CN106909897A - A kind of text image is inverted method for quick - Google Patents

Abstract

The present invention relates to a text image inversion rapid detection method, including preprocessing the input text image to obtain the binarization processing result B; performing effective text line detection to obtain an effective text line sequence; and performing text line classification, the method is as follows 1) For each effective text line s of the effective text line sequence, fill the blank space between the adjacent characters of the effective text line s; 2) Calculate the projection value of each effective text line s in the vertical direction, expressed by V(c) , wherein c represents the column number; 3) obtains the left boundary and the right boundary of the effective text line s; 4) obtains the left boundary and the right boundary of the effective text line sequence; 5) judges "left indentation text line" and "right indentation text line" and "non-indented text line"; text image inversion detection.

Description

A fast detection method for text image inversion

technical field

The invention relates to a text image enhancement technology, in particular to a direction inversion detection technology for a scanned text image.

Background technique

With the continuous development of computer technology, text image digitization technology based on OCR (Optical Character Recognition) has been widely used. In the process of completing OCR, the text direction in the text image is very important to the performance of character recognition. When the text is tilted, if it is not corrected, it will seriously affect the recognition rate of the text. Especially when the text is inverted (that is, it deviates from the normal direction by about 180°). Therefore, before performing OCR, it is necessary to judge whether there is an inversion of the text image. For the inversion situation, it should be considered to perform rotation processing first to ensure the normal execution of the subsequent recognition process.

For text images with skew, existing skew correction algorithms can be used to detect the skew and correct accordingly. However, most of the existing text image correction methods assume that the slope of the input text image is within a certain range, first obtain the slope angle information, and then complete the slope correction. However, when the input text image is completely inverted, the existing tilt angle detection methods basically fail. Zeng Fanfeng et al. proposed a fast detection method for text image inversion based on punctuation marks. This method first detects text characters; then combines the structural characteristics of Chinese characters and punctuation marks to screen out punctuation marks in text images, and judges the type of punctuation marks according to the distribution characteristics of punctuation marks pixels; finally, combines the usage habits of punctuation marks to judge whether Chinese text images inverted. Zhu Min et al. (Patent Publication No. CN102831421A) proposed a text up-down direction detection method based on punctuation marks. The method proposed in this patent judges the direction of the text based on the relative position attributes of punctuation marks and text lines, and its basic idea is similar to the method proposed by Zeng Fanfeng. Such punctuation-based methods rely entirely on punctuation features and are not effective for text images with fewer punctuation marks. Therefore, such methods have limited scope of application and are not universal.

Contents of the invention

The purpose of the present invention is to overcome the above-mentioned deficiencies of the prior art, and provide a fast detection method for orientation inversion oriented to text images. The technical scheme is as follows:

A text image inversion fast detection method comprises the following steps:

Step 1: Preprocess the input text image to obtain the binarized processing result B;

Step 2: Perform valid text line detection to obtain a valid text line sequence;

Step 3: Carry out text line classification, the method is as follows:

1) For each effective text line s of the effective text line sequence, use the rectangular structure operator to perform expansion operation, and fill the blanks between the adjacent characters of the effective text line s;

2) Calculate the projection value of each valid text line s in the vertical direction, represented by V(c), where c represents the column number;

3) Count the values of c that satisfy the condition V(c)>0.5×R _hei (s), and record the minimum value of c as c _min , which is called the left boundary of the effective text line s; record the maximum value as c _max is called the right boundary of the effective text line s, and the length of the scan line is R _leg =c _max -c _min ;

4) Count c _min (m) and c _max (m) corresponding to each effective text line in the same effective text line sequence, and the minimum value of c _min (m) is called the left boundary of the effective text line sequence, which is denoted as c _lef ; the maximum value of c _max (m) is called the right boundary of the effective text line sequence, which is denoted as c _rgt ;

5) For a valid text line m, if 0.6<|c _min (m)-c _lef |/|c _max (m)-c _min (m)|<0.9 is satisfied, the valid text line m is judged as " Left-indented text line"; if 0.6<|c _rgt -c _max (m)|/(c _max (m)-c _min (m))<0.9 is satisfied, the effective text line m will be judged as "right-indented Enter text line”; if the above two conditions are not met, the text line will be judged as “non-indented text line”;

