Optical Character Recognition (OCR) detects and recognizes the printed and handwritten text from an image and converts it to editable text. The editable version of the image is usually used for further processing operations like indexing, searching, analyzing, and modification. The OCR has many applications such as data entry [1], vehicle license plate recognition [2], postal address reading [3], bank cheque reading [4], intelligent driving systems [5], and invoice reading [6]. The OCR systems are either printed-based or handwritten-based OCRs. People have different writing styles; therefore, recognizing handwritten scripts is more challenging than recognizing printed text and the expected accuracy is low. There are online and offline OCR systems. The online OCR performs the recognition result in real-time and only uses handwritten scripts. Most OCR approaches are designed to extract the printed script from a single image document. However, few approaches are proposed for video-based OCR.

Most of the existing Arabic OCR systems are either handcrafted or learned-based OCR systems. Deep learned features showed superior results compared with handcrafted features during the last decade in terms of recognition accuracy and speed in many fields of image processing [7,8]. Deep learned-based OCR systems use Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNN) for feature extraction and classification [9,10]. The CNN architecture consists of a sequence of layers. The CNN layers include convolutional, max pooling, sampling, and a fully connected layer. Arabic handcrafted OCR systems usually extract features from contour features [11], statistical features [12], topological features [13], geometrical features [14], from feature descriptors such as SIFT descriptor [15], Speeded Up Robust Features (SURF) [16] and Histograms of Oriented Gradients (HOGs) descriptor [17].

Most of the existing OCR systems work on Latin, Japanese, Chinese, and Korean scripts [18]. Many authors mentioned that few efforts were made in Arabic OCR compared with Latin and English scripts [19,20]. During the last decade, there has been dramatic advancement in the development of Graphics Processing Units (GPUs). This significantly speeds up the computation time and allows scientists to design deep CNNs that can extract and process more features. The Deep CNNs methods achieved state-of-the-art results in many fields of computer vision and image processing. CNNs extract features directly from raw pixels without any preprocessing operations, are invariant to object class changes, handles inputs with high dimension, and have distinct features [21]. Tian et al. [22] proposed OCR approach to recognize English characters from natural images. The method was inspired by the success of using deep CNNs for object recognition. The architecture is based on the VGG16 CNN model followed by a Bi-Directional Long Short-Term Memory (LSTM). The method achieved state-of-the-art on ICDAR2013 and ICDAR2015 benchmarks. Ye et al. [23] developed an OCR approach called TextFuseNet to recognize English texts with irregular shapes from natural scenes. The CNN architecture uses the ResNet model to fuse and extract characters, words, and global text features. The method achieved state-of-the-art on the following datasets: CTW-1500, ICDAR2015, ICDAR2013, and Total-Text. Mackay et al. [24] proposed OCR to recognize real images that contain English words. The approach is called Rotational You Only Look Once (R-YOLO). The system is based on the YOLO4 CNN object detector architecture. Non-Maximum Suppression is used to eliminate the redundant bounding boxes. The method detects text with arbitrary rotation angles. The method results outperformed the state-of-the-art in the following datasets ICDAR2017-MLT, ICDAR2013, ICDAR2015, and MSRA-TD500.

Each language has different writing structures and styles. The Arabic language is written cursive, from right to left, and diacritics are used and can change the word meaning accordingly. Recognizing Arabic script is more challenging than English script due to many facts [25–27]: the cursive nature of the Arabic language, the high similarity between the Arabic letters, the Arabic language is very rich morphologically, the diacritics in the Arabic language can change the word meaning. Most of the existing Arabic OCR approaches in the literature are tested under a controlled environment with some constraints, such as datasets with high-quality images, isolated characters, or isolated words, and few OCRs were tested on a page-level script. Some approaches were used with specific font types and sizes. Most of the existing Arabic OCR approaches ignore diacritics. According to an experiment [28] that evaluated four well-known OCR systems, Sakhr, ABBYY, RDI, and Tesseract, the recognition accuracies are 51.56%, 75.19%, 46.00%, and 48.61%, respectively (font type was Arabic transparent). The experiments were conducted on the KAFD dataset, a page-level printed text dataset. Hegghammer [29] evaluated the performance of Google Document AI and Tesseract OCR systems in English and Arabic. The systems’ performance was lower in the Arabic language compared with English. Despite the evaluated English dataset was challenging and included scripts from old books and different font styles. In contrast, the Arabic dataset includes simple font styles and writings from the internet like Wikipedia. The word recognition rates for Document AI and Tesseract are 85.8% and 65.2% for English scripts and 80.2% and 58.9% for Arabic scripts, respectively. The OCR for Arabic script is still an unsolved problem for printed and handwritten scripts, especially for page-level scripts [28–30].

