CN101567049B - Method for processing noise of half tone document image - Google Patents

Abstract

The invention relates to the fields of computer vision, graphics and image processing, and discloses a noise processing method for halftone document images. It includes the following steps: firstly, halftone document images are divided into four categories; secondly, for halftone document images with light background and dark text, the method based on connected region labeling is used to remove background noise; for light background and light For halftone document images with dark text, the method based on Gaussian smoothing filtering is used to remove background noise; for halftone document images with dark background and dark text, halftone document images with dark background and light text, Wiener-based Filtering method to remove background noise; finally, observe the halftone document image with background noise removed, construct text burr removal template and defect repair template, and remove text burrs and repair text defects through template matching.

Description

A Noise Processing Method for Halftone Document Image

technical field

The present invention relates to the fields of computer vision, graphics and image processing, in particular to a noise processing method for halftone document images, which can be applied to text denoising and text extraction of various halftone document images such as scanned document images and fax images, and can be used It is further applied to the fields of text recognition and full-text retrieval of halftone image documents.

Background technique

With the popularization of printers, scanners and facsimile machines, halftone document images are widely used in various fields of social life. Therefore, in order to realize text recognition and full-text retrieval of halftone document images, text denoising and text extraction technologies for halftone document images are very necessary. A traditional method is to inverse halftone the halftone image and convert it into the corresponding grayscale image. However, after this method, the image usually produces serious blurring, making the details of the text part blurred. Lost, thereby reducing the text recognition rate. Another traditional method is to directly process the halftone dots in the image, which can effectively avoid the blurring of text. However, how to remove the image part and graphic part in the halftone dot image while ensuring that the image is not lost as much as possible? The details of the text part become a key problem to be solved in this technology.

Contents of the invention

The object of the present invention is to provide a noise processing method for a halftone document image, which can remove background noise in the halftone document image, eliminate burr noise of characters, and maintain the integrity and smoothness of character strokes.

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

A noise processing method for a halftone document image, comprising the following steps:

First, divide halftone document images into four categories: light background and dark text, light background and light text, dark background and dark text, and dark background and light text;

Secondly, remove the background noise of the halftone document image: for the halftone document image with light background and dark text, the method based on connected region labeling is used to remove the background noise; for the halftone document image with light background and light text, Use a method based on Gaussian smoothing filtering to remove background noise; for halftone document images with dark background and dark text, and halftone document images with dark background and light text, use a method based on Wiener filtering to remove background noise;

Finally, observe the halftone document image with background noise removed, construct text deburring templates and defect repair templates, and remove text burrs and repair text defects through template matching.

Further improvements and features of the present invention are:

(1) The described halftone document image is divided into four classes, specifically: carry out Gaussian smoothing filter to halftone document image, calculate the Otsu threshold value and the gray level histogram of halftone document image after Gaussian smoothing filter, estimate according to Otsu threshold value The halftone dot density of the background, the halftone dot density of the text is estimated according to the gray histogram, and the halftone document image is divided into four categories according to the halftone dot density of the background and the halftone dot density of the text, namely light background and dark text, light background and light text, dark background and dark text, dark background and light text.

(2) The deburring template for the text is expressed in a table, including:

Among them, a cell represents a pixel, and the positional relationship between cells represents the positional relationship between pixels; in a cell, 1 represents a text pixel, 0 represents a background pixel, and ① represents a text pixel that needs to be removed, that is, the text pixel 1 needs to be changed to is background pixel 0; empty cells are represented as background pixel 0 or text pixel 1.

(3) The defect repair template is represented by a table, including the following table:

表示需要修复的文字像素，即需要把背景像素0变为文字像素1；空单元格表示为背景像素0或文字像素1。Among them, the cell represents the pixel, and the positional relationship between the cells represents the positional relationship between the pixels; in the cell, 1 represents the text pixel, 0 represents the background pixel,

Indicates the text pixel that needs to be repaired, that is, the background pixel 0 needs to be changed to text pixel 1; an empty cell is represented as background pixel 0 or text pixel 1.

