KERTAS: dataset for automatic dating of ancient Arabic manuscripts

Islamic civilization contributed significantly to modern civilization; the period from the 8th to 14th century is known as the Islamic golden age of knowledge. This period marked an era in history when culture and knowledge thrived in the Middle East, Africa, Asia and parts of Europe. Arabic was the language of science and the Arab world was the center of knowledge [1]. Millions of Arabic manuscripts from that era on a wide variety of topics are scattered in different collections across the world. Many efforts have been made by numerous contributors to preserve this valuable heritage. Unfortunately, due to physical degradation of the paper and the ink, processing and studying these documents has proven to be a challenging process. Consequently, these documents are actively being digitized to preserve them. Historians and paleographers are encouraged to work with these digitized versions of the manuscripts. These digital copies are very attractive to researchers because they allow quick and easy access to these historical manuscripts, which in turn provides a way to evaluate, analyze and research these documents without physically handling the delicate and precious works.

The publication or writing date of a historical manuscript has always been important for historians. It can help them understand the sub-textual context of the document and also help in understanding the cultural and historical references that are presented in the text. Knowing when the manuscript was written can also help researchers catalogue and categorize historical documents more accurately and efficiently. Traditionally, historians and paleographers have used invasive methods such as identifying the texture and composition of the paper or components used to make the ink to estimate the age of the document [2]. Some even try to find clues such as dates of historical events within the written content as well as the handwriting and punctuation in order to find the age of the document [3]. A few researchers have also studied ornamentation and watermarks in the documents in order to determine the age of these manuscripts [4]. As mentioned earlier, a large number of ancient manuscripts have been scanned and digitized by libraries and museums. These scanned images have enticed the pattern recognition community as a whole and image processing researchers in particular to try and solve the problem of document age detection using noninvasive techniques [5].

Classifying ancient documents based on writing styles is one of the techniques used to date these documents. System for paleographic Inspection (SPI) [6] is one of the earliest researches that employs writing style-based techniques for ancient documents dating. SPI uses tangent distance and statistical based algorithms to build models of all characters. Afterward, SPI uses the models to measure similarity of the letters in their dataset with the letters of the tested document. Moreover, He et al. in [7] proposed an approach where global and local support vector regression is used with writing style-based features (hinge and fraglets to estimate the date of historical documents. Alternative research on dating ancient manuscript [8], suggests using histogram of orientation of strokes as a feature descriptor to represent the image documents. The descriptor is later sent to self-organizing map clustering system to match the image with a date label. Similarly, Wahlberg et al. used a method based on shape context and stroke width transformation to create a statistical structure for dating ancient Swedish characters [9]. Whereas Howe et al. at [10] applied the Inkball models of isolated character for dating ancient Syriac characters.

While there are quite a few online libraries with datasets in various languages that possess thousands of manuscripts. However, most researchers had to develop their own datasets and find the authorship and age information for verification before they could test and verify their algorithms. A brief review on some existing online dataset is studied in Sect. 4.

The next section provides a brief history of Arabic handwriting over the centuries and its distinguishing characteristics in each period of Islamic history. The design process and description of KERTAS are provided in Sect. 3. Section 4 focuses on a comparison of KERTAS dataset with currently available digitized manuscript resources. Section 5 presents the proposed features to identify the age of historical handwritten Arabic manuscripts. Results and discussion is elaborated in Sect. 6. Then, conclusions are presented in Sect. 7.

While Arabic scripts existed before Islam, Arab was an oral society in that period. Only a few inscriptions were found that go back to that time. Figure 1 shows one of the pre-Islamic inscriptions found in south of Damascus, Syria, in the 6th century C.E. The inscription was written with al Jazm, one of the earliest known styles of the modern Arabic scripts [11].

Islamic calendar (Hijri) started at 622 C.E. Dates in Islamic calendar are denoted A.H. (Latin: Anno Hegirae). In the first two Islamic centuries, most of the historical documents were Qur’anic manuscripts written in Kufic and Hijazi scripts with no signs for the short vowels and no dots to differentiate similar letters as shown in Fig. 2. After the Islamic world expanded, many non-Arab Muslims found it difficult to read the Quran and distinguish between Arabic letters. Arab grammarians adjusted the text of the Quran to avoid distortion; therefore, dots were introduced.

The second half of the second A.H. (8th C.E.) century, the world had seen a growth in the translation movement, and this period was the starting point of the new phase of manuscript history. The arrival of paper in the region contributed to the increase in material documented during that time. Paper was known in China five centuries before Islam. Arabs were introduced to paper in 751 C.E. The number of writing scripts increased. In the 3rd A.H (9th C.E.) century, six scripts were introduced. They were thuluth, naskh, tawāqīʿ, rayḥān, muḥaqqaq and riqā [12]. The dominate scripts in the period of the 4^th–7th Islamic century (10th to 13th C.E) were thuluth, naskh, tawāqīʿ, rayḥān, muḥaqqaq, and riqā. Figure 3 shows a manuscript from the 5th A.H. (11th C.E.) century.

After 8th A.H. (14th C.E.) century, manuscript topics presented less interest in science and more in literature and poetry alongside Islamic subjects. Toward the 13th Islamic century, many of the scripts used earlier had disappeared and the following were used in a wide scale: kūfī, thuluth, naskh, ruqʿah, faressy and tawāqīʿ [13]. There was also a special interest in Arabic calligraphy in that period. Figure 4 shows an example of scripts used during 13th Islamic century.

