1 Introduction

Recently, there is a substantial growth in the number of photo-to-text software, and many organizations are trying to digitize all their documents to get rid of their paper-based archived files. To this end, after scanning the documents, the scanned images are given to an image-to-text system to extract the text. Text-line detection usually uses for improving the accuracy of optical character recognition (OCR) techniques [9, 11, 26, 38, 43, 50]. Also, it is one of the pre-processing steps in document layout analysis [9, 27, 38], text restoration [22, 43], binarization [10], and camera-captured images [12, 17, 38, 53, 61].

One of the main requirements in designing text-line detection methods is having high accuracy in extracting lines with a low computational cost [30]. Image distortions like the blurring of images, image acquisition by the low-resolution camera, and the light of mobile camera flash are the major obstacle to achieve high accuracy in such methods [11, 26]. These distortions may happen in pages scanned by scanners or mobile phone cameras. The OCR methods try to remove these obstacles as much as possible in the pre-processing step [9, 26].

Another major problem that we face in the segmenting of a text line is non-linear warping. Warping decreases the readability of the text as well as reduces the OCR accuracy [26]. In the last decade, several methods have been proposed to overcome non-linear warping, which are known as the de-warping methods [8, 19, 57, 58]. These methods can divide into two categories: 1) 2D image processing techniques [19, 58], 2) 3D document shape restoration techniques [8, 57]. Several methods [19, 41, 58, 60, 64] as well as the proposed method fall into the first category. The second category needs image capture with some special camera; also, for displaying the document surface, the method needs to use a 3D shape model [26].

In addition to the problems mentioned above, in languages like Arabic and Persian, the letters divide into two groups: connectors and non-connectors. Each connector letter connects to the next letter in a word, which makes the segmentation more difficult. Also, in these languages, some words have diacritic. The existence of diacritic has a reverse effect on the accuracy in extracting lines [4]. In recent years, several text-line detection techniques proposed for different languages like English, Chinese, French, Italian, and German. On the other hand, there is still much work to be done for languages like Persian [2].

In this paper, we present a novel approach for text line detection. The method starts by binarizing the input image, then removes all small objects from it. In the next step, the method measures the diameter of each CC in the image and defines the final font size of the image. Using the final font size of the image as the searching length, the method finds the nearest horizontal neighbor of each CC, then connects all these CCs that are horizontally in the same direction, and extracts all text lines.

We make the following contributions in this paper. The proposed method works on three-channel RGB and grayscale images. The method is based on the size of the font extracted from the image, which plays an important role in extracting curved lines. We have also introduced a new dataset, which contains curved lines.

The remaining of the paper is organized as follows: a brief description of previous text-line detection approaches is presented in Section 2. The proposed method is described in Section 3. Performance evaluation and experimental results are given in Section 4, followed by conclusions in Section 5.

2 Related works

As we mentioned, most of the existing text-line segmentation techniques are mainly proposed for pre-processing steps of the document layout analysis, optical character recognition (OCR), text restoration, binarization and camera-captured images. Several text-line segmentation approaches have been proposed in the last decades. These approaches classified as techniques based on: 1) CCs based methods [30, 37, 38, 43], 2) project profile-based methods [1, 5, 16, 25, 33, 36, 54, 56, 63], 3) smearing-based methods [3, 4, 7, 23, 33, 44, 56], 4) bounding-based methods [56], 5) Hough transform-based methods [46], 6) combined methods [51, 56], 7) tree-based methods [45], and 8) deep learning-based methods [15, 34, 42]. These techniques are discussed in this section.

Mahmood and Srivastava [43] proposed a text line segmentation technique for Urdu typewritten text based on edge information of the CCs. This technique is tested over two benchmark datasets that collected, compiled and organized by themselves with an accuracy of 87.36% and 84.75% respectively.

