Offline Omni Font Arabic Optical Text Recognition System using Prolog Classification Technique

Editor's Notes

• #2: A commonly used term in conjunction with OCR software. Omnifont recognition refers to the capability of computer software, usually OCR software, to read (or recognize) virtually any font that maintains fairly standard character shapes
• #4: The field of Optical Character Recognition (OCR) is a branch of technology that deals with automatic reading of a text. The ultimate goal of OCR is to simulate the human ability to read both machine-printed and hand-written texts. Currently, a variable system can read faster than a human, but it cannot reliably read such a wide variety of texts. Humans are also better at reading from highly distorted text or noisy (unclear) media. Therefore, a great deal of intensive research is still needed to narrow the gap between the humans’ and machines’ reading capabilities. OCR has been used in many practical areas that are independent of the language to which it is applied. One of the earliest and most successful applications was sorting checks in banks, where the volume of checks circulating daily proved to be too enormous to be handled by a manual entry method. Reading of handwritten and printer postal codes, text archiving and retrieving, reading of customers’ handwritten forms and aiding visually impaired people to read are a few other examples. Most of the work on OCR has been on Latin and Chinese characters. Work on Arabic character recognition has only started recently and had advanced relatively slowly due to the complexity of recognizing Arabic text, which has characters that are cursive in nature. Arabic character recognition is still an open and challenging field of research.
• #5: This thesis proposes a complete system that classifies and recognizes machine-printed Arabic text. The input to the system is a clean, high-resolution Tag Image File Format (.TIFF) that contains Arabic text to be recognized; the output is simply the generated Arabic text saved in a Microsoft Word Document (.DOC) file of the recognized Arabic text. The technique is based on cleverly describing the text in terms of shape primitives derived from Freeman chain codes. A rule-based data enhancement technique is used to improve recognized features as much as possible. The recognized features are processed by a Prolog feature-matching engine to classify character classes as well as diacritic information as three separate streams (character class stream, diacritic stream and corners information stream). In addition to the three provided streams, estimated font size is also provided as a fourth input. Characters are finally determined by processing a permutation of the three streams using Definite Clause Grammar (DCG).
• #7: The extant literature includes a considerable number of studies focused on recognition of writing in languages such as Latin, Chinese and Hebrew. Unfortunately, limited research has been done on the recognition of Arabic writing. This might be attributed to the peculiar aspects of the writing of Arabic characters . For example, Arabic writing is from right to left. This may not be considered a major technical point, but it becomes an issue when an Arabic text contains some foreign text such as Latin or French, which is written from left to right, and vice versa. Arabic writing is cursive. Arabic characters are generally tailored to each other. This requires a process of segmentation of the Arabic words and characters before any recognition step can be taken. In fact, segmentation and character isolation is the most difficult problem in the character recognition schemes of Arabic language. An Arabic word might be comprised of both cursive or separated characters, Twenty-two of the 29 sets of Arabic characters assume different shapes and sizes depending on other positions within the words. Some characters such as “ ع ” take four shapes, which are all different from each other: “ ع , ع , ع ,ع ”. The character “ ”ج takes three shapes: “ .” ج ,ج ,جThere are only seven characters, which, regardless of their positions within the words,have only one shape: “ و, ذ, د, ظ, ط, ز, ر ”. In general, characters, even if they are of the same group, are of different sizes and they require different rectangular boxes that can enclose them.
• #9: This thesis proposes a complete system that classifies and recognizes machine-printed Arabic text. The input to the system is a clean, high-resolution Tag Image File Format (.TIFF) that contains Arabic text to be recognized; the output is simply the generated Arabic text saved in a Microsoft Word Document (.DOC) file of the recognized Arabic text. The technique is based on cleverly describing the text in terms of shape primitives derived from Freeman chain codes. A rule-based data enhancement technique is used to improve recognized features as much as possible. The recognized features are processed by a Prolog feature-matching engine to classify character classes as well as diacritic information as three separate streams (character class stream, diacritic stream and corners information stream). In addition to the three provided streams, estimated font size is also provided as a fourth input. Characters are finally determined by processing a permutation of the three streams using Definite Clause Grammar (DCG).
• #10: Having conducted a thorough literature review, we started designing our system by experimenting with prior researchers’ techniques, adopting or modifying some of them if they met our requirements, but otherwise developing our own techniques. Consequently, the components of our system are either due to the work of others, the result of our improvement of others’ work, or our own completely new techniques. During the development phase of the system, many of the investigated techniques were rejected due to their ineffectiveness. This ineffectiveness may be due to inherent deficiency or due to incomplete description specified by their text sources.
• #11: Preprocessing PhaseThis phase is implemented in C language to perform some image processing with the help of open source image processing libraries such as LibTiff and OpenCV.In our proposed system, Arabic text images have been obtained by optical scanning of the character images on the plain paper. The input data obtained by scanning of printed text is almost contaminated with noise and contains redundant information. Preprocessing includes digitalization, scaling, word-level segmentation, noise removal and elimination of redundant information as far as possible.Image information retrievalIn this process, we load and read the input Tagged Image File Format (TIFF) image file as binary; retrieve the image properties (size, width, height, pixel resolution, image channels and image alignment; and create memory storage for system intermediate processing.Image digitalizationThe image digitalization process digitizes the TIFF image in order to apply fixed-level thresholding and to convert the gray-scale and bitmapped image to a binary (0’s and 1’s representation) scale image. It does the vertical and horizontal histograms to retrieve the number of lines per page and number of components (words) per each line. In this phase, we calculate the font baseline and size. The font baseline can be calculated by finding the maximum horizontal histogram of each line per page. This enables the dots or other special characters such as Shadda, Madda, and Tanween to be classified as upper or lower components related to this baseline. Text line detection: Text line detection has been performed by scanning the input page image horizontally. Frequency of black pixels in each row is counted in order to construct the row histogram. The position between two consecutive lines, where the number of pixels in a row is zero, denotes a boundary between the lines. Here it is assumed that the text block contains only single column of textWord segmentation: After a line has been detected, it is scanned vertically. In order to find the column histogram, the number of black pixels in each column is calculated. If there exists n consecutive scans that find no black pixel, we denote it to be a marker between two words. The value of n is taken experimentally. Figure 5 andFigure 6show the preprocessing phase.
• #20: It is suggested from our practical experimentation to apply some rules to the retrieved Freeman chain contours. We introduced an algorithm to remove noisy pixels which come within any straight line, and to convert Arabic characters to approximately straight lines. These enhancement rules, which are derived from testing Arabic characters multiple times, reduce the time required for character recognition. The result of the algorithm is illustrated in Figure 10.