KERTAS is a term which is applied to paper in Arabic literature records, it is also called waraq [14]. Scholars of Arabic history, scripts and texts consider division on the basis of Islamic centuries to be a better fit. Therefore, manuscripts in KERTAS dataset are categorized based on the Islamic century of their publication/writing. In the development of KERTAS dataset, special care has been taken to find and gather a significant number of manuscripts from each century. We have also attempted to properly verify the writing date of each manuscript in the database. This was not a simple task especially since there are quite a few manuscripts with ambiguous or incorrect writing dates. We checked the writing dates of these manuscripts from multiple sources before adding them to the database While collecting the database, we faced another very complicated issue regarding the writing date of certain manuscripts. In the later centuries, some manuscripts were copied by hand from earlier works. In this case, what would the manuscript writing date be and who would be the author? Ideally, the date should be the writing date of the original manuscript and the author should be the author of the original manuscript. For the algorithms attempting to automatically identify the author or writing date based on the style of writing, the name of the copier and the date of copying should be selected. In this dataset, we have solved this issue by recording the name of the copier and date of copying with the manuscript. This is done in order to provide a consistent view of the data for the author and writing date recognition algorithms. Table 1 shows the sources of KERTAS dataset and the number of manuscripts that were collected from each source. KERTAS dataset not only contains a large number of images (over 2000), but they are quite nicely divided over the 14 Islamic centuries.

Table 1 also shows that Qatar National Library is the main contributor with significant contributions from Cambridge Digital Library, University of Tokyo Library, University of Tubingen, Berlin, and Princeton University Library. The storage structure of the dataset is shown in Fig. 5a. Manuscript is the main directory with a subdirectory for each century. Within each century subdirectory, there are additional directories for each manuscript. These individual manuscript directories contain images of multiple pages from these manuscripts. Figure 5b presents the content of the first manuscript in the first century. Each manuscript subdirectory also contains an XML file, which provides metadata about the manuscript including information about the century in which it was created, source, description, ID, manuscript name and writer’s name if available. Figure 5c shows a sample XML file.

Deciding on the most suitable feature extraction method is perhaps the most crucial step to achieve a high recognition rate. Age detection of historical manuscripts is a very challenging problem. We have to contend with the complexities inherent in working with noisy images, and there is an additional challenge we have to tackle. The class boundaries between the documents written in two adjacent centuries are highly nonconvex and nonlinear. This means that for documents written in two centuries and for documents written by two different authors, a very high interclass similarity will exist. In addition, handwritten documents by different authors are present in a single century, thus providing very high interclass variability.

The proposed approach presented above is evaluated on KERTAS dataset in order to show its performance. The experiment is set up by keep around 32% of the images selected randomly from the dataset as unseen test data. Every attempt has been made to keep this selection process random with only check been to make sure that all the classes are properly represented. The remaining images are kept as part the evaluation dataset.

Different image sizes are created by scaling the image to smaller sizes and no cropping takes place. We used the whole image because we are interested in using both the writing style and the layout style for document age detection as both elements are able to provide some information about the age of the document and it makes sense to utilize them both in order to improve the performance.

Different image sizes are tested to select the optimal feature set size. The rationale behind reducing image size is two folds: Firstly, reducing the image size reduces the dimensionality of the data thus producing a better underdetermined linear system which in term provided a better optimization solution. Secondly, the smaller variations such as noise and blemished are automatically removed thus providing an inherent robustness to the whole process. There is an issue with reducing the image size, i.e., if a manuscript image becomes too small, a lot of handwriting features that are useful in detecting the age of the manuscript may be lost. This reduces the possibility of a correct match. On the flip side, increasing the image size increases the dimensionality and thus reducing the chance of a better optimization solution. In order to identify the optimal size for manuscript age detection, we tested the algorithm multiple images sizes starting from 12 × 12 (which was deemed optimal by authors in [26] for facial recognition) to 25 × 25, 50 × 50, 100 × 100, 200 × 200 and d 250 × 250. We also tested the algorithm with two different splits for the dataset. The first split is with predefined training and testing samples (95% training and 5% testing). This split enables us to train the algorithm with maximum variance of data. The second split of the dataset uses another approach where two-third random samples used for training (68%) and one-third for testing (32%). The results of the experiments are provided in Table 3.

The results show that as we increase the image size, initially the accuracy tends to improve and this is because more discriminative features are made available with the increase in size. The highest accuracy is obtained at the image size of 50 × 50. However, if we continue to increase the image size the accuracy tends to drop.

To evaluate the previous approach, we compare the results from sparse representation-based with some of state-of-the-art methods for writing style detection. First, the dataset was preprocessed to be used with writing style-based features. Text area was segmented and binarized using Otsu threshold method [27]. Afterward, features were extracted by run length, edge hinge and edge direction methods. We evaluated the performance of the features using k-Nearest Neighbor (k-NN) with k = 3. Table 4 presents the results of using writing style-based features on KERTAS dataset.

In this paper, we presented a dataset (KERTAS Dataset) accumulated specifically to assist researchers working on designing solutions and algorithms for digital paleography. The dataset consists of over 2000 high-quality, high-resolution digital images acquired from multiple historical handwritten Arabic manuscripts from multiple sources. Detailed metadata are provided for each image to assist in testing and verification of manuscript author detection and manuscript age detection algorithms. In addition, we presented a sparse representation-based approach to detect the age of manuscripts in order to highlight the suitability of the dataset. The algorithm also provides a baseline accuracy measure that can be compared with other algorithms developed in the future by using KERTAS dataset. Furthermore, we employed some writing style-based features to compare with the proposed approach and to study the consistency of writing style in each century in KETRAS dataset. The dataset will be made publicly available for research purpose through competition and IAPR TC-11 site.