The present invention divides the halftone document image into four categories, and then adaptively selects the most suitable background noise removal method for these four categories, and can keep as many characters as possible while performing preliminary denoising of the halftone document image. Then, through the construction of a series of text deburring templates and defect repair templates, through template matching, deburring and defect repair are performed on the text of the preliminary denoised halftone document image, so that further OCR recognition of text becomes easier. more easily.

The present invention is more effective for halftone document images based on error diffusion. On the one hand, the halftone technology based on error diffusion can be simulated by an approximate linear model, so the method of Wiener filtering is more effective for such halftone images; On the one hand, the halftone technology based on error diffusion uses the density of halftone dots instead of the size of halftone dots to convert grayscale images into halftone images, so the method based on Unicom region marking is more effective for this type of halftone images. efficient.

Description of drawings

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments.

FIG. 1 is an overall flowchart of a noise processing method for a halftone document image.

FIG. 2 is a grayscale histogram of a halftone document image after Gaussian smoothing, wherein the abscissa is the grayscale level ranging from 0 to 255, and the ordinate is the cumulative value of the histogram.

Fig. 3 is a flowchart of background denoising of a halftone document image based on connected region labeling.

FIG. 4 is a flow chart of background denoising of a halftone document image based on Gaussian smoothing filtering.

FIG. 5 is a flow chart of background denoising of a halftone document image based on Wiener filtering.

Figure 6 is a graph of color depth combinations of background and text in a halftone document image, (a) dark background and dark text, (b) dark background and light text, (c) light background and dark text , (d) light background and light text.

FIG. 7 is a background denoising result map corresponding to FIG. 6 .

Fig. 8(a) is an inverse image of a halftone document image after removing background noise (mainly including glitches), Fig. 8(b) is an enlarged (800%) schematic diagram of part A in Fig. 5(a), Fig. 8( c) is an enlarged (800%) schematic diagram of part B of Fig. 8(a).

Figure 9(a), Figure 9(b), and Figure 9(c) are the effect diagrams corresponding to Figure 8(a), Figure 8(b), and Figure 8(c), respectively, for text noise removal.

Figure 10(a) is an inverse enlarged (300%) image of a halftone document image after background noise has been removed (contains both burrs and defects); Figure 10(b) is the rendering of Figure 10(a) with text noise removed .

Detailed ways

Referring to FIG. 1 , the method for processing noise of a halftone document image according to the present invention mainly includes the following steps.

(1), Gaussian smoothing filtering is performed on the halftone document image, the Otsu threshold and the grayscale histogram of the halftone document image after the Gaussian smoothing filter are calculated, the halftone point density of the background is estimated according to the Otsu threshold, and the text is estimated according to the grayscale histogram halftone dot density, and according to the halftone dot density of the background and the halftone dot density of the text, halftone document images are divided into four categories: light background and dark text, light background and light text, dark Background and dark text, dark background and light text.

Specifically, a Gaussian smoothing filter of a 5×5 (pixel) template is used for the halftone image, and an Otsu threshold and a grayscale histogram of the image after the Gaussian smoothing filter are calculated.

The halftone dot density of the background is estimated according to the Otsu threshold, and the halftone document image is divided into dark background and light background according to the color depth of the background. For the Otsu threshold and the binarization of grayscale images using the Otsu threshold, refer to the following literature: N.Otsu, "A Threshold Selection Method from Gray Level Histograms", IEEE TRANSACTIONS ON SYSTREMS, MAN, AND CYBERNETICS, VOL.SMC-9, NO.1, JANUARY 1979. The inventor has verified through a large number of experiments, and obtained the relationship between the background color depth described in Table 1 and the Otsu threshold of binarization:

Table 1

Background gray value (0-255) 128 160 192 225 240 250 Otsu Threshold 65 86 107 134 145 153

The preferred Otsu judgment threshold is 105.

