Erosion text data set forming method for simulating real effect

Abstract

The invention provides an erosion text data set forming method for simulating a real effect. The specific process sequentially comprises the steps of 1, data acquisition, 2, data preprocessing, 3, local erosion extraction, 4, data enhancement, 5, erosion superposition, 6, combination and the like. The invention aims to provide an improved inscription repairing data set, the quality and reality sense of an eroded image are enhanced through a preprocessing and synthesizing method, and more reliable and efficient data support is provided for model training and practical application in the character repairing field. Development of the inscription repairing technology is promoted, the inscription repairing accuracy and efficiency are improved, and then ancient inscription cultural heritage is protected and inherited. Through the application of the method, remarkable progress and achievement are expected to be achieved in the inscription repairing field, and important impetus is brought to research and practice in related fields.

Description

Technical Field

[0001] The invention relates to a data set synthesis method, in particular to an eroded text data set synthesis method simulating real effects. Technical Background

[0002] Erosion refers to the process in which anything gradually becomes blurred, worn or disappears due to the influence of the external environment, materials or other natural or man-made factors over a certain period of time. The scope of influence of erosion in a broad sense covers all areas, from the natural environment to cultural heritage, all of which are under the threat of erosion. Specifically speaking, the erosion of text carriers refers to the process in which the specific carriers of the text (such as paper, stone tablets, etc.) are affected by external factors over a certain period of time, causing the text to gradually become blurred, worn or disappear. These factors can be natural, such as humid climates causing paper to mold, sunlight exposure causing pigments to fade, etc.; they can also be man-made, such as frequent reading causing paper wear.

[0003] The restoration of text carrier erosion is of great significance for the preservation of cultural relics, archaeology and cultural heritage. The innovation of the analysis and restoration methods of text carrier erosion can help preserve historical information and pass on cultural heritage. With the development of computer technology, related restoration technologies represented by Chinese characters have become the research focus in the field of restoration. However, most of the current text restoration methods are limited by the problem of datasets. Most of the existing datasets currently obtain only a small number of fonts from a single source, which limits their generalization in the restoration of real damaged characters. This limitation means that these datasets cannot adapt well to various real-world erosion conditions. Furthermore, existing restoration methods usually use a small number of masks that are completely unrelated to the real erosion to expand and synthesize the dataset according to the needs of downstream tasks. However, this type of mask has a low restoration effect on real erosion and often only plays a small role. To solve this problem, it is necessary to develop more generalized datasets that can cover various real-world erosion conditions and provide sufficient training samples for different types of restoration tasks. At the same time, for the specific needs of downstream tasks, smarter and more fine-grained mask generation methods should be adopted to ensure that the generated data can effectively improve the performance of the restoration model. In this way, we can better cope with the challenges of character erosion repair in the real world and provide more reliable solutions for related application areas.

[0004] At present, the methods for synthesizing mask datasets can be divided into two categories:

[0005] The first category is to use noise and other damage methods to synthesize data sets; the second category is to use image transformation methods to synthesize data sets.

[0006] The first type of implementation method based on noise damage means is mainly based on the basic idea of ​​using various typical noise and other damage means to blur the existing data images. Specifically, first collect a large number of image samples as the original data set, and then use Gaussian noise, salt and pepper noise, Perlin noise, ink blur, Gaussian blur, snow, binary mask algorithm and other methods to add noise or blur. During the process of adding noise and blurring, the original image data set will generate the corresponding task data set. However, the disadvantage of this method is that this type of generation method may be very different from the data set generation method required by the actual downstream task, and cannot simulate the damage effect in the real situation.

[0007] The second type is to process the original image based on image transformation to obtain a recovery model from damaged characters to repaired characters. First, a large number of image samples are collected as the original data set, and then the image transformation method is used to process directly on the original image. The relevant image transformation methods include image masking, image rotation, image cropping, image scaling, Fourier transform, or migration, etc. By using these implementation algorithms multiple times and cross-linking, the original data set can be effectively expanded and synthesized. However, the disadvantage of this method is that this type of generation method is too mechanical and does not meet actual needs, so it cannot simulate the damage effect in real situations.