This study reviews the main contributions in Arabic optical character recognition during the last decade.

The review will be useful for all readers interested in Arabic OCR: non-expert people can have an overview of the Arabic OCR techniques. At the same time, the expert readers can extend the existing methods and datasets or adapt them to their applications. The contributions of this article are summarized as follows:

Explain the main characteristics of the Arabic language.

The Arabic language is the official language in all Arab countries. According to World Bank statistics [37], the population in Arab countries will be more than four hundred million people in 2020. Muslims worldwide use the Arabic language to read the Holy Quran written in the Arabic language. The Arabic alphabet is used in many other languages such as Persian, Pashto, Kurdish, Urdu, Punjabi, Brahui, Baluchi, and Kashmiri. The Arabic language consists of twenty-eight letters in addition to Hamza (). The Arabic letter shape is changed according to its location in the world. At most, there are four shapes for each letter beginning, middle, end, and alone, as shown in Table 1. Dots are used in the Arabic language and change the letter meaning according to their locations, as shown in Fig. 3. Fig. 4 shows how the shape of the letter Ayan (“”) changes according to its location in the word. Unlike Latin text, the Arabic language is written in a cursive way and from right to left direction, and there are no capital and small letters. The cursive property makes it difficult to segment the Arabic word into characters. The Arabic language has many diacritics that can change the word meaning accordingly. Arab can understand the word meaning from the sentence context without writing the word’s diacritics. Table 2 shows the main Arabic diacritics with their pronunciation. Table 3 shows examples of how diacritics change the word meaning. The Arabic language has a rich and complex morphological structure [25,26]. Table 4 shows some but not all variations of the root Katab (). According to [27], recognizing Arabic script is more challenging than recognizing English script. The Character Error Rate (CER) for Arabic is 3.3% vs. 1.1% for English. According to the authors, that is due to the high similarity between Arabic characters, the Arabic language is written in a cursive way, and it has many ligatures and a variety of Arabic fonts and styles.

The OCR system includes 4 stages: preprocessing, segmentation, feature extraction and classification as shown in Fig. 5.

The APTI dataset was developed in 2009 [112]. It represents a Large-scale printed Arabic dataset for OCR evaluation. The dataset contains 45,313,600 images. Each image contains one Arabic word, with about 250 million Arabic characters. The dataset is synthetic and generated from a distinct 113,284 words. It includes ten font types, ten font sizes (6 pt--24 pt), and four font styles. The dataset is divided into five sets. Set number 5 is used for testing, and the other sets are used for training. The author publicly published the first four sets. Fig. 19 shows a sample of 15 images from the APTI dataset. The dataset is available at https://diuf.unifr.ch/main/diva/APTI/download.html.

The MMAC is a printed Arabic text dataset developed in 2010 [113]. The number of unique words and PAWS are 282,593 and 66,725, respectively. The number of images is increased by a factor of three by skewing and adding noise to the images. The dataset was collected from old books, Arabic research, and the Holy Quran. Fig. 20 shows sample images from the MMAC dataset. The dataset is available at http://www.ashrafraouf.com/mmac.

The HACDB is an Arabic handwritten dataset [114]. Fifty writers collected it from ages 14 and 50. Each writer writes 132 shapes of characters. The shapes represent the way of writing the Arabic characters at different locations in the word (beginning, middle, end, and isolated). The total number of characters is 6,600. Fig. 21 shows sample images from the HACDB dataset.

The AcTiV is a video-based OCR dataset [115]. It includes 80 videos (850,000 frames) collected from four news Arabic channels: Aljazeera, France 24, Russia Today, and EI Wataniya. The dataset includes texts with different sizes, colors, positions, and fonts. Additionally, the background is complex and has many objects with shapes like Arabic characters. Fig. 22 shows sample images from the AcTiV dataset.