The halftone dot density of the text is estimated according to the gray histogram, and the halftone document image is divided into dark text and light text according to the color depth of the text. Through a large number of experimental verifications, the inventor found that in the range of 0-255, the gray scale judgment value is preferably 20, which can satisfy the classification of most halftone document images. The grayscale histogram of the halftone document image after Gaussian smoothing filtering is shown in Figure 2. It can be seen that there are two significant peaks: the peak with a smaller grayscale value corresponds to the text part in the image (ie, the text peak), And the peak with larger gray value corresponds to the background part of the image (background peak). The approximate color depth of the text in the half-tone image can be obtained by calculating the position of the peak value of the text in the half-tone document image after the Gaussian smoothing filter on the gray level.

To sum up, for a specific halftone document image, the depth of the background color is judged according to the Otsu threshold. If the Otsu threshold is greater than 105, the background is considered to be a light background, otherwise it is a dark background; the depth of the text color is determined according to the gray If the gray value corresponding to the peak value is less than 20, the text is considered to be dark, otherwise the text is considered to be light. Combining the two classification methods, halftone document images are divided into four categories: light background and dark text, light background and light text, dark background and dark text, dark background and light text.

(2), remove the background noise of the halftone document image: for the halftone document image with the light background and dark text, use the method based on the connected region mark to remove the background noise; for the halftone document with the light background and light text For images, use a method based on Gaussian smoothing filtering to remove background noise; for halftone document images with dark background and dark text, and halftone document images with dark background and light text, use a method based on Wiener filtering to remove background noise .

Referring to Figure 3, for a halftone document image with a light background and dark text, a method based on connected region marking is used to remove background noise. Firstly, the halftone document map is reversed, and the connected regions are marked, and then the areas of all connected regions are sequentially extracted and counted. If the area of the connected region is greater than the specified threshold (set to 3 pixels), keep the connected region, otherwise remove the connected region, and finally reverse the image, that is, remove the halftone document with light background and dark text Image background noise.

Referring to FIG. 4 , for a halftone document image with a light background and light text, a method based on Gaussian smoothing filtering is used to remove background noise. Firstly, the Gaussian smoothing filter of 5×5 (pixel) template is performed on the halftone document image, and then the Otsu threshold is calculated, and the Otsu method is used for image binarization, that is, the background of the halftone document image with light background and light text is removed noise.

Referring to FIG. 5 , for a halftone document image with a dark background and dark text, and a halftone document image with a dark background and light text, a method based on Wiener filtering is used to remove background noise. First, carry out Gaussian smoothing filtering of a 3×3 (pixel) template on the halftone document image, and then use the Gaussian smoothed image as a reference image to perform Wiener filtering on the original halftone document image, and then perform Wiener filtering on the Wiener filtered image. The image is subjected to Gaussian smoothing filtering of the 3×3 (pixel) template again, and finally the Otsu threshold is calculated, and the image is binarized by the Otsu method, that is, the halftone document image, dark background and light text are removed from the dark background and dark text. Background noise for halftone document images with colored text.

After a lot of experiments, it is found that for halftone images with dark background and dark text, and halftone images with dark background and light text, the method based on Wiener filtering is better. For halftone images with light background and dark text, It is more appropriate to use the method based on connected region marking, and for halftone images with light background and light text, it is more appropriate to use the method based on Gaussian smoothing filter.

Referring to Figures 6 and 7, according to the color depth of the halftone document image background and text, different background denoising methods are adaptively selected. It can be seen that the halftone background noise is basically removed.