[0008] Related technologies:

[0009] The existing multiple datasets and synthesis methods fill the gap in text datasets and lay the foundation for many character studies. However, they also have some shortcomings, such as fewer categories and fewer samples, which are a great obstacle to learning and processing involving other tasks. The restoration of Chinese inscriptions is in such a situation: these Chinese character recognition-specific datasets do not meet the needs of restoration tasks, so it is particularly important to choose a suitable dataset synthesis method.

[0010] Due to the lack of a large-scale dataset of ancient Chinese inscription restoration that matches the real erosion effect, most of the current methods for Chinese inscription restoration tasks are still limited. These datasets use common image drawing masks, such as rectangular block masks, but they fail to achieve the generalization of repairing real damaged characters, resulting in poor repair results. At the same time, most dataset collection and expansion methods only obtain a small number of fonts from a single source, so they cannot provide extensive coverage of real eroded character repair. Finally, these methods use masks that are completely unrelated to real erosion to train the repair model, which cannot meet the needs of real tasks and results in poor repair results. Summary of the invention

[0011] The present invention aims to solve the problems of insufficient data sets and poor restoration effects faced by the current field of ancient Chinese stele restoration, and to achieve the task of ancient Chinese stele restoration.

[0012] A method for synthesizing an eroded text dataset that simulates real effects includes two key steps: first, it is proposed to collect real ancient character images from multiple data sources. By collecting inscription images from different sources, the diversity and representativeness of the dataset are ensured. Secondly, the real erosion effect is simulated by extracting key information from the real erosion process. This process involves the analysis and local extraction of different erosion factors. The present invention extracts these erosion features from real inscriptions and applies them to the data set synthesis process to simulate the effect of the real erosion process.

[0013] Technical Solution

[0014] A method for synthesizing eroded text data sets that simulates real effects, the specific process sequentially includes step one, data collection, step two, data preprocessing, step three, extracting local erosion, step four, data enhancement, step five, erosion superposition, step six, combination, etc.

[0015] The present invention aims to provide an improved inscription restoration dataset, which enhances the quality and realism of eroded images through preprocessing and synthesis methods, and provides more reliable and efficient data support for model training and practical applications in the field of text restoration. This will help promote the development of inscription restoration technology, improve the accuracy and efficiency of inscription restoration, and further protect and inherit the cultural heritage of ancient inscriptions. Through the application of the present invention, it is expected to achieve significant progress and results in the field of inscription restoration, and bring important impetus to research and practice in related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is the main flow chart of the present invention.

[0017] Figure 2 For the embodiments, eroded and uneroded versions of Chinese characters from real inscriptions were collected.

[0018] Figure 3 This is a schematic diagram of the comparison before and after pretreatment in the embodiment.

[0019] Figure 4 It is a schematic diagram showing that the local performance of erosion in the embodiment is consistent.

[0020] Figure 5 In this embodiment, representative local erosions are extracted from the collected erosion images as an initial local mask map.

[0021] Figure 6 It is a schematic diagram of the local erosion coverage of the embodiment and the effect presented after multiple accumulation and superposition.

[0022] Figure 7 It is a schematic diagram of the complete process of combined operation in the embodiment.

[0023] Figure 8 This is an example of the effect of the Chinese ancient inscriptions dataset. DETAILED DESCRIPTION

[0024] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0025] A method for synthesizing eroded text data sets that simulates real effects, the specific process includes step 1 data collection, step 2 data preprocessing, step 3 extracting local erosion, step 4 data enhancement, step 5 erosion superposition, step 6 combination, etc. Detailed description of the following steps

[0026] Step 1: Dataset Collection

[0027] First, we collected a large number of inscription images from multiple dynasties and different sources as the original dataset, covering texts of various styles and periods. Secondly, we divided the original dataset into two parts according to whether it was eroded: the real erosion dataset (E1, E2, E3, …, E p , p is a positive integer) and the uneroded data set (U1,U2,U3,…,U q , q is a positive integer).

[0028] Step 2: Data preprocessing

[0029] Data preprocessing includes contrast enhancement, double threshold binarization, and median filtering steps to ensure that all single-character images meet the standards of clear contours, obvious black-and-white color contrast, and low noise pollution.

[0030] First, a contrast enhancement operation is used, which enhances the contrast of the image by adjusting the brightness and color distribution of the image, making the details and edges clearer. Enhancing the contrast can make the brightness difference between different areas in the image more obvious, making the image clearer and fuller, which is convenient for subsequent extraction of erosion. c Set to 1.25.