H. I. Koo [38] introduced a text-line detection method for the camera-capture document images which developed by incorporating state estimation into the connected-components. They extracted connected-components with the maximally stable extremal region (MSER) method and then estimated the scales and orientations of connected-components after they segmented the text-lines.

Yandong Guo et al. [30] proposed a text line detection method based on estimated optimization cost for text line direction; in the first step, they extracted the CC from the image. Then they appraised the direction of the text line in several local regions and tried to optimize their estimation.

Koichi Kise et al. [37] proposed a text line segmentation method based on the Voronoi diagram based on the CC analysis. They use the Voronoi diagram to get the approximate coordinates of each line. The page segmentation process can be considered as selecting the appropriate sections of the line as the boundaries of the document components. Also, Lawrence O’Gorman [47] proposed a new text line segmentation method based on the nearest-neighborhood clustering of CCs extracted from the document image.

Projection profile is a technique for text-line segmentation, classified into two main types: 1) vertical projection profile and 2) horizontal projection profile [1, 5, 16, 25, 56, 63]. The horizontal projection profile finds the interline gaps between lines and considers it a separator between two consequent lines. This approach is used for printed-text, where there is no overlapping or touching between lines [16, 33, 36, 54, 56]. Also, the techniques based on the strip can deal with overlapping and touching text lines [25].

The smearing methods use for text-line segmentation; these methods are trying to segment the text area by making similar regions along with the text locations [4, 44]. This approach uses both types of text (handwritten texts and printed texts), but it has some weaknesses; it fails in case of overlapping or touching between lines [3, 4, 7, 23, 33, 56].

Another solution for text line segmentation is detecting the bounding box. First of all, the method generates the histogram of the image, then specifies lines with a lesser number of pixels, then tries to find the centroid of each line and determines the boundaries for each line [56].

The Run-Length Smoothing method is used in smearing [51, 56]. This approach, at first, smears the adjacent black pixels along the horizontal direction. If the distance of white space between black pixels becomes less than the predefined threshold, the white pixels turn black. The bounding boxes of the CCs surround text lines. Finally, using the Run-Length Smoothing method, the strength of the histogram will be increased [51].

One of the practical methods used in text-line segmentation is the Hough Transform (HT) [46]. Two of the most efficient methods of Hough Transform used for line segmentation are Standard Hough Transform (SHT) and Progressive Probabilistic Hough Transform (PPHT) [46]. Hough peaks are determined, and then, according to these peaks, the lines are extracted [4].

One of the tree-based segmentation methods is XY-Cut [45], where each page is located at the root of the tree and the final segmented areas are in the leaves of the tree. The method recursively divides the image into two smaller rectangular areas showing tree nodes.

Recently, researchers use deep learning techniques to segmenting text lines to reach more accuracy. For extracting the document area, Lyu Bing et al. [42] use ARU-Net [28] to segment the text line. ARU-Net is a deep learning model for text line segmentation in historical documents, which applies A-Net and RU-Net as base networks. In the extraction process, some text lines may miss. Lyu Bing et al. use LeNet, a simple deep learning method [42], to classify extracted articles and handwritten pictures.

Deep learning is one of the most effective techniques for text line segmentation, but to gain a high accuracy, it needs to be trained by very large datasets. Due to the lack of a Persian OCR dataset, we can not use these methods.

As mentioned before, one of the problems in the text line detection field is detecting the curve/skewed text lines. There are several approaches proposed in the literature to mitigate this problem. Hassan El Bahi et al. [18] proposed a new method to recognize the texts taken from smartphones. In the pre-processing step, the method detects all contours of the image, then by dilation technique removes the background of the image, finally by connecting the CCs in each line, segments the region into text line images.

Rituraj Soni et al. [55] introduced a new method for the classification of the text regions in natural scene images. Achieving this aim depends on choosing optimal-classifiers and optimal-features. Using the improved MSER method, their method detects the possible text regions; then, extracts eleven features from these regions. Using the CfsSubsetEval and BFS parameters of the Weka tool [32], the method chooses an optimal-feature set from these eleven features to discriminate the text and non-text components. In the end, with the help of the Weka tool on the ICDAR 2013 training set, the method train many classifiers using these optimal-features.