The Hijja dataset represents a handwritten Arabic dataset developed in 2020 [116]. 591 children write it under 12 from Riyadh, Saudi Arabia. The dataset contains 47,434 characters that can be used with real-life applications to teach children spelling and handwriting skills. The author reported that the dataset is complex, and it is challenging to train a model to fit the data. Fig. 23 shows sample images from the Hijja dataset. The dataset is available at https://github.com/israksu/Hijja2.

The KAFD dataset was developed by king Fahd University and Qassim University in 2014 [117]. It includes 15,068 images and 2,576,024 lines of printed text. Images have different resolutions 100 dpi, 200 dpi, 300 dpi, and 600 dpi. The dataset includes four different Arabic fonts, ten font sizes ranging from 8 to 24 points, and four font styles; Normal, Bold, Italic, and Bold Italic. The dataset images are divided into training, testing, and validation. A sample picture from the dataset is shown in Fig. 24. The dataset is available at http://kafd.ideas2serve.net/.

The KHATT is a handwritten text dataset developed in 2013 [118]. The dataset contains one thousand forms written by one thousand different writers. The dataset includes two thousand paragraphs with 9,327 lines taken from forty-six sources. It includes three image resolutions 200 dpi, 300 dpi, and 600 dpi. Seventy percent of the dataset is used for training, 15% for testing, and 15% for validation. Fig. 25 shows a sample of three images from the KHATT dataset. The dataset is available at http://khatt.ideas2serve.net/KHATTDownload.php.

The IFN/ENIT database is one of the oldest handwritten text datasets developed in 2002 [119]. The database includes 2,200 images with 300 dpi resolution for Tunisian cities, the images contain 26,459 words, and the total of Arabic characters is 212,211. The dataset achieved the highest number of citations (640) compared with the existing Arabic OCR database. Fig. 26 shows a sample of three images from the dataset. The dataset is available at http://www.ifnenit.com/.

The Yarmouk Arabic OCR dataset is a printed Arabic text dataset developed in 2018 [120]. It includes 8,994 images with 300 dpi resolution. The images contain 436,921 words extracted from the Wikipedia website. Fig. 27 shows a sample image from the dataset. The dataset is available at https://drive.google.com/drive/folders/0B4Kx3iMuktgsdC12Ui1neklnMzQ?resourcekey=0-dX3YkFT4xArRrT81wQ2wSw.

The APTID/MF is a printed Arabic text dataset developed in 2013 [69]. It includes 1,845 images with 300 dpi resolution. The images contain 27,402 characters. The images are taken from 387 pages of Arabic documents. The images include ten font types, two font styles (normal and bold), and four font sizes (12 pt, 14 pt, 16 pt, and 18 pt). Fig. 28 shows a sample of three images from the dataset. The dataset is available upon request.

The ARABASE dataset is a printed and handwritten Arabic text dataset developed in 2005 [121]. More than 400 writers wrote handwritten images, most of them from Tunisia. The printed text was obtained from daily newspapers and the book published by the Tunisian national library on the internet. The image’s resolution ranged from 200 dpi to 600 dpi. Fig. 29 shows a sample of two images from the dataset. The dataset is available upon request.

This section summarizes the techniques and dataset proposed for Arabic OCR during the last decade. Initially, Google Scholar was used to find related articles. A broad set of keywords were used to identify the list of related articles. The used articles are peer-reviewed and have an excellent citation number. The selected articles are relevant to the Arabic OCR, and the methodology of the proposed approaches and datasets were clearly described. In addition to the Google Scholar database, the snowball method was used to find related articles using references from some relevant articles. The search stage includes the following terms: Arabic OCR, Arabic optical character recognition, Arabic OCR dataset, Arabic OCR Database, handwritten Arabic characters, printed Arabic recognition, CNN Arabic OCR, Handcrafted Arabic OCR, and deep learning Arabic OCR.

Rosenberg et al. [15] proposed OCR called SOCR to recognize printed Arabic text. The approach used sliding window techniques with SIFT descriptor to segment a line of Arabic text into paws and letters. In addition to SIFT, a set of features are extracted: Mass Center, Black color Histogram, Crosshair, Ratio, and Outline Features. The extracted features are classified using Hidden Markov Model (HMM). The authors reported that the training set is small compared with previous methods. The approach evaluated the PATS dataset and outperformed the PATS [119] results on five out of eight fonts. Additionally, the approach evaluated six font sizes (6, 8, 10, 12, 18, 204) of the set4 of the APTI dataset and achieved a 99.6%-character recognition rate.

