CN105160300A - Text extraction method based on level set segmentation - Google Patents

Abstract

The invention discloses a text extraction method based on level set segmentation. The method includes reading image data information and determining a border curve; performing graying on the read image; extracting graying feature values; dividing the image into two areas by adopting a level set function according to the graying feature value; performing binaryzation on the two areas obtained through division; performing connection element calibration on the two areas subjected to binaryzation; filtering connection elements calibrated in the two areas; performing polarity judgment on the areas subjected to filtering and determining a text pixel area and a background pixel area; filtering the text area and performing filtering for background noise removal; and outputting a text extraction result. By adopting the method provided by the invention, text information in a complex background can be extracted. Besides, extraction of image text containing outline letters is also very accurate. The method has certain universality and practicability.

Description

A kind of text abstracting method based on level-set segmentation

Technical field

The present invention relates to the text abstracting method in image processing field, particularly relate to a kind of text abstracting method based on level-set segmentation.

Background technology

Along with the development of network and computer technology, increasing information occurs with the multimedia form such as image or video.Contain abundant text message in image or video, these text messages play a part to illustrate and annotate to image or video.Extract and identify that these text messages have great significance to aspects such as image understanding, video content analysis, intelligent transportation, machine vision, Based Intelligent Control.But because text message is in complex background usually, general OCR system is difficult to identify text message.Thus text detection out also needed process and a text extraction process removing background before submitting to OCR system.Therefore, how from complex background image, to extract text message, becoming with text message is that clue is to understand a mission critical of picture material.

Existing image text extraction technique is mainly divided into the method for the method based on threshold value, the method based on cluster and Corpus--based Method model.Method based on threshold value mainly utilizes the segmentation of text and background color, and setting threshold value is by text and background separation.Threshold value chosen overall threshold values and local threshold values two kinds.The effect that this kind of method extracts depends on the discrimination of threshold values to image background and text, is generally applicable to the situation that image background is more single.Text block image is divided into K class by the method general colouring information based on cluster, then according to the threshold values of a certain clustering algorithm and setting by legal Type of Collective, the number of categories of minimizing color progressively.The last corresponding class wherein of text pixel, all the other are all kinds of is background.These class methods but when in background containing identical or close with textcolor composition time, these become branches to be divided into text class by mistake, thus produce a large amount of residual background, affects OCR identification.The method of Corpus--based Method model sets up probability model to all pixels in text block, then sets the parameter in rational probability model, then determines whether each pixel belongs to text pixel according to maximum likelihood rule.In probability model approach, model parameter generally needs statistical learning to obtain, and needs a large amount of learning samples.

Above-mentioned various text abstracting method, only make use of gray scale or the chromatic information of image bottom local, when extracting the text in complex background image or outline letters, often there is residual background, it is bad that text extracts effect.

Summary of the invention

Object of the present invention is exactly to solve the problem, and provides a kind of text abstracting method based on level-set segmentation.First adopt level set function that image is divided into two regions, then polarity judgement is carried out to two territories, judge text filed and background area, finally to text filed filtering, remove ground unrest.This process employs the full figure information of image, the text message in complex background can not only be extracted, and also very good to the extraction effect of outline letters image.There is certain versatility and practicality.

To achieve these goals, the present invention adopts following technical scheme:

Based on a text abstracting method for level-set segmentation, comprising:

Reads image data information, determines boundary curve; Gray processing is carried out to the image read; Extract gray feature value; Adopt level set function that image is divided into boundary curve inner region and boundary curve exterior domain according to gray feature value; Binaryzation is carried out to two regions be partitioned into; Respectively connected member demarcation is carried out to two regions of binaryzation; Filtering is carried out to the connected member demarcated in two regions; Polarity judging is carried out to filtered region, judges text pixel region and background pixel region; Filtering is carried out, filter out background noise to text filed; Export text and extract result.

