CN105160300B - A kind of text abstracting method based on level-set segmentation - Google Patents

Abstract

The invention discloses a text extraction method based on level set segmentation, which includes: reading image data information, determining the boundary curve; graying the read image; extracting gray feature values; The set function divides the image into two areas; binarizes the two divided areas; calibrates the connected elements of the two areas after binarization; filters the marked connected elements in the two areas; The polarity of the filtered area is judged, and the text pixel area and the background pixel area are judged; the text area is filtered to filter out the background noise; and the text extraction result is output. The invention not only can extract the text information in the complex background, but also can extract the image text containing outline characters very accurately, and has certain versatility and practicability.

Description

A Text Extraction Method Based on Level Set Segmentation

technical field

The invention relates to a text extraction method in the field of image processing, in particular to a text extraction method based on level set segmentation.

Background technique

With the development of network and computer technology, more and more information appears in the form of multimedia such as images or videos. Images or videos contain rich text information, which can explain and explain the images or videos. Extracting and recognizing these text information is of great significance to image understanding, video content analysis, intelligent transportation, machine vision, intelligent control, etc. However, since text information is usually in a complex background, it is difficult for general OCR systems to recognize text information. Therefore, before the text is detected and submitted to the OCR system, a process of removing the background is required, that is, the text extraction process. Therefore, how to extract text information from complex background images has become a key task to understand image content using text information as clues.

Existing image text extraction techniques are mainly divided into threshold-based methods, cluster-based methods and statistical model-based methods. The threshold-based method mainly uses the segmentation of text and background color, and sets the threshold to separate the text and background. There are two types of threshold selection: global threshold and local threshold. The extraction effect of this method depends on the degree of discrimination between the image background and the text by the threshold value, and is generally applicable to situations where the image background is relatively single. The clustering-based method generally uses color information to divide the text block image into K categories, and then aggregates the categories that meet the rules according to a certain clustering algorithm and a set threshold, gradually reducing the number of color categories. The text pixels finally correspond to one of them, and the rest are backgrounds. However, when the background contains components with the same or similar color as the text, these components will be misclassified into the text class, resulting in a large amount of residual background, which affects OCR recognition. The method based on the statistical model establishes a probability model for all pixels in the text block, then sets reasonable parameters in the probability model, and then determines whether each pixel belongs to a text pixel according to the maximum likelihood rule. In the probabilistic model method, the model parameters generally need statistical learning, which requires a large number of learning samples.

The various text extraction methods mentioned above only use the local grayscale or color information of the bottom layer of the image. When extracting text or outline characters in a complex background image, there is often a residual background, and the text extraction effect is not good.

Contents of the invention

The object of the present invention is to provide a text extraction method based on level set segmentation in order to solve the above problems. First, the level set function is used to divide the image into two regions, and then the polarity of the two regions is judged to determine the text region and the background region. Finally, the text region is filtered to remove the background noise. This method utilizes the full image information of the image, not only can extract the text information in the complex background, but also has an ideal extraction effect on the outline image. It has certain versatility and practicality.

In order to achieve the above object, the present invention adopts the following technical solutions:

A text extraction method based on level set segmentation, comprising:

Read the image data information and determine the boundary curve; grayscale the read image; extract the grayscale feature value; divide the image into the area inside the boundary curve and the area outside the boundary curve according to the gray level feature value using the level set function; Perform binarization on the two divided areas; perform connected element calibration on the two binarized areas; filter the marked connected elements in the two areas; determine the polarity of the filtered area to determine the text The pixel area and the background pixel area; filter the text area to filter out the background noise; output the text extraction result.

Specific steps include:

Step (1): Given an image u ₀ (x, y), (x, y) ∈ Ω, Ω is the image area, ω is the open subset of Ω, and C is the boundary curve of ω, read the image information;

Step (2): Grayscale the read image;

Step (3): extracting the grayscale feature value of the image;

Step (4): using the level set function to segment the image into regions inside the boundary curve and regions outside the boundary curve;

Step (5): judge whether the segmentation is completed, if it is completed, then enter step (6), otherwise, return to step (4);

Step (6): Binarize the two divided regions, that is, the region inside the curve is represented by black pixels, and the region outside the curve is represented by white pixels;