(3) Observing the halftone document image with background noise removed, constructing a deburring template and defect repairing template for text, removing text burrs and repairing text defects through template matching.

a. After careful observation and a large number of experimental verifications, the following preferred text deburring templates are used for template matching to remove text burrs, which has a good denoising effect. The preferred text deburring templates are expressed in tables, including:

Among them, a cell represents a pixel, and the positional relationship between cells represents the positional relationship between pixels; in a cell, 1 represents a text pixel, 0 represents a background pixel, and ① represents a text pixel that needs to be removed, that is, the pixel needs to be changed from 1 to becomes 0; an empty cell indicates that the pixel is 0 or 1, that is, it can be either 0 or 1, in other words, it can be a text pixel or a background pixel.

b. After careful observation and a large number of experimental verifications, the following preferred text defect repair templates are used for template matching to repair text defects, which has a good denoising effect. The preferred text defect repair templates are expressed in tables, including the following tables:

表示需要修复的文字像素，即需要把该像素由0变为1；空单元格表示像素为0或1，即：既可为0也可为1，换言之，可以是文字像素，也可以为背景像素。Among them, the cell represents the pixel, and the positional relationship between the cells represents the positional relationship between the pixels; in the cell, 1 represents the text pixel, 0 represents the background pixel,

Indicates the text pixel that needs to be repaired, that is, the pixel needs to be changed from 0 to 1; an empty cell indicates that the pixel is 0 or 1, that is, it can be either 0 or 1, in other words, it can be a text pixel or a background pixels.

Referring to Figure 8 and Figure 9, for the halftone document image with background noise removed, there are many burrs around the text, and the strokes of the text itself in the figure are relatively thin (the thinnest stroke width is 2 pixels), and the burrs are used to remove them. The template processes it to obtain the text in Figure 6 with the burrs removed, while at the same time, the details of the text are preserved as much as possible.

Referring to Figures 10(a) and 10(b), for the halftone document image with background noise removed, there are not only many burrs around the text, but also many defects, which are processed with the burr removal template and the defect repair template to obtain Remove burrs and complete the text in the incomplete area, and the strokes of the text become smooth.

Claims (3)

1. A noise processing method of halftone document image, is characterized in that, comprises the following steps: First, divide halftone document images into four categories: light background and dark text, light background and light text, dark background and dark text, and dark background and light text; Secondly, remove the background noise of the halftone document image: for the halftone document image with light background and dark text, the method based on connected region labeling is used to remove the background noise; for the halftone document image with light background and light text, Use a method based on Gaussian smoothing filtering to remove background noise; for halftone document images with dark background and dark text, and halftone document images with dark background and light text, use a method based on Wiener filtering to remove background noise; Finally, observe the halftone document image with background noise removed, construct a text burr removal template and a defect repair template, and remove text burrs and repair text defects through template matching; The halftone document image is divided into four categories, specifically: performing Gaussian smoothing filtering on the halftone document image, calculating the Otsu threshold and grayscale histogram of the halftone document image after Gaussian smoothing filtering, and estimating the halftone of the background according to the Otsu threshold Tone dot density, estimate the halftone dot density of the text according to the grayscale histogram, and divide the halftone document image into four categories according to the halftone dot density of the background and the halftone dot density of the text, namely light background and dark color Text, light background and light text, dark background and dark text, dark background and light text. 2. the noise processing method of a kind of halftone document image according to claim 1, is characterized in that, the deburring template of described text is represented by table, comprises:

Among them, a cell represents a pixel, and the positional relationship between cells represents the positional relationship between pixels; in a cell, 1 represents a text pixel, 0 represents a background pixel, and ① represents a text pixel that needs to be removed, that is, the text pixel 1 needs to be changed to is background pixel 0; empty cells are represented as background pixel 0 or text pixel 1. 3. The noise processing method of a kind of halftone document image according to claim 1, characterized in that, the defect repair template is represented by a table, comprising the following table: Among them, the cell represents the pixel, and the positional relationship between the cells represents the positional relationship between the pixels; in the cell, 1 represents the text pixel, 0 represents the background pixel,

Indicates the text pixel that needs to be repaired, that is, the background pixel 0 needs to be changed to text pixel 1; an empty cell is represented as background pixel 0 or text pixel 1.

Images