Concrete steps comprise:

Step (1): Given Graph is as u ₀(x, y), (x, y) ∈ Ω, Ω are image-region, and ω is the open subset of Ω, and C is the boundary curve of ω, reading images information;

Step (2): to the image gray processing read;

Step (3): the gray feature value of abstract image;

Step (4): adopt level set function Iamge Segmentation to become boundary curve inner region and boundary curve exterior domain;

Step (5): judge whether segmentation completes, if completed, enters step (6), otherwise, return step (4);

Step (6): carry out binaryzation to two regions of segmentation, namely curve inner region black picture element represents, extra curvature region white pixel represents;

Step (7): adopt region growth method to carry out connected member demarcation respectively to the region of two after binaryzation;

Step (8): judge that connected member is demarcated and whether complete, enter step (9) if completed, otherwise, return step (7);

Step (9): filtering is carried out to the connected member in two regions;

Step (10): judge whether the filtering of two regional connectivity units completes, and enters step (11) if completed, otherwise, return step (9);

Step (11): carry out polarity judging to filtered two regions, to judge in two regions, which region is text filed; By comparing the number of connected member in two regions, it is text filed for getting the many regions of connected member number, and getting the few region of connected member number is background area;

Step (12): text filed to what determine, residual background is removed in further filtering;

Step (13): export text and extract result.

In described step (4), the energy function of level-set segmentation is:

Wherein, μ, v, λ ₁, λ ₂all normal numbers, c ₁, c ₂image u respectively ₀in (x, y), curved boundary C is inner puts down with outside gray scale

Average, H (z) and δ (z) represents Heaviside function H (z) and Dirac function δ (z) of regularization respectively; Wherein,

$\begin{array}{cc}H\left(z\right)=\left\{\begin{array}{cc}1,& z\&GreaterEqual;0\\ 0,& z<0\end{array}\right.;& \mathrm{\&delta;}\left(z\right)=\frac{d}{dz}H\left(z\right)\end{array}.$

Concrete grammar in described step (4) is:

Step (4-1): by boundary curve curve C level set function replace, as fruit dot (x, y) is inner in curve C, then as fruit dot (x, y) is outside in curve C, then if fruit dot (x, y) is in curve C, then

Step (4-2): initialization level set function, order k=0; for constant value;

Step (4-3): the energy function minimizing level set fixing be the K time iteration value, calculate c ₁ ^kand c ₂ ^kvalue;

Step (4-4): the energy function minimizing level set fixing c ₁ ^kand c ₂ ^k, calculate wherein when representing kth time iteration value;

Step (4-5): judge solution whether tend towards stability, if not tending towards stability, then another k=k+1, returns step (4-3), continues interative computation, otherwise stops iteration entering step (4-6);

Step (4-6): output level set function segmentation result.

C is calculated during described step (4-3) kth time iteration ₁and c ₂the method of value is:

Wherein, u ₀(x, y) is the point on Given Graph picture, for the Heaviside function of regularization.

Calculate concrete grammar be:

Utilize the c calculated in step (4-3) ₁ ^kand c ₂ ^k, first calculate according to the following formula then integration is obtained

Wherein, div represent divergence operator, represent gradient operator, μ, v, λ ₁, λ ₂all normal numbers, c ₁, c ₂image u respectively ₀the average gray that in (x, y), curved boundary C is inner and outside.

In described step (7) to the method that the region of two after binaryzation adopts region growth method to carry out connected member demarcation be respectively:

Step (7-1): search for from top to bottom, left to right respectively the pixel in region, marks if search pixel, then compose the mark number that this pixel is new;

Step (7-2): with the pixel newly marked for starting point carries out 8 neighborhood search, if in its 8 neighborhood search to unlabelled pixel, unmarked pixel then for searching composes identical label, and with the pixel newly marked for starting point carries out 8 neighborhood search;

Step (7-3): if do not search unlabelled pixel in 8 neighborhoods, then terminate this search;

Step (7-4): judge that all pixels have marked whether; Step (7-5) is entered if completed; Enter step (7-1) if do not completed, unlabelled pixels all in region are marked, until complete all pixels mark;

Step (7-5): will there is the pixel of identical label as a connected member.

In described step (9) to the method for connected member filtering be:

Judge the number of pixel in the position of connected member in two regions and connected member respectively, if connected member is connected with border or in connected member, pixel number is less than setting threshold value, then this connected member is deleted.

In described step (11), to the method that polarity judging is carried out in filtered two regions be:

Step (11-1): will there is the pixel of identical label as a connected member in two regions after filtering;

Step (11-2): the number adding up connected member in two regions respectively, if the number of connected member is respectively n in two regions ₁and n ₂;

Step (11-3): compare n ₁and n ₂if, n ₁> n ₂, then n ₁corresponding region is text filed, otherwise n ₂corresponding region is text filed.

In described step (12), text filed to what determine, the method removing residual background is further:

By the average gray of connected member each in statistical regions, and by the average gray of each connected member by order arrangement from small to large, then the difference of neighboring gradation mean value is calculated, then successively the threshold value of gray scale difference value and setting is compared, if gray scale difference value is greater than setting threshold value, then using this difference as segmentation position, after all difference judgements terminate, obtain N number of segmentation position, get that section that in each segmentation, corresponding pixel number is maximum and be text filed section, connected member corresponding to text area segments is text connected member, corresponding to text connected member, position is text filed, other region in image is background area.