Step (7): The region growth method is used to calibrate the connected elements of the two regions after binarization;

Step (8): Determine whether the calibration of connected elements is completed, if it is completed, enter step (9), otherwise, return to step (7);

Step (9): filter the connected elements in the two regions;

Step (10): Judging whether the filtering of the connected elements of the two regions is completed, if completed, enter step (11), otherwise, return to step (9);

Step (11): Determine the polarity of the two regions after filtering to determine which of the two regions is the text region; by comparing the number of connected elements in the two regions, take the region with the largest number of connected elements as the text region , take the area with a small number of connected elements as the background area;

Step (12): For the determined text area, further filter to remove the residual background;

Step (13): output the text extraction result.

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

Among them, μ, v, λ ₁ , λ ₂ are all positive constants, c ₁ , c ₂ are the gray level of the inside and outside of the curve boundary C in the image u ₀ (x, y) respectively.

The mean, H(z) and δ(z) represent the regularized Heaviside function H(z) and Dirac function δ(z) respectively; where,

The concrete method in the described step (4) is:

Step (4-1): use the level set function of the boundary curve curve C Instead, if the point (x,y) is inside the curve C, then If the point (x,y) is outside the curve C, then If the point (x,y) is on the curve C, then

Step (4-2): Initialize the level set function, let k=0; is a constant value;

Step (4-3): Minimize the energy function of the level set fixed for the Kth iteration The value of , calculate the value of c ₁ ^k and c ₂ ^k ;

Step (4-4): Minimize the energy function of the level set Fixing c ₁ ^k and c ₂ ^k , calculate in Indicates that when the kth iteration value;

Step (4-5): Judgment Whether the solution tends to be stable, if it is not stable, then another k=k+1, return to step (4-3), and continue iterative operation, otherwise stop iteration and enter step (4-6);

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

The method for calculating _c1 and _c2 values during the kth iteration of the step (4-3) is:

Among them, u ₀ (x, y) is a point on the given image, is the regularized Heaviside function.

calculate The specific method is:

Using the c ₁ ^k and c ₂ ^k calculated in step (4-3), first calculate according to the following formula Then integrate to find

Among them, div represents the divergence operator, Represents the gradient operator, μ, v, λ ₁ , λ ₂ are all positive constants, c ₁ , c ₂ are the average gray levels of the inside and outside of the curve boundary C in the image u ₀ (x, y) respectively.

In the step (7), the method of using the region growth method to calibrate the connected elements in the two regions after binarization is as follows:

Step (7-1): Search the pixels in the area in order from top to bottom and from left to right. If the searched pixel point is not marked, assign a new label number to the pixel point;

Step (7-2): Use the newly marked pixel as the starting point to perform 8-neighborhood search. If an unmarked pixel is found in its 8-neighborhood, assign the same label to the searched unmarked pixel, and Perform 8-neighborhood search with the newly marked pixel as the starting point;

Step (7-3): If no unmarked pixel is found within the 8 neighborhoods, then end the search;

Step (7-4): Determine whether all pixel points are marked; if completed, proceed to step (7-5); if not complete, proceed to step (7-1), mark all unmarked pixels in the area until completed until all pixels are marked;

Step (7-5): Take the pixels with the same label as a connected element.

In described step (9), the method to connected element filtering is:

The position of the connected element in the two regions and the number of pixels in the connected element are respectively judged. If the connected element is connected to the boundary or the number of pixels in the connected element is less than the set threshold, the connected element is deleted.

In described step (11), the method for polarity judgment is carried out to two regions after filtering is:

Step (11-1): After filtering, use the pixels with the same label in the two regions as a connected element;

Step (11-2): count the number of connected elements in the two areas respectively, and set the number of connected elements in the two areas to be _n1 and _n2 respectively;

Step (11-3): compare n ₁ and n ₂ , if n ₁ >n ₂ , then the area corresponding to n ₁ is a text area, otherwise the area corresponding to n ₂ is a text area.

In the step (12), for the determined text region, the method for further removing the residual background is:

By counting the average gray value of each connected element in the area, and arranging the average gray value of each connected element in ascending order, then calculating the difference between the adjacent gray average values, and then sequentially Values are compared with the set threshold, if the gray difference is greater than the set threshold, then this difference is used as the segment position, after all the differences are judged, N segment positions are obtained, and the corresponding The segment with the largest number of pixels is the text region segment, the connected elements corresponding to the text region segment are the text connected elements, the corresponding position of the text connected elements is the text region, and the other regions in the image are the background regions.