A hierarchical recursive text detection method is proposed in [59] to detect texts in complex scenes. Besides, to detect these kinds of texts in complex scenes, [14] proposed an Adaptive Convolution and Path Enhancement Pyramid Network (ACPEPNet), which can more accurately locate the text instances with arbitrary shapes. In [31], The method combines the saliency model with the text detection approach in a natural scene to generate text saliency. The proposed method in [49] detects text regions from camera captured images by using the Fuzzy Distance Transform that is based on an adaptive stroke filter. Finally, In [35], a random forest classifier is used to recognize the newspaper texts. The inputs of the random forest classifier are the features of characters extracted by different kinds of feature extraction techniques.

OCRopus is a neural network-based approach that was first introduced in 2008 and presents good pre-processing in OCR [6]. This method has given several updates to date. The latest update of this method in 2017, which is available in [39], includes the following sections: line detection, optical character recognition using LSTM, and using statistical linguistic models.

The specific type of OCR software that we employed in our tests is an open-source OCR program called Kraken, developed by Benjamin Kiessling at Leipzig University’s Alexander von Humboldt Chair for Digital Humanities. Kraken detects each line, then uses a neural network to recognize letters in each line. The latest Kraken update in the line recognition field was released in 2021, which is available in [40].

3 Proposed approach

In this section, we describe our text line segmentation method. Our proposed method has sub-processes of pre-processing, recognizing the final font size, extracting the main body of the line, and building the box of each line. These sub-processes are shown in Fig. 1. The proposed method removes noises and excess lines from the image in the “pre-processing” step. The proposed method has a bottom-up design. In this way, in the “recognizing final font size” step, it first extracts all CCs (CCs) of the image, then calculates the diameter of each of them, then the final font size calculates according to the definition. In the “extracting the main body of the line” step, Using the final font size of the image extracted from the previous step, the proposed method performs a search on the horizon axis around each of the CCs. If any other CCs are in that radius, the method connects the two neighboring CCs, thus extracting the main body of each line. In the last step, the proposed method tries to find any point or diacritic that does not stick to the main body of the line, and by using the projection technique, extracts the whole line.

Fig. 1
figure 1

Text line segmentation process

Full size image

3.1 Pre-processing

Pre-processing is the first step in line segmentation. Algorithm 1 shows the steps of producing a clean image that facilitates the processing of the next steps. In this step, firstly, our method converts the 3-channel input image shown in Fig. 2a into a grayscale channel. It then obtains a threshold for the image to converts the grayscale image into binary. This threshold is calculated by the Otsu method [48]. Secondly, our method removes the salt-and-pepper noises from the image [29]. Initial labeling and measuring are the other steps of pre-processing.

Fig. 2
figure 2

The pre-processing

Full size image

For de-noising, the method performs median filtering of the input image in two dimensions [24]. As shown in Fig. 2b, each output pixel contains the median value in a 3-by-3 neighborhood around the corresponding pixel in the input image. For initial labeling and measuring, the method tags every CC with a separate label to become recognizable from other CCs. To separate CCs from each other and assign a label to each, we used the MATLAB function bwlabel. As shown in Fig. 3, in this method, if the value of two adjacent pixels is one and these two pixels connect along the horizontal, vertical, or diagonal direction, these two pixels are part of the same object.