[0031] Secondly, the double threshold binarization operation is used to convert the image into a black and white binary image. By using two thresholds, the double threshold binarization can reduce the influence of grayscale changes in the image on the final result and improve the stability and accuracy of image processing. The present invention sets the threshold lower limit Th l is 100, the upper threshold Th u The value is 150, and the areas above the lower limit and below the upper limit are extracted.

[0032] Finally, the median filter is used to remove the noise in the image. The preprocessed image is further smoothed by replacing the gray value of each pixel with the median of the neighboring pixels. The principle of median filtering is based on the theory of sorting statistics. It can retain the edge information of the image because it does not blur the edges. Median filtering is particularly effective in processing images containing speckle noise or salt and pepper noise, because these noises can cause sudden changes in pixel values, and median filtering can replace these abnormal values ​​with the median of the surrounding pixel values, thereby reducing the impact of noise. The size K of the median blur kernel of the present invention is s Set to 3 to reduce the interference of noise on subsequent processing steps and improve the quality and analyzability of the image.

[0033] The data preprocessing work standardizes the images collected from different sources by performing operations such as contrast enhancement, double threshold binarization and median filtering on the images. After preprocessing, the processed real erosion text set can be obtained ( p is a positive integer) and the processed uneroded text set ( q is a positive integer).

[0034] Step 3: Extract local erosion

[0035] After the data is prepared, the following method is used to extract local erosion. Erosion is a cumulative process of local erosion. At the same time, erosion from different sources shows similar local erosion patterns. Therefore, the present invention extracts similar local erosions and subsequently increases the number and diversity of these local erosions through specific expansion methods. Specifically, similar local erosions are extracted from the processed real erosion text set to form an initial local erosion set S LM =(LM1,LM2,LM3,…,LM m , m is a positive integer).

[0036] Step 4: Data Enhancement

[0037] Since the number of extracted local erosions is small, the following method is used to expand the initial local erosion set to obtain the local erosion set (LM1, LM2, LM3, …, LM n , n is a positive integer and n>m). The present invention firstly starts from the initial local erosion set S LM A local erosion LM is randomly selected as the object of the current operation. Then a series of transformation operations are performed on the selected LM to simulate various situations that may occur in the real erosion process. These transformation operations include random rotation, random scaling, random horizontal translation, and random vertical translation.

[0038] The random rotation operation used first is to simulate the various angle changes that may occur in local erosion during real erosion, and set the angle lower bound θ lis -180°, the upper limit of the angle is θ u is 180°. Secondly, a random scaling operation is used to simulate the scale changes that may occur in local erosion during the erosion process, and the lower bound of the scaling rate is set 0.1, the upper limit of the scaling rate is 0.3. Finally, random horizontal translation and random vertical translation are used to simulate the changes in horizontal and vertical positions of local erosion during the erosion process, and the maximum absolute value of the horizontal translation H is set. t is 0.9, the maximum absolute value of vertical translation H v is 0.9.

[0039] Step 5: Erosion Overlay

[0040] The erosion process is subsequently simulated. In each superposition process, the present invention superimposes a transformed local erosion onto M, gradually forming a professional synthetic erosion M with diversity, until the number of superpositions reaches the set T. s This operation ensures that the generated set of professional synthetic erosions ( s is a positive integer) has sufficient diversity and richness.

[0041] Step 6: Combination

[0042] Then the preprocessed real uneroded dataset ( q is a positive integer) and the erosion set ( s is a positive integer) to generate a synthetic erosion dataset ( q is a positive integer).

[0043] Finally, the synthetic erosion dataset ( q is a positive integer), professional synthetic erosion set ( s is a positive integer), the processed uneroded text set ( q is a positive integer) and the uneroded data set (U1, U2, U3, ..., U q , q is a positive integer) to obtain the ancient Chinese stele dataset (C1, C2, C3, …, C t , t is a positive integer). The Chinese ancient stele inscription dataset includes a synthetic erosion dataset, a corresponding professional synthetic erosion dataset, a pre-processed un-eroded text dataset, and an original un-eroded text dataset, thus expanding the original eroded text image.