The beneficial effects of the present invention are:

According to the characteristics of the text information in the complex background image, the invention firstly uses the level set function to segment the image, and then performs polarity judgment and background filtering on the segmented area to obtain the text extraction result. This method utilizes the global information of the text image, not only can extract the text information in the complex background image, but also the text extraction effect of the outline characters is very accurate, has certain versatility and practicability, and avoids the influence of the residual background on the extraction results. influences. The achievements of this invention can be directly applied to image understanding, video content analysis, intelligent transportation, machine vision, intelligent control and other fields, and have broad application prospects.

Description of drawings

FIG. 1 is a flowchart of a text extraction method based on level set segmentation in the present invention.

Detailed ways:

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

The basic hardware condition needed to realize the system structure of the present invention is: a main frequency is 2.4GHZ, and the internal memory is a computer with 1G, and the required software condition is: the programming environment is Visual C++.

A text segmentation method based on the level set active contour model, as shown in Figure 1, the specific steps are as follows:

Step (1): start, read the image;

Step (2): Grayscale the read image;

Step (3): extracting the grayscale feature value of the image;

Step (4): using the level set function to divide the image into two regions;

Given an image u ₀ (x, y), (x, y) ∈ Ω, Ω is called the image region, ω is an open subset of Ω, C is the boundary curve of ω, and the curve C can use the level set function Instead, if the point (x,y) is inside the curve C, then If the point (x,y) is outside the curve C, then If the point (x,y) is on the curve C, then

The level set energy function can be expressed as:

Among them, μ, v, λ ₁ , λ ₂ are positive constants, c ₁ , c ₂ are the average gray values of the inside and outside of the curve boundary C in the image u ₀ (x, y), H(z) and δ(z ) represent the regularized Heaviside function H(z) and Dirac function δ(z) respectively

Minimize the energy function, fixed The values of c ₁ and c ₂ can be estimated,

Then, fix c ₁ , c ₂ , and minimize the energy function, we can get

The specific implementation steps are:

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

Step (4-2): Calculate c ₁ ^k and c ₂ ^k according to formulas (4) and (5);

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

Step (4-4): judge whether the solution tends to be stable, if not, otherwise k=k+1, go to step (4-2), continue iterative operation, otherwise stop iteration and enter step (4-5);

Step (4-5): Output level set segmentation results.

Step (5): Judging whether the segmentation is completed, if completed, then enter step (6), if not, then enter step (4);

Step (6): Binarize the two divided regions, that is, the region inside the curve is represented by black pixels, and the region outside the curve is represented by white pixels;

Step (7): Perform 8-connected element calibration on the two segmented regions using the region growth method;

The specific steps are:

Step (7-1): Search the pixels in the area in order from top to bottom and from left to right. If the searched pixel point is not marked, assign a new label number to the pixel point;

Step (7-2): Use the newly marked pixel as the starting point to perform 8-neighborhood search. If an unmarked pixel is found in its 8-neighborhood, assign the same label to the searched unmarked pixel, and Perform 8-neighborhood search with the newly marked pixel as the starting point;

Step (7-3): If no unmarked pixel is found within the 8 neighborhoods, then end the search;

Step (7-4): Judging whether all pixel markings are completed. If it is completed, enter step (7-5); if it is not completed, enter step (7-1), mark all unmarked pixels in the area until all pixel points are marked;

Step (7-5): Take the pixels with the same label as a connected element.

Step (8): Determine whether the calibration of connected elements is completed, if completed, enter step (9), if not complete, return to step (7);

Step (9): Filter the connected elements in the two regions, respectively judge the position of the connected elements in the two regions and the number of pixels in the connected elements, if the connected elements are connected to the boundary or the number of pixels in the connected elements is small If it is lower than a given threshold, the connected element is deleted.

