**2. Related Works**

Arabic document page segmentation has also been studied by using traditional machine learning (ML) techniques. Hesham et al. [7] proposed an automatic layout analysis scheme for Arabic manuscripts. They further appended a line segmentation support. Text and non-text areas were differentiated by using the Support Vector Machine (SVM) algorithm. They also identified words and lines.

Artificial Neural Networks were further tested on Arabic document layout analysis schemes. Bukhari et al. [14] differentiated the central body and the side manuscript by applying the Multilayer Perceptron (MLP) classifier. A dataset is created which includes 38 historical document images and they achieved 95% classification accuracy. Long Short Term Memory (LSTM) and CNN are employed for document page segmentation of scientific manuscripts written in English in [15,16]. Amer et al. developed a CNN-based document page segmentation scheme for Arabic newspapers and Arabic printed manuscripts. They obtained approximately 90% accuracy in detecting text and non-text areas. CNNs have also been employed for historical document layout analysis [2,3,17]. The page segmentation algorithms are important because they could be applied prior to keyword spotting, HTR and OCR techniques in some studies (as in our work) and, therefore, their performance is critical.

There are very few Arabic handwriting keyword spotting studies in the literature [6]. Some QbE studies ([18–20]) are proposed for the historical Arabic documents and used a matching method adjusted to the Arabic script. QbS approaches [21,22] used the HMM technique for keyword spotting in handwritten Arabic manuscripts. They were standard HMM KWS applications without taking the particular properties of the Arabic script into account. A spotting scheme is developed specifically for Arabic handwritten digits/symbols achieved an overall precision of 80% and 83.3% recall [10]. Another prominent keyword spotting research conducted on both historical Arabic dataset VML and George Washington datasets. Barakat et al. [23] applied a convolutional siamese network that uses two identical convolutional networks to rank the similarity between two word images. In this way, they developed a system which is more robust against different writing styles and is able to recognize out of vocabulary words.

After spotting the numerals, Arabic digits should be recognized for information retrieval from the historical manuscripts. Arabic digit recognition is a well-studied topic in the literature [13] (see Table 1). Melhaoui et al. proposed an Arabic digit recognition scheme that used multi-layer perceptron and K-nearest neighbor classifiers [24]. They run tests on the dataset include 600 Arabic digits with 200 testing images and 400 training images. They achieved 99% recognition accuracy on this small database. The HODA dataset was used for testing Persian (which is based on Arabic scripts) handwritten digit recognition systems in the literature [25–27]. Takruri et al. [28] proposed a three-level classifier that uses Support Vector Machine, Fuzzy C Means, and Unique Pixels for the classification of handwritten Arabic digits. They achieved 88% accuracy on the dataset containing 3510 images. Sawy et al. also achieved 88% accuracy by using CNN on the public ADBase dataset [13]. Kateeb et al. used the same dataset (ADBase) and applied the Dynamic Bayesian Network technique for digit recognition. They achieved 85.26% accuracy. Ashiquzzaman et al. achieved 97.3% accuracy by using MLP with appropriate activation and regularization functions on the public CMATERDB 3.3.1 Arabic handwritten digit dataset [29]. They further improved their system accuracy by using data augmentation and dropout to 99.4% [12]. However, as mentioned before, these studies were carried out on modern open datasets, and they did not need to alleviate the low-quality data issues of historical manuscripts. To the best of our knowledge, our study is the first to develop a CNN-based Arabic numeral spotting and handwritten digit recognition system for historical documents by using deep transfer learning methods.


**Table 1.** The comparison of our study with the Arabic handwritten digit recognition studies on different datasets.