[0044] Example

[0045] The technical scheme of the embodiment includes data collection process, data preprocessing process, erosion extraction, data enhancement, erosion superposition, and combination operation, such as Figure 1 As shown;

[0046] The details are as follows:

[0047] Step 1) Data Collection

[0048] Most of the ancient Chinese characters were recorded on stone tablets. Over time, these characters have been completely eroded, and the original uneroded copies are often damaged or lost, with only the final text version preserved. This makes the current character collection difficult in that it is difficult to collect both the uneroded and eroded versions of the same work or text at the same time. If the collection method is used alone, the data set will be very small.

[0049] The method proposed in the present invention first collects the eroded and uneroded versions of Chinese characters in real inscriptions (such as Figure 2 ). Then, key information is extracted from the real eroded text and used to synthesize erosion. These synthetic erosions will be overlaid on the uneroded image to simulate the real erosion effect. This method can generate a dataset with high similarity to meet the needs of the inscription restoration task.

[0050] Step 2) Data preprocessing

[0051] In order to standardize all single-character images collected from different sources, the present invention performs appropriate preprocessing on these text images. Specifically, given the collected image I, the preprocessing can be defined as:

[0052] I p =Data preprocessing operations (I,E c ,Th l ,Th u ,K s ) (1)

[0053] Among them, I and I p are the unprocessed target image and the preprocessed result image, and the contrast enhancement ratio is E c , double threshold binarization threshold lower bound Th l and the upper bound Th u , the kernel size K of the median blur s The specific processing principle and process are as follows.

[0054] Contrast enhancement refers to enhancing the visual effect and quality of an image by adjusting the contrast of the image. Contrast refers to the degree of brightness difference between different areas in an image. In image preprocessing, contrast enhancement can enhance the details and edges in an image by adjusting the brightness and color distribution of the image, making the image clearer, easier to analyze and understand. Contrast enhancement is an important step in image processing, which can help improve the quality of the image and make the image more suitable for subsequent analysis and processing. The present invention uses contrast enhancement operations to adjust the brightness and color distribution of the input image I←I×E c, where the contrast enhancement ratio E acting on a single text image is c is 1.25.

[0055] Double threshold binarization is a method in image preprocessing, which is used to convert an image into a black and white binary image. Double threshold binarization uses two thresholds to divide the image into three parts: the part below the first threshold is set to black, the part above the second threshold is set to white, and the part between the two thresholds is assigned a grayscale level according to its grayscale value. Its advantage is that it can better preserve the detail information in the image, especially in areas with large grayscale changes. Double threshold binarization can help distinguish areas with different grayscale levels in the image, so as to better perform subsequent image analysis and processing. The present invention uses double threshold binarization to distinguish areas with different grayscale levels in the image and set the lower bound Th l is 100, the upper bound Th u is 150, extracting values ​​above the lower limit and extract values ​​below the upper limit and use Extract the area between these two intervals.

[0056] The median filtering method is used to remove noise from an image. Its principle is to replace the grayscale value of each pixel in the image with the median of its neighboring pixels. This method works well for removing salt and pepper noise and speckle noise because it is not affected by the intensity of the noise pixels. The principle of median filtering is based on the theory of sorting statistics, which can retain the edge information of the image because it does not blur the edges. Median filtering is particularly effective in processing images containing speckle noise or salt and pepper noise because these noises can cause sudden changes in pixel values, and median filtering can replace these outliers with the median of the surrounding pixel values, thereby reducing the impact of noise. When preprocessing an image, selecting a median blur kernel with an appropriate kernel size can retain the detailed information of the image while removing noise, thereby improving the quality and analyzability of the image. The present invention uses median filtering to further smooth the preprocessed image. p ←Median filter operation (I p ,K s ) and set the median blur kernel K s The value is 3 to reduce the impact of noise on subsequent processing steps and improve the quality and analyzability of the image.

[0057] Figure 3 The difference before and after pretreatment is shown

[0058] The specific Chinese algorithm pseudocode is as follows:

[0059]

[0060]

[0061] Step 3) Erosion Extraction

[0062] Erosion manifests itself in various forms in ancient artifacts such as inscriptions, which may include ink stains, noise, interference patterns, and damage caused by excavation. These elements are often presented in a randomly mixed manner, which makes it difficult to accurately simulate the real erosion effect even with the use of multiple irregular masks and mixed Gaussian noise. In addition, directly extracting the complete erosion mask may result in inaccurate results, which will directly affect the generalization effect of the model and lead to poor performance in practical applications. Unlike simply synthesizing overall erosion, the present invention focuses more on extracting similar local erosions and adding these local erosions to the existing data set through specific expansion methods. In this process, it is necessary to comprehensively consider the interaction between different erosion elements, as well as their distribution and density changes in the image.