Fig. 3
figure 3

Using bwlabel

Full size image

In the next step, the algorithm measures some properties of each label. These properties are the length (in pixels) of the major axis and the minor axis of the CC (label). The major Axis is the longest diameter of the CC. It goes from one side of the CC, through the center, to the other side, at the widest part of the CC. The minor axis is the shortest diameter (at the narrowest part of the CC). Finally, as shown in Fig. 2c, the method removes some labels that have this condition:

$$ \frac{MajorAxisLength}{MinorAxisLength}>\varepsilon\ $$
(1)

By using this condition, the method measures the diameter of the label. If MajorAxisLength is much bigger than MinorAxisLength, the method removes that label. The resulting image of the pre-processing phase is shown in Fig. 2d.

figure a

3.2 Final font size recognition

This phase aims to find the final font size of the image. The method needs the final font size for connecting the CCs that are horizontally in the same direction.

3.2.1 CC’s diameter calculation

Our method starts by measuring the diameters of all CCi’s pixels (where: 0 ≤ i ≤ Number Of CCs). Consider Fig. 4a as the image of CCi. In each pixel of CCi, for example, the pixelj,k shown in Fig. 4b, the method measures the distance between two edges of CCi at four different angles: horizontally (Fig. 4c), vertically (Fig. 4d), and obliquely in two directions (Fig. 4e and f). The Euclidean distance was used to calculate each of these distances. The Euclidean distance between two points like p and q is shown in (2).

$$ d\left( p,q\right) = \sqrt {\left( q_{x}-p_{x}\right)^{2} + \left( q_{y}-p_{y}\right)^{2}} $$
(2)
Fig. 4
figure 4

CC’s diameter calculation

Full size image

Filters such as canny filters [13] can be used to extract CC edges. However, there is no need to extract edges because the pixels’ values are binary. Therefore, any pixel that has at least one black pixel in its neighborhood is an edge. By doing this, the speed of the algorithm increases.

Then it sets the minimum distance as the CCi’s diameter in that pixel. This is according to the definition of CCi’s diameter in this paper. At any point of CCi, the closest distance between the two edges considered being the CCi’s diameter in that pixel. After completing this step, the diameter of the CCi in all its pixels is specified. Then the diameter with the most repetitions among the CCi’s pixels is introduced as the final CCi’s diameter. Finally, the diameter value of all CCi’s pixels changes to this diameter value. The proposed method chooses the largest number of repetitions of the diameter, as the final diameter. For example, consider the CCi pixels that are visible with the red box around them in Fig. 5. Table 1 shows the diameters extracted from these pixels. As shown in Table 1, most parts of the CCi have the same diameter. Algorithm 2 summarizes this step.

Fig. 5
figure 5

Pixels’ diameter

Full size image
Table 1 Some pixel’s diameter in CCi
Full size table
figure b

3.2.2 Final font size

In the next step, the method gets all the different CC’s diameters with their number of repetitions. The method chooses the diameter with the most number of repetitions as the first font size. To improve method’s performance, we need to remove all points and diacritics before calculating the number of repetitions. Otherwise, due to the existence of these objects in documents, the algorithm may make errors in its calculations, and the final font size may extract based on the diameter of these points and diacritics.

For removing all points and diacritics, we used the MATLAB function bwareaopen. But before using this function, the algorithm needs to extract the number of pixels per CC. If the number of pixels is less than a predetermined value, that CC is removed. This function is used to remove small objects from binary images. Therefore, first, the proposed method finds the largest number of repetitions of the diameter in all CCis, then calculates the number of pixels of each CCi. Finally, the proposed method uses the bwareaopen function to remove CCs whose number of pixels is less than ten times the diameter with the most repetition. This ten times the largest number of repetitions of the diameter in CCi has been achieved with many trial and error on different font types in various sizes and styles. By removing a CC, the proposed method removes CC’s diameters from the diameters table. After eliminating all points and diacritics, the method sorts the remaining diameters on the image by number, then choosing the diameter with the greatest number of repetitions as the second font size.

