the Chinese character segmentation result figure that figure (b) is noisy an inscription on a tablet image;

Fig. 8 illustrates the schematic diagram of a kind of Chinese character extracting method towards an inscription on a tablet of the present invention.

Embodiment

By reference to the accompanying drawings the embodiment of the present invention is described in detail.

Implementation process of the present invention is mainly divided into four steps: structure, the Voronoi area of Area Voronoi figure merge, an inscription on a tablet image column is split, individual character cutting.

The structure of step one, Area Voronoi figure, is mainly divided into six basic steps:

(1) first Region dividing is carried out to image, obtain the connected domain of n non-overlapping copies, be designated as G={g ₁, g ₂..., g _n, the corresponding closed polygon of each connected domain, its mark value be k (k=1,2 ..., n).The value of the background pixel be not divided in image in connected domain is set to BACKGROUND_VALUE simultaneously.

(2) be arranged in each pixel of connected domain inside, arrange mark value k that its pixel value is this connected domain (k=1,2 ..., n).

(3) utilize Freeman coding strand to represent the profile of each connected domain, be designated as IC _k.And the pixel value of the point be positioned on profile is set to k+M, wherein M is a constant.

(4) each connected domain g of initialization _kstate variable F _k:=true, wherein k=1,2 ..., n.

(5) for each connected domain g _kjudge its state variable, if F _k==false, then export this connected domain; If F _k==true, carries out following steps.

(5.1) according to Freeman coding strand IC _k, obtain the extended contour chain EC of this profile _k.

(5.2) for profile chain IC _kon each pixel, if its value is not EDGE_VALUE, its value is set to k (replacing original k+M).

(5.3) for extended contour chain EC _keach pixel, compare the size of its value, if between M and M+k, so this pixel is the point be positioned on other connected domain profiles, be designated as a Voronoi edge point, and the value arranging this point is EDGE_VALUE.Otherwise if value is BACKGROUND_VALUE, illustrate that this point is a background dot, arranging its value is corresponding profile value k+M.

(5.4) if the institute on extended contour is not a little background dot, then F is made _k==false.

(5.5)IC _k:＝EC _k。

(6) structure of Area Voronoi figure is completed.

Step 2, Voronoi area merge.

As shown in Figure 1, each connected domain in image Voronoi area represents, but each region is divided into Chinese character connected domain and non-Chinese character connected domain according to whether comprising Chinese character assembly, Chinese character zone location is and merges Chinese character connected domain, removes non-Chinese character connected domain.The standard deviation of element width in adjacent connected domain is adopted whether to belong to Chinese character region as measurement two Voronoi area in the present invention, whether similar, later the need of merging.

Its key step is:

(1) Euclidean distance transformation matrix is set up, obtain the skeleton of each connected domain, then the minor increment of each point on same skeleton to connected domain border is calculated, using the width thickness (g of the average of 2 of the average of these distances times as assembly in connected domain _k), be specially:

$thickness\left(g_k\right)=\frac{1}{n}\underset{1}{\overset{n}{\Σ}}2\×min\left(d\right(s_i,C_k\left)\right)---\left(1\right)$

Wherein, s _kfor a bit on skeleton, n is the number that skeleton is put, C _krepresent the profile of connected domain, min (d (s _i, C _k)) represent and skeleton puts s _kto the bee-line on profile border.

(2) choose be positioned at picture centre Voronoi area as prime area, calculate the standard deviation between the connected domain adjacent with this connected domain.Wherein, any two connected domain g _iand g _jbetween the standard deviation of element width be defined as:

$std(g_i,g_j)=\left|\frac{thickness\left(g_i\right)-thickness\left(g_j\right)}{2}\right|---\left(2\right)$

(3) size of standard of comparison difference, because standard deviation occurs that larger value mainly concentrates between non-Chinese character connected domain, therefore set suitable threshold value, the size namely by standard deviation determines whether Chinese character connected domain.This threshold value is defined as:

$T_{std}=\frac{{\max }_{i,j}std(g_i,g_j)}{2}---\left(3\right)$

Step 3, an inscription on a tablet image column are split

After orienting the Chinese character region in an inscription on a tablet image, need to carry out binary conversion treatment to image, the convenient operation with carrying out column split.But because in image, itself also has certain noise, directly carry out the segmentation result that binaryzation can make noise effect follow-up, therefore first adopt TV smoothing algorithm to carry out pretreatment operation to image in the present invention, filter out partial noise in image and carry out binaryzation operation again.Why choose TV smoothing algorithm, therefore the method can be good at the structural information retaining Chinese character, and the situation of other smoothing algorithm blurred pictures can not occur.

Arrange in a column direction according to an inscription on a tablet calligraphy work and there is certain regularity, adopt the method for histogram projection, add up the number of picture black pixel in each row, owing to there is certain background area (i.e. white pixel) between two row Chinese characters, make to occur local minimum in histogram projection, present the phenomenon of " trough ", can thus as the foundation of row cutting.

Simultaneously in order to avoid the outshot of some Chinese character is for the interference of cutting, the method adopting both connected domain and Contour extraction to combine processes.Main process is as follows:

(1) when row a certain in cutting image run into target pixel points, to the corresponding connected domain of image zooming-out near this point, frontier point X identical with this horizontal ordinate in connected domain is extracted.

(2) carry out Contour tracing operation, with the shortest distance join target pixel points and frontier point X, this shortest profile is designated as cut-point sequence.

Step 4, individual character cutting.