[0063] By analyzing large-scale actual erosion from different angles, conclusions can be drawn about erosion extraction. Erosion from different sources shows similar local erosion patterns. This shows that although the specific sources and environments of erosion may be different, there are certain similarities in the patterns and characteristics of local erosion. This similarity may be due to common erosion mechanisms or environmental conditions, which make erosion from different sources consistent in local performance.

[0064] (like Figure 4 )

[0065] According to the above characteristics of real large-scale erosion, the present invention extracts representative local erosion from the collected erosion image as the initial local mask. Figure 5 )

[0066] Step 4) Data Augmentation

[0067] Since the number of local erosions extracted with similarity is small, data enhancement is performed on the collected initial local masks with similarity. Specifically, given a set of local masks S LM , then the method of data enhancement in the algorithm is:

[0068]

[0069] Among them, θ l and θ u are the lower and upper bounds of the random rotation angle, and are the lower and upper bounds of the random scaling rate, H t and H v is the maximum absolute value of the horizontal and vertical translation. In this process, the present invention first selects the local mask set SLM A local mask LM is randomly selected as the object of the current operation. Then a series of transformation operations are performed on the selected local mask LM to simulate various situations that may occur in the real erosion process. These transformation operations include random rotation, random scaling, random horizontal translation, and random vertical translation.

[0070] The random rotation operation is intended to simulate various angle changes that may occur in the local mask during the actual erosion process. In the present invention, the angle lower bound θ is set l is -180°, the upper limit of the angle is θ u The random scaling operation simulates the scale change that may occur in the local mask during the erosion process. In the present invention, the scaling rate lower bound is set 0.1, the upper limit of the scaling rate Then, the random horizontal translation operation and the random vertical translation operation respectively simulate the changes in the horizontal position and vertical position of the local mask that may occur during the erosion process. In the present invention, the maximum absolute value H of the horizontal translation is set. t is 0.9, and the maximum absolute value of the vertical translation H v is 0.9.

[0071]

[0072] Step 5) Erosion Overlay

[0073] By analyzing large-scale actual erosion from different angles, we can also draw conclusions about erosion synthesis. Erosion is often a cumulative process of local erosion. This means that new erosion will not only cover the existing erosion, but also interact with it and gradually overlap, eventually forming a more complex erosion effect. In other words, erosion does not achieve the final effect in a single action, but requires multiple accumulations and superpositions to present the true state. (e.g. Figure 6 )

[0074] Therefore, the expanded local mask set can be overlaid to form the final erosion mask. Specifically, given a set of local masks S LM , the number of superpositions is T s , then the method of data enhancement in the algorithm is:

[0075]

[0076] The remaining parameters are the same as those of the data enhancement operation. After each data enhancement operation is completed, the present invention superimposes the transformed local mask LM on the synthetic mask M. In this way, each iteration will superimpose a transformed local mask on M. Repeat the above steps until the number of superpositions reaches the set T. s, and collect all synthesized masks into the synthesized erosion set M e This method can ensure that the generated synthetic erosion set has sufficient diversity and richness, wherein the present invention sets the number of superpositions T s The value is 50.

[0077] The Chinese algorithm pseudocode for data enhancement and erosion superposition is as follows:

[0078]

[0079]

[0080] Step 6) Assembly

[0081] According to the above method, the present invention can simulate various changes of local erosion in the real process, thereby generating a synthetic erosion set M with realism and diversity. e , and the synthetic eroded and uneroded datasets are superimposed to obtain the synthetic eroded dataset, which is used for dataset expansion and synthesis in the inscription restoration task. The complete process of the combined operation is shown in Figure 7 .

[0082] The synthetic erosion dataset, the professional synthetic erosion set M e The processed un-eroded text set is merged with the un-eroded data set to obtain the ancient Chinese stele data set. The ancient Chinese stele data set includes a synthetic erosion data set, a corresponding professional synthetic erosion data set, a pre-processed un-eroded text set, and an original un-eroded data set, thus expanding the original eroded text image.