Sabbour et al. [11] proposed a Printed Arabic OCR approach called Naboc. It was trained to recognize two types of fonts: Arabic Naskh and Urdu Nastaleeq. A dataset called UPTI was created for Printed Urdu Text. The system’s main steps are to segment the page script into lines. After that, the lines are segmented into ligatures. A descriptor is used to describe the features of ligatures. Finally, classify the ligatures into a predefined class label. Contour pixels’ intensity and Shape Context are used to describe the ligatures. K-Nearest Neighbor is used for classification. The recognition error rate for Arabic ligature is 16.1%, whereas Tesseract’s error rate is 16.2%.

Talaat et al. [13] used preprocessing operations to enhance image quality. The preprocessing operation includes Binarization, Slant Correction, Normalization, Statistical Noise removal, and Morphological operations (filling, Dilation Bridging). The extracted features include Lower and Upper image profiles, Vertical and Horizontal profiles, connected components, and Topological features—Neural Networks (NNs) area used for classification. The input vector for NNs is 133 elements, whereas the output is 28 neurons (number of Arabic letters). The approach achieved 88% accuracy on the CENPRMI dataset. Characters ( and ) achieved low accuracy rates (61% and 66%).

Hafiz et al. [122] performed preprocessing operations: binarized, slant removal resizing, and dilation to improve image quality. For feature extraction, the images are divided vertically into 6-pixel widths. Then, Images are manually segmented into paw images. The following features are extracted: foreground color density, the transition of white and black pixels, the sum of pixels in a vertical column, and concavity features. A combination of HMMS and K-NN is used for classification. The approach achieved 82.67%, 86%, and 94% on splits A, B, and C of the IFN/ENIT database.

El-Sawy et al. [123] developed CNNs to recognize isolated handwritten Arabic letters. The CNN architecture contains two Conv2D and two max pooling with Relu Activation, a fully connected layer, and SoftMax Activation. The input is an image of 32 × 32 pixels, and the output is a fully connected layer with 28 numbers, representing the probabilities for 28 Arabic letters. The approach achieved 94.9% accuracy on a private dataset.

The OpenITI team developed OCR for printed classical Arabic scripts collected from the old book “Ibn al-Faqīh’s Kitāb al-Buldān [124].” The approach is based on a customized Kraken open-source OCR software. The overall recognition rate of the method is 97.56%.

Mudhsh et al. [125] proposed deep CNNs to recognize handwritten characters. The Alphanumeric method is based on the VGGNet architecture. The number of filters was reduced by 8, reducing the time complexity to run the VGGNet. The approach achieved 97.32% and 99.66% accuracies on HACDB and ADBase datasets.

Nashwan et al. [67] proposed a holistic Arabic OCR. The whole world is recognized without the need to segment it into letters. Compared with previous works, the approach used many vocabularies and reduced the recognition time. Clustering techniques cluster similar words’ shapes to reduce the recognition time. Two features are extracted: Discrete Cosine Transform (DCT) and local block features. Then the features are clustered, and Euclidean distance is used to find the best possible matched word. The accuracy of evaluating the approach on 3465 words is 84.8%.

Doush et al. [110] reported that their proposed approach represents the first end-to-end Arabic post-processing approach. For many Arabic OCR systems, the OCR output does not match the ground truth text for some characters due to high similarity between Arabic characters. Therefore, the author proposed using post-processing to correct the word spelling. The method is based on: Language Model (LM) to check whether the word is correct or not, the Error Model (EM), and Google to correct the wrong words. The proposed approach reduced the error rate from 14.95 to 14.42. the approach was evaluated on a dataset of 500 images.

Radwan et al. [78] developed an OCR system to recognize printed Arabic words. The architecture is based on three CNNs. The CNNs were used for three tasks: font size detection, character segmentation, and character recognition. The authors reported that OCR accuracy is 94.38% on APTI synthetic dataset.