The invention has the beneficial effects as follows:

The present invention, according to the feature of complex background image Chinese version information, first adopts level set function to Image Segmentation Using, then carries out polarity judgement, background filtering to cut zone, obtain text and extract result.This process employs the global information of text image, the text message in complex background image can not only be extracted, and it is also very accurate to extract effect to the text of outline letters, has certain versatility and practicality, avoids residual background to the impact extracting result.The achievement of this invention can directly apply to the fields such as image understanding, video content analysis, intelligent transportation, machine vision, Based Intelligent Control, has broad application prospects.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of a kind of text abstracting method based on level-set segmentation of the present invention.

Embodiment:

Below in conjunction with accompanying drawing and embodiment, the present invention will be further described:

The basic hardware condition realized needed for system architecture of the present invention is: a dominant frequency is 2.4GHZ, inside saves as the computing machine of 1G, and required software condition is: programmed environment is visual c++.

Based on a text segmenting method for level set movable contour model, as shown in Figure 1, concrete steps are as follows:

Step (1): start, reading images;

Step (2): to the image gray processing read;

Step (3): the gray feature value of abstract image;

Step (4): be used in level set function and Iamge Segmentation become two regions;

Given Graph is as u ₀(x, y), (x, y) ∈ Ω, Ω is called as image-region, and ω is the open subset of Ω, and C is the boundary curve of ω, curve C available horizontal set function replace, as fruit dot (x, y) is inner in curve C, then as fruit dot (x, y) is outside in curve C, then if fruit dot (x, y) is in curve C, then

Level set energy function can be expressed as:

Wherein, μ, v, λ ₁, λ ₂normal number, c ₁, c ₂image u ₀the average gray that in (x, y), curved boundary C is inner and outside, H (z) and δ (z) represents Heaviside function H (z) and Dirac function δ (z) of regularization respectively

$H\left(z\right)=\left\{\begin{array}{cc}1,& z\&GreaterEqual;0\\ 0,& z<0\end{array}\right.---\left(2\right)$

$\mathrm{\&delta;}\left(z\right)=\frac{d}{dz}H\left(z\right)---\left(3\right)$

Minimization of energy function, fixing c can be estimated ₁, c ₂value,

Then, fixing c ₁, c ₂, minimization of energy function, can obtain

Specific implementation step is:

Step (4-1): initialization level set function, order k=0, chooses 5 circles as level set initialization curve in the present invention;

Step (4-2): according to formula (4), (5) calculate c ₁ ^kand c ₂ ^k;

Step (4-3): according to the c calculated ₁ ^kand c ₂ ^k, calculate according to formula (6)

Step (4-4): judge whether solution tends towards stability, if do not had, another k=k+1, forwards step (4-2) to, continues interative computation, otherwise stops iteration entering step (4-5);

Step (4-5): output level collection segmentation result.

Step (5): judge whether segmentation completes, if completed, enters step (6), if do not completed, enters step (4);

Step (6): carry out binaryzation to two regions of segmentation, namely curve inner region black picture element represents, extra curvature region white pixel represents;

Step (7): 8 connected member demarcation are carried out to employing region, two regions growth method be partitioned into;

Concrete steps are:

Step (7-1): search for from top to bottom, left to right respectively the pixel in region, marks if search pixel, then compose the mark number that this pixel is new;

Step (7-2): with the pixel newly marked for starting point carries out 8 neighborhood search, if in its 8 neighborhood search to unlabelled pixel, unmarked pixel then for searching composes identical label, and with the pixel newly marked for starting point carries out 8 neighborhood search;

Step (7-3): if do not search unlabelled pixel in 8 neighborhoods, then terminate this search;

Step (7-4): judge that all pixels have marked whether.Step (7-5) is entered if completed; Enter step (7-1) if do not completed, unlabelled pixels all in region are marked, until complete all pixels mark;

Step (7-5): will there is the pixel of identical label as a connected member.

Step (8): judge that connected member is demarcated and whether complete, enter step (9) if completed, return step (7) if do not completed;

Step (9): filtering is carried out to the connected member in two regions, judge the number of pixel in the position of connected member in two regions and connected member respectively, if connected member is connected with border or in connected member, pixel number is less than given threshold value, then this connected member is deleted.

