Studies On Offline Arabic Handwriting Optical Character Recognition Method

The increasing availability of online digital libraries has rekindled interest in the problems of how to automatically digitize and index those handwriting documents and how to efficiently browse, search and manipulate them. Traditionally, Optical Character Recognition (OCR) provides an effective solution for converting machine-printed documents into digitized form, it is not, however, a solved problem for handwriting paper documents, which requires accurate and robust offline handwriting OCR algorithm.From perspective of modern theory of pattern recognition and computer vision, it is valuable and important if breakthrough of offline handwritten OCR technology can be achieved. At the same time, the research of handwritten document recognition algorithm is helpful in various document related fields, such as improving the recognition accuracy of degraded documents, solving the text detection and recognition in video images, identifying the fonts, size, and machine-printed/handwriting of characters, recognizing Arabic handwritten document as well as other image recognition technology and application. The research result might even be used in improving on-line handwriting.Arabic handwriting has its unique characteristic and challenges which requires innovative solutions, since existing methods only achieved limited success in constrained cases in this field. In recent years more and more research institutes started to investigate the recognition of offline handwriting Arabic document. Domestic and international researchers have performed a wide investigation from different perspective in this field, and have achieved some success. Overall, the challenges of Arabic handwriting OCR are listed as follows:1) Style:Arabic text, both handwritten and machine printed, is cursive in nature.2) Text Flow:Arabic script is written from right to left, rather than left to right, in contrast to Latin text or other languages, such as Chinese, Japanese or Korean. Letters are connected to form a "baseline". For handwriting, the baseline is an ideal concept and a simplification of actual writing. In practice connections occur near, but not necessarily on a baseline. This factor improves the difficulty of recognition. 3) Letter Shape:The shapes of the letters differ depending on where in the word they are found. The same letter at the beginning and end of a word can have a completely different shape. Together with the dots and marks representing vowels, this makes the effective size of the alphabet about 160 characters. Another challenge is the letter shape is very variable as horizontal stroke elongations frequently occur in letter ligatures.4) Word Shape:Even for the same word, the different writers may have different writing styles, resulting in different word shape.5) Overlap:Letters in a word often overlap due to the writing styles, making the segmentation difficult. In some cases the letters may be written in reverse orders by some writers.Up until now, there has no commercial product of offline handwriting Arabic recognition system. The accurate and the choice of testing set of existing system should be improved. It has distinguished distance between the effect and reality.In this thesis, we have implemented an end-to-end offline handwriting Arabic recognition system. It includes the acquisition of document images, pre-processing, feature extraction and designing of classifiers. The existing algorithms have been evaluated. The main research work is presented as follows:â‘ Arabic words are characterized by a baseline where most of the pixels in the words are concentrated. It is a virtual line over which the writer writes the word. Ideally the modified forms of alphabets join on the baseline. The ascender and descender positions are determined by this virtual reference line. A common technique for global reference line detection is based on the vertical histogram of pixels. The method is effective if the text is written horizontally and scanned without skew, but fails when text is skewed. Unfortunately, baseline skew is present to varying degrees in most freeform handwriting. Our proposed baseline detection algorithm is based on a two-step linear regression:Locate the local minima points on the contour of the word (the minima points where the contour changes direction from downwards to upwards);Calculate an approximate baseline by the linear regression (least squared sum) on minima points;Locate the points that are close to the first approximation; Refine the line by a second regression only on the minima points that are close to the first approximation;â‘¡Feature extraction. Human beings can recognize characters through extracting features such as the structure of a character, strokes, etc. For the computer feature extraction is also required. The common method is to analyze the stroke, characteristic point, information of projection, distribution of point region and so on. We often use structural and statistical analysis methods to extract features. In this thesis, GSC and Principal Component Analysis and compression are used to describe handwriting Arabic document images.â‘¢Classifier design. The research of classifiers usually chooses BP net, SVM and HMM. Because the characteristic of Arabic ligature is the same as the recognition of speech, we investigate the application of HMM based on the non-segmentation method, such as the designing and choosing of arguments, optimizing the experimental value. We implemented the system through automatic study, automatic choosing, automatic optimizing and applying hybrid classifier techniques so that the recognition accuracy can be improved.