For removing all points and diacritics, we used we used the MATLAB function bwareaopen. First, the method finds the largest number of repetitions of the diameter in CCi, then calculates the number of pixels of the CCi. Finally, the method removes CCi’s diameters from diameter’s table if the number of pixels of the CCi is less than ten times the largest number of repetitions of the diameter in CCi. This ten times the largest number of repetitions of the diameter in CCi has been achieved with a lot of trial and error on different font types in different sizes and styles. After removing all points and diacritics, the method sorts the remaining diameters on the image by number, then choosing the diameter with the most number of repetitions as the second font size.

Sometimes the font-size of the title is extremely big (Fig. 6a), and the method segments the title lines incorrectly. To solve this problem, first, the method separates the titles and saves them as a new image. Thus, we will have two images with different font sizes. As shown in Algorithm 3, the proposed method finds all unique diameters after removing all points and diacritics. If any of the unique diameters is more than twice the second font size, the method detects the title in the image and chooses the second font size as the body’s font size (the final font size of the body image). As a result, any diameter after this distance, considered as the diameter of the title. The method considers the most repeated value of remaining unique diameters as the title’s font size (the final font size of the title image). As shown in Fig. 6b, the CCs whose their diameter is in the range of the body text diameter, are placed in one image, and the rest of the CCs, are placed in another image (Fig. 6c).

Fig. 6
figure 6

Image with big title

Full size image

The important point in the proposed method is to use the original image in the next step. This means, after removing the small objects and calculating the final font size, the remaining steps will use the original image, which also has all of the small objects.

figure c

3.3 Extracting the main body of the line

The method has to connect any CCs that have been horizontally in the same direction to get the approximate lines’ locations of the image. To reach this goal, every CC in the image tries to find the nearest horizontal neighbor by using a searching method in the horizontal direction. In this step, the line segmentation method uses the final font size as an input parameter and searches in the horizontal direction in the specified length to find another CC, if the searching method finds any white pixel that belongs to another CC, it connects these two CCs. The searching length depends on the size of the final font size. We set an adaptive coefficient for controlling the searching length. Using this adaptive coefficient and multiplying it by the final font size, the best search length is obtained.

In the next step, we used one of the morphological structuring elements called strel. Its basic syntax is:

$$ se = strel(Shape,parameters) $$
(3)

Which we used one of the special types:

$$ se = strel('line', Len, Deg) $$
(4)

This type of strel, creates a flat, linear structuring element, where Len specifies the length, and Deg specifies the angle (in degrees) of the line, as measured in the counterclockwise direction from the horizontal axis. Finally, with another morphological operation called closing, it connects CCs within the search radius of each other. And the lines are detected correctly. Algorithm 4 shows how to calculate this adaptive coefficient.

figure d

3.4 Building the box of each line

Despite implementing all the previous steps to segment lines correctly, the method detects some of the up and down word’s points as a new label (Fig. 7a). To handle this problem, we had to describe the last part of our proposed method in this step. Firstly, the method finds min(x), min(y), max(x) and max(y) for every label. As shown in Fig. 7b, if the position of the point’s label is near to the main label, the method connects these labels. Otherwise, the point’s label will be considered as another CC.

Fig. 7
figure 7

Bounding Box steps

Full size image

3.5 Time complexity

In Algorithm 1, the time complexity of the first two functions is in O(N), in which N is the number of pixels in the image. Also, the time complexity of the third function is in O(N + w), in which w is the number of labels in the image. And because N is so larger than W, we considered it as O(N). Therefore, the time complexity of Algorithm 1 is equal to O(N). In Algorithm 2, the time complexity is in \(O({N_{w}^{2}})\), in which Nw is the number of pixels labeled in the image. In Algorithm 3, the time complexity both choosing the diameter with the most number of repetitions, and calculating the number of repetitions of unique diameters are in O(N). Both for loops in the Main function are in O(w). The final-font-sizes function is in O(d), in which d is the number of unique diameters in the image. Therefore, the time complexity of Algorithm 3 is equal to O(N). Finally, in Algorithm 4, the time complexity is in O(N).