Step (10): Judging whether the filtering of connected elements in two regions is completed, if completed, enter step (11), if not completed, enter step (9);

Step (11): Polarity determination is performed on the filtered two regions to determine which of the two regions is the text region. Compare the number of pixels contained in the two areas, take the area with more pixels as the text area, and the area with fewer pixels as the background area;

The specific steps are:

Step (11-1): After filtering, use the pixels with the same label in the two regions as a connected element;

Step (11-2): count the number of connected elements in the two areas respectively, and set the number of connected elements in the two areas to be _n1 and _n2 respectively;

Step (11-3): compare n ₁ and n ₂ , if n ₁ >n ₂ , then the area corresponding to n ₁ is a text area, otherwise the area corresponding to n ² is a text area.

Step (12): For the determined text area, further filter to remove the residual background;

The specific steps are:

Step (12-1): Calculate the average gray value of each connected element in the region;

Step (12-2): Arrange the average gray value of each connected element in ascending order;

Step (12-3): Calculate the difference between each gray-scale average value and the adjacent gray-scale average value;

Step (12-4): The difference obtained in step (12-3) is compared with the set threshold respectively, if the difference is greater than the set threshold, then use this difference as the segmentation position;

Step (12-5): judge whether all difference and threshold value comparisons are completed, if complete, enter step (12-6), if not complete, enter step (12-4);

Step (12-6): After the comparison, a total of N segmentation positions are obtained, and the N segmentation positions divide each connected element into N+1 segments;

Step (12-7): Count the number of pixels contained in the connected elements corresponding to each segment in the N+1 segment respectively. The connected element corresponding to the segment with the largest number of pixels is a text connected element, and the text connected element corresponds to The area is the text area, and the area corresponding to the rest of the segments is the background area.

Step (12-8): Delete the background area.

Step (13): output the text extraction result.

Although the specific implementation of the present invention has been described above in conjunction with the accompanying drawings, it does not limit the protection scope of the present invention. Those skilled in the art should understand that on the basis of the technical solution of the present invention, those skilled in the art do not need to pay creative work Various modifications or variations that can be made are still within the protection scope of the present invention.

Claims (8)

1. A text extraction method based on level set segmentation is characterized by comprising the following steps:

reading image data information and determining a boundary curve; graying the read image; extracting a gray characteristic value; dividing the image into an area inside the boundary curve and an area outside the boundary curve by adopting a level set function according to the gray characteristic value; carrying out binarization on the two divided regions; respectively calibrating connected elements of the two binarized areas; filtering the calibrated connected elements in the two areas; judging the polarity of the filtered area to judge a text pixel area and a background pixel area; filtering the text region to filter background noise; outputting a text extraction result;

the method comprises the following specific steps:

step (1): given image u₀(x, y), (x, y) belongs to omega, omega is an image area, omega is an open subset of omega, C is a boundary curve of omega, and image information is read;

step (2): graying the read image;

and (3): extracting a gray characteristic value of the image;

and (4): dividing the image into an area inside the boundary curve and an area outside the boundary curve by adopting a level set function;

and (5): judging whether the segmentation is finished or not, if so, entering the step (6), and otherwise, returning to the step (4);

and (6): binarizing the two divided regions, namely representing the region inside the curve by using black pixels and representing the region outside the curve by using white pixels;

and (7): respectively adopting a region growing method to calibrate the connected elements of the two binarized regions;

and (8): judging whether the calibration of the connected element is finished, if so, entering the step (9), and otherwise, returning to the step (7);

and (9): filtering connected elements in the two regions;

step (10): judging whether the filtering of the two area connected elements is finished, if so, entering the step (11), and otherwise, returning to the step (9);

step (11): judging the polarity of the two filtered areas to judge which area of the two areas is a text area; comparing the number of the connected elements in the two regions, taking the region with a large number of the connected elements as a text region, and taking the region with a small number of the connected elements as a background region;

step (12): further filtering the determined text area to remove residual background;

step (13): outputting a text extraction result;

in the step (4), the energy function of the level set segmentation is as follows:

wherein, mu, v, lambda₁,λ₂Are all normal numbers, c₁,c₂Are respectively an image u₀(x, y) mean gray levels inside and outside the boundary curve C, h (z) and δ (z) representing regularized Heaviside function h (z) and Dirac function δ (z), respectively;representing a level set function, (x, y) belongs to omega, and omega is an image area; wherein,