[0083] The effect example of the Chinese ancient stele inscription dataset of the present invention. Figure 8 )

[0084] Dataset Evaluation

[0085] The present invention uses the 2AFC forced selection method to evaluate the synthetic erosion effect and professionalism. Among them, the 2AFC forced selection method is a method for evaluating a product or experiment by asking participants to choose the better one from two options. Specifically, the present invention randomly selects 60 synthetic eroded text images and 60 real eroded text images covered on uneroded images from the ancient Chinese inscription data set, and shuffles the order. Each person participating in the evaluation is presented with the combination one by one and asked: "Which one do you think is the real damaged image?". Participants are allowed to observe the image for 3 seconds and make a choice within 5 seconds. Finally, the proportion of synthetic images mistaken for real eroded images is calculated to directly reflect the deception rate of the participants and indirectly reflect the effect of synthetic erosion of the present invention. In this process, the embodiment invited a total of 37 artificial restoration experts and 200 graduate students of related majors to participate in the test. The final results show that an average of 63% of the synthetic eroded images of the present invention were recognized by restoration experts and 55% were recognized by graduate students, which shows that the synthesis method of the present invention is effective and consistent with the real erosion effect.

Claims

1. A method for synthesizing eroded text datasets that simulates real effects. It is characterized in that The method includes step one, data collection; step two, data preprocessing; step three, local erosion extraction; step four, data enhancement; step five, erosion superposition; and step six, combination.

2. A method for synthesizing eroded text data sets simulating real effects according to claim 1, It is characterized in that Step 1: Dataset Collection First, we collected a large number of inscription images from multiple dynasties and different sources as the original dataset, covering texts of various styles and periods. Secondly, the original dataset is divided into two parts according to whether it is eroded or not: the real erosion dataset (E 1 , E 2 , E 3 ,...,E p , p is a positive integer) and the uneroded dataset (U 1 , U 2 , U 3 ,...,U q , q is a positive integer).

3. The method for synthesizing eroded text data set simulating real effects according to claim 1, It is characterized in that Step 2: Data preprocessing Data preprocessing includes contrast enhancement, double threshold binarization, and median filtering steps, so that the single character image can meet the standards of clear outline, obvious black and white color contrast, and low noise pollution. After preprocessing, the processed real eroded text set is obtained ( p is a positive integer) and the processed uneroded text set ( q is a positive integer).

4. The method for synthesizing an eroded text data set simulating a real effect according to claim 1, It is characterized in that Step 3: Extract local erosion Extract similar local erosions from the processed real erosion text set to form an initial local erosion set S LM =(LM 1 , L.M. 2 , L.M. 3 ,...,LM m , m is a positive integer).

5. The method for synthesizing eroded text data set simulating real effects according to claim 1, It is characterized in that Step 4: Data Enhancement First, from the initial local erosion set S LM A local erosion LM is randomly selected as the object of the current operation; then a series of transformation operations are performed on the selected LM to simulate various situations that may occur in the real erosion process; the transformation operations include random rotation, random scaling, random horizontal translation and random vertical translation; the above method is used to expand the initial local erosion set to obtain the local erosion set (LM 1 , L.M. 2 , L.M. 3 ,...,LM n , n is a positive integer and n>m).

6. The method for synthesizing eroded text data set simulating real effects according to claim 1, It is characterized in that Step 5: Erosion Overlay In each superposition process, a transformed local erosion is superimposed on the erosion M, gradually forming a professional synthetic erosion M with diversity, until the number of superpositions reaches the set T s ; A collection of professional synthetic erosions generated ( s is a positive integer).

7. The method for synthesizing eroded text data set simulating real effects according to claim 1, It is characterized in that Step 6: Combination The preprocessed real uneroded dataset ( q is a positive integer) and the erosion set ( s is a positive integer) to generate a synthetic erosion dataset ( q is a positive integer); Finally, the synthetic erosion dataset ( q is a positive integer), professional synthetic erosion set ( s is a positive integer), the processed uneroded text set ( q is a positive integer) and the uneroded data set (U 1 , U 2 , U 3 ,...,U q , q is a positive integer) to obtain the Chinese ancient stele dataset (C 1 , C 2 , C 3 ,...,C t , t is a positive integer).

Images