4 Results and evaluation

In this section, we explain and discuss the experimental results. We also provide information on how to make the dataset and the contents of them. Finally, we compare the results of the proposed method with related works on our dataset. To validate the effectiveness of the proposed system, we have conducted experiments with three different datasets. We have used our own system to implement and evaluate the proposed method. The CPU of our system is Intel(R)-Core(TM) i7-9750H, and our RAM is 16 GB.

4.1 Datasets

Due to the lack of a Persian OCR dataset, we made a standard dataset. The generated dataset is publicly available. We collected the documents’ text of our dataset from various sources like newspapers, magazines, and books. The collected text is made in different sizes, fonts, and styles and saved as images in png format. Some of the words in these documents have diacritic.

To create datasets in different fonts, styles, and sizes from a text, we have used the python-docx library. Using this library, you can easily produce text in various sizes, styles, and fonts. The latest version of this library is for 2019. python-docx saves input text as a Microsoft Word file. For more details on this library, refer to the documentation provided for the python-docx library. Finally, each page of the Microsoft Word file will be saved as an image in png format. The size of each page is 1653 × 2339. This dataset is available in [21].

Our advantage over other existing datasets is the presence of different font types in various sizes and styles of Persian in a relatively significant number of lines. We have created images from six famous Persian font types in three styles and six different sizes.

We used six distinct font types to test our proposed method, including Arial, Nazanin, Ziba, IranNastaliq, Tahoma, and XB-Niloofar. These font types are common in writing magazines, books, and newspapers. We used all three types of bold, italic, and regular fonts in our dataset. We have also used font sizes 8, 10, 14, 18, and 22 in our dataset. Also, the total number of lines in this dataset is more than 24,000 lines. Table 2 shows a summary of the dataset information.

Table 2 A statistic of the font types and sizes used in building our data-set
Full size table

To further evaluate the performance of the proposed method, we used two more datasets to test the proposed method. The first dataset is in Arabic. This dataset which is called the Arabic OCR dataset is available in [62]. The images in this dataset are in different sizes. Arabic OCR dataset aims to solve a more straightforward problem of OCR with images that contain only Arabic characters. Figure 8 shows an example of the images in this dataset.

Fig. 8
figure 8

Sample of Arabic OCR dataset

Full size image

To evaluate the performance of the proposed method, we randomly selected 50 images from this dataset and tested the proposed method on these images. The images in this dataset are single columns lines, but lines have a small amount of rotation, which increases the complexity of the dataset.

We prepared the second database ourselves. The images in this dataset are provided by the HP Scanjet 4890 scanner with default settings. We have scanned 80 images from various sources such as magazines, newspapers, and books with this scanner, which is available in [20]. This dataset’s images have more complexity than the images of the previous two datasets. One of the most considerable complexities of the images in this dataset is the pages’ rotation, which creates skew lines. As shown in Fig. 9, which is one of the images in this dataset, the images in this dataset have rotation, making line recognition difficult.

Fig. 9
figure 9

Sample of dataset 3

Full size image

Images’ quality in dataset 3 is 300, 200, or 120 dpi. The maximum number of lines in the images of this dataset is 35 lines. Some images in this database were rotated up to about 10 degrees, and some other pages were curved, which increase the complexity of the dataset.

4.2 Parameter tuning

In the previous section, we defined some parameters, coefficients, and constants with specific values. In this subsection, we will explain in detail how to obtain these values. We said before the proposed method removes CCs which, the number of pixels is less than ten times the diameter with the most repetition. This Coefficient value has been achieved with many trial and error.

We obtained the best value for this coefficient among several values of 5, 10, 15, and 20 by performing a test on 50 images of the first dataset. Figure 10 shows part of a page that its text is in XB-Niloofar font type, regular font style, and font size 18. Figure 11 shows the result of using each of these coefficients in Fig. 10. As shown in Fig. 11a, some points are not omitted if the diameter coefficient is low. On the other hand, if the diameter coefficient is high (Fig. 11c, and d), some letters of the alphabet are omitted incorrectly. Therefore, choosing the suitable diameter coefficient is very important (Fig. 11b).