Step 4: Text image inversion detection, the method is as follows:

Count the number of left-indented text lines and right-indented text lines in a single text image, represented by N _lef and N _rgt respectively; use the following formula to determine whether there is an inversion of the text image:

Preferably, the method of the second step is as follows:

1) Calculate the projection value of each row in B in the horizontal direction, denoted by H(r), where r represents the row number;

2) Calculate the maximum value of H(r), represented by H _max ;

3) For the r-th scanning line, if H(r)>0.5×H _max is satisfied, the line is judged as a valid scanning line;

4) Statistical distribution of each effective scanning line, if detected that continuous m lines are judged as effective scanning lines, and satisfy m>M/100, then form an effective text line sequence by these continuous m effective scanning lines;

Determine the line numbers of the top and bottom effective scanning lines in this effective text line sequence, represent the upper and lower boundaries of this effective text line sequence respectively with R _top (s) and R _bot (s), define the height of this effective text line sequence R _hei (s)=|R _top (s)-R _bot (s)|, the symbol |·| represents the absolute value symbol, where s is the serial number of a valid text line.

Description of drawings

Fig. 1 is the flowchart of proposed method of the present invention

Fig. 2 is a schematic diagram of important definitions used in the present invention

Fig. 3 schematic diagram of text line type defined by the present invention

Figure 4 is a schematic diagram of the number of left and right indented text lines in the text image used in the experiment of the present invention

detailed description

Firstly, preprocessing operations such as grayscale, bilateral filtering, contrast enhancement, and binarization are performed on the input text color image to improve the visual quality of the document image; then, with the help of horizontal projection analysis, effective text lines in the text image are detected and combined with text lines Classify the text lines based on the position and length features of the text; finally, according to the relative number of left-indented text lines and right-indented text lines, it is judged whether there is an inversion of the text image. Figure 1 shows the block diagram of the proposed method.

First some useful definitions are given. A text image is composed of multiple paragraphs, and the character fonts and formats in each paragraph are basically the same. In the present invention, the leftmost and rightmost positions where each character can appear in each paragraph are called "paragraph left boundary" and "paragraph right boundary" respectively. Figure 2 shows a diagram of the left and right borders of a paragraph. Each paragraph may include one or more text lines. For any text line, the left side of the leftmost character and the right side of the rightmost character are called the "line left boundary" and "line right boundary" of the text line, respectively. boundary". Figure 2 shows a schematic diagram of the row left boundary and the row right boundary.

For a text line whose left and right borders basically coincide with the left and right borders of the paragraph to which it belongs, the text line is said to be a "complete text line". If for a certain text line, its left border is 2 to 4 characters away from the left border of the paragraph to which it belongs, and at the same time, the right border of the text line basically coincides with the right border of the paragraph to which it belongs, then the text behavior is called "left indented text line". For a text line, if its right border is 2 to 4 characters away from the right border of the paragraph to which it belongs, and at the same time, the left border of the text line basically coincides with the left border of the paragraph to which it belongs, then the text behavior is called a "right indented text line". Figure 3 shows a schematic diagram of the above three types of text lines.

According to the writing habits of Chinese and English, the characters in the first line of each paragraph are generally indented to the right by 2 to 4 characters, that is, for a paragraph containing two or more text lines, there must be a left indented text line. If the text image is positive, multiple left-indented text lines must be detected. Conversely, multiple right-indented text lines can be detected if the text image is inverted. The present invention judges whether there is an inversion of the text image by detecting and judging the relative numbers of left-indented text lines and right-indented text lines in the text image.

The specific processing process of the method proposed in the present invention includes four main steps: preprocessing, text line detection, text line classification, and text direction inversion detection.

1. Pretreatment

The purpose of preprocessing is to improve the visual quality of the document image, which mainly includes steps such as grayscale, smoothing filter, contrast enhancement and binarization.

(1) Gray scale:

Determine whether the input text image is a grayscale image, if it is a grayscale image, it remains unchanged; if it is a color image, use C _R , C _G and C _B to represent the three color channels of red, green and blue respectively, using the formula ( 1) Calculate the grayscale image, denoted by I.