Step (10): judge whether the filtering of two regional connectivity units completes, and enters step (11), enter step (9) if do not completed if completed;

Step (11): carry out polarity judging to latter two region of filtering, to judge in two regions, which region is text filed.The relatively number of pixel contained by two regions, the region that capture vegetarian refreshments number is many is text filed, and the region that pixel is few is background area;

Concrete steps are:

Step (11-1): will there is the pixel of identical label as a connected member in two regions after filtering;

Step (11-2): the number adding up connected member in two regions respectively, if the number of connected member is respectively n in two regions ₁and n ₂;

Step (11-3): compare n ₁and n ₂if, n ₁> n ₂, then n ₁corresponding region is text filed, otherwise n ²corresponding region is text filed.

Step (12): text filed to what determine, residual background is removed in further filtering;

Concrete steps are:

Step (12-1): the average gray asking each connected member in region;

Step (12-2): each connected member average gray is arranged according to order from small to large;

Step (12-3): calculate the difference between each average gray and average gray adjacent thereafter;

Step (12-4): difference step (12-3) obtained compares, if difference is greater than the threshold value of setting, then using this difference as segmentation position with the threshold value of setting respectively;

Step (12-5): judge more all differences complete with threshold value, enters step (12-6) if completed, and enters step (12-4) if do not completed;

Step (12-6): obtain N number of segmentation position after comparing end altogether, each connected member is divided into N+1 section by this N number of segmentation position;

Step (12-7): the number of adding up pixel contained by connected member corresponding to each section in N+1 section respectively, connected member corresponding to the segmentation that number of pixels is maximum is text connected member, the region that text connected member is corresponding is text filed, and region corresponding to all the other segmentations is background area.

Step (12-8): delete background area.

Step (13): export text and extract result.

By reference to the accompanying drawings the specific embodiment of the present invention is described although above-mentioned; but not limiting the scope of the invention; one of ordinary skill in the art should be understood that; on the basis of technical scheme of the present invention, those skilled in the art do not need to pay various amendment or distortion that creative work can make still within protection scope of the present invention.

Claims (10)

1., based on a text abstracting method for level-set segmentation, it is characterized in that, comprising:

Reads image data information, determines boundary curve; Gray processing is carried out to the image read; Extract gray feature value; Adopt level set function that image is divided into boundary curve inner region and boundary curve exterior domain according to gray feature value; Binaryzation is carried out to two regions be partitioned into; Respectively connected member demarcation is carried out to two regions of binaryzation; Filtering is carried out to the connected member demarcated in two regions; Polarity judging is carried out to filtered region, judges text pixel region and background pixel region; Filtering is carried out, filter out background noise to text filed; Export text and extract result.

2. a kind of text abstracting method based on level-set segmentation as claimed in claim 1, it is characterized in that, concrete steps comprise:

Step (1): Given Graph is as u ₀(x, y), (x, y) ∈ Ω, Ω are image-region, and ω is the open subset of Ω, and C is the boundary curve of ω, reading images information;

Step (2): to the image gray processing read;

Step (3): the gray feature value of abstract image;

Step (4): adopt level set function Iamge Segmentation to become boundary curve inner region and boundary curve exterior domain;

Step (5): judge whether segmentation completes, if completed, enters step (6), otherwise, return step (4);

Step (6): carry out binaryzation to two regions of segmentation, namely curve inner region black picture element represents, extra curvature region white pixel represents;

Step (7): adopt region growth method to carry out connected member demarcation respectively to the region of two after binaryzation;

Step (8): judge that connected member is demarcated and whether complete, enter step (9) if completed, otherwise, return step (7);

Step (9): filtering is carried out to the connected member in two regions;

Step (10): judge whether the filtering of two regional connectivity units completes, and enters step (11) if completed, otherwise, return step (9);

Step (11): carry out polarity judging to filtered two regions, to judge in two regions, which region is text filed; By comparing the number of connected member in two regions, it is text filed for getting the many regions of connected member number, and getting the few region of connected member number is background area;

Step (12): text filed to what determine, residual background is removed in further filtering;

Step (13): export text and extract result.