Analyze the formation of Chinese character, be mainly made up of basic stroke (point, horizontal, vertical, slashs, right-falling stroke etc.), each stroke is interconnected, and forms the radical of Chinese character, also can be described as the parts of Chinese character.By analyzing the relation in Chinese character between parts, nine kinds of situations as shown in Figure 3 can be obtained, be summarised as up-down structure (such as prosperous, morning etc.), tiled configuration (such as you, say), comprise structure (such as scheme, be stranded), semi-surrounding structure (such as sentence, huge etc.).When picture quality is poor, when dividing connected domain, by the different component clusterings of same Chinese character in different connected domains, also by the different component clusterings of different Chinese character in same connected domain, can cause the situation of " adhesion " simultaneously.Using the boundary rectangle of connected domain as the feature of connected domain, nine of Chinese character kind of modular construction can be reduced to three kinds of relations of connected domain and external matrix: upper and lower relation, left-right relation and overlapping relation.

As shown in Figure 4, two connected domain g are illustrated _iand g _jthere is overlapping relation, with (L _i, U _i) and (R _i, D _i) be respectively g _ithe corresponding upper left corner of boundary rectangle and the coordinate in the lower right corner represent.The related definition of connected domain is as follows:

Wide Width (g _i)=R _i-L _i

High Height (g _i)=D _i-U _i

The ratio of width to height $Ratio\left(g_i\right)=\frac{Width\left(g_i\right)}{Height\left(g_i\right)}$

Area Square (g _i)=Width (g _i) × Height (g _i)

Adjacent connected domain carries out merging later related definition:

Width Uwidth (g _i, g _j)=max (R _i, R _j)-min (L _i, L _j)

Height Uheight (g _i, g _j)=max (D _i, D _j)-min (U _i, U _j)

The ratio of width to height $Uratio(g_i,g_j)=\frac{Uwidth(g_i,g_j)}{Uheight(g_i,g_j)}$

Overlapping width Owidth (g _i, g _j)=Width (g _i)+Width (g _j)-Uwidth (g _i, g _j)

Overlapping height Oheight (g _i, g _j)=Height (g _i)+Height (g _j)-Uheight (g _i, g _j)

Overlapping area Osquare (g _i, g _j)=Owidth (g _i, g _j) × Oheight (g _i, g _j)

Unit defined above carries out representing with the number of pixel.

After quantization means is carried out for the overlapping relation of connected domain, just can carry out the operation of next step individual character cutting.

After obtaining single-row Chinese character region, and each row image zooming-out connected domain wherein, be designated as C _k(k=1,2 ..., M), wherein M is the number of connected domain.The individual Chinese character that the expression that this M connected domain has is complete, has plenty of, be different Chinese character different parts between combination, need further to segment it, merge the different assemblies of same Chinese character.The false code of its main segmentation algorithm as shown in Figure 5.

Utilize above-mentioned implementation process, to the operation of the Chinese character extraction that muting an inscription on a tablet image and noisy an inscription on a tablet image carry out simultaneously, as shown in Figures 6 and 7, result shows that the present invention has certain validity and noise immunity to its result.

Claims (2)

1., towards a Chinese character extracting method for an inscription on a tablet, it is characterized in that comprising the following steps:

Step (1), the Area Voronoi adopting the method for extended boundary to build an inscription on a tablet image scheme, and each Voronoi area represents a connected domain, and each connected domain can be divided into Chinese character connected domain and non-Chinese character connected domain according to wherein whether comprising Chinese character; Due to the discreteness of Chinese character, the separation between individual Chinese character stroke can be caused in the process building Area Voronoi figure, thus belong in different connected domains, therefore for Chinese character connected domain, only comprise part Chinese character, be called the assembly of Chinese character;

Step (2), according to the otherness between connected domain element width standard deviation, remove non-Chinese character connected domain, and merge all Chinese character connected domains, thus the whole Chinese character region in an inscription on a tablet image can be oriented;

Step (3), TV smooth operation is carried out to Chinese character area image remove certain noise spot, carry out binaryzation afterwards, make image only comprise black and white two kinds of pixels and facilitate subsequent operation; Calculate the row projection histogram in bianry image, according to the intermittent rule of an inscription on a tablet calligraphy column distribution, be partitioned into single-row Chinese character region;

Step (4), the connected domain extracted in Chinese character image row, this connected domain may comprise single complete Chinese character, also may be the combination of the different piece of different Chinese character, therefore needs to recombinate to connected domain; Analyze nine kinds of common structural characteristics of Chinese character, the relation between connected domain boundary rectangle is summarized as three kinds of situations: upper and lower relation, left-right relation and overlapping relation, effectively extract Chinese character according to these three kinds of relations.

2. as claimed in claim 1 towards the Chinese character extracting method of an inscription on a tablet, it is characterized in that: the step building Area Voronoi figure in described step (1) is as follows:

Step (11), first connected domain division is carried out to image, the corresponding closed polygon of each connected domain, and each connected domain and the background dot that is not divided into connected domain are marked;

Step (12), adopt Freeman coding strand to represent the profile of each connected domain, and on profile and the pixel of contoured interior carry out initialization, initial value is the mark of corresponding profile;

The mode of step (13), employing spread F reeman coding strand expands the border of connected domain; Judge each point on extended contour, if this point is also positioned on another one connected domain profile simultaneously, so this pixel can be considered a Voronoi frontier point, and arranging its value is EDGE_POINT simultaneously; This point is a background dot else if, then arrange the mark that its value is corresponding profile;

Step (14), the expansion of iteration, judgement, each connected domain of mark, when pixels all in plane all no longer comprises background pixel, namely represent all pixels or belong to some Voronoi area, for the frontier point in region, this completes the structure of the Area Voronoi figure of image.