I(x,y)=min{C _R (x,y),C _G (x,y),C _B (x,y)} (1)

In the formula, x=0, 1, 2,..., M-1, y=0, 1, 2,..., N-1, M and N are the height and width of the text image respectively, that is, the scanning total line and scan the total number of columns.

(2) smoothing filter

Considering that the text image is polluted by noise in the process of acquisition and digitization, the grayscale image I is filtered by bilateral filtering technology to reduce the influence of noise. G represents the image processed by bilateral filtering.

(3) Contrast enhancement

Due to the influence of lighting and other reasons, the contrast of the text image may be low, and the histogram equalization technology is used to enhance the filtered image G, and the processing result is represented by E.

(4) Binary processing

The global threshold corresponding to E is calculated using the classic Otsu method, denoted by T _h . Use T _h to binarize E, and the processing result is represented by B. The specific method is:

Among them, the points with a value of 1 in B represent text points, and the points with a value of 0 represent background points.

2. Valid text line detection

Valid text line detection is done using the following algorithm:

Valid text line detection algorithm:

1) Calculate the projection value of each row in B in the horizontal direction, denoted by H(r), where r represents the row number.

2) Calculate the maximum value of H(r), denoted by H _max .

3) For the r-th scanning line, if H(r)>0.5×H _max is satisfied, the line is judged as a valid scanning line.

4) Count the distribution of each effective scanning line. If it detects that m consecutive lines are judged as effective scanning lines and satisfies m>M/100, then a valid text line is composed of these m continuous effective scanning lines.

5) Determine the line numbers of the top and bottom effective scanning lines in the effective text line, represent the upper and lower boundaries of the text line with R _top (s) and R _bot (s), define the height of the text line as R _hei (s)=|R _top (s)-R _bot (s)|, the symbol |·| represents the absolute value symbol, where s is the sequence number of a valid text line.

3. Classification of text lines

Text line classification is done using the following algorithm:

Text line classification algorithm:

1) For a valid text line s, use the rectangular structure operator to perform expansion operation on the text line, and fill the blank space between the adjacent characters of the text line. The height of the rectangular structure operator is 2 pixels, and the width is 50% of the height of the text line.

2) Calculate the projection value of the text line in the vertical direction, denoted by V(c), where c represents the column number.

3) Count the values of c that satisfy the condition V(c)>0.5×R _hei (s), record the minimum value of c as c _min , which is called the left boundary of the text line; record the maximum value as c _max , Called the right boundary of the text line, the length of the scan line is R _leg =c _max - c _min .

4) Count c _min (m) and c _max (m) corresponding to each effective text line in the same paragraph, and the minimum value of c _min (m) is called the left and left boundary of the paragraph, which is denoted as c _lef ; c The maximum value of _max (m) is called the right boundary of the paragraph, denoted as c _rgt .

5) For a valid text line m, if 0.6<|c _min (m)-c _lef |/|c _max (m)-c _min (m)|<0.9 is satisfied, the text line is judged as "left indentation enter the text line"; if 0.6<|c _rgt -c _max (m)|/(c _max (m)-c _min (m))<0.9 is satisfied, the text line is judged as "right indented text line"; If the above two conditions are not satisfied, then judge the text line as "non-indented text line".

4. Text image inversion detection

Count the number of left-indented text lines and right-indented text lines in a single text image, denoted by N _lef and N _rgt respectively. Use formula (3) to judge whether there is an inversion of the text image:

Examples are as follows:

Matlab2015a under Windows 10 Professional Edition system is used as the experimental simulation platform, and the hardware platform is Intel i5-6200U CPU, 8G memory.

90 text images collected by the patent applicant are selected as the test set, including 78 inverted text images and 12 positive text images. Among all 90 text images, there are 56 Chinese text images, accounting for 62%, and 34 English text images, accounting for 38%. The test image is processed by the method proposed by the invention, and 100% of the inverted images are detected normally. Figure 4 shows the distribution of the number of left-indented text lines and right-indented text lines in 90 document images. It can be seen from the figure that for a forward text image, the number of left-indented text lines is significantly greater than the number of right-indented text lines; conversely, for an inverted text image, the number of right-indented text lines is greater than the number of left-indented text lines. It is obviously divided into two categories, that is, the inverted text image category (identified by the symbol "*" in the figure) and the forward text image category (identified by the symbol "o" in the figure).