Darwish et al. [104] developed printed Arabic OCR. The approach performed preprocessing operations: converting image into grayscale, median filter, morphological operations, correct rotation, and image resizing. Gray Level Co-Occurrence Matrix (GLCM) was used for feature extraction. A genetic algorithm was used to choose the best features and reduced the running time by two. K-Nearest Neighbor classifier is used for classification. The approach was evaluated on PATS-A01 (650 images) and APTI (550 images) datasets and achieved 98.69% and 95.37% recognition rates, respectively.

Fasha et al. [126] proposed a hybrid OCR approach for printed Arabic text. The approach includes five CNNs layers followed by Bi-Directional Short LSTM. The CNN architecture includes five Conv2D layers with Relu for activation and five max pooling layers. The output from the CNN is passed to Bi-Directional LSTM (BDLSTM). The BDLSTM consists of two layers of LSTM. Each layer contains two cells, and each cell has 256 hidden units. The approach was evaluated on the APTI dataset and a chevied 85.15%-character recognition rate and 76.3%-word recognition rate.

Shams et al. [127] proposed OCR for handwritten Arabic text. The Proposed CNNs architecture consists of three convolutional layers and three max pooling layers to recognize handwritten Arabic isolated characters. CNNs extract features, dropout operation, is applied to reduce the running time, and SVM is used to classify the features into 28 classes (number of Arabic letters). The approach achieved 95.07% accuracy.

Altwaijry et al. [116] proposed CNNs architecture to recognize isolated handwritten Arabic characters. The CNNs include three Conv2D layers, each followed by ReLU activation and max pooling. The output is flattened into a fully connected layer followed by two fully connected layers with an 80% dropout rate. The approach was evaluated on AHCD and Hijja datasets and achieved 97% and 88% recognition rates.

Balaha et al. [21] proposed 14 different CNNs architectures for handwritten Arabic OCR. The proposed architecture includes three Conv2D layers and three max pooling layers. The difference between the different architectures is in the number of filters and the fully connected layers. Additionally, combinations of VGG16, VGG19, and MobileNetV2 are evaluated. The accuracy of the proposed CNN-5 architecture is 91.96% on the HMBD dataset, and it requires less memory and processing time compared with VGG16, which achieved 92.74%.

Ahmed et al. [128] proposed CNNs architecture that includes 9 Conv2D layers with 3 × 3 kernels and five max-pooling layers with 2 × 2 kernels. Batch normalization is used after each Conv2D layer. Dropout with rates from 0.1 to 0.4 is used after each max pooling layer. The tensor flattens by a fully connected layer followed by another fully connected layer. The authors reported that the proposed method achieved a super result (99.94%) compared with VGGNet-19.

Jbrail et al. [129] developed four CNNs architectures to recognize isolated handwritten Arabic characters. The architectures use a different number of layers (3, 9, 13 layers). Different activations functions (Relu and Softmax) and optimizations (Gradient descent and Adam). It includes a deep neural network with nine hidden layers. The layers contain Conv2D with 3 × 3 kernel and max pooling with 3 × 3 kernel with Relu and SoftMax activation. The approach achieved 99.3% accuracy on the Hijja dataset.

Table 6 summarizes the mentioned above Arabic OCR approaches and Table 7 summarizes the Arabic datasets. The citations column is obtained in May 2022 from Google Scholar. In the last two rows of Table 6, ABBYY FineReader Engine and Tesseract open-source library were used in MATLAB and evaluated on a sample of 1,000 images from the APTI dataset.

This section summarizes the leading commercial and open-source software for Arabic OCR.

Table 8 shows the recognition rates for some well-known OCR software that support the Arabic language. Four software, Tesseract, Abbyy FineReader, Sakhr, and Readiris, are used in the comparison. The authors used different datasets to evaluate the four software. It is clear from the table that the recognition rates for the software vary from one dataset to another, and no one software outperformed all software on all tested datasets. The average accuracy rate for these softer ranges from 70% to 80%.

Tesseract [131] is an open-source OCR. It supported more than 100 languages and was developed by Hewlett-Packard (HP) as a PhD project from 1985 to 1994. In 2006, it was sponsored and redeveloped by Google till 2018. The latest available version of Tesseract is version 5. It supports different operating systems Windows, Linux, and Mac. Tesseract converted the input image into a binary image. Fuzzy space is used to divide the text into words. The connected component is used to identify the layout of the characters. Finally, a classifier is trained to find the labels of the character.