2. the method for extracting text based on level set segmentation as claimed in claim 1, wherein the specific method in step (4) is as follows:

step (4-1): using the boundary curve C as a level set functionInstead, if the point (x, y) is inside the boundary curve C, thenIf the point (x, y) is outside the boundary curve C, thenIf the point (x, y) is on the boundary curve C, then

Step (4-2): initialize level set function, orderk＝0；Is a constant value;as a function of the level setAn initial value of (1);

step (4-3): minimizing energy function of level setFixingFor the k-th iterationA value of (c) is calculated₁ ^kAnd c₂ ^kA value of (d); c. C₁ ^kIs the mean value of the gray levels inside the boundary curve C at the k-th iteration, C₂ ^kThe gray level average value outside the boundary curve C in the k iteration is obtained;

step (4-4): minimizing energy function of level setFastening of c₁ ^kAnd c₂ ^kCalculatingWhereinWhen representing the (k + 1) th iterationA value of (d);

step (4-5): judgment ofIf not, returning to the step (4-3) to continue the iterative operation, otherwise, stopping the iteration and entering the step (4-6);

step (4-6): and outputting a level set function segmentation result.

3. The method as claimed in claim 2, wherein c is calculated at the k-th iteration of the step (4-3)₁And c₂The method of the value is:

wherein u is₀(x, y) is the number of given images,is a regularized Heaviside function.

4. The method of claim 2, wherein the computing is based on a level set segmentation text extraction methodThe specific method comprises the following steps:

using c calculated in step (4-3)₁ ^kAnd c₂ ^kFirst, according to the following formulaThen, the integral is calculated

Wherein div represents the divergence operator, ▽ represents the gradient operator, μ, v, λ₁,λ₂Are all normal numbers, c₁,c₂Are respectively an image u₀The gray level average value inside and outside the boundary curve C in (x, y).

5. The method for extracting text based on level set segmentation as claimed in claim 1, wherein the method for performing connected component calibration on the two binarized regions by using the region growing method in step (7) comprises:

step (7-1): searching pixels in the region from top to bottom and from left to right respectively, and assigning a new mark number to the pixel if the pixel is not marked;

step (7-2): carrying out 8 neighborhood search by taking the newly marked pixel point as a starting point, if an unmarked pixel point is searched in the 8 neighborhood, assigning the same label to the searched unmarked pixel point, and carrying out 8 neighborhood search by taking the newly marked pixel point as the starting point;

step (7-3): if the unmarked pixel point is not searched in the 8-neighborhood, the search is finished;

step (7-4): judging whether all the pixel point marks are finished or not; entering step (7-5) if the step is finished; if the step (7-1) is not finished, marking all unmarked pixel points in the area until all pixel point marks are finished;

step (7-5): and taking the pixel points with the same label as a connected element.

6. The method as claimed in claim 1, wherein the method for extracting text based on level set segmentation in step (9) comprises the following steps:

and respectively judging the positions of the connected elements in the two regions and the number of pixel points in the connected elements, and deleting the connected elements if the connected elements are connected with the boundary or the number of the pixel points in the connected elements is less than a set threshold value.

7. The method for extracting text based on level set segmentation as claimed in claim 1, wherein in the step (11), the method for determining the polarity of the two filtered regions comprises:

step (11-1): after filtering, taking the pixel points with the same label in the two areas as a connected element;

step (11-2): respectively counting the number of connected elements in the two regions, and respectively setting the number of the connected elements in the two regions as n₁And n₂；

Step (11-3): comparison of n₁And n₂If n is₁＞n₂Then n is₁The corresponding region is a text region, otherwise n₂The corresponding region is a text region.

8. The method for extracting text based on level set segmentation as claimed in claim 1, wherein in the step (12), for the determined text region, the method for further removing the residual background comprises:

the gray level average value of each connected element in the region is counted, the gray level average values of the connected elements are arranged from small to large, then the difference value of the adjacent gray level average values is calculated, the gray level difference value is compared with a set threshold value in sequence, if the gray level difference value is larger than the set threshold value, the difference value is used as a segmentation position, after all the difference values are judged, N segmentation positions are obtained, the section with the largest number of corresponding pixel points in each segmentation is taken as a text region section, the connected element corresponding to the text region section is taken as a text connected element, the position corresponding to the text connected element is taken as a text region, and other regions in the image are taken as background regions.