3. a kind of text abstracting method based on level-set segmentation as claimed in claim 2, it is characterized in that, in described step (4), the energy function of level-set segmentation is:

Wherein, μ, v, λ ₁, λ ₂all normal numbers, c ₁, c ₂image u respectively ₀the average gray that in (x, y), curved boundary C is inner and outside, H (z) and δ (z) represents Heaviside function H (z) and Dirac function δ (z) of regularization respectively; Wherein,

$\begin{array}{cc}H\left(z\right)=\left\{\begin{array}{cc}1,& z\&GreaterEqual;0\\ 0,& z<0\end{array}\right.;& \mathrm{\&delta;}\left(z\right)=\frac{d}{dz}H\left(z\right)\end{array}.$

4. a kind of text abstracting method based on level-set segmentation as claimed in claim 3, it is characterized in that, the concrete grammar in described step (4) is:

Step (4-1): by boundary curve curve C level set function replace, as fruit dot (x, y) is inner in curve C, then as fruit dot (x, y) is outside in curve C, then if fruit dot (x, y) is in curve C, then

Step (4-2): initialization level set function, order k=0; for constant value;

Step (4-3): the energy function minimizing level set fixing be the K time iteration value, calculate c ₁ ^kand c ₂ ^kvalue;

Step (4-4): the energy function minimizing level set fixing c ₁ ^kand c ₂ ^k, calculate wherein when representing kth time iteration value;

Step (4-5): judge solution whether tend towards stability, if not tending towards stability, then another k=k+1, returns step (4-3), continues interative computation, otherwise stops iteration entering step (4-6);

Step (4-6): output level set function segmentation result.

5. a kind of text abstracting method based on level-set segmentation as claimed in claim 3, is characterized in that, calculates c during described step (4-3) kth time iteration ₁and c ₂the method of value is:

Wherein, u ₀(x, y) is the point on Given Graph picture, for the Heaviside function of regularization.

6. a kind of text abstracting method based on level-set segmentation as claimed in claim 3, is characterized in that, calculates concrete grammar be:

Utilize the c calculated in step (4-3) ₁ ^kand c ₂ ^k, first calculate according to the following formula then integration is obtained

Wherein, div represent divergence operator, represent gradient operator, μ, v, λ ₁, λ ₂all normal numbers, c ₁, c ₂image u respectively ₀the average gray that in (x, y), curved boundary C is inner and outside.

7. a kind of text abstracting method based on level-set segmentation as claimed in claim 2, is characterized in that, in described step (7) to the method that the region of two after binaryzation adopts region growth method to carry out connected member demarcation be respectively:

Step (7-1): search for from top to bottom, left to right respectively the pixel in region, marks if search pixel, then compose the mark number that this pixel is new;

Step (7-2): with the pixel newly marked for starting point carries out 8 neighborhood search, if in its 8 neighborhood search to unlabelled pixel, unmarked pixel then for searching composes identical label, and with the pixel newly marked for starting point carries out 8 neighborhood search;

Step (7-3): if do not search unlabelled pixel in 8 neighborhoods, then terminate this search;

Step (7-4): judge that all pixels have marked whether; Step (7-5) is entered if completed; Enter step (7-1) if do not completed, unlabelled pixels all in region are marked, until complete all pixels mark;

Step (7-5): will there is the pixel of identical label as a connected member.

8. a kind of text abstracting method based on level-set segmentation as claimed in claim 2, is characterized in that, in described step (9) to the method for connected member filtering be:

Judge the number of pixel in the position of connected member in two regions and connected member respectively, if connected member is connected with border or in connected member, pixel number is less than setting threshold value, then this connected member is deleted.

9. a kind of text abstracting method based on level-set segmentation as claimed in claim 2, is characterized in that, in described step (11), to the method that polarity judging is carried out in filtered two regions be:

Step (11-1): will there is the pixel of identical label as a connected member in two regions after filtering;

Step (11-2): the number adding up connected member in two regions respectively, if the number of connected member is respectively n in two regions ₁and n ₂;

Step (11-3): compare n ₁and n ₂if, n ₁> n ₂, then n ₁corresponding region is text filed, otherwise n ₂corresponding region is text filed.

10. a kind of text abstracting method based on level-set segmentation as claimed in claim 2, is characterized in that, in described step (12), text filed to what determine, and the method removing residual background is further:

By the average gray of connected member each in statistical regions, and by the average gray of each connected member by order arrangement from small to large, then the difference of neighboring gradation mean value is calculated, then successively the threshold value of gray scale difference value and setting is compared, if gray scale difference value is greater than setting threshold value, then using this difference as segmentation position, after all difference judgements terminate, obtain N number of segmentation position, get that section that in each segmentation, corresponding pixel number is maximum and be text filed section, connected member corresponding to text area segments is text connected member, corresponding to text connected member, position is text filed, other region in image is background area.