Abbyy FineReader [130] is commercial software that supports more than 200 languages. It supports different operating systems Windows, Linux, and Mac. It accepts documents in different formats such as PDF, TIFF, and JPEG, printed and handwritten. The image is preprocessed to enhance its quality. The preprocessing operations include banalization, rotation, and deskewing. The document analysis stage is used to identify the image structure and the formats of its elements, such as the location of the header, footer, tables, diagrams, and text fields. Since its commercial software, the details of the recognition approach are unknown. Abbyy OCR has an online trial version with 1000 pages for each registered user.

Sakhr [137] is commercial software it supports the Arabic language or the languages that use Arabic characters such as Farsi, Urdu Pashto, and Jawi. It supports both printed and printed scripts and provides online and offline recognition. It runs on Windows operating systems, and there is no trial version. Sakhr claimed to be the best available OCR for the Arabic system, according to US government evaluators. Sakhr claimed 99.8% accuracy for the documents with high-quality images.

Readiris [138] is commercial software. It supports around 130 languages (Including Arabic, Russian, and East Asia languages). It accepts images, PDF files, and document folders and converts them into editable text. It supports Windows and Mac OS.

Many evaluation metrics are used to evaluate the performance of the Arabic OCR system. Character Error Rate (CER) and Word Error Rate (WER) are the most used metrics. According to Eq. (3), CER is calculated where iC,dCandsC denote the minimal number of character insertion, deletion, and substitution operations (Edit distance), respectively, to transform the OCR output to the ground truth script (see Fig. 30). nC denotes the total number of characters in the text. WER is calculated according to Eq. (4), where iw,dwansw denote the minimal number of word insertion, deletion, and substitution operations, respectively, to transform the OCR output to the ground truth script. nw indicates the total number of words in the text. The Character Recognition Rate (CRR) and Word Recognition Rate (WRR) are computed according to Eqs. (5) and (6), respectively [28]. Some dataset contains isolated characters for OCR evaluation. Therefore Eq. (7) is used [123], where cC denotes the total number of correctly recognized characters, and nC denotes the total number of tested characters. Eq. (8) is used to find the word accuracy rate, where wC denotes the total number of correctly recognized words, and nw denotes the total number of tested words. Text Recognition Rate (TRR) considers the whole image text as one unit. The image could contain one word, a line of words, or a paragraph. TRR measures the percentage of correctly recognized text images according to Eq. (9), where tC is number of images that are correctly recognized, nt indicates the total number of text images [71]. (3)CER=iC+dC+sCnC∗100% (4)WER=iw+dw+swnw∗100% TER=it+dt+stnt∗100% (5)CRR=100−CER (6)WRR=100−WER (7)CA=cCnC∗100% (8)WA=wCnw∗100% (9)TRR=tCnt

Developing an accurate and fast Arabic Optical Character Recognition system will be helpful for many people in the Arab and Muslim regions. However, the accuracy of the existing printed-Arabic commercial OCR software does not exceed 75%, according to some studies, when tested on a page-level image. Additionally, most current approaches work offline and do not recognize the Arabic script in real-time speed. Recognizing Arabic text is a challenging task due to many reasons. Therefore, Arabic character recognition is still an open research area, and there is a range for enhancing and improving the existing systems. Many approaches are evaluated on a private dataset, a word, or a paragraph level, making it difficult to know their performance in real-world Arabic scripts. Convolutional Neural Networks have been used wildly during the last decade in Arabic optical recognition and showed significant results compared with handcrafted approaches. A comprehensive review of the latest advances during the previous decade in Arabic Optical character recognition is introduced in this paper. This paper reviews the following: the characteristics of the Arabic language; different types of OCR systems; the main stages of the Arabic OCR system, the techniques used in each step, and the researchers’ contributions; comparisons between the existing Arabic OCR methods, commercial and open-source software; the current datasets for Arabic OCR and their characteristics; evaluation metrics for the OCR system. Future works include the followings: customizing and training popular CNNs models such as DenseNet, AlexNet, VGGNet, MobileNet, SqueezeNet, ResNet, and GoogLeNet to develop a new Arabic OCR system; using a large dataset for training the CNNs model to achieve a higher recognition rate and using RNNs with CNNs for handwritten and printed text; evaluating the proposed approach on a well-known benchmark to measure the actual performance of Arabic OCR systems.