An example of a Persian word consists of two PAWs, where p represents pen width. The red line is the visualization of the imaginary baseline. 

Context in source publication

Context 1

... languages with Roman alphabet and East Asian scripts. Although there has been a great attempt in producing omni-font OCR systems for Persian/Arabic language, the overall performance of such systems are far from perfect. Persian written language, which uses modified Arabic alphabet, is written cursively, and this intrinsic feature makes it difficult for automatic recognition. An essential part of a document understanding process, which seems trivial in human's character recognition, is how we perceive the written text parts and how we distinguish between different sections of a document. To eye of an educated person, the layout of a document is discerned once it came to sight but to a machine, this seemingly simple task will become burdensome. To a large extend, the layout of written materials appear in rectangular columns with all text and shapes aligned to the margins. This arrangement is mostly referred to as Manhattan style documents. In addition to this style, there are more sophisticated styles where printed text and figures in a document are not aligned. Thus, majority of efforts for layout analysis of documents are focused on Manhattan style, but recently, there is a growing attention toward the complicated styles as well [3]. Persian/Arabic printed documents are treated in the same way in layout analysis however; there is an absence of solid resources in the literature for these languages. Once the layout of a document is extracted, the system has access to each line of text. Thereafter, the problem of understanding of the script reduces into recognition of the words. There are two main approaches to automatic understanding of cursive scripts: holistic and segmentation-based [4]. In the first approach, each word is treated as a whole and the recognition system does not consider it as a combination of separable characters. Very similar to the speech recognition systems, in almost all significant results of holistic methods, Hidden Markov Models have been used as the recognition engine [5, 6]. The second strategy, which owns the majority in the literature, segments each word to containing characters as the building blocks, and recognizes each character then. In comparison, the first strategy usually outperforms the second, but it needs a more detailed model of the language, which its complexity grows as the vocabulary set gets larger. In addition, in this method, the number of recognition classes is far more than similar number in segmentation-based methods. Recently, there is also a trend toward hybrid methods which incorporates the segmentation and recognition systems to obtain overall results; these methods are usually called segmentation-by-recognition [7, 8]. One of the main concerns of designing every OCR system is to make it robust to the font variations. Thus, successful examples are omni-font recognition systems with ability to learn new fonts from some tutor. In holistic methods, as the character recognition problem is viewed from a different perspective, and the system collectively uses learning mechanisms for few connected characters, the transformation of the system into an omni-font learning system would be smooth. On the other hand, the segmentation-based systems mainly use learning methods only in recognition process, and to the best of our knowledge, the learning systems are never used for the segmentation process in the literature [9]. Usually, human recognizes unfamiliar words by segmenting them and recognizing each character separately to understand the whole word. With this perspective, in this research, the whole task is broken down into two separate learning systems to gain from reduction of complexity in hierarchy as well as adaptability of learning systems. The layout of this paper is as follows: Section 2 emphasizes on the characteristics of Persian script that were crucial for the design of OCR systems. In Section 3, we will discuss the proposed algorithm, which includes the layout analysis, segmentation, and recognition modules in separate subsections. In Section 4, implementation details and results are discussed which is entailed with conclusive remarks and acknowledgements. In this section, we will briefly describe some of the main characteristics of Persian/Arabic script to point out the main difficulties which an OCR system should overcome. As one of the main properties, the script consists of separated words which are aligned by a horizontal virtual line called "baseline". Words are separated by long spaces and each word consists of one or more isolated segments each of them is called Piece of a Word (PAW). On the contrary PAWs are separated by short spaces and each PAW includes one or more characters. If one PAW has more than one character, each of them will be connected to its neighbors along the baseline. Figure 1 shows a sample Persian/Arabic script, where a represents the space between two different words, and b is the short space between PAWs of a word. In Figure 2, the first PAW on the right comprises three characters and the second one, on the left, consists of only a single character, and p denotes the pen width value which is heuristically equal to the most frequent value of the vertical projection in each line. The overall block diagram of the system is presented in Figure 3 which depicts layout analysis, post-process, and natural language processing (NLP) subsystems in addition to recognition and segmentation blocks. This paper presents the design of the layout analysis system in addition to the segmentation, feature extraction, and recognition sections (Figure 3). The design of the NLP module is out of scope of this paper. With respect to the popularity of Manhattan style in Persian documents, this research was focused on documents with aligned columns without any graphical figure inside; therefore, the automatic exclusion of the graphical material from text is left for future works. The designed method includes a two-step skew detection with use of a combination of basic Hough transform method [10] and the method proposed in [11] that uses fuzzy run-lengths. The layout analysis section will process a document and adaptively segments it into paragraphs and then into the separated lines. In the proposed system, the segmentation-based approach is exploited, and some measures are considered to overcome the main weaknesses of it. The best results are achieved with aid of artificial neural networks (ANN) for performing segmentation with some extended features (Section 3.2). In recognition section, we obtained a definite set of features from each segmented symbol, which was fed to a support vector machine (SVM) classification engine to obtain the recognized symbol. Using large margin classifiers enables us to achieve high recognition rates which are in coherence with the best results in the literature [2]. We also decomposed each character of Persian script to more primitive symbols called graphemes. This novel decomposition has decreased the complexity of the recognition and segmentation procedures and has improved the overall result. Few different characters could share a single grapheme, and additionally, several joint graphemes could build a single character. In addition, Persian language includes many characters which the only difference they have is the number of dots and placement of them. To finalize the character recognition task, a post- processing section is implemented to combine the result of grapheme recognition and the number of dots. Besides, this section corrects some common grapheme recognition errors using an embedded confusion matrix. Figure 4 shows the combination of grapheme recognition and post-processing processes with dot recognition module. Before proceeding further, we provide concepts of some frequently used terms in this paper for clarification: Grapheme : In this research, we refer grapheme to any graphical image that would be a character or a part of it which acts as a fundamental building block of words. This resembles the concept of phonemes in speech, but we don't directly choose them in relation to real phonemes. Pen tip : The vertical position of the pen in the skeleton of a PAW image. Junction points : The horizontal position of the grapheme boundary. Thus, cutting the word at junction points results separated graphemes. The layout analysis section is responsible for segmenting document images into lines. The input of this system is clean black and white images with low noise, and the output is images of the separated lines. In this research, we considered a 300 DPI scanned document as input since it is a regular resolution for Persian/Arabic OCR. All of the other internal parameters are fixed for this resolution and for other scanning resolutions, all we need is to scale them proportional to the ratio they have to this predefined DPI. Thus, parameters of the proposed system only depend on the easily defined scanning resolution. Since after scanning a regular size paper such a detailed resolution creates a huge file size, we need to consider some measures to avoid the slow down of the algorithm. Hence, in line segmentation block, most of the time-consuming processes are computed on a downsampled version of the documents image, and only the final processes are performed on the full-size image with some slight adjustment to insure the accuracy. Figure 5 illustrates the block diagram of the layout analysis subsystem. The downsampling is performed with the ratio of 1/5, so that will result a very rough version of the original image. In Figure 5, H and V smear refer to the process of black painting every neighboring pixels of each back pixel in the boundary of 1 in horizontal or vertical direction respectively. In the skew detection module, a method similar to [11] was performed to estimate skew angle, and with this estimation, the exhaustive search of skew angle with Hough transform of the full-size image was narrowed ...