The size of the test image is 1944×2592 with a resolution of 50 million pixels. The average speed of processing an image is about 2300ms. If the algorithm is written in C language with higher execution efficiency, the processing speed will be faster and can meet real-time processing requirements.

It can be seen from the experimental results that the method of the present invention can quickly and effectively judge whether the input scanned text image is inverted, and can process text images in multiple languages including Chinese and English.

The steps of the present invention are summarized as follows:

Step 1: Judging the type of the input scanned text image, if it is a grayscale image, it remains unchanged; if it is a color image, then use formula (1) to convert it into a grayscale image, and use I to represent the grayscale image.

Step 2: Use bilateral filtering technology to filter the grayscale image I, and the filtering result is represented by G.

Step 3: Use the histogram equalization technique to enhance the filtering result image G, and denote the processing result by E.

Step 4: Use the Otsu method to calculate the global threshold of the enhanced result image, and combine Equation (2) to binarize E, and the processing result is denoted by B.

Step 5: Use a valid text line detection algorithm to detect valid text lines in the scanned text image.

Step 6: Use the text line classification algorithm to classify each valid text line, and determine the numbers N _lef and N _rgt of left-indented text lines and right-indented text lines.

Step 7: Combining formula (3), determine whether the scanned text image is inverted.

Claims (2)

1. a kind of text image is inverted method for quick, comprises the following steps：

The first step：Text image to being input into is pre-processed, and obtains binary conversion treatment result B；

Second step：Effective line of text detection is carried out, effective line of text sequence is obtained；

3rd step：Line of text classification is carried out, method is as follows：

1) for each effective line of text s of effective line of text sequence, dilation operation is carried out using rectangular configuration operator, filling should Blank between effective line of text s adjacent characters；

2) projection values of each effective line of text s in vertical direction is calculated, is represented with V (c), wherein c represents row sequence number；

3) statistics meets condition V (c)>0.5×R_heiS the c values of (), c is designated as by the minimum value of c_min, referred to as effective line of text The left margin of s；Maximum is designated as c respectively_max, the right margin of referred to as effective line of text s, the length of the scan line is R_leg= c_max-c_min；

4) the corresponding c of each effective line of text in same effective line of text sequence is counted_min(m) and c_max(m), by c_minM () most Small value is referred to as the left margin of effective line of text sequence, is designated as c_lef；By c_maxM the maximum of () is referred to as effective line of text sequence Right margin, be designated as c_rgt；

5) for certain effective line of text m, if meeting 0.6<|c_min(m)-c_lef|/|c_max(m)-c_min(m)|<0.9, then this is had Effect line of text m is judged to " left indent text row "；If meeting 0.6<|c_rgt-c_max(m)|/(c_max(m)-c_min(m))<0.9, then will Effective line of text m is judged to " right indent text row "；If above-mentioned two condition is all unsatisfactory for, this article one's own profession is judged to " non- Indent text row "；

4th step：Text image is inverted detection, and method is as follows：

The number of left indent text row and right indent text row, uses N respectively in statistics single width text image_lefAnd N_rgtRepresent；Make Judge text image with the presence or absence of inversion with following formula：

2. text image according to claim 1 is inverted method for quick, it is characterised in that the method for second step is such as Under：

1) each row projection value in the horizontal direction in B is calculated, is represented with H (r), wherein r represents line number sequence number；

2) maximum of H (r) is calculated, H is used_maxRepresent；

3) for r scan lines, if meeting H (r)>0.5×H_max, then the row is judged to an active line；

4) distribution situation of each active line is counted, if detecting continuous m rows is judged to active line, and m is met>M/ 100, then an effective line of text sequence is constituted by this continuous m active line；

Determine the line number of the top and bottom active line in effective line of text sequence, use R_top(s) and R_bot(s) difference The up-and-down boundary of effective line of text sequence is represented, the height for defining effective line of text sequence is R_hei(s)=| R_top(s)- R_bot(s) |, symbol | | the symbol that takes absolute value is represented, s is the sequence number of effective line of text in formula.