Fig. 10
figure 10

Sample of first dataset image

Full size image
Fig. 11
figure 11

Removing points and small objects from image

Full size image

Table 3 shows in more detail than Fig. 11 the error rate of each coefficient in separating points and main text. As shown in Table 3, the value 10 is the best possible coefficient among coefficients and has the lowest error rate among coefficients. Besides, we calculated the confusion matrix for the coefficient with the value of 10. Table 4 shows the confusion matrix for 50 images of the first dataset, which contains nearly 50,000 labels. The number of correct and incorrect predictions are summarized with count values and broken down by each class. In addition, the ROC curve of these four coefficients is shown in Fig. 12. The recall and precision calculation are as follows:

$$ Precision_{i}=\frac{M_{ii}}{\sum\nolimits_{i} M_{ij}} $$
(5)
$$ Recall_{i}=\frac{M_{ii}}{\sum\nolimits_{j} M_{ij}} $$
(6)
Table 3 The error rate of each coefficient in separating points and main text
Full size table
Fig. 12
figure 12

ROC curves for coefficient with the value of 5, 10, 15, and 20

Full size image
Table 4 Confusion matrix for the coefficient with the value of 10
Full size table

One of the challenges we have been facing in this step was finding the best searching length, especially when the image has more than one column containing text (Fig. 13a). As shown in Fig. 13c, because of the short searching length, sometimes the searching method cannot find any CCs around the main CC to connect them. As a result, the searching algorithm cannot detect the line correctly. As shown in Fig. 13b, the opposite can also happen, especially if the image has multiple columns.

Fig. 13
figure 13

Searching length problem

Full size image

As we said in the last paragraph, by increasing or decreasing the searching length, the method encountered different kinds of problems. These problems are exacerbated in curve lines (Fig. 14a). Short searching length poses the same problem as previously described and could not recognize all the CCs of a line correctly (Fig. 14c). But, as shown in Fig. 14b, by increasing the searching length, the method could not separate two consecutive lines in a column correctly and connects them. Using this adaptive coefficient and multiplying it by the final font size, the method does not need to do anything more to detect curve lines (Fig. 14d).

Fig. 14
figure 14

Searching length problem

Full size image

4.3 Experimental result

Table 5 shows the performance results of our method versus five widely used methods for text-line extractions: i)Smearing [23], ii)XY cut [45], iii)Voronoi diagram-based [37], iv)Seam curving [52], vi)Docstrum [47]. In this part, we obtain these results when tested on our dataset with six different font types and Bold styles. As shown in Table 5, our method has the best performance in all cases. We have also used formula (7) to calculate the accuracy in each case:

$$ accuracy_{i},_{j}=\frac{\#of lines correctly segmented}{totlal \#of lines},\ i \in font\ styles, j \in font\ types $$
(7)
Table 5 The Bold style accuracy of proposed method and five states of the art methods for different font types
Full size table

Also, Tables 6 and 7 show the performance results of our method versus five widely used methods for text-line extractions. As shown in Tables 6 and 7, our method has the best performance in all cases except the “Ziba” font. As shown in Fig. 15a and b, even though the “Ziba” and “Tahoma” font sizes are the same, the diameter of the “Ziba” font is smaller than the “Tahoma” font. As a result, the final-font-size of the “Ziba” font, and the searching-length, are smaller. As shown in these tables, as much the font size is larger, the easier it is to recognize text lines, as the number of words within a line decreases and the final font size increases.

Table 6 The Italic style accuracy of proposed method and five states of the art methods for different font types
Full size table
Table 7 The Regular style accuracy of proposed method and five states of the art methods for different font types
Full size table
Fig. 15
figure 15

Ziba font problem

Full size image

Tables 8, 9, and 10 show the performance results of our method versus five widely used methods for text-line extractions: i)Smearing [23], ii)XY cut [45], iii)Voronoi diagram-based [37], iv)Seam curving [52], vi)Docstrum [47]. In this part, we obtain these results when tested on our dataset with five different font sizes and three styles: Bold, Regular, and Italic. As shown in Tables 8, 9 and 10, our method has the best performance in most cases.

Table 8 The Bold style accuracy of proposed method and five states of the art methods for different font sizes
Full size table
Table 9 The Italic style accuracy of proposed method and five states of the art methods for different font sizes
Full size table
Table 10 The Regular style accuracy of proposed method and five states of the art methods for different font sizes
Full size table

As mentioned beforehand, we used two more datasets to evaluate the performance of the proposed method further. In addition to the proposed method, we test the smearing method, which is the best method among the related works, and the OCRopus and the Kraken methods, which are more up-to-date than other methods, on these two datasets. Tables 11 and 12 show the accuracy with average runtime of the proposed method and other methods on the Arabic OCR dataset and the dataset 3. As shown in Tables 11 and 12, the proposed method has the best performance among all methods.

Table 11 The text lines recognition accuracy with an average runtime of the proposed method and other methods on the Arabic OCR dataset
Full size table
Table 12 The text lines recognition accuracy with an average runtime of the proposed method and other methods on dataset 3
Full size table

Due to the high complexity of the third dataset compared to the other two datasets, the accuracy of the methods in this dataset is lower. As shown in formula (7), the accuracy of each method represents the percentage of the number of lines that have been extracted correctly. The OCRopus method, despite its good accuracy in detecting lines, also has many disadvantages.

One of the main problems with the OCRopus method is its high runtime, which can take up to 30 seconds in images with many CCs. This high runtime is one of the main problems of this method, but the problems do not end here. One of the most critical problems of this method is removing small objects and points in the image. Figure 16a is one of the output lines in the proposed method. Figure 16b shows the output of the same line in the OCRopus method. As can be seen, most of the points were removed in this method. Failure to run the program in images with high resolution (more than 300 dpi) is another problem of this method. The proposed method is specific to the Persian language. In Persian, the number of dots is much higher than in languages such as English. Therefore, in the proposed method, the dots are well kept.

Fig. 16
figure 16

Problem of removing small objects in OCRopus method

Full size image

Same as the OCRopus method, the Kraken method has high runtime. The Kraken method has low accuracy on images that are skewed or curved. For this reason, it does not work well on dataset 3. Figure 17 shows an example of the output of the Kraken method on dataset3. As shown in Fig. 17, most lines overlap, which reduces the performance of the OCR system.

Fig. 17
figure 17

The output of the Kraken method on dataset 3

Full size image

Both OCRopus and Kraken methods have difficulty detecting images that are skewed or curved. While the skewness of the lines or their curvature has less effect on our method. Our method is resistant to curvature and skew due to the use of the final font size.

5 Conclusion

In this paper, we introduced a new text-line detection method for printed text images by using the final font size. The proposed method measures the diameters of each CC from any direction, then chooses the maximum of these diameters as the font size of that CC. The algorithm tries to select one of the font sizes as the final font size by defining an adaptive parameter. Finally, the algorithm tries to connect any CCs that are horizontally in the same direction. The proposed method is tested on more than 24000 lines with different font types, sizes, and styles. The proposed method can process document images with the following points: i) Persian text with and without diacritics; ii) non-linear warping text lines; iii) skew text lines; iv) six Persian font types: Nazanin, Ziba, IranNastaliq, XB-Niloofar, Arial, and Tahoma; v) Bold, italic and regular style; vi) and different font sizes 8,10,14,18,22 including the mix between them. As the method has good results and reached 99.3% accuracy on our dataset, we still need to do more evaluation on other Persian fonts and extend this to Arabic and other languages. Also, text line segmentation for handwriting documents and scene text detection are some other